PLoS Biology

doi: 10.1371/journal.pbio.2005628pmid: 29608559

The tyranny of temperature in ecological systems Life at all scales—from E. coli to elephants—is powered by metabolism. Metabolic processes convert resources and energy to do the work of life. Warming temperatures accelerate metabolic rates by increasing the kinetic energy of biochemical systems. Despite the bewildering complexity of metabolism, two metabolic processes—aerobic respiration, oxygenic photosynthesis—are shared across diverse groups of animals, plants, and many microbes and fungi. The importance and constancy of these metabolic processes across the diversity of life lends some intriguing (and promising) predictability to how living systems respond to temperature change. This signal of temperature on the pace of life across ecological systems and species has been described in the metabolic theory of ecology (MTE) [1–5]. Remarkably, the effect of temperature on vital rates, such as growth or development, and on biological energy and material fluxes (e.g., respiration) is quantitatively consistent with how temperature affects key metabolic processes of respiration or photosynthesis at the subcellular level [2,6,7] (Fig 1). Consequently, the flux of energy and materials driven by metabolic processes increases predictably with warming at scales of cells [2], organisms [8,9], populations [4,10,11], communities [12,13], and ecosystems [3,14] and even along biogeographic gradients [7]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. The MTE posits that temperature constrains rates of metabolic processes within cells, and these constraints emerge at higher levels of biological organization, such as individuals, populations, and species interactions. Within individuals, constraints imposed by temperature on cellular respiration and associated biological processes (A) can be estimated as the exponential increase of metabolic rate over a temperature gradient (described by the activation energy parameter Ea). Very similar Ea values characterize the relationship between mass-normalized organismal respiration rate and temperature across species from a wide range of taxonomic diversity (B) (for ease of interpretation, lower horizontal axes are shown in reversed 1/kT, where k is Boltzmann’s constant and T is temperature in Kelvin, while upper horizontal axes are in °C; data from [2]). The temperature dependence of respiration constrains demographically important rates, such as development rate and its inverse, development time, which decreased exponentially with increasing temperature in 72 marine animals (C). The exponential effects of temperature shown in the left panel are often log transformed for analysis, allowing Ea to be described as a slope on an Arrenhius plot (right panel) (data from [9]). Temperature-dependent performance influences the outcomes of species interactions, including consumer–resource and host–parasite dynamics (D). Ea, activation energy; MTE, metabolic theory of ecology. https://doi.org/10.1371/journal.pbio.2005628.g001 The inescapable effects of temperature on fundamental metabolic rates provide context and constraints for myriad adaptations of traits that minimize the otherwise extreme consequences of excessively cold or warm environments [15–17]. For example, within hundreds of species examined so far, development times of marine larvae, zooplankton, and fish increase exponentially as temperatures decline [8,9], but negative effects of slow development, such as increased vulnerability to predation or starvation, have been overcome by some species in cold climates through evolution of life histories that involve increased parental care of offspring [18]. Evolutionary processes have only somewhat modified the temperature dependence of photosynthesis or respiration across different life forms but instead have acted on other, more labile, traits to generate patterns in growth and resource-use traits across thermal gradients, which can compensate for the effects of temperature on metabolic performance [19,20]. In plants, adaptation of leaf and tree traits may erase the signal of temperature-dependent photosynthesis on canopy production at certain spatial and temporal scales, even though the underlying photosynthetic processes are still predictably sensitive to temperature [19]. Evolution plays out in the context of resource limitation, competition, facilitation, and selection by predation and parasitism, leading to complex ecological systems whose function and structure may not be clearly attributable to temperature variation alone. However, MTE offers an approach to understanding ecological systems that begins with the highly repeatable temperature dependence of metabolism and then considers how evolutionary and ecological processes play out, given this constraint imposed by temperature, providing a multi-scale framework for understanding how ecological systems change with temperature. Metabolic scaling and the outcomes of species interactions One area in which the importance of temperature-dependent metabolism has been more difficult to understand is in the domain of population-level processes at local scales. This is because, while temperature-dependent metabolic rates affect vital rates of births and deaths directly, these vital rates are also affected by predation, disease, parasitism, resource supply, competition, or allocation. Metabolic scaling as a unifying principle that includes the dynamics of one or a few species has been more elusive, and complicated by evidence that general metabolic temperature dependence is potentially overwhelmed by the complexity, contingency, and context dependence of how temperature affects physiological traits, demographic processes, and their interactions [21–23]. Compounding the problem, empirical tests that actually measure demographic rates over temperature gradients under conditions that meet the assumptions of MTE are few and far between [24]. When the temperature dependence of fundamental metabolic rates is used in mathematical models to predict demographic vital rates and the outcomes of species interactions, the outcomes suggest nonintuitive shifts in abundance and persistence of populations with warming [10,25], predator–prey interactions [26], and host–parasite interactions [27]. These outcomes are not directly proportional to the effects of temperature on metabolism because population dynamics mediate the relationship between temperature, metabolism, and abundance. However, MTE and associated empirical tests [10,28] suggest that under simple conditions (minimal stress or mortality), temperature-dependent metabolism underlies processes at the population level in a manner consistent with models of general metabolic scaling. This is one of the most promising approaches ecological science has right now to understand how environmental temperature affects the dynamics and future persistence of ecological systems in a changing world. A new look at a host–parasite interaction in a warming world Kirk et al. [29] combined mathematical and experimental approaches to determine whether metabolic scaling theory helped to understand how temperature affects the cost of parasitism. Parasites require a host for all or part of their life cycle, using energy or resources of their hosts and often costing the host its life or a portion of its ability to pass on genes to the next generations. All living organisms are vulnerable to parasites, and changes to host–parasite interactions can have devastating consequences for plant or animal populations. Changes to the cost of parasitism resulting from environmental change can be a critical component of the evolutionary trajectory of a population and even affect its ability to persist. In their experiment, Kirk et al. [29] exposed healthy water fleas (Daphnia magna) to a microsporidian parasite (Ordospora colligata) at a range of 9 temperatures. They measured the effect of temperature on key attributes of the host–parasite species interaction. Importantly, their 9 temperature levels allowed them to estimate the functional form of the temperature dependence of these vital rates. Both Daphnia and parasites were affected by temperature; Daphnia life spans declined as temperatures increased over most of the temperature range, except for increases in life span with warming at the lowest temperatures [29], and parasites were able to infect Daphnia at higher rates at higher temperatures. The effect of temperature on parameters that influence host–parasite dynamics, such as parasite population growth rate and Daphnia survival rate, were consistent with expectations for how temperature affects respiratory metabolism [2,4] and differed over the temperature gradient (Fig 2). Kirk et al. [29] compared 2 approaches to describing the effects of temperature on growth and mortality rates: first, they estimated these parameters at each temperature separately, and secondly, they fit a Sharpe-Schoolfield model with an Arrhenius function describing the increase in performance with temperature to all responses across all temperatures. They found that the model with the Arrhenius function, as predicted by MTE, was a more accurate and efficient description of how temperature affected Daphia and its parasite over the temperature gradient. Kirk et al. [29] then used this function to model population dynamics and estimate the cost of parasitism. The temperature dependences of other population attributes, abundance, and declines in life spans under physiologically stressful conditions were not well aligned with predictions based on the temperature dependence of respiration. It was expected that parasite abundance would be negatively related to the temperature dependence of respiration [4] and potentially modified by changes in parasite phenotype. For the curvature parameter that describes the decline in Daphnia performance at high temperatures, MTE does not predict a relationship between this and the temperature dependence of respiration. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Metabolic scaling in the context of host–parasite dynamics. Using a combination of experiments and mathematical modeling, Kirk et al. [29] show that host survival, parasite growth, and the cost of infection in Daphnia magna can be predicted based on the temperature dependence of metabolic processes (A). Parasite growth rate increased with temperature, and the temperature dependence of parasite growth reflected the temperature dependence of cellular respiration, consistent with MTE. A mechanistic model of within-host parasite population dynamics, based on parameters derived from MTE (e.g., parasite growth rate), accurately predicted host life span, which was highest at intermediate temperatures (B). MTE, metabolic theory of ecology. https://doi.org/10.1371/journal.pbio.2005628.g002 Kirk and colleagues also report that the demographic consequences of temperature-dependent vital rates in this host–parasite interaction explain a maximal cost of infection at intermediate temperatures, even though population growth rates of the parasite are greatest at warmer temperatures (Fig 2). Such relationships between demographic outcomes and individual metabolic rates are no surprise to many population ecologists focusing on bioenergetics [30]. However, dynamic outcomes that reflect nonlinear changes to individual performance and population dynamics are not always considered when results of global change experiments are extended to predictions for effects of climate change. Kirk et al. [29] offer an important reminder that comparing experimental outcomes with metabolic theory predictions and their potential implications requires consideration of the dynamics that link temperature and an emergent response, such as abundance. When these researchers took this approach, they found that the metabolic theory of ecology can explain how temperature affects species interactions and their outcomes. Kirk et al. [29] show that MTE leads to accurate predictions that can be applied over a continuous range of temperatures (in contrast to other models that can only be applied at discrete temperatures, for which empirical data are available). Their approach, based on theory, complements another leading framework for predicting global change impacts, the multiple stressor framework. Kirk and colleagues’ finding of a high cost of infection at intermediate temperatures but a small cost of infection at high and low temperatures might not have been predicted by the multiple stressor framework, which emphasizes responses to temperature at extreme temperatures. This finding highlights the importance of considering biological responses to temperature as continuous functions that reflect both stressful and nonstressful metabolic processes. Conclusions The metabolic theory of ecology has provided new opportunities to understand how ecological systems grow and change across scales of space, time, and biological organization. This framework has challenged existing paradigms historically restricted to a narrower ecological scale but has also suggested that energetic constraints on metabolism are highly general and constrain demographic and evolutionary outcomes. Building on this, we are in a better position to project future ecological states from a theoretical framework with a broad conceptual and empirical domain, rather than making predictions by extrapolating data or models of restricted scope. Kirk et al. [29] demonstrate that MTE can be a powerful framework for predicting disease dynamics over gradients of temperature and for making predictions in the context of climate change.

doi: 10.1371/journal.pbio.2005707pmid: 29668680

Introduction Mitochondrial function is key to normal cellular physiology, given the many different biochemical process that occur in the organelle [1]. A tremendous amount of effort over the past several decades has been dedicated to understanding how mitochondrial dysfunction impacts the cellular environment and organismal health. This has been largely based on studies of rare mitochondrial diseases that share many molecular mechanisms with more common disorders that also present with mitochondrial dysfunction, e.g., Parkinson’s disease, cancer, and diabetes. Despite these efforts, fundamental aspects of how mitochondria function impacts cellular physiology remain ill defined. For instance, how the mitochondria communicate with and impact reactions within the nucleus is poorly understood. The gene expression program(s) and metabolic rewiring that change in response to mitochondrial dysfunction are not clear. It is also not known whether mitochondrial-driven epigenetic changes impact gene transcription. Finally, despite the fact that mitochondrial dysfunction can increase DNA methylation [2,3], a mechanistic link between these effects is still missing. We recently described a novel cell culture model of progressive mitochondrial DNA (mtDNA) depletion in the human embryonic kidney 293 (HEK293) background [4]. This model relies on the inducible expression of a mutant mtDNA polymerase gamma that works as a dominant negative, herein called DN-POLG (dominant-negative DNA polymerase gamma transgene). Upon addition of doxycycline, DN-POLG is expressed, and over a period of 9 days, the mtDNA is completely depleted. Because the mtDNA encodes critical components of the electron transport chain (ETC)—which generates ATP using tricarboxylic acid (TCA) cycle–derived nicotinamide adenine dinucleotide reduced (NADH) or flavin adenine dinucleotide hydroquinone (FADH2)—by depleting the mitochondrial genome, we can regulate electron transport, ATP production, and the flux through the TCA cycle. Additionally, because a membrane potential (ΔΨm) is generated as a byproduct of the ETC, producing reactive oxygen species (ROS), we can also modulate redox signaling. Using this model, we showed that complete loss of mtDNA and of ETC function (achieved at day 9) led to severe mitochondrial dysfunction and loss of cell proliferation. We also showed that, concomitant to loss of mtDNA, histone acetylation was decreased in the nucleus. Using isogenic cells that ectopically express 2 nonmammalian proteins—NADH dehydrogenase-like 1 (NDI1) and alternative oxidase (AOX)—we found that maintenance of NADH oxidation with a pseudo-ETC was sufficient to preserve levels of histone acetylation but had no impact on cellular proliferation in the absence of mtDNA. Conversely, by deleting ATPase inhibitory factor subunit 1 (ATPIF1), a regulatory subunit of the mitochondrial ATPase, we showed that maintenance of the ΔΨm without rescue of histone acetylation or ATP production sustained cell proliferation under conditions of complete mtDNA loss, seemingly by restoring redox signaling [4]. The unique progressive nature of mtDNA depletion and mitochondrial dysfunction of this cell culture system provides an exceptional opportunity to fill some of the gaps of knowledge in the field. In this study, we took advantage of this model to gain insights into how cells respond to stepwise mitochondrial dysfunction from transcriptomic, metabolic, and epigenetic perspectives. Our approach revealed that metabolic, epigenetic, and gene expression changes initiate prior to detectable signs of mitochondrial dysfunction, primarily centering on an amino acid response that also aims to sustain the TCA. Unexpectedly, we found that polyamine metabolism is significantly changed upon mitochondrial dysfunction; this, in turn, impacts the methionine cycle and DNA methylation in ways that are independent of serine-driven one-carbon (1C) remodeling or transsulfuration. Results MtDNA depletion drives dynamic transcriptional changes centered on acetyl coenzyme A (acetyl-CoA) metabolism The progressive nature of mtDNA depletion in the DN-POLG system provided us with a unique opportunity to address fundamental questions about the nuclear response to progressive mitochondrial dysfunction. We took an integrative approach that involved transcriptomic, metabolomics, and epigenetic analyses at each time point (days 0, 3, 6, and 9) during the course of complete mtDNA depletion. Furthermore, we utilized the cells expressing NDI1/AOX to define the responses to complete mtDNA loss that were specifically linked to NADH oxidation in the mitochondria. A schematic representation of this integrative approach is shown in S1A Fig. The reproducibility of the results is shown in S1B–S1G Fig. We started by performing RNA sequencing (RNA-seq) at days 0, 3, 6, and 9 in the DN-POLG cells, and found 2,854 genes (S1 Data), including all mtDNA-encoded transcripts, whose expression was changed (adjusted p ≤ 0.05) at any given day compared to day 0 (Fig 1A). When stratifying by day, we found 236 differentially expressed genes (DEGs) at day 3, which were mostly (78%) upregulated, while we found 2,135 DEGs at day 6 (1,064 upregulated and 1,071 downregulated). At day 9 we found 1,272 DEGs, most of which (64%) were upregulated (Fig 1B). Common to all time points were 121 DEGs, including POLG that was upregulated at least 7-fold relative to day 0 (S1 Data). The identification of over 200 DEGs at day 3 was surprising given the lack of significant changes in mitochondrial function at this time [4]. Our model also revealed progressive upregulation of fibroblast growth factor 21 (FGF21) and growth differentiation factor 15 (GDF15) relative to day 0 (S1 Data). FGF21 and GDF15 are metabolic cytokines induced in patients and mouse models of mtDNA or protein translation defects, which have been proposed as biomarkers of mitochondrial dysfunction [1]. Validation of randomly selected genes by quantitative real-time PCR can be found in S2A Fig. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Differential regulation of transcriptional and metabolic profiles in DN-POLG cells in the course of inducible mtDNA depletion. RNA-seq was performed from DN-POLG cells at days 0, 3, 6 and 9 of doxycycline supplementation; N = 3 biological replicates per time point. Transcriptional profiling was based on DEGs identified using log2-transformed fold-changes in RPKM versus the mean at day 0 (Log2FC) in two-way ANOVA tests (gene × time) at an adjusted Benjamini-Hochberg p ≤ 0.05. (A) Unsupervised clustering analysis of log2-fold expression changes versus average of day 0 for DEGs observed in DN-POLG cells by RNA-seq; increased (red) or decreased (green) expression. (B) Number of DEGs identified in DN-POLG cells at days 3, 6 and 9 of doxycycline supplementation compared to day 0 (circle plots, top). (C) Schematic representation of the main transcriptional responses associated to DEGs in DN-POLG cells at days 3, 6 and 9 of dox-inducible mtDNA depletion in DN-POLG cells, per IPA. (D) Schematic representation of interconnected amino acid and methionine metabolic pathways associated to differentially enriched metabolites at days 3, 6, and 9 of dox-inducible mtDNA depletion in DN-POLG cells (N = 4 per time point), per IPA. (E–H) Box plots of relative content versus day 0 (IQR-outlier format) depict distribution of individual replicates in DN-POLG cells within statistical groups for (E) IMP, Xan, XMP, Cho, Met, Leu, Put, 4-HPPA, (F) Ser, (G) f-Met, and (H) Cys; partial gray backgrounds highlight metabolite relative contents at day 3 in (E). Underlying data are reported in S1 Data for (A) and (B); S2 Data for (C) and (D); and S3 Data for (E–H). 4-HPPA, 4-hydroxyphenylpyruvate; Cho, choline; Cys, cysteine; DEGS, differentially enriched genes; DN-POLG, dominant-negative DNA polymerase gamma transgene; ETC, electron transport chain; f-Met, N-formylmethionine; GABA, γ-aminobutyric acid; IMP, inosine 5'-monophosphate; IPA, Ingenuity Pathway Analysis; Leu, leucine; Met, methionine; mtDNA, mitochondrial DNA; OXPHOS, oxidative phosphorylation; Put, putrescine; RNA-seq, RNA sequencing; ROS, reactive oxygen species; RPKM, reads per kilobase per million; Ser, serine; TCA, tricarboxylic acid; Xan, xanthine; XMP, xanthosine 5'-monophosphate. https://doi.org/10.1371/journal.pbio.2005707.g001 The use of Ingenuity Pathway Analysis (IPA) revealed that, globally, the genes modulated during the course of mtDNA depletion enriched for tRNA charging, cholesterol biosynthesis, glutamate (via 4-aminobutyrate) and putrescine degradation, and D-myo-inositol-5-phosphate metabolism (top 10, S2 Data). These results are consistent with the loss of mtDNA having a negative impact on oxidative TCA flux, since glutamate degradation can contribute to the cycle as precursor for succinate. Likewise, putrescine can contribute to the TCA as succinate via γ-aminobutyric acid (GABA) catabolism [5], known to occur in cells beyond the central nervous system [6]. To our knowledge, this is the first report to identify polyamine metabolism as responsive to loss of mitochondrial function. Inositols are sensitive to glucose and NADH levels [7], the metabolism of which is also impacted by loss of mtDNA. To understand how DN-POLG cells respond to progressive loss of mtDNA, we next stratified the analysis based on time point and the directionality of gene changes. We reasoned that the genes modulated at day 3 would reflect early responses to mtDNA depletion that take place prior to signs of mitochondrial dysfunction [4]. Those DEGs at day 6 would reveal responses to significant loss of ETC function and the pathways directly linked to this process—e.g., ATP production and the oxidative TCA—while the ones detected at day 9 would reveal adaptive changes within the cell. A summary of the main transcriptional responses identified at each time point is shown in Fig 1C. Most genes detected at day 3 (184) were upregulated and enriched for methionine and cysteine degradation (S2 Data). Alterations in the methionine cycle have not been directly associated with mitochondrial dysfunction, although channeling of 1C units toward transsulfuration of homocysteine to cysteine, a branch point in the methionine cycle, has been recently reported [8,9]. However, the degradation of methionine can ultimately input into the TCA cycle by contributing pyruvate (through cysteine) and succinyl-CoA through aminobutanoate [10–12]; this would provide a link between mtDNA depletion and methionine degradation. Nevertheless, it was surprising to identify these changes at day 3 when no significant alterations in mitochondrial function were identified [4]. Concomitant to the degradation of methionine, we observed DEGs involved in the recycling (or salvage) of this amino acid through betaine, which was likely an attempt to maintain methionine levels (S1 and S2 Data). The 52 genes downregulated at day 3 were enriched for lipid metabolism, presumably to spare acetyl-CoA, and endothelial nitric oxide synthase (eNOS) signaling (S2 Data). The 2,135 DEGs identified at day 6 enriched for pathways involved in cell signaling, cell cycle regulation, and inositol metabolism, which were driven by the upregulated genes (S2 Data). The 1,071 downregulated genes enriched primarily for inhibition of cholesterol biosynthesis, which is in line with the suppression of fat metabolism initiated at day 3 (S2 Data). Because cell proliferation is affected at day 6 [4] and cholesterol has roles in membrane structures, inhibition of cholesterol biosynthesis—in addition to sparing acetyl-CoA—may be a response to loss of cell division. When mtDNA was fully depleted at day 9, the 805 upregulated genes enriched for serine and glycine metabolism, as was recently reported by others [8,9], and for methionine salvage through betaine. The use of betaine to recycle methionine, which is a folate-independent pathway, may reflect serine-associated folate being channeled to purine metabolism [9]. We also identified changes in tRNA charging, which suggests an attempt by the cell to preserve cytosolic protein synthesis (S2 Data). Inhibition of cholesterol biosynthesis, as found at day 6, was the top category identified with the 467 DEGs that were downregulated (S2 Data). We found at day 9 that the degradation of several proteinogenic amino acids was inhibited and that the TCA cycle was suppressed (S2 Data); this is consistent with their utilization for protein synthesis rather than, for instance, supplying precursors for the TCA. IPA can also predict upstream regulators involved in driving the transcriptional programs identified. Activating transcription factor 4 (ATF4) is a transcription factor recently linked to a mitochondrial stress response [13] and was predicted only when evaluating the genes that were upregulated under our experimental conditions, irrespective of the degree of mitochondrial dysfunction (S2B Fig). Conversely, several upstream regulators were predicted to be associated with the downregulated genes, including tumor protein p53 (TP53), MYC, and peroxisome proliferator activated receptor alpha (PPARα). The only gene consistently predicted to play a role in the inhibitory responses at all times was the major facilitator superfamily domain-containing protein 2a (MFSD2A) (S2B Fig), which has been recently linked to fatty acid oxidation [14]. Changes in purine nucleotides, the methionine cycle, and the TCA drive the early metabolic response to mtDNA depletion We previously performed a metabolomics analysis in DN-POLG cells at days 0, 3, 6, and 9 and showed that many metabolites were changed during mtDNA depletion [4]. To gain more insights into the progressive remodeling of the metabolome as a function of mtDNA depletion, and to explore the relationship with the transcriptome changes, we used the metabolite data to identify the pathways that were enriched over time. We started by determining those metabolites that were statistically different at any given point relative to day 0, using adjusted p ≤ 0.05 and an effect size of 1.15-fold (for more information, see Methods). We found a total of 459 metabolites using these statistical criteria, of which 231 were significantly different at day 3, 396 at day 6, and 345 at day 9 (S3A Fig and S3 Data); common to all time points were 179 metabolites (S3A Fig and S3 Data). We then performed pathway enrichment analysis using the 459 metabolites, which revealed the dynamic nature of the metabolic changes over time. For example, most pathways progressively enriched between days 3–9, while some initiated at day 6, and others decreased by day 9 (S3B Fig). The top enriched pathways involved purine nucleotides and the superpathway of methionine degradation, which was also the most significantly enriched pathway across the experimental time course (S3B Fig). The fact that methionine degradation was captured at the transcriptional level already at day 3 (S1 and S2 Data) and was also the highest significant metabolic pathway engaged over time revealed an unexpected connection between methionine metabolism and loss of mtDNA. This finding was consistent with the overall amino acid response identified from the transcriptome data. The metabolite analysis showed the engagement of both catabolic and biosynthetic amino acid pathways; a summary of the main pathways is schematically represented in Fig 1D. Many of enriched pathways for amino acid degradation involved those that can input into the TCA to make acetyl-CoA, like leucine, valine and lysine, or other intermediates such as malate, succinyl-CoA, or α-ketoglutarate (Figs 1D and S3B). Biosynthesis of other amino acids—such as serine, cysteine, and glutamate—was also observed (Figs 1D and S3B). Consistent with amino acid degradation, the urea cycle that recycles ammonia derived from amino acid catabolism was enriched; linked to it was the biosynthesis of citrulline (S3B Fig). The degradation of putrescine, which can input into the TCA as succinate, was also identified (S3B Fig); this was in line with the transcriptome data (S2 Data). It is noteworthy that the urea cycle provides ornithine, the precursor of putrescine, thus offering a constant supply of these metabolites in the DN-POLG cells. The urea cycle, while mostly connected with the liver, occurs partially in the kidneys [15]. Several (although not all) genes involved in this pathway are expressed in different tissues [16,17]. The identification of the urea cycle as enriched in HEK293 cells is likely a reflection of activation of components of the pathway to recycle ammonia, rather than the canonical liver urea cycle, under our experimental conditions. Also, increased degradation of choline—the precursor of betaine—and glycine/betaine metabolism were enriched (S2 Data), which is in agreement with the RNA-seq findings that suggested that methionine levels were maintained through salvage pathways. Various examples of the relationship between the transcriptome and metabolic remodeling can be found in S3C–S3G Fig. We assumed that the changes found at day 3 would reveal the drivers of the global metabolic response to mtDNA depletion. To define those drivers, we ranked the relevance of the pathways based on the ones most significantly enriched at day 3, focusing arbitrarily only on the ones with a p ≤ 10−7. What we found were 3 main nodes that essentially centered around purines, the TCA, and redox reactions (S3H Fig). The levels of some metabolites involved in these pathways are shown in Fig 1E. These data suggest, despite the lack of detectable changes in mitochondrial function, that the level of mtDNA depletion achieved at day 3 remodels metabolism in a way that prepares the cells to adjust nucleic acid metabolism (transcription, DNA repair, and replication), cell cycle, protein translation, methylation reactions, and redox homeostasis. This analysis also revealed 6 pathways that were not significantly enriched at day 3 but that were identified at later time points (S3B Fig). These pathways were associated with overt mitochondrial dysfunction and included pyrimidine ribonucleotide interconversion, biosynthesis of cysteine, glutathione, glutamine, and the polyamines spermidine and spermine (S3B Fig). It was surprising that the biosynthesis of cysteine and glutathione was not engaged at day 3, since mtDNA depletion was recently shown to induce serine biosynthesis (also shown here, at day 3 p = 10−2; S3B Fig and Fig 1F), channeling 1C metabolism to cysteine and glutathione production through transsulfuration [8,9]. It is worth noting that serine is also involved in the formation of formyl-methionine by feeding into the mitochondrial folate cycle [1]. Formyl-methionine is the unique amino acid used to initiate translation of mtDNA-encoded proteins [18]. Despite significant loss of mtDNA at day 3 (S4A Fig), levels of mtRNA transcripts were stable (S4B Fig), and mtDNA-encoded proteins were not significantly affected [4]. Thus, we hypothesized that serine biosynthesis at day 3 serves to maintain mitochondrial protein translation and sustain organellar function; at later time points, it likely supports cysteine and glutathione production, as shown by others [8,9]. In agreement with this hypothesis, levels of formyl-methionine were higher at day 3 compared to days 6 or 9 (Fig 1G), whereas that of cysteine followed the opposite trend (Fig 1H). The reason why the serine biosynthetic pathway is activated upon mtDNA depletion remains unclear. Loss of mtDNA leads to DNA hypermethylation through increased SAM Carbon units derived from folate-1C metabolism are used for the synthesis of purines and the generation of S-adenosyl-methionine (SAM), which is considered the universal methyl donor for DNA, RNA, lipids, and proteins [19,20]. Levels of SAM are also influenced by polyamine synthesis, which uses decarboxylated SAM for the production of spermidine and spermine from putrescine, generating 5-methyl-thioadenosine (MTA). MTA is recycled back into the methionine cycle through a salvage pathway that also produces adenine, thus feeding into the purine pool [21]. Interestingly, MTA has been shown to be the major source of de novo adenine in human cells [22]. Our transcriptomic and metabolic data suggest that the progressive mtDNA depletion achieved over 9 days significantly affects the methionine cycle in various ways, including (i) by channeling homocysteine to transsulforation, (ii) by increasing the utilization of betaine as a folate-independent methionine precursor, (iii) by promoting the degradation of methionine, and (iv) by altering polyamine synthesis and degradation that, in turn, affects MTA recycling (Fig 2A). However, whether these changes impact the levels of SAM, influencing methylation reactions, remains unknown. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Metabolic response to mitochondrial dysfunction also affects DNA methylation patterns in DN-POLG cells. Metabolomics was performed in DN-POLG cells at days 0, 3, 6 and 9; N = 4 per time point. Differentially enriched metabolites were identified based on log2-transformed fold-changes in arbitrary detection units versus the mean at day 0 (y-axis) in each time point during doxycycline treatment (x-axis) by a two-way ANOVA test (metabolite × time) at an adjusted Benjamini-Hochberg p ≤ 0.05. (A) Schematic representation of interconnected polyamine synthesis, purine metabolism and methionine salvage pathways associated to differentially enriched metabolites at days 3, 6 and 9 of dox-inducible mtDNA depletion in DN-POLG cells, per Ingenuity Pathway Analysis. (B, E) Box plots of relative content versus day 0 (IQR-outlier format) depict distribution of individual replicates in DN-POLG cells within statistical groups for (B) Hom, Gly, Bet, MTA, SAM, SAH, Spd, Spm, and (E) 2-HG. (C) Unsupervised clustering analysis of Δ%mCG for DML observed in DN-POLG cells by HM-450K BeadArrays; hypermethylation (red) or hypomethylation (green). N = 3 per time point. DML were identified based methylation beta-values versus the mean at day 0 (Log2FC) by a two-way ANOVA test (probe × time) at an adjusted Benjamini-Hochberg p ≤ 0.05. (D) Average content ratios for Succ and Fum to αKG at days 3, 6, and 9 in DN-POLG cells normalized to the mean at day 0, based on metabolomics output (N = 4 per time point). (F) DNMT activity was gauged in 143B rho0 and rho+ cells by following the transfer of radiolabeled methylated substrate onto poly-dIC oligonucleotide. N = 3 per cell model; data are presented relative to detected activity in rho+ cells (bar plot: mean ± SEM). Underlying data are reported in S3 Data for (B), (D), and (E); S4 Data for (C); and S9 Data for (F). 2-HG, 2-hydroxyglutarate; αKG, α-ketoglutarate; AMP, adenosine monophosphate; Bet, betaine; BHMT, betaine homocysteine-methyltransferase; Δ%mCG, DNA methylation changes versus average of day 0; DML, differentially methylated genomic loci; DNMT, DNA methyltransferase; DN-POLG, dominant-negative DNA polymerase gamma transgene; Fum, fumarate; Gly, glycine; Hom, homocysteine; IMP, inosine 5'-monophosphate; MTA, 5-methylthioadenosine; mtDNA, mitochondrial DNA; poly-dIC, poly DNA inosinic-polycytidylic acid; SAH, S-adenosylhomocysteine; SAM, S-adenosylmethionine; Spd, spermidine; Spm, spermine; Succ, succinate. https://doi.org/10.1371/journal.pbio.2005707.g002 We examined the metabolites associated with the methionine cycle (Fig 2A) and found that while homocysteine levels decreased over time (Fig 2B), levels of methionine (Fig 1E), serine (Fig 1F), glycine (Fig 2B), and cysteine (Fig 1H) increased. Choline (Fig 2B), betaine, SAM, and MTA levels were maximal at day 6, returning to levels closer to basal at day 9 (Fig 2B). Levels of S-adenosyl-homocysteine (SAH), the byproduct of SAM metabolism, increased at day 6 and decreased at day 9 below basal levels (Fig 2B). A high SAM/SAH ratio is favorable to methylation reactions since SAH inhibits the methyltransferases [10,23]. Steady state levels of the polyamines putrescine, spermidine, and spermine followed an interesting trend. While putrescine decreased by day 6 and increased by day 9 (Fig 1E), the levels of spermidine and spermine decreased over time (Fig 2B). This effect on the steady state levels of the polyamines is also reflective of an increased catabolism of spermine and spermidine through spermine/spermidine N-acetyl-transferase (SAT1), which is upregulated at the transcriptional level in the DN-POLG (S1 Data) and whose net product is putrescine [24]. Since previous observations that DNA methylation is influenced by mtDNA depletion and mitochondrial dysfunction in cultured cells and animal models [2,3], we hypothesized that the changes in SAM we observed could drive this effect. Specifically, we predicted that the DNA would be hypermethylated, with maximal levels at day 6. To test this hypothesis, we evaluated whole genome DNA methylation status at a single nucleotide resolution using the Illumina 450K platform. We found that mtDNA depletion progressively increased DNA methylation in promoters, gene bodies, or intergenic regions (S5A Fig), with hypermethylation peaking at day 6 and decreasing at day 9 (Fig 2C). Although the changes we detected were somewhat modest (full range of Δ%mCG: −30% to +40%; see Fig 2C) we reasoned they reflected the short time frame of the experiments. Indeed, when evaluating DNA methylation using the same approach in cells chronically depleted of mtDNA (rho0) in the 143B background, we found that methylation changes were more prominent, ranging between −60% and +60% with respect to cells with endogenous mtDNA levels (rho+) in the same 143B background (S5B Fig). The increased methylation of the DNA is consistent with the increased levels of SAM and with the kinetics of availability of SAM/SAH amounts over time. However, changes in other TCA metabolites could also play a role in this phenotype. For example, α-ketoglutarate is a cofactor of the Ten-eleven translocation (TET) enzymes, which are involved in the DNA demethylation reactions. Also, succinate, fumarate, and 2-hydroglutarate (2-HG) can compete with α-ketoglutarate in the active site of the TETs, inhibiting their function [25]. Thus, decreased α-ketoglutarate, increased succinate, fumarate, and/or 2-HG could also lead to hypermethylation of the DNA. However, no changes in the levels of α-ketoglutarate were observed (S3 Data), and no increases in the succinate or fumarate to α-ketoglutarate ratios were identified over the time course of the experiments (Fig 2D). Despite the fact that 2-HG increased as mtDNA was depleted, only a small change was observed at day 6, and maximal accumulation was observed at day 9 (Fig 2E), which is inconsistent with the kinetics of DNA hypermethylation (Fig 2C). Levels of methylated cytosines (5meC) were increased, while no changes in the levels of 5-hydroxy-methyl-cytosine (5hmeC)—the product of TET reaction—were identified in cells chronically depleted of mtDNA (S5C and S5D Fig). We also showed enhanced DNA methyltransferase (DNMT) activity (Fig 2F). Collectively, these data are in support of DNA hypermethylation resulting from increased DNA methylation and not from inhibition of the demethylases. Changes in DNA methylation occur prevalently in DEGs In order to determine whether the changes in global methylation influenced gene expression, we cross-referenced the coordinates of the promoters differentially methylated at days 3, 6, or 9 with those of the DEGs. We found that 1,627 (approximately 57%) of the DEGs showed significant alterations in their promoter methylation when compared to day 0 (S4 Data). The number of differentially methylated DEGs increased over time from 63 at day 3 (27%), 978 (46%) at day 6, and 879 (70%) at day 9 (S5E Fig). The odds ratio (OR) of a gene being differentially expressed and having a change in its promoter methylation was OR = 0.81, p < 0.01 (S5F Fig), suggesting that incidence of promoter DNA methylation changes is different for DEGs and genes not differentially expressed. To better understand the relationship between differential methylation, gene expression, and mitochondrial dysfunction, we performed IPA on the DEGs that were differentially methylated. This analysis revealed that genes involved in key pathways that responded to mtDNA depletion were targets of differential methylation. For instance, at day 3, the 63 differentially methylated and expressed genes enriched for methionine degradation; at day 6, for cholesterol biosynthesis; and at day 9, for the metabolism of several amino acids, cholesterol, and the TCA (S5 Data). Similar findings were observed when evaluating the 143B rho0 cells chronically depleted of mtDNA that also showed hypermethylation of the DNA. In those cells, 621 DEGs were also differentially methylated (S6 Data) and enriched for pathways involved, for instance, in folate transformations (S5G Fig). While DNA methylation is not the only parameter governing gene expression, we attempted to define the level of concordance between the changes in DNA methylation status over time with the directionality of expression of the DEGs harboring those changes. Whether we combined the entire methylation profile of genes or considered only promoter marks, the concordance ranged from 30%–50% over the 9 days of mtDNA depletion (S5H Fig). Taken together, these findings suggest a correlation between DNA methylation changes and the expression of a fraction of DEGs responding to progressive mitochondrial dysfunction. NADH oxidation in the mitochondria links polyamine and methionine metabolism to the TCA cycle and DNA methylation It is possible that the mechanism connecting mtDNA depletion to SAM and DNA hypermethylation involves serine biosynthesis and 1C-folate remodeling, which in turn can affect the methionine cycle [8,9]. While this is feasible, the fact that choline/betaine are engaged in maintaining methionine salvage independent of folate would argue against this possibility. Alternatively, the methionine cycle may be directly affected by mtDNA depletion through changes in both methionine and polyamine metabolism. These molecules are not only linked in the regulation of SAM levels [10], but they can provide intermediates such as pyruvate, succinyl-CoA (a precursor of succinate), and succinate to the TCA in their catabolic pathways. An increase in their degradation to feed the TCA could set a cascade of compensatory changes that impacts the SAM pool. To test this hypothesis, we took advantage of the DN-POLG cells overexpressing NDI1/AOX, which are cells that have the ability to oxidize NADH and maintain TCA flux despite the complete loss of mtDNA [4]. We reasoned that if the methionine cycle is directly impacted by the TCA, in these cells methionine-associated intermediates should not be changed. We reanalyzed the metabolomics data that we previously generated with the NDI1/AOX cells [4] using the same criteria as for the DN-POLG cells (S3 Data). We then focused on the intermediates associated with the methionine, serine, folate, and polyamine pathways. We found that in the NDI1/AOX cells, the levels of SAM, SAH, MTA, and the polyamines were maintained over time (Fig 3A); the levels of serine, cysteine, methionine, betaine, choline, and folate followed the same pattern as was observed with the DN-POLG cells (compare Figs 3A and 1E–1G). Most notably, levels of succinate, which were increased at day 9 in the DN-POLG, were decreased in the NDI1/AOX cells (Fig 3B). Taken together, these results support the hypothesis that polyamine and methionine metabolism are directly responding to changes in TCA flux, likely as contributors of succinate. Furthermore, these data suggest that serine biosynthesis and folate-1C remodeling caused by mtDNA depletion are not responding to changes in NADH oxidation or TCA flux. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. NDI1/AOX ectopic expression maintains DNA methylation while not completely rescuing metabolism. Metabolomics was performed in NDI1/AOX cells at days 0, 3, 6 and 9; N = 4 per time point. Differentially enriched metabolites were identified based on log2-transformed fold-changes in arbitrary detection units versus the mean at day 0 (y-axis) in each time point during doxycycline treatment (x-axis) by a two-way ANOVA test. (A) Box plots of relative content versus day 0 (IQR-outlier format) depict distribution of individual replicates in NDI1/AOX cells within statistical groups for SAM, SAH, MTA, Put, Spd, Spm, Ser, Cys, Met, Bet, Cho, and 5-MeTHF. (B) Box plots of relative content versus day 0 (IQR-outlier format) depict distribution of individual replicates within statistical groups for Succ in DN-POLG cells (left) and their NDI1/AOX counterparts (right). (C) Δ%mCG in NDI1/AOX cells at days 0 and 9 for the same probes identified as DML in DN-POLG cells by HM-450K BeadArrays [see Fig 2C]; hypermethylation (red) or hypomethylation (green). N = 3 per time point. (D) Average normalized read counts (bar plot: mean ± SEM) of mtDNA fragments obtained by next-generation sequencing of whole-cell DNA for NDI1/AOX cells; N = 2 per timepoint. (E) Box plots (IQR-outlier format) of average DNA methylation differences between day 0 and day 9 at independently identified DML in NDI1/AOX cells overlapping the genomic range (overall), only promoters, or only bodies of DN-POLG DEGs; separate box plots depict measurements from DN-POLG DEGs that are DMG or not in DN-POLG cells also. Underlying data are reported in S3 Data for (A) and (B); S4 Data for (C) and (E); and S1 Data for (D). (F) Model for the cross talk between methionine salvage, polyamine synthesis and the TCA cycle with DNA methylation: loss of mtDNA decreases TCA flux, which in turn sets a cascade of transcriptional and metabolic changes centered largely on amino acid degradation to maintain TCA cycle output. Degradation of methionine and Put, both of which can feed into the TCA cycle, are among the first changes detected. Put levels are regulated by ornithine, which is provided by the recycling of NH3 resulting from amino acid degradation. Put is also a precursor of Spd and Spm, both of which require dSAM for their synthesis. The main byproduct is MTA, which needs to be quickly recycled, given its accumulation is toxic; the salvage of MTA recycles methionine, a cycle that is also maintained by folate-independent Cho/Bet when mtDNA is depleted. Recycling of MTA also generates adenine, which can enter the purine pool, and α-ketoglutarate that can feed the TCA. By maintaining NADH oxidation in the mitochondria, flux through the TCA cycle is largely normalized, “turning off” the polyamine/MTA salvage response that in turn decreases levels of SAM. Decrease in degradation of amino acids to feed the TCA diminishes flux through the urea cycle, decreasing the input of ornithine to Put biosynthesis. 5-MeTHF, 5-methyltetrahydrofolate; acetyl-CoA, acetyl coenzyme A; AOX, alternative oxidase; Bet, betaine; Cho, choline; Cys, cysteine; Δ%mCG, DNA methylation differences versus average of day 0; DEG, differentially expressed gene; DMG, differentially methylated gene; DML, differentially methylated loci; DN-POLG, dominant-negative DNA polymerase gamma transgene; dSAM, decarboxylated S-adenosylmethionine; ETC; electron transport chain; Met, methionine; MTA, 5-methylthioadenosine; NADH, nicotinamide adenine dinucleotide reduced; NDI1, nicotinamide adenine dinucleotide reduced dehydrogenase-like 1; OXPHOS, oxidative phosphorylation; Put, putrescine; RPM, reads per million reads; SAH, S-adenosylhomocysteine; SAM, S-adenosylmethionine; Ser, serine; Spd, spermidine; Spm, spermine; Succ, succinate; TCA, tricarboxylic acid. https://doi.org/10.1371/journal.pbio.2005707.g003 We also evaluated whole genome methylation using the Illumina 450K platform in NDI1/AOX cells. We used cells at days 0 and 9, since we gauged that mtDNA would be fully depleted at this latter time and would provide the largest effect. Remarkably, no significant changes in DNA methylation were observed in the cells expressing NDI1/AOX, despite complete loss of mtDNA (Fig 3C and 3D). We then focused specifically on the coordinates of the 1,626 DEGs that were differentially methylated in the DN-POLG cells at day 9 (S5A Fig). However, we found that average DNA methylation change in those sites was only approximately 2% in the NDI1/AOX cells (Fig 3E). Hence, we conclude that changes in polyamine synthesis and the MTA salvage pathway, which in turn affect SAM levels, seem to be critical for differential DNA methylation in the nucleus of DN-POLG cells. Lack of changes in DNA methylation are associated with the prevention of differential gene expression, even in the context of complete mtDNA loss We performed gene expression analysis in the NDI1/AOX cells using microarrays in order to determine whether the promoter methylation status has the potential to impact the differential expression of the 879 genes identified in the DN-POLG cells at day 9. Unexpectedly, we found no significant DEGs in the NDI1/AOX cells between days 0 and 9 when adjusting for false discovery rate (FDR; S7 Data). Relaxing statistical thresholds based on pairwise comparisons without multiple testing corrections revealed 23 genes that were differentially expressed between days 0 and 9 (S6 Data), 4 of which were also differentially expressed in the DN-POLG at day 9, as gauged by RNA-seq. To rule out that these results were due to a lack of sensitivity of microarrays to detect the relatively small changes in gene expression identified by RNA-seq, we performed microarrays in DN-POLG cells at days 0 and 9. We found 1,408 genes with adjusted p ≤ 0.05 that were differentially expressed between days 0 and 9 in this cellular background (S7 Data). These results indicate that it is the maintenance of NADH oxidation in the mitochondria, in the context of mtDNA depletion, that prevents the differential expression of genes. To better understand the effects of NDI1/AOX expression in the presence of mtDNA, we next compared the microarray data from DN-POLG cells with those from NDI1/AOX cells at day 0. Again, no DEGs were detected when adjusting for FDR. Using unadjusted p-values, we found 842 genes that were differently expressed between the 2 cell types at day 0 (S8 Data). However, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis identified little to no overlap to the findings obtained when utilizing the DEGs identified in the DN-POLG when mtDNA was depleted (S8 Data). Thus, we conclude that expression of NDI1/AOX does not cause significant off-target effects. Nevertheless, the maintenance of NADH oxidation provided by these enzymes is sufficient to prevent the DNA methylation and transcriptomic changes that result from mtDNA depletion, independent of mitochondrial ATP production or the ΔΨm, which were not rescued by NDI1/AOX expression [4]. MtDNA depletion drives dynamic transcriptional changes centered on acetyl coenzyme A (acetyl-CoA) metabolism The progressive nature of mtDNA depletion in the DN-POLG system provided us with a unique opportunity to address fundamental questions about the nuclear response to progressive mitochondrial dysfunction. We took an integrative approach that involved transcriptomic, metabolomics, and epigenetic analyses at each time point (days 0, 3, 6, and 9) during the course of complete mtDNA depletion. Furthermore, we utilized the cells expressing NDI1/AOX to define the responses to complete mtDNA loss that were specifically linked to NADH oxidation in the mitochondria. A schematic representation of this integrative approach is shown in S1A Fig. The reproducibility of the results is shown in S1B–S1G Fig. We started by performing RNA sequencing (RNA-seq) at days 0, 3, 6, and 9 in the DN-POLG cells, and found 2,854 genes (S1 Data), including all mtDNA-encoded transcripts, whose expression was changed (adjusted p ≤ 0.05) at any given day compared to day 0 (Fig 1A). When stratifying by day, we found 236 differentially expressed genes (DEGs) at day 3, which were mostly (78%) upregulated, while we found 2,135 DEGs at day 6 (1,064 upregulated and 1,071 downregulated). At day 9 we found 1,272 DEGs, most of which (64%) were upregulated (Fig 1B). Common to all time points were 121 DEGs, including POLG that was upregulated at least 7-fold relative to day 0 (S1 Data). The identification of over 200 DEGs at day 3 was surprising given the lack of significant changes in mitochondrial function at this time [4]. Our model also revealed progressive upregulation of fibroblast growth factor 21 (FGF21) and growth differentiation factor 15 (GDF15) relative to day 0 (S1 Data). FGF21 and GDF15 are metabolic cytokines induced in patients and mouse models of mtDNA or protein translation defects, which have been proposed as biomarkers of mitochondrial dysfunction [1]. Validation of randomly selected genes by quantitative real-time PCR can be found in S2A Fig. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Differential regulation of transcriptional and metabolic profiles in DN-POLG cells in the course of inducible mtDNA depletion. RNA-seq was performed from DN-POLG cells at days 0, 3, 6 and 9 of doxycycline supplementation; N = 3 biological replicates per time point. Transcriptional profiling was based on DEGs identified using log2-transformed fold-changes in RPKM versus the mean at day 0 (Log2FC) in two-way ANOVA tests (gene × time) at an adjusted Benjamini-Hochberg p ≤ 0.05. (A) Unsupervised clustering analysis of log2-fold expression changes versus average of day 0 for DEGs observed in DN-POLG cells by RNA-seq; increased (red) or decreased (green) expression. (B) Number of DEGs identified in DN-POLG cells at days 3, 6 and 9 of doxycycline supplementation compared to day 0 (circle plots, top). (C) Schematic representation of the main transcriptional responses associated to DEGs in DN-POLG cells at days 3, 6 and 9 of dox-inducible mtDNA depletion in DN-POLG cells, per IPA. (D) Schematic representation of interconnected amino acid and methionine metabolic pathways associated to differentially enriched metabolites at days 3, 6, and 9 of dox-inducible mtDNA depletion in DN-POLG cells (N = 4 per time point), per IPA. (E–H) Box plots of relative content versus day 0 (IQR-outlier format) depict distribution of individual replicates in DN-POLG cells within statistical groups for (E) IMP, Xan, XMP, Cho, Met, Leu, Put, 4-HPPA, (F) Ser, (G) f-Met, and (H) Cys; partial gray backgrounds highlight metabolite relative contents at day 3 in (E). Underlying data are reported in S1 Data for (A) and (B); S2 Data for (C) and (D); and S3 Data for (E–H). 4-HPPA, 4-hydroxyphenylpyruvate; Cho, choline; Cys, cysteine; DEGS, differentially enriched genes; DN-POLG, dominant-negative DNA polymerase gamma transgene; ETC, electron transport chain; f-Met, N-formylmethionine; GABA, γ-aminobutyric acid; IMP, inosine 5'-monophosphate; IPA, Ingenuity Pathway Analysis; Leu, leucine; Met, methionine; mtDNA, mitochondrial DNA; OXPHOS, oxidative phosphorylation; Put, putrescine; RNA-seq, RNA sequencing; ROS, reactive oxygen species; RPKM, reads per kilobase per million; Ser, serine; TCA, tricarboxylic acid; Xan, xanthine; XMP, xanthosine 5'-monophosphate. https://doi.org/10.1371/journal.pbio.2005707.g001 The use of Ingenuity Pathway Analysis (IPA) revealed that, globally, the genes modulated during the course of mtDNA depletion enriched for tRNA charging, cholesterol biosynthesis, glutamate (via 4-aminobutyrate) and putrescine degradation, and D-myo-inositol-5-phosphate metabolism (top 10, S2 Data). These results are consistent with the loss of mtDNA having a negative impact on oxidative TCA flux, since glutamate degradation can contribute to the cycle as precursor for succinate. Likewise, putrescine can contribute to the TCA as succinate via γ-aminobutyric acid (GABA) catabolism [5], known to occur in cells beyond the central nervous system [6]. To our knowledge, this is the first report to identify polyamine metabolism as responsive to loss of mitochondrial function. Inositols are sensitive to glucose and NADH levels [7], the metabolism of which is also impacted by loss of mtDNA. To understand how DN-POLG cells respond to progressive loss of mtDNA, we next stratified the analysis based on time point and the directionality of gene changes. We reasoned that the genes modulated at day 3 would reflect early responses to mtDNA depletion that take place prior to signs of mitochondrial dysfunction [4]. Those DEGs at day 6 would reveal responses to significant loss of ETC function and the pathways directly linked to this process—e.g., ATP production and the oxidative TCA—while the ones detected at day 9 would reveal adaptive changes within the cell. A summary of the main transcriptional responses identified at each time point is shown in Fig 1C. Most genes detected at day 3 (184) were upregulated and enriched for methionine and cysteine degradation (S2 Data). Alterations in the methionine cycle have not been directly associated with mitochondrial dysfunction, although channeling of 1C units toward transsulfuration of homocysteine to cysteine, a branch point in the methionine cycle, has been recently reported [8,9]. However, the degradation of methionine can ultimately input into the TCA cycle by contributing pyruvate (through cysteine) and succinyl-CoA through aminobutanoate [10–12]; this would provide a link between mtDNA depletion and methionine degradation. Nevertheless, it was surprising to identify these changes at day 3 when no significant alterations in mitochondrial function were identified [4]. Concomitant to the degradation of methionine, we observed DEGs involved in the recycling (or salvage) of this amino acid through betaine, which was likely an attempt to maintain methionine levels (S1 and S2 Data). The 52 genes downregulated at day 3 were enriched for lipid metabolism, presumably to spare acetyl-CoA, and endothelial nitric oxide synthase (eNOS) signaling (S2 Data). The 2,135 DEGs identified at day 6 enriched for pathways involved in cell signaling, cell cycle regulation, and inositol metabolism, which were driven by the upregulated genes (S2 Data). The 1,071 downregulated genes enriched primarily for inhibition of cholesterol biosynthesis, which is in line with the suppression of fat metabolism initiated at day 3 (S2 Data). Because cell proliferation is affected at day 6 [4] and cholesterol has roles in membrane structures, inhibition of cholesterol biosynthesis—in addition to sparing acetyl-CoA—may be a response to loss of cell division. When mtDNA was fully depleted at day 9, the 805 upregulated genes enriched for serine and glycine metabolism, as was recently reported by others [8,9], and for methionine salvage through betaine. The use of betaine to recycle methionine, which is a folate-independent pathway, may reflect serine-associated folate being channeled to purine metabolism [9]. We also identified changes in tRNA charging, which suggests an attempt by the cell to preserve cytosolic protein synthesis (S2 Data). Inhibition of cholesterol biosynthesis, as found at day 6, was the top category identified with the 467 DEGs that were downregulated (S2 Data). We found at day 9 that the degradation of several proteinogenic amino acids was inhibited and that the TCA cycle was suppressed (S2 Data); this is consistent with their utilization for protein synthesis rather than, for instance, supplying precursors for the TCA. IPA can also predict upstream regulators involved in driving the transcriptional programs identified. Activating transcription factor 4 (ATF4) is a transcription factor recently linked to a mitochondrial stress response [13] and was predicted only when evaluating the genes that were upregulated under our experimental conditions, irrespective of the degree of mitochondrial dysfunction (S2B Fig). Conversely, several upstream regulators were predicted to be associated with the downregulated genes, including tumor protein p53 (TP53), MYC, and peroxisome proliferator activated receptor alpha (PPARα). The only gene consistently predicted to play a role in the inhibitory responses at all times was the major facilitator superfamily domain-containing protein 2a (MFSD2A) (S2B Fig), which has been recently linked to fatty acid oxidation [14]. Changes in purine nucleotides, the methionine cycle, and the TCA drive the early metabolic response to mtDNA depletion We previously performed a metabolomics analysis in DN-POLG cells at days 0, 3, 6, and 9 and showed that many metabolites were changed during mtDNA depletion [4]. To gain more insights into the progressive remodeling of the metabolome as a function of mtDNA depletion, and to explore the relationship with the transcriptome changes, we used the metabolite data to identify the pathways that were enriched over time. We started by determining those metabolites that were statistically different at any given point relative to day 0, using adjusted p ≤ 0.05 and an effect size of 1.15-fold (for more information, see Methods). We found a total of 459 metabolites using these statistical criteria, of which 231 were significantly different at day 3, 396 at day 6, and 345 at day 9 (S3A Fig and S3 Data); common to all time points were 179 metabolites (S3A Fig and S3 Data). We then performed pathway enrichment analysis using the 459 metabolites, which revealed the dynamic nature of the metabolic changes over time. For example, most pathways progressively enriched between days 3–9, while some initiated at day 6, and others decreased by day 9 (S3B Fig). The top enriched pathways involved purine nucleotides and the superpathway of methionine degradation, which was also the most significantly enriched pathway across the experimental time course (S3B Fig). The fact that methionine degradation was captured at the transcriptional level already at day 3 (S1 and S2 Data) and was also the highest significant metabolic pathway engaged over time revealed an unexpected connection between methionine metabolism and loss of mtDNA. This finding was consistent with the overall amino acid response identified from the transcriptome data. The metabolite analysis showed the engagement of both catabolic and biosynthetic amino acid pathways; a summary of the main pathways is schematically represented in Fig 1D. Many of enriched pathways for amino acid degradation involved those that can input into the TCA to make acetyl-CoA, like leucine, valine and lysine, or other intermediates such as malate, succinyl-CoA, or α-ketoglutarate (Figs 1D and S3B). Biosynthesis of other amino acids—such as serine, cysteine, and glutamate—was also observed (Figs 1D and S3B). Consistent with amino acid degradation, the urea cycle that recycles ammonia derived from amino acid catabolism was enriched; linked to it was the biosynthesis of citrulline (S3B Fig). The degradation of putrescine, which can input into the TCA as succinate, was also identified (S3B Fig); this was in line with the transcriptome data (S2 Data). It is noteworthy that the urea cycle provides ornithine, the precursor of putrescine, thus offering a constant supply of these metabolites in the DN-POLG cells. The urea cycle, while mostly connected with the liver, occurs partially in the kidneys [15]. Several (although not all) genes involved in this pathway are expressed in different tissues [16,17]. The identification of the urea cycle as enriched in HEK293 cells is likely a reflection of activation of components of the pathway to recycle ammonia, rather than the canonical liver urea cycle, under our experimental conditions. Also, increased degradation of choline—the precursor of betaine—and glycine/betaine metabolism were enriched (S2 Data), which is in agreement with the RNA-seq findings that suggested that methionine levels were maintained through salvage pathways. Various examples of the relationship between the transcriptome and metabolic remodeling can be found in S3C–S3G Fig. We assumed that the changes found at day 3 would reveal the drivers of the global metabolic response to mtDNA depletion. To define those drivers, we ranked the relevance of the pathways based on the ones most significantly enriched at day 3, focusing arbitrarily only on the ones with a p ≤ 10−7. What we found were 3 main nodes that essentially centered around purines, the TCA, and redox reactions (S3H Fig). The levels of some metabolites involved in these pathways are shown in Fig 1E. These data suggest, despite the lack of detectable changes in mitochondrial function, that the level of mtDNA depletion achieved at day 3 remodels metabolism in a way that prepares the cells to adjust nucleic acid metabolism (transcription, DNA repair, and replication), cell cycle, protein translation, methylation reactions, and redox homeostasis. This analysis also revealed 6 pathways that were not significantly enriched at day 3 but that were identified at later time points (S3B Fig). These pathways were associated with overt mitochondrial dysfunction and included pyrimidine ribonucleotide interconversion, biosynthesis of cysteine, glutathione, glutamine, and the polyamines spermidine and spermine (S3B Fig). It was surprising that the biosynthesis of cysteine and glutathione was not engaged at day 3, since mtDNA depletion was recently shown to induce serine biosynthesis (also shown here, at day 3 p = 10−2; S3B Fig and Fig 1F), channeling 1C metabolism to cysteine and glutathione production through transsulfuration [8,9]. It is worth noting that serine is also involved in the formation of formyl-methionine by feeding into the mitochondrial folate cycle [1]. Formyl-methionine is the unique amino acid used to initiate translation of mtDNA-encoded proteins [18]. Despite significant loss of mtDNA at day 3 (S4A Fig), levels of mtRNA transcripts were stable (S4B Fig), and mtDNA-encoded proteins were not significantly affected [4]. Thus, we hypothesized that serine biosynthesis at day 3 serves to maintain mitochondrial protein translation and sustain organellar function; at later time points, it likely supports cysteine and glutathione production, as shown by others [8,9]. In agreement with this hypothesis, levels of formyl-methionine were higher at day 3 compared to days 6 or 9 (Fig 1G), whereas that of cysteine followed the opposite trend (Fig 1H). The reason why the serine biosynthetic pathway is activated upon mtDNA depletion remains unclear. Loss of mtDNA leads to DNA hypermethylation through increased SAM Carbon units derived from folate-1C metabolism are used for the synthesis of purines and the generation of S-adenosyl-methionine (SAM), which is considered the universal methyl donor for DNA, RNA, lipids, and proteins [19,20]. Levels of SAM are also influenced by polyamine synthesis, which uses decarboxylated SAM for the production of spermidine and spermine from putrescine, generating 5-methyl-thioadenosine (MTA). MTA is recycled back into the methionine cycle through a salvage pathway that also produces adenine, thus feeding into the purine pool [21]. Interestingly, MTA has been shown to be the major source of de novo adenine in human cells [22]. Our transcriptomic and metabolic data suggest that the progressive mtDNA depletion achieved over 9 days significantly affects the methionine cycle in various ways, including (i) by channeling homocysteine to transsulforation, (ii) by increasing the utilization of betaine as a folate-independent methionine precursor, (iii) by promoting the degradation of methionine, and (iv) by altering polyamine synthesis and degradation that, in turn, affects MTA recycling (Fig 2A). However, whether these changes impact the levels of SAM, influencing methylation reactions, remains unknown. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Metabolic response to mitochondrial dysfunction also affects DNA methylation patterns in DN-POLG cells. Metabolomics was performed in DN-POLG cells at days 0, 3, 6 and 9; N = 4 per time point. Differentially enriched metabolites were identified based on log2-transformed fold-changes in arbitrary detection units versus the mean at day 0 (y-axis) in each time point during doxycycline treatment (x-axis) by a two-way ANOVA test (metabolite × time) at an adjusted Benjamini-Hochberg p ≤ 0.05. (A) Schematic representation of interconnected polyamine synthesis, purine metabolism and methionine salvage pathways associated to differentially enriched metabolites at days 3, 6 and 9 of dox-inducible mtDNA depletion in DN-POLG cells, per Ingenuity Pathway Analysis. (B, E) Box plots of relative content versus day 0 (IQR-outlier format) depict distribution of individual replicates in DN-POLG cells within statistical groups for (B) Hom, Gly, Bet, MTA, SAM, SAH, Spd, Spm, and (E) 2-HG. (C) Unsupervised clustering analysis of Δ%mCG for DML observed in DN-POLG cells by HM-450K BeadArrays; hypermethylation (red) or hypomethylation (green). N = 3 per time point. DML were identified based methylation beta-values versus the mean at day 0 (Log2FC) by a two-way ANOVA test (probe × time) at an adjusted Benjamini-Hochberg p ≤ 0.05. (D) Average content ratios for Succ and Fum to αKG at days 3, 6, and 9 in DN-POLG cells normalized to the mean at day 0, based on metabolomics output (N = 4 per time point). (F) DNMT activity was gauged in 143B rho0 and rho+ cells by following the transfer of radiolabeled methylated substrate onto poly-dIC oligonucleotide. N = 3 per cell model; data are presented relative to detected activity in rho+ cells (bar plot: mean ± SEM). Underlying data are reported in S3 Data for (B), (D), and (E); S4 Data for (C); and S9 Data for (F). 2-HG, 2-hydroxyglutarate; αKG, α-ketoglutarate; AMP, adenosine monophosphate; Bet, betaine; BHMT, betaine homocysteine-methyltransferase; Δ%mCG, DNA methylation changes versus average of day 0; DML, differentially methylated genomic loci; DNMT, DNA methyltransferase; DN-POLG, dominant-negative DNA polymerase gamma transgene; Fum, fumarate; Gly, glycine; Hom, homocysteine; IMP, inosine 5'-monophosphate; MTA, 5-methylthioadenosine; mtDNA, mitochondrial DNA; poly-dIC, poly DNA inosinic-polycytidylic acid; SAH, S-adenosylhomocysteine; SAM, S-adenosylmethionine; Spd, spermidine; Spm, spermine; Succ, succinate. https://doi.org/10.1371/journal.pbio.2005707.g002 We examined the metabolites associated with the methionine cycle (Fig 2A) and found that while homocysteine levels decreased over time (Fig 2B), levels of methionine (Fig 1E), serine (Fig 1F), glycine (Fig 2B), and cysteine (Fig 1H) increased. Choline (Fig 2B), betaine, SAM, and MTA levels were maximal at day 6, returning to levels closer to basal at day 9 (Fig 2B). Levels of S-adenosyl-homocysteine (SAH), the byproduct of SAM metabolism, increased at day 6 and decreased at day 9 below basal levels (Fig 2B). A high SAM/SAH ratio is favorable to methylation reactions since SAH inhibits the methyltransferases [10,23]. Steady state levels of the polyamines putrescine, spermidine, and spermine followed an interesting trend. While putrescine decreased by day 6 and increased by day 9 (Fig 1E), the levels of spermidine and spermine decreased over time (Fig 2B). This effect on the steady state levels of the polyamines is also reflective of an increased catabolism of spermine and spermidine through spermine/spermidine N-acetyl-transferase (SAT1), which is upregulated at the transcriptional level in the DN-POLG (S1 Data) and whose net product is putrescine [24]. Since previous observations that DNA methylation is influenced by mtDNA depletion and mitochondrial dysfunction in cultured cells and animal models [2,3], we hypothesized that the changes in SAM we observed could drive this effect. Specifically, we predicted that the DNA would be hypermethylated, with maximal levels at day 6. To test this hypothesis, we evaluated whole genome DNA methylation status at a single nucleotide resolution using the Illumina 450K platform. We found that mtDNA depletion progressively increased DNA methylation in promoters, gene bodies, or intergenic regions (S5A Fig), with hypermethylation peaking at day 6 and decreasing at day 9 (Fig 2C). Although the changes we detected were somewhat modest (full range of Δ%mCG: −30% to +40%; see Fig 2C) we reasoned they reflected the short time frame of the experiments. Indeed, when evaluating DNA methylation using the same approach in cells chronically depleted of mtDNA (rho0) in the 143B background, we found that methylation changes were more prominent, ranging between −60% and +60% with respect to cells with endogenous mtDNA levels (rho+) in the same 143B background (S5B Fig). The increased methylation of the DNA is consistent with the increased levels of SAM and with the kinetics of availability of SAM/SAH amounts over time. However, changes in other TCA metabolites could also play a role in this phenotype. For example, α-ketoglutarate is a cofactor of the Ten-eleven translocation (TET) enzymes, which are involved in the DNA demethylation reactions. Also, succinate, fumarate, and 2-hydroglutarate (2-HG) can compete with α-ketoglutarate in the active site of the TETs, inhibiting their function [25]. Thus, decreased α-ketoglutarate, increased succinate, fumarate, and/or 2-HG could also lead to hypermethylation of the DNA. However, no changes in the levels of α-ketoglutarate were observed (S3 Data), and no increases in the succinate or fumarate to α-ketoglutarate ratios were identified over the time course of the experiments (Fig 2D). Despite the fact that 2-HG increased as mtDNA was depleted, only a small change was observed at day 6, and maximal accumulation was observed at day 9 (Fig 2E), which is inconsistent with the kinetics of DNA hypermethylation (Fig 2C). Levels of methylated cytosines (5meC) were increased, while no changes in the levels of 5-hydroxy-methyl-cytosine (5hmeC)—the product of TET reaction—were identified in cells chronically depleted of mtDNA (S5C and S5D Fig). We also showed enhanced DNA methyltransferase (DNMT) activity (Fig 2F). Collectively, these data are in support of DNA hypermethylation resulting from increased DNA methylation and not from inhibition of the demethylases. Changes in DNA methylation occur prevalently in DEGs In order to determine whether the changes in global methylation influenced gene expression, we cross-referenced the coordinates of the promoters differentially methylated at days 3, 6, or 9 with those of the DEGs. We found that 1,627 (approximately 57%) of the DEGs showed significant alterations in their promoter methylation when compared to day 0 (S4 Data). The number of differentially methylated DEGs increased over time from 63 at day 3 (27%), 978 (46%) at day 6, and 879 (70%) at day 9 (S5E Fig). The odds ratio (OR) of a gene being differentially expressed and having a change in its promoter methylation was OR = 0.81, p < 0.01 (S5F Fig), suggesting that incidence of promoter DNA methylation changes is different for DEGs and genes not differentially expressed. To better understand the relationship between differential methylation, gene expression, and mitochondrial dysfunction, we performed IPA on the DEGs that were differentially methylated. This analysis revealed that genes involved in key pathways that responded to mtDNA depletion were targets of differential methylation. For instance, at day 3, the 63 differentially methylated and expressed genes enriched for methionine degradation; at day 6, for cholesterol biosynthesis; and at day 9, for the metabolism of several amino acids, cholesterol, and the TCA (S5 Data). Similar findings were observed when evaluating the 143B rho0 cells chronically depleted of mtDNA that also showed hypermethylation of the DNA. In those cells, 621 DEGs were also differentially methylated (S6 Data) and enriched for pathways involved, for instance, in folate transformations (S5G Fig). While DNA methylation is not the only parameter governing gene expression, we attempted to define the level of concordance between the changes in DNA methylation status over time with the directionality of expression of the DEGs harboring those changes. Whether we combined the entire methylation profile of genes or considered only promoter marks, the concordance ranged from 30%–50% over the 9 days of mtDNA depletion (S5H Fig). Taken together, these findings suggest a correlation between DNA methylation changes and the expression of a fraction of DEGs responding to progressive mitochondrial dysfunction. NADH oxidation in the mitochondria links polyamine and methionine metabolism to the TCA cycle and DNA methylation It is possible that the mechanism connecting mtDNA depletion to SAM and DNA hypermethylation involves serine biosynthesis and 1C-folate remodeling, which in turn can affect the methionine cycle [8,9]. While this is feasible, the fact that choline/betaine are engaged in maintaining methionine salvage independent of folate would argue against this possibility. Alternatively, the methionine cycle may be directly affected by mtDNA depletion through changes in both methionine and polyamine metabolism. These molecules are not only linked in the regulation of SAM levels [10], but they can provide intermediates such as pyruvate, succinyl-CoA (a precursor of succinate), and succinate to the TCA in their catabolic pathways. An increase in their degradation to feed the TCA could set a cascade of compensatory changes that impacts the SAM pool. To test this hypothesis, we took advantage of the DN-POLG cells overexpressing NDI1/AOX, which are cells that have the ability to oxidize NADH and maintain TCA flux despite the complete loss of mtDNA [4]. We reasoned that if the methionine cycle is directly impacted by the TCA, in these cells methionine-associated intermediates should not be changed. We reanalyzed the metabolomics data that we previously generated with the NDI1/AOX cells [4] using the same criteria as for the DN-POLG cells (S3 Data). We then focused on the intermediates associated with the methionine, serine, folate, and polyamine pathways. We found that in the NDI1/AOX cells, the levels of SAM, SAH, MTA, and the polyamines were maintained over time (Fig 3A); the levels of serine, cysteine, methionine, betaine, choline, and folate followed the same pattern as was observed with the DN-POLG cells (compare Figs 3A and 1E–1G). Most notably, levels of succinate, which were increased at day 9 in the DN-POLG, were decreased in the NDI1/AOX cells (Fig 3B). Taken together, these results support the hypothesis that polyamine and methionine metabolism are directly responding to changes in TCA flux, likely as contributors of succinate. Furthermore, these data suggest that serine biosynthesis and folate-1C remodeling caused by mtDNA depletion are not responding to changes in NADH oxidation or TCA flux. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. NDI1/AOX ectopic expression maintains DNA methylation while not completely rescuing metabolism. Metabolomics was performed in NDI1/AOX cells at days 0, 3, 6 and 9; N = 4 per time point. Differentially enriched metabolites were identified based on log2-transformed fold-changes in arbitrary detection units versus the mean at day 0 (y-axis) in each time point during doxycycline treatment (x-axis) by a two-way ANOVA test. (A) Box plots of relative content versus day 0 (IQR-outlier format) depict distribution of individual replicates in NDI1/AOX cells within statistical groups for SAM, SAH, MTA, Put, Spd, Spm, Ser, Cys, Met, Bet, Cho, and 5-MeTHF. (B) Box plots of relative content versus day 0 (IQR-outlier format) depict distribution of individual replicates within statistical groups for Succ in DN-POLG cells (left) and their NDI1/AOX counterparts (right). (C) Δ%mCG in NDI1/AOX cells at days 0 and 9 for the same probes identified as DML in DN-POLG cells by HM-450K BeadArrays [see Fig 2C]; hypermethylation (red) or hypomethylation (green). N = 3 per time point. (D) Average normalized read counts (bar plot: mean ± SEM) of mtDNA fragments obtained by next-generation sequencing of whole-cell DNA for NDI1/AOX cells; N = 2 per timepoint. (E) Box plots (IQR-outlier format) of average DNA methylation differences between day 0 and day 9 at independently identified DML in NDI1/AOX cells overlapping the genomic range (overall), only promoters, or only bodies of DN-POLG DEGs; separate box plots depict measurements from DN-POLG DEGs that are DMG or not in DN-POLG cells also. Underlying data are reported in S3 Data for (A) and (B); S4 Data for (C) and (E); and S1 Data for (D). (F) Model for the cross talk between methionine salvage, polyamine synthesis and the TCA cycle with DNA methylation: loss of mtDNA decreases TCA flux, which in turn sets a cascade of transcriptional and metabolic changes centered largely on amino acid degradation to maintain TCA cycle output. Degradation of methionine and Put, both of which can feed into the TCA cycle, are among the first changes detected. Put levels are regulated by ornithine, which is provided by the recycling of NH3 resulting from amino acid degradation. Put is also a precursor of Spd and Spm, both of which require dSAM for their synthesis. The main byproduct is MTA, which needs to be quickly recycled, given its accumulation is toxic; the salvage of MTA recycles methionine, a cycle that is also maintained by folate-independent Cho/Bet when mtDNA is depleted. Recycling of MTA also generates adenine, which can enter the purine pool, and α-ketoglutarate that can feed the TCA. By maintaining NADH oxidation in the mitochondria, flux through the TCA cycle is largely normalized, “turning off” the polyamine/MTA salvage response that in turn decreases levels of SAM. Decrease in degradation of amino acids to feed the TCA diminishes flux through the urea cycle, decreasing the input of ornithine to Put biosynthesis. 5-MeTHF, 5-methyltetrahydrofolate; acetyl-CoA, acetyl coenzyme A; AOX, alternative oxidase; Bet, betaine; Cho, choline; Cys, cysteine; Δ%mCG, DNA methylation differences versus average of day 0; DEG, differentially expressed gene; DMG, differentially methylated gene; DML, differentially methylated loci; DN-POLG, dominant-negative DNA polymerase gamma transgene; dSAM, decarboxylated S-adenosylmethionine; ETC; electron transport chain; Met, methionine; MTA, 5-methylthioadenosine; NADH, nicotinamide adenine dinucleotide reduced; NDI1, nicotinamide adenine dinucleotide reduced dehydrogenase-like 1; OXPHOS, oxidative phosphorylation; Put, putrescine; RPM, reads per million reads; SAH, S-adenosylhomocysteine; SAM, S-adenosylmethionine; Ser, serine; Spd, spermidine; Spm, spermine; Succ, succinate; TCA, tricarboxylic acid. https://doi.org/10.1371/journal.pbio.2005707.g003 We also evaluated whole genome methylation using the Illumina 450K platform in NDI1/AOX cells. We used cells at days 0 and 9, since we gauged that mtDNA would be fully depleted at this latter time and would provide the largest effect. Remarkably, no significant changes in DNA methylation were observed in the cells expressing NDI1/AOX, despite complete loss of mtDNA (Fig 3C and 3D). We then focused specifically on the coordinates of the 1,626 DEGs that were differentially methylated in the DN-POLG cells at day 9 (S5A Fig). However, we found that average DNA methylation change in those sites was only approximately 2% in the NDI1/AOX cells (Fig 3E). Hence, we conclude that changes in polyamine synthesis and the MTA salvage pathway, which in turn affect SAM levels, seem to be critical for differential DNA methylation in the nucleus of DN-POLG cells. Lack of changes in DNA methylation are associated with the prevention of differential gene expression, even in the context of complete mtDNA loss We performed gene expression analysis in the NDI1/AOX cells using microarrays in order to determine whether the promoter methylation status has the potential to impact the differential expression of the 879 genes identified in the DN-POLG cells at day 9. Unexpectedly, we found no significant DEGs in the NDI1/AOX cells between days 0 and 9 when adjusting for false discovery rate (FDR; S7 Data). Relaxing statistical thresholds based on pairwise comparisons without multiple testing corrections revealed 23 genes that were differentially expressed between days 0 and 9 (S6 Data), 4 of which were also differentially expressed in the DN-POLG at day 9, as gauged by RNA-seq. To rule out that these results were due to a lack of sensitivity of microarrays to detect the relatively small changes in gene expression identified by RNA-seq, we performed microarrays in DN-POLG cells at days 0 and 9. We found 1,408 genes with adjusted p ≤ 0.05 that were differentially expressed between days 0 and 9 in this cellular background (S7 Data). These results indicate that it is the maintenance of NADH oxidation in the mitochondria, in the context of mtDNA depletion, that prevents the differential expression of genes. To better understand the effects of NDI1/AOX expression in the presence of mtDNA, we next compared the microarray data from DN-POLG cells with those from NDI1/AOX cells at day 0. Again, no DEGs were detected when adjusting for FDR. Using unadjusted p-values, we found 842 genes that were differently expressed between the 2 cell types at day 0 (S8 Data). However, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis identified little to no overlap to the findings obtained when utilizing the DEGs identified in the DN-POLG when mtDNA was depleted (S8 Data). Thus, we conclude that expression of NDI1/AOX does not cause significant off-target effects. Nevertheless, the maintenance of NADH oxidation provided by these enzymes is sufficient to prevent the DNA methylation and transcriptomic changes that result from mtDNA depletion, independent of mitochondrial ATP production or the ΔΨm, which were not rescued by NDI1/AOX expression [4]. Discussion Our understanding of how changes in mitochondrial function can impact the epigenetic control of gene expression in the nucleus is still incomplete. Despite the fact that mitochondrial dysfunction has been shown to affect histone modifications and DNA methylation, mechanistic links between these processes have not been elucidated, particularly in terms of methylation reactions. Earlier studies proposed a prominent role for ROS or 2-HG as inhibitors of the DNA or histone demethylases [26–29]. However, a direct demonstration that the demethylases are inhibited as mitochondria become dysfunctional is still lacking. In this study, we used a novel cell culture system of progressive mtDNA depletion, an isogenic counterpart cell line that was engineered to maintain NADH oxidation, despite loss of mtDNA, and a widely used cell line that is chronically depleted of mtDNA to demonstrate that (i) the methionine cycle responds to loss of TCA function; (ii) salvage pathways of methionine, including through MTA, are engaged in the context of mtDNA loss; (iii) DNA hypermethylation is associated with higher SAM concentrations, 5meC levels, and DNMT activity; (iv) changes in DNA methylation occur predominantly in genes differentially expressed as a result of mtDNA depletion; and (v) genes involved in key pathways responding to mtDNA depletion are targets of differential methylation. Collectively, these findings provide new mechanistic insights that connect mitochondrial function and epigenetics in a way that is likely to have broader relevance to health and disease. In view of their potential impact, further studies to validate and explore these findings in different cell types and in vivo are warranted. Our transcriptomic analysis revealed a dynamic relationship between the loss of mtDNA and the genes that respond to the resulting mitochondrial dysfunction. Our results confirm the serine response that was recently reported by others [8,9,13], but we found that it involves a series of additional amino acids, including methionine. We also identified that inhibition of fat metabolism, perhaps as an additional means of sparing acetyl-CoA, started early in the progression of mitochondrial dysfunction. The identification of methionine degradation as an early response to mtDNA depletion was unexpected. To our knowledge, no direct link between these processes have been previously reported. Our data with the NDI1/AOX cells clearly suggest that resuming TCA flux can turn off the methionine response involving the salvage pathway through MTA. Methionine salvage from MTA can also contribute to the generation of α-ketoglutarate, and because it is intimately linked to polyamine synthesis, where putrescine can generate succinate, we propose a model in which the methionine cycle responds to mtDNA depletion based on changes in TCA intermediates (Fig 3F). At the same time, MTA can also contribute to the purine pool, whose imbalance seems to be an early response to mtDNA loss, perhaps as an additional means to maintain homeostasis. It had been previously shown that, at least in cancer cells, serine contributes to SAM, DNA, and RNA methylation by de novo ATP synthesis [30]. However, maintenance of TCA flux in the NDI1/AOX cells had no impact on serine biosynthesis under our experimental conditions, which rules out that serine is limiting for SAM levels under conditions of mtDNA depletion. It is currently unclear how (and whether it is that) changes in flux through the TCA cycle or changes in the levels of particular TCA metabolites are sensed outside of the mitochondria. Given that acetyl-CoA is an important intermediate of the TCA cycle, which is used both for biosynthetic purposes and for posttranslational protein modifications [31,32], it is possible that its levels are sensed by other parts of the cell. Changes in these levels initiate an entire cascade to ultimately preserve a level of acetyl-CoA that is compatible with maintenance of key cellular functions. We suggest that a sensitive but yet-unidentified pathway that senses and regulates TCA intermediates, such as acetyl-CoA (or flux), within the mitochondria must exist and efficiently communicate to the rest of the cell. The activation of serine biosynthesis and the remodeling of mitochondrial folate and 1C metabolism toward transsulfuration to produce cysteine and glutathione have recently been the subject of various studies. It was shown that this pathway is consistently engaged upon in vitro and in vivo mtDNA depletion, by the inhibition of the ETC with agents such as rotenone or antimycin C, or by compounds that alter protein homeostasis [8,9,13,33]. Our data revealed that the remodeling of 1C metabolism toward transsulfuration is a response to overt mitochondrial dysfunction but does not take place when the organelle is still able to maintain ETC function. It remains unclear why the mitochondrial folate pool and serine biosynthesis are engaged upon mitochondrial dysfunction, although it has recently been shown that mammalian target of rapamycin (mTOR) and the transcription factor ATF4 are involved in these processes [8,13,33]. While we envision that alterations in 1C metabolism can effectively alter nucleotide pools, redox, and methylation reactions—thus potentially impacting cell cycle, transcription, replication, and signaling concomitantly—the exact signal that arises from mitochondria remains to be identified. It has been proposed that ATF4-dependent serine biosynthesis arises from oxidative stress due to a stalled respiratory chain [8]. However, our data do not support this hypothesis, since we found no changes in respiratory function at day 3 [4] and, most importantly, that maintenance of NADH oxidation in the NDI1/AOX cells does not turn off the serine biosynthesis response. Further studies will be required to resolve this issue. Our finding that there is an extensive cellular metabolic rewiring associated with mitochondrial dysfunction centered on amino acid and lipid metabolism was not fully surprising. What was unexpected was that the maintenance of NADH oxidation in NDI1/AOX cells prevented, most significantly, the changes in methionine metabolism and polyamine synthesis. To our knowledge, this is the first report that directly connects these pathways to mitochondrial function. The roles of polyamines in cell biology are still poorly understood and are linked to effects on cell proliferation, chromatin configuration, gene transcription, and even mitochondrial calcium homeostasis [34,35]. Whether the changes in polyamine levels we identified to be associated with mitochondrial dysfunction are sufficient to affect any of these processes requires further studies. It is worth noting that a mouse completely deficient in spermine synthase activity was identified in a cohort of female irradiated offspring [36,37]. The complete loss of polyamines is embryonic lethal in vivo [38]. Most interestingly, these mice have many phenotypes that resemble mitochondrial diseases, including deafness, sterility, neurological abnormalities, and reduced life span [39,40], which supports a potential link between polyamine synthesis and mitochondrial health. Finally, our data revealed that many of the genes that are differentially expressed in response to mtDNA depletion are also differentially methylated. Moreover, we showed these changes in methylation and gene expression can be prevented by maintaining NADH oxidation in the NDI1/AOX cells. These findings support the notion that the epigenetic changes caused by mtDNA depletion are intimately associated with the differential expression of genes in the DN-POLG cells. It is possible that posttranscriptional modifications of specific proteins, and not transcriptional changes of specific genes, account for the lack of differential gene expression in the cells overexpressing the NDI1/AOX transgenes. However, we think this is unlikely to be the case. Alternatively, it can be argued that it is the metabolic rescue provided by NADH oxidation in the mitochondria that is directly responsible for turning off the transcriptional response in these cells. Given the continued activation of the metabolism of serine, folates, and others, despite maintenance of TCA flux, we do not favor this possibility. More studies are required to better address these issues. The remarkable parallels between our in vitro results and the data obtained with in vivo mouse models and patient samples of mitochondrial disorders [9,33] suggest that our findings may be relevant to human health. In fact, careful analysis of the metabolic data from heart and muscle of the Deletor mouse strain indicate that the methionine salvage pathway and polyamine synthesis are altered based on increased steady state levels of choline, betaine, ornithine, and MTA (S6 Fig) [9]. Similar findings were observed in the liver of another mouse model of mtDNA abnormalities driven by a thymidine kinase mutation [41] in which MTA and adenosine levels are higher than wild-type controls (S6 Fig). However, whether the DNA is also hypermethylated in those models and which genes may be affected by it remain to be addressed. It also remains to be determined the extent to which the flux from polyamines to the methionine cycle end up in methylation reactions and whether these changes are drivers or contributors of the overall response to mtDNA depletion. It is also unclear whether these same effects occur in the context of other types of mitochondrial dysfunction. This will be especially important for studies to reveal how environmental toxicants that target the mitochondria change the biology of the cell. Irrespective of these limitations, our findings have the potential to fundamentally change our understanding about the role and impact of mitochondrial metabolism in health and disease. Materials and methods Cells HEK293T cells carrying a tetracycline (Tet)-on inducible DN-POLG from [42] were used to generate derivatives also ectopically expressing AOX and NDI1 and cultured as described previously [4]. The osteosarcoma cell line 143B and its rho0 derivative, graciously obtained from Dr. Eric Schon at Columbia University, were routinely grown in DMEM high glucose (4.5 g/L) supplemented with 10 mM pyruvate, 50 μg/mL of uridine, 10% FBS, and 1% penicillin/streptomycin under 37°C and 5% CO2. Except for ChIP-seq experiments (N = 2) all experiments were performed on N = 3 independent biological replicates. RNA extraction RNA was extracted from 3 independent cell cultures of DN-POLG cells at days 0, 3, 6, and 9 using RNAeasy Mini and QIAshredder kits (QIAGEN) and was used for both RNA-seq and microarrays; RNA from NDI1/AOX cells was obtained at days 0 and 9 and used for microarrays only. In all cases, samples from 3 independent experiments were collected (N = 3 per time point per cell model). Quantification of DNMT activity in 143B cells DNMT activity was assayed using radioactive filter-binding assay, as previously described [43]. Briefly, 143B cells from rho+ or rho0 cells were lysed and the nuclear fraction enriched using differential centrifugation. Then, cell lysates were used to monitor the incorporation of tritiated (3H) methyl groups into a poly-IC duplex DNA oligonucleotide. The unreacted (methyl-3H) was separated from the radiolabeled DNA using filter binding. The 3H–CH3-containing duplex DNA was then quantified by liquid scintillation. Data were normalized to protein content and presented relative to the detected activity in rho+ cells. Samples from 3 independent replicates were collected (N = 3 per 143B derivative). Gene expression experiments by RNAseq in DN-POLG cells RNA from DN-POLG cells at days 0, 3, 6, and 9 (N = 3 each timepoint) was poly-A-selected and sequenced with a HiSeq 2000 system (Illumina). Following 3′ adapter trimming and base-calling filtering (phred score > 20), we obtained approximately 100 million 126-nt paired-end reads per individual sample (7–8 flow cell lanes/sample), which were aligned to the hg19 human reference genome (Genome Reference Consortium GRCh37 from February 2009) [44] with TopHat-Fusion function [45]. Composite RPKM counts within genomic coordinates of 20,304 nonhaplotype HGNC-annotated genes were used to calculate gene expression differences based on log2-transformed fold-change (log2FC) relative to average gene RPKM at day 0 at a significance level p < 0.05 adjusted for multiple comparisons [46]. Detection of DEGs by RNA-seq DEGs were detected using weighed two-way ANOVA (gene × time) of log2-transformed expression fold change measurements with respect to the average composite RPKM at day 0 (log2FC); N = 12 (3 biological replicates per time point). Gene-wise log2FC values were weighed by a relative metric of sequencing representation (cumulative hazard of significance scores from gene-wise RPKM rate modeling with an exponential distribution and inverse link function). A total 2,854 HGNC-annotated DEGs were detected in DN-POLG cells at a significance level p < 0.05 adjusted for multiple comparisons [46], filtering against a minimum gene-wise effect size δlog2FC > 0.3 × σlog2FC, and post hoc pairwise significance (Student t test p < 0.05) between log2FC values at days 3, 6, or 9 versus day 0. For gene-level effect size filtering, δlog2FC = 0.3 × σSSR is 5% of the 6σ-spread log2FC regression error with respect to a gene’s grand mean (where [σSSR]2 = [SSRlog2FC] / [N– 1]) compared to 5% of the 6σ-spread in measurement error about the mean log2FC at each time point in the gene (where [σlog2FC]2 = [SSElog2FC] / [N-1]). Gene expression experiments by microarray technology and data analyses For microarrays analysis of gene expression, the Affymetrix Human Genome U133 Plus 2.0 GeneChip arrays were used. Samples were prepared as per manufacturer’s instructions. Arrays were scanned in an Affymetrix Scanner 3000 and data were obtained using the GeneChip Command Console and Expression Console Software (AGCC, Version 3.2; Expression Console, Version 1.2) using the MAS5 algorithm to generate CHP-extension files. ANOVA was used to identify statistical differences between means of groups at α < 0.05 level among HG-U133 Plus 2.0 probe sets unambiguously mapped to UCSC known gene transcripts. Detection of differentially enriched metabolites Differentially enriched metabolites were detected for DN-POLG and NDI1/AOX cells separately using two-way ANOVA (metabolite × time) of log2-transformed relative changes in abundance versus untreated cells (log2RC); N = 16 (4 biological replicates per time point in each DN-POLG and NDI1/AOX); refer to previously published original data elsewhere [4]. Data with significance level p < 0.05 adjusted for multiple comparisons [46] were then filtered against minimum effect size δlog2RC ≥ 0.3 × σlog2RC = 0.20 (approximately 1.15-fold change). The δlog2RC corresponds to the smallest estimate between DN-POLG and NDI1/AOX cells of metabolome-wide effects at 5% of the 6σ-spread in log2RC measurement error across all metabolite × time groups (where [σlog2RC]2 = [SSElog2RC] / [N– 1]). DNA methylation arrays and data analyses Genomic DNA was extracted from 3 independent cell cultures of DN-POLG cells at days 0, 3, 6, and 9 and NDI1/AOX cells at days 0 and 9 and bisulfite-converted using an EZ DNA Methylation kit (Zymo Research) following the manufacturer’s protocol. Differential methylation at the CpG dinucleotide level was conducted using Human Methylation450 v1 BeadChip arrays (Illumina) following the InfiniumHD methylation protocol. Data were obtained using Illumina’s GenomeStudio software (version 2011.1) using no background subtraction and no normalization parameters. Probes in a probe × cell × time block with >1 failed reads or >1 outliers (1.5 IQR rule on residuals around probe × cell × time sample means) were discarded from analysis. Probe-level groups with N = 2 after failed read or outlier filtering were brought to N = 3 by substitution with probe × cell × time trimmed mean values adjusted by imputed array-wise residual estimates using the pairwise correlation matrix of cell × time statistical groups. Probe methylation percentage was quantified from fluorometric signal intensities of methylation (mCG) and unmethylation (CG) in terms of ß = [mCG/(mCG+CG)] × 100. ANOVA was used to identify statistical differences between the means of groups at a significance level p < 0.05 adjusted for multiple comparisons [46] using JMP software (Version 11) followed by post hoc pairwise significance testing (p < 0.05) with respect to day 0 in DN-POLG or NDI1/AOX cells. Cells HEK293T cells carrying a tetracycline (Tet)-on inducible DN-POLG from [42] were used to generate derivatives also ectopically expressing AOX and NDI1 and cultured as described previously [4]. The osteosarcoma cell line 143B and its rho0 derivative, graciously obtained from Dr. Eric Schon at Columbia University, were routinely grown in DMEM high glucose (4.5 g/L) supplemented with 10 mM pyruvate, 50 μg/mL of uridine, 10% FBS, and 1% penicillin/streptomycin under 37°C and 5% CO2. Except for ChIP-seq experiments (N = 2) all experiments were performed on N = 3 independent biological replicates. RNA extraction RNA was extracted from 3 independent cell cultures of DN-POLG cells at days 0, 3, 6, and 9 using RNAeasy Mini and QIAshredder kits (QIAGEN) and was used for both RNA-seq and microarrays; RNA from NDI1/AOX cells was obtained at days 0 and 9 and used for microarrays only. In all cases, samples from 3 independent experiments were collected (N = 3 per time point per cell model). Quantification of DNMT activity in 143B cells DNMT activity was assayed using radioactive filter-binding assay, as previously described [43]. Briefly, 143B cells from rho+ or rho0 cells were lysed and the nuclear fraction enriched using differential centrifugation. Then, cell lysates were used to monitor the incorporation of tritiated (3H) methyl groups into a poly-IC duplex DNA oligonucleotide. The unreacted (methyl-3H) was separated from the radiolabeled DNA using filter binding. The 3H–CH3-containing duplex DNA was then quantified by liquid scintillation. Data were normalized to protein content and presented relative to the detected activity in rho+ cells. Samples from 3 independent replicates were collected (N = 3 per 143B derivative). Gene expression experiments by RNAseq in DN-POLG cells RNA from DN-POLG cells at days 0, 3, 6, and 9 (N = 3 each timepoint) was poly-A-selected and sequenced with a HiSeq 2000 system (Illumina). Following 3′ adapter trimming and base-calling filtering (phred score > 20), we obtained approximately 100 million 126-nt paired-end reads per individual sample (7–8 flow cell lanes/sample), which were aligned to the hg19 human reference genome (Genome Reference Consortium GRCh37 from February 2009) [44] with TopHat-Fusion function [45]. Composite RPKM counts within genomic coordinates of 20,304 nonhaplotype HGNC-annotated genes were used to calculate gene expression differences based on log2-transformed fold-change (log2FC) relative to average gene RPKM at day 0 at a significance level p < 0.05 adjusted for multiple comparisons [46]. Detection of DEGs by RNA-seq DEGs were detected using weighed two-way ANOVA (gene × time) of log2-transformed expression fold change measurements with respect to the average composite RPKM at day 0 (log2FC); N = 12 (3 biological replicates per time point). Gene-wise log2FC values were weighed by a relative metric of sequencing representation (cumulative hazard of significance scores from gene-wise RPKM rate modeling with an exponential distribution and inverse link function). A total 2,854 HGNC-annotated DEGs were detected in DN-POLG cells at a significance level p < 0.05 adjusted for multiple comparisons [46], filtering against a minimum gene-wise effect size δlog2FC > 0.3 × σlog2FC, and post hoc pairwise significance (Student t test p < 0.05) between log2FC values at days 3, 6, or 9 versus day 0. For gene-level effect size filtering, δlog2FC = 0.3 × σSSR is 5% of the 6σ-spread log2FC regression error with respect to a gene’s grand mean (where [σSSR]2 = [SSRlog2FC] / [N– 1]) compared to 5% of the 6σ-spread in measurement error about the mean log2FC at each time point in the gene (where [σlog2FC]2 = [SSElog2FC] / [N-1]). Gene expression experiments by microarray technology and data analyses For microarrays analysis of gene expression, the Affymetrix Human Genome U133 Plus 2.0 GeneChip arrays were used. Samples were prepared as per manufacturer’s instructions. Arrays were scanned in an Affymetrix Scanner 3000 and data were obtained using the GeneChip Command Console and Expression Console Software (AGCC, Version 3.2; Expression Console, Version 1.2) using the MAS5 algorithm to generate CHP-extension files. ANOVA was used to identify statistical differences between means of groups at α < 0.05 level among HG-U133 Plus 2.0 probe sets unambiguously mapped to UCSC known gene transcripts. Detection of differentially enriched metabolites Differentially enriched metabolites were detected for DN-POLG and NDI1/AOX cells separately using two-way ANOVA (metabolite × time) of log2-transformed relative changes in abundance versus untreated cells (log2RC); N = 16 (4 biological replicates per time point in each DN-POLG and NDI1/AOX); refer to previously published original data elsewhere [4]. Data with significance level p < 0.05 adjusted for multiple comparisons [46] were then filtered against minimum effect size δlog2RC ≥ 0.3 × σlog2RC = 0.20 (approximately 1.15-fold change). The δlog2RC corresponds to the smallest estimate between DN-POLG and NDI1/AOX cells of metabolome-wide effects at 5% of the 6σ-spread in log2RC measurement error across all metabolite × time groups (where [σlog2RC]2 = [SSElog2RC] / [N– 1]). DNA methylation arrays and data analyses Genomic DNA was extracted from 3 independent cell cultures of DN-POLG cells at days 0, 3, 6, and 9 and NDI1/AOX cells at days 0 and 9 and bisulfite-converted using an EZ DNA Methylation kit (Zymo Research) following the manufacturer’s protocol. Differential methylation at the CpG dinucleotide level was conducted using Human Methylation450 v1 BeadChip arrays (Illumina) following the InfiniumHD methylation protocol. Data were obtained using Illumina’s GenomeStudio software (version 2011.1) using no background subtraction and no normalization parameters. Probes in a probe × cell × time block with >1 failed reads or >1 outliers (1.5 IQR rule on residuals around probe × cell × time sample means) were discarded from analysis. Probe-level groups with N = 2 after failed read or outlier filtering were brought to N = 3 by substitution with probe × cell × time trimmed mean values adjusted by imputed array-wise residual estimates using the pairwise correlation matrix of cell × time statistical groups. Probe methylation percentage was quantified from fluorometric signal intensities of methylation (mCG) and unmethylation (CG) in terms of ß = [mCG/(mCG+CG)] × 100. ANOVA was used to identify statistical differences between the means of groups at a significance level p < 0.05 adjusted for multiple comparisons [46] using JMP software (Version 11) followed by post hoc pairwise significance testing (p < 0.05) with respect to day 0 in DN-POLG or NDI1/AOX cells. Supporting information S1 Fig. Experimental design and measures of reproducibility in DN-POLG and NDI1/AOX model of dox-inducible mtDNA depletion. (A) Schematic representation of experimental procedure and data integration. Culture splits were performed every 3 days, corresponding to observed interval needed for noninduced DN-POLG or NDI1/AOX starter cultures to return to confluence after 1:2 subcultivation. As depicted, each cell culture round going from one single starter culture to individual cultures at day 0 (1 flask), 3 (1 flask), 6 (2 flasks, pooled), or 9 (4 flasks, pooled) of continuous supplementation with doxycycline (10 ng/mL) represents a single biological replicate. Independent biological replicates per cell model and per timepoint were produced for different assays: transcriptomics, N = 3; metabolomics, N = 4; DNA methylation, N = 3. (B–D) Reproducibility of biochemical output among DN-POLG biological replicates as shown by pairwise concordance and Pearson’s correlation r within different assays: (B) RNA-seq (Log10[RPKM], poly[A]-enriched RNA, uniquely aligned reads to hg19 reference genome; days 0, 3, 6, and 9); (C) Illumina HM-450K DNA Methylation BeadArrays (% mCG/CG, bisulfite-converted DNA; days 0, 3, 6, and 9); and (D) Metabolon mass spectrometry (Log10[Content], arbitrary units; days 0, 3, 6, and 9). (E–G) Reproducibility of biochemical output among NDI1/AOX biological replicates, as shown by pairwise concordance and Pearson’s correlation r within different assays: (E) Affymetrix HG-U133 Plus 2.0 Microarrays (Log10[Intensity], Total RNA; days 0 and 9); (F) Illumina HM-450K DNA Methylation BeadArrays (%mCG/CG, bisulfite-converted DNA; days 0 and 9); and (G) Metabolon mass spectrometry (Log10[Content], arbitrary units; days 0, 3, 6, and 9). Underlying data are reported in S1 Data for (B); S4 Data for (C) and (F); S3 Data for (D) and (G); and S7 Data for (E). %mCG, percentage of DNA methylation; AOX, alternative oxidase; DN-POLG, dominant-negative DNA polymerase gamma transgene; mtDNA, mitochondrial DNA; NDI1, nicotinamide adenine dinucleotide reduced dehydrogenase-like 1; RNA-seq, RNA sequencing; RPKM, reads per kilobase per million. https://doi.org/10.1371/journal.pbio.2005707.s001 (TIF) S2 Fig. Validation of RNA-seq measurements in DN-POLG cells by quantitative PCR. (A) Left upper panel depicts concordance of relative expression estimates normalized to housekeeping ACTB between RNA-seq and qPCR experiments for 12 nuclear-encoded DEGs and 3 mtDNA-encoded genes; increasingly darker shades of brown (nuclear-encoded) and gray (mtDNA-encoded) depict relative gene expression values at days 3, 6, and 9 each versus day 0. qPCR was performed from cDNA templates derived using the same total RNA extracts for both techniques (N = 3 per timepoint) and were performed in technical triplicates for each of the 3 biological replicates. Right top panels show graphs for mtDNA-encoded genes while the bottom panels depict data from nuclear DNA-encoded transcripts. Underlying data are reported in S1 Data. (B) Top 5 significantly enriched upstream regulators for upregulated and downregulated DEGs in DN-POLG cells at days 3, 6, and 9 of dox-inducible mtDNA depletion, per Ingenuity Pathway Analysis. ACTB, ß-actin; DEGs, differentially expressed genes; DN-POLG, dominant-negative DNA polymerase gamma transgene; mtDNA, mitochondrial DNA; qPCR, quantitative polymerase chain reaction; RNA-seq, RNA sequencing. https://doi.org/10.1371/journal.pbio.2005707.s002 (TIF) S3 Fig. Differentially enriched metabolites and significantly associated metabolic pathways in DN-POLG cells. Metabolomics was performed in DN-POLG cells at days 0, 3, 6, and 9; N = 4 per time point. Differentially enriched metabolites were identified based on log2-transformed fold-changes in arbitrary detection units versus the mean at day 0 by a two-way ANOVA test (metabolite × time) at an adjusted Benjamini-Hochberg p ≤ 0.05, and are detailed in S3 Data. (A) Number of differential metabolites identified in DN-POLG cells at days 3, 6, and 9 of doxycycline supplementation compared to day 0 (circle plots, top), and their overlap between time points (Venn diagram, bottom). (B) Heatmap of significantly represented canonical metabolic pathways per Ingenuity Pathway Analysis [−log(p) > 1.3] based on differentially enriched metabolites in DN-POLG cells at days 3, 6, and 9 of dox-inducible mtDNA depletion. Coloring of listed pathway names correspond to clades of pathways inferred by unsupervised clustering of enrichment scores across timepoints. (C–G) Data integration of average log2-fold changes relative to day 0 in gene expression (down: green, up: red; numerical values reported in S1 Data) and metabolite enrichment (down: blue, up: red; numerical values reported in S3 Data) with PathVisio engine for (C) nucleotide metabolism, (D) Methionine De Novo and Salvage Pathway, (E) One-Carbon Metabolism and Related Pathways, (F) Amino Acid Interconversion, and (G) Amino Acid Metabolism. (H) Depiction of three metabolic nodes identified as the main drivers of the response to mtDNA depletion in DN-POLG cells by day 3, per Ingenuity Pathway Analysis. acetyl-Coa, acetyl coenzyme A; DN-POLG, dominant-negative DNA polymerase gamma transgene; mtDNA, mitochondrial DNA; TCA, tricarboxylic acid. https://doi.org/10.1371/journal.pbio.2005707.s003 (PDF) S4 Fig. Doxycycline supplementation induces loss of mtDNA copy number and expression in DN-POLG cells over 9 days. (A) Average normalized read counts (RPM; bar plots: mean ± SEM) of mtDNA fragments obtained by next-generation sequencing of whole-cell DNA for DN-POLG cells; N = 2 per timepoint. (B) Average normalized read counts (reads per kilobase per million reads, RPKM; bar plots: mean ± SEM) of mtDNA transcripts (mtRNA) obtained by RNA-seq for DN-POLG cells; N = 3 per timepoint. Underlying data are reported in S1 Data. DN-POLG, dominant-negative DNA polymerase gamma transgene; mtDNA, mitochondrial DNA; RNA-seq, RNA sequencing; RPKM, reads per kilobase per million reads; RPM, reads per million reads. https://doi.org/10.1371/journal.pbio.2005707.s004 (TIF) S5 Fig. DNA methylation levels in DN-POLG cells are changed in the course of mtDNA depletion. (A) Violin plots depict Δ%mCG for DML observed in DN-POLG cells by HM-450K BeadArrays broken down by probe annotated location: IGR, TSS 1500 and TSS 200, 5'UTR, 1st exon (exon 1 as per annotated reference genome), body (gene body), and 3'UTR. Number of DML at each time point is shown at the top of each plot frame within each time point and probe location class. (B) Unsupervised hierarchical clustering showing the percentage of methylation of individual locus, presented in each row, in either rho+ or rho0 cells based on Illumina 450K methylation arrays; N = 3. (C) Immunofluorescence using antibodies again 5meC or 5hmeC in both cell types; images taken with confocal microscope; blue—DAPI-stained nucleus; controls containing only primary or only secondary antibodies or background fluorescence are omitted. (D) Dot blots probing levels of 5meC or 5hmeC were performed in technical triplicates; N = 12. (E) Number of DMEGs in DN-POLG cells at each time point is shown above each panel. All identified differentially methylated probes relative to day 0 associated with DEGs in DN-POLG cells are shown. X-axis depicts the degree of methylation change while the y-axis shows the expression change for each respective DEG. (F) Expected versus observed number of genes that bear or do not bear changes in DNA methylation relative to DEGs in DN-POLG cells; p-value was calculated using chi-squared test. (G) Bar graph shows metabolic pathways identified enriched in 143B rho0 cells chronically depleted of mtDNA based on 621 DEGs that were also differentially methylated, per Ingenuity Pathway Analysis; the top x-axis depicts log10-tranformed significance score of enriched pathways, whereas the bottom ratio x-axis (bottom) and line plot (orange) reflects the proportion of members within each pathway that are present in the dataset. (H) Concordance between gene expression directionality (y-axis) and change in DNA methylation levels (x-axis) are presented for each time point. The number of DEGs in each quadrant is also indicated; colors depict quantile density for “gene overall” plots. Values of Δ%mCG for “gene overall” plots equal the average of all probes inside genes, probes in their 5'–or 3' untranslated regions, and promoters; Δ%mCG values are also split into averages of promoter only [P, blue] or within-gene-body [B, red] probes. Underlying data are reported in S4 Data for (A), (E), (F), and (H); in S6 Data for (B); and in S1 Data for (E), (F), and (H). 5hmeC, 5-hydroxy-methyl-cytosine; 5meC, 5-methyl-cytosine; 5'UTR, 5- untranslated region; Δ%mCG, DNA methylation differences versus average of day 0; DEGs, differentially expressed genes; DMEGs, differentially methylated and expressed genes; DML, differentially methylated loci; DN-POLG, dominant-negative DNA polymerase gamma transgene; IGR, intergenic region; mtDNA, mitochondrial DNA; TSS 1500, transcription start site 1,500 https://doi.org/10.1371/journal.pbio.2005707.s005 (PDF) S6 Fig. Metabolite increased in the DN-POLG cells are also changed in mice models of mtDNA depletion. Graphs depict levels of metabolites in Deletor mice (skeletal muscle or heart) based on the fold-change relative to wild-type littermates, as reported in [9]. TK2 data were calculated based on normalized metabolite values from the KOs versus wild-type counterparts, as reported in [41]. DN-POLG, dominant-negative DNA polymerase gamma transgene; KO, knockout; MTA, 5-methyl-thioadenosine; mtDNA, mitochondrial DNA; TK2, timidine kinase 2 https://doi.org/10.1371/journal.pbio.2005707.s006 (TIF) S1 Data. Differentially expressed genes detected in DN-POLG cells by RNA-seq during 9-day time course of dox-inducible mtDNA depletion. ACTB, ß-actin; DN-POLG, dominant-negative DNA polymerase gamma transgene; mtDNA, mitochondrial DNA; RNA-seq, RNA sequencing; RPKM, reads per kilobase per million; ChrM, chromosome M (mitochondrial). https://doi.org/10.1371/journal.pbio.2005707.s007 (XLSX) S2 Data. Significantly enriched pathways in DN-POLG at days 3, 6, and 9 of dox-inducible mtDNA depletion for upregulated and downregulated DEGs via Ingenuity Pathway Analysis. DEGs, differentially expressed genes; DN-POLG, dominant-negative DNA polymerase gamma transgene; mtDNA, mitochondrial DNA. https://doi.org/10.1371/journal.pbio.2005707.s008 (XLSX) S3 Data. Differentially enriched metabolites (log2-ratio versus day 0) in DN-POLG and NDI1/AOX cells by metabolomics assays (Metabolon) during 9-day time course of dox-inducible mtDNA depletion. AOX, alternative oxidase; DN-POLG, dominant-negative DNA polymerase gamma transgene; mtDNA, mitochondrial DNA; NDI1, nicotinamide adenine dinucleotide reduced dehydrogenase-like 1 https://doi.org/10.1371/journal.pbio.2005707.s009 (XLSX) S4 Data. DEGs with differential DNA methylation in DN-POLG cells per time point during mtDNA depletion. AOX, alternative oxidase; Δ%mCG, DNA methylation differences versus average of day 0; DEGs, differentially expressed genes; DN-POLG, dominant-negative DNA polymerase gamma transgene; mtDNA, mitochondrial DNA; NDI1, nicotinamide adenine dinucleotide reduced dehydrogenase-like 1 https://doi.org/10.1371/journal.pbio.2005707.s010 (XLSX) S5 Data. Significantly enriched pathways at days 3, 6, and 9 of dox-inducible mtDNA depletion for DEGs with differential DNA methylation in DN-POLG cells via IPA. DEGs, differentially expressed genes; DN-POLG, dominant-negative DNA polymerase gamma transgene; IPA, Ingenuity Pathway Analysis; mtDNA, mitochondrial DNA. https://doi.org/10.1371/journal.pbio.2005707.s011 (XLSX) S6 Data. DEGs with differential DNA methylation in 143B rho0 cells. Δ%mCG, DNA methylation differences versus average of day 0; DEGs, differentially expressed genes. https://doi.org/10.1371/journal.pbio.2005707.s012 (XLSX) S7 Data. Differentially expressed genes detected at day 9 of dox-inducible mtDNA depletion in NDI1/AOX cells by microarray (Affymetrix HG-U133 Plus 2.0). AOX, alternative oxidase; mtDNA, mitochondrial DNA; NDI1, nicotinamide adenine dinucleotide reduced dehydrogenase-like 1 https://doi.org/10.1371/journal.pbio.2005707.s013 (XLSX) S8 Data. Differentially expressed genes detected at day 9 of dox-inducible mtDNA depletion in DN-POLG cells by microarray (Affymetrix HG-U133 Plus 2.0). DN-POLG, dominant-negative DNA polymerase gamma transgene; mtDNA, mitochondrial DNA. https://doi.org/10.1371/journal.pbio.2005707.s014 (XLSX) S9 Data. Original data from Fig 2F. https://doi.org/10.1371/journal.pbio.2005707.s015 (XLSX) Acknowledgments We thank the staff at the Core Facilities at the National Institute of Environmental Health Sciences (NIEHS) and National Institute of Health (NIH; Epigenetics and Genomics) and Drs. Raja Jothi and Paul Wade (NIEHS) for comments on the manuscript.

doi: 10.1371/journal.pbio.2005473pmid: 29621251

Introduction Rapid and reliable protein localisation is critical for the functional characterisation of any protein of interest (POI). Traditionally, this has been achieved through antibody-mediated methods or tagging with a fluorescent protein, such as GFP. The recent emergence of nanobodies (small, single-domain antibodies amenable to cellular expression) has allowed the development of new biotechnological tools based on the detection of epitopes in living cells [1,2], although the availability of defined variable domains for antigen binding remains limiting. At the same time, the use of enzymatic tags such as the soybean ascorbate peroxidase (APEX) for ultrastructural detection of proteins provides an alternative to the use of traditional antibody labelling in electron microscopy (EM) [3], with the advantage of protein localisation throughout the depth of whole cells or tissues making it compatible with the latest revolutionary 3D EM methods [4]. We have previously generated expression plasmids that encode a GFP-nanobody/binding peptide (GBP) for high-resolution detection of GFP-tagged proteins by electron microscopy. To achieve this, we genetically fused the GBP nanobody to the well-characterized soybean-derived enzyme APEX. When APEX–GBP is expressed in the presence of any GFP-tagged POI, its localisation can be determined by transmission EM following processing [5,6]. Here, we have developed and characterized a new suite of APEX/nanobody-mediated tools. As GFP and mCherry are the most broadly used fluorescent proteins in cell biology, we used cell expression to screen a library of putative mCherry-binding peptides (ChBPs) by single-molecule coincidence detection. We demonstrate the utility of a single mCherry nanobody for high-resolution, EM-based analysis of protein distribution and use this probe for correlative analyses. Furthermore, we generate conditionally stable (cs) nanobodies for both GFP and mCherry fused to APEX and show that degradation of unbound cs nanobodies by the proteasomal system reduces background APEX signals and results in an increased signal-to-noise ratio. Finally, we show that the new suite of APEX nanobody tools opens up entirely new avenues for EM localisation through the application of the csAPEX-nanobody system to bimolecular fluorescence complementation, allowing the detection and localisation of intracellular protein-protein interactions at the ultrastructural level. Results and discussion To date, no modular systems exist to sensitively detect mCherry-tagged POIs to high-resolution for transmission electron microscopy. Therefore, we initially sought to generate a modular APEX-ChBP expression vector. We screened six sequences previously shown to have affinity for mCherry [7] by fluorescence cross-correlation spectroscopy in Leishmania tarentolae cell-free lysate [8]. Each peptide was first expressed fused to the open reading frame of GFP and assayed for self-association or cross-reactivity with GFP (S1A–S1F Fig). ChBP1 and ChBP2 behaved as monomeric proteins (S1A and S1B Fig), whereas ChBP3, ChBP4, ChBP6, and ChBP8 demonstrated bursts of GFP signal above baseline monomeric protein behaviour (S1C–S1F Fig), indicating a propensity for self-association. We next performed single-molecule coincidence detection after co-expression of mCherry-Caveolin1 (Cav1) [9]. mCherry-Cav1 was selected as it generates stable, uniform, and membrane-associated oligomeric Cav1, resulting in highly clustered mCherry tags within the confocal volume. Co-expression of GFP-tagged ChBP1, ChBP3, ChBP4, and ChBP6 with mCherry-Cav1 did not result in significant coincidence between mCherry-Cav1 and GFP-tagged ChBP1, suggesting that these peptides are inefficient at binding the mCherry tag in this context (S1A and S1C–S1E Fig second and third panels). However, ChBP2-GFP and ChBP8-GFP demonstrated a considerable coincidence between the GFP-tagged ChBP and mCherry-Cav1, with a coincidence ratio of Cherry to Cherry and GFP of approximately 0.5, indicating a 1:1 binding ratio of GFP to Cherry (S1B and S1F Fig second and third panel). We selected ChBP2 as the best-performing peptide in our analysis and incorporated this into our modular expression system (Fig 1A; mammalian expression vector hitherto termed APEX-ChBP). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Modular detection of mCherry-tagged proteins using APEX-tagged ChBPs. A) Schematic of cell-based transfection of modular APEX-ChBP and any mCherry-tagged POI. B-D) Electron micrographs of BHK cells co-expressing APEX-ChBP and B) mCherry, C) mCherry-Cavin1, and D) 2xFYVE-mCherry; arrows highlight areas of enriched electron density. Note the increased density in the cytoplasm compared to mitochondria. Scale bars: lower magnification = 1 μm; insets = 500 nm. E-H) CLEM-based detection nls-mCherry–transfected cells using APEX-ChBP. E) 10x magnification of stacked bright field and epifluorescent images of live BHK cells transfected with H2B-mCherry and APEX-ChBP. The grid coordinate (7K) can be resolved in the bright field image. White box = region of interest. F) Bright field image of flat-embedded cells after removal of the coverslip and tissue culture dish (corresponds to the region of interest from [E]). Significant DAB reaction product can be resolved in the nucleus of cells transfected with the higher expression of the H2B-mCherry. Eight different cells were selected for higher-resolution EM analysis. G) Montaged electron micrographs of the region of interest correlated with red channel epifluorescence image from (E). H) High-resolution transmission electron micrographs of transfected cells (regions 1 to 8, respectively) demonstrated restricted electron density within the nuclei of high-expressing cells (regions 1 to 5) and low-expressing cells (region 6) and no increased electron density above background in untransfected cells (regions 7 and 8). Scale bars: E = 100 μm, F–G = 50 μm, H = 5 μm. DAB, 3,3′-Diaminobenzidine; APEX, ascorbate peroxidase; BHK, baby hamster kidney; Cav, caveolae; CLEM, correlative light and electron microscopy; ChBP, mCherry-binding peptide; Cyto, cytoplasm; EM, electron microscopy; End, endosome; ER, endoplasmic reticulum; H2B, Histone 2B; Mito, Mitochondria; nls-mCherry, nuclear localized mCherry; Nuc, nucleus; PM, plasma membrane; POI, protein of interest. https://doi.org/10.1371/journal.pbio.2005473.g001 To verify that this construct could be used for high-resolution EM, we co-transfected baby hamster kidney (BHK) cells with APEX-ChBP and three different subcellular markers: (i) mCherry to denote the cytoplasm, (ii) mCherry-Cavin1 to denote caveolae on the plasma membrane (PM), and (iii) 2xFYVE-mCherry to denote early endosomes. Co-expression of the soluble APEX-ChBP and mCherry (with subsequent DAB reaction in the presence of H2O2 and post-fixation with osmium tetroxide [OsO4]) resulted in the accumulation of electron density in the cytoplasm of transfected cells (Fig 1B). This observation closely mirrored the expression of GFP with APEX-GBP [6]. Cavin1 is a critical structural component of plasma membrane microdomains termed ‘caveolae’ and, when present at the PM, resides only within these domains [10]. When APEX-ChBP was co-transfected with mCherry-Cavin1, the electron density generated by the APEX tag and the DAB reaction was restricted to the plasma membrane at structures with morphologies consistent with caveolae (Fig 1C). Finally, we attempted to localize the phosphoinositide (PI) probe 2xFYVE-mCherry (a marker of PI(3)P lipids), which are highly enriched within early endosomes [11]. Co-expression of 2xFYVE-mCherry and APEX-ChBP resulted in the specific accumulation of electron density surrounding structures consistent with early endosomal morphology (Fig 1D). These data demonstrate that our APEX-ChBP vector can be used to localize mCherry-tagged proteins at ultrastructural resolution. As shown in Fig 1E–1H, use of the APEX-ChBP system is compatible with efficient correlative light and EM. Because the APEX2 probe is visible under both light and EM, this represents a simple alternative to more complex and currently widely used CLEM methods. The modular system for EM detection of fluorescently tagged POIs involves recruitment of APEX-tagged binding peptides to the fluorescent protein (FP). Any unbound APEX nanobody will produce a diffuse cytosolic pool that will hinder detection of the POI and reduce the signal-to-noise ratio, particularly for low-abundance antigens. Recent work using the GBP nanobody has shown that manipulation of specific conserved residues produces a cs protein that is rapidly degraded by the proteasomal system in the unbound state [2]. We used this knowledge to generate csAPEX-GBP; (schematically depicted in Fig 2A) and introduced the analogous residue changes to APEX-ChBP (generating csAPEX-ChBP). Expression of csAPEX-GBP in cells lacking GFP co-expression resulted in only negligible cytosolic APEX signal (Fig 2B); however, in a small number of cells, restricted electron density was observed in a punctate distribution (Fig 2B inset). We hypothesise that this signal represents the residual expression of APEX-GBP in the process of proteasomal degradation. In contrast, co-expression of GFP produced a strong cytosolic signal (Fig 2C, quantitated in S2A Fig) and a complete loss of the punctate distribution observed in the csAPEX-GBP alone. The csAPEX-GBP protein showed efficient recruitment to different cellular compartments, including the plasma membrane, endosomes, and caveolae, showing the functionality of the csAPEX-GBP construct for detection of any GFP-tagged protein (Fig 2D–2F). Consistent results were obtained with csAPEX-ChBP- and mCherry-tagged markers (Fig 2G–2J). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Conditional stabilisation of GBP and ChBP, and detection of protein–protein interactions using bimolecular fluorescence complementation. A) Schematic illustrating detection of GFP-tagged POIs using csAPEX-GBP. The probe is degraded by the proteasome unless stabilized by interactions with a GFP-tagged protein, resulting in loss of any nonspecific, electron-dense APEX signal when csAPEX-GBP does not bind to its target. B) csAPEX-GBP shows minimal signal when expressed in cells lacking GFP-tagged proteins; only a low level of labelling is detectable in specific regions of a subset of cells (inset, arrows). In contrast, cells co-expressing soluble GFP together with csAPEX-GBP show a strong cytosolic signal (C, quantitated in S2 Fig. A). D-F) Examples of subcompartment-specific labelling in cells expressing GFP-tagged POIs associating with the PM, the early endosomes, and caveolae, respectively. G-H) Examples of subcompartment-specific labelling in cells expressing mCherry-tagged POIs associating with the PM, nucleus, early endosomes, and caveolae, respectively. K-P) Co-transfection of BHK cells with constructs tagged with each half of split YFP along with csAPEX-GBP gives strong and specific labelling at sites of protein–protein interactions. K) Schematic illustrating detection of interactions between two POIs tagged with different halves of a split YFP. csAPEX-GBP is able to bind only when the YFP pair is fully reconstituted and folded. In the absence of a correctly folded GFP derivative, csAPEX-GBP is degraded by the proteasome. L) Cavin1-YFP-N and Cavin3-YFP-C co-expression gives specific labelling associated with PM pits and vesicular profiles characteristic of caveolae. Note the specificity of the labelling, which allows identification of Cavin1/Cavin3 complexes associated with both surface caveolae and putative endocytic caveolar carriers associated with intracellular compartments (arrow). Further examples are shown in S2B and S2C Fig. M) Reciprocal experimental conditions with specific fragments of YFP switched between constructs gives consistent labelling. N) Cells with an abnormally high transfection level show intracellular aggregates of Cavin (compare with caveolar labelling in L and M). O) Control cells transfected with just one split GFP half and csAPEX-GBP show no labelling in the majority of cells. P) APEX positive inclusions are seen in a small percentage of control cells. These are clearly distinguishable from the specific staining of the recombined protein complex (L-M). Further examples are shown in S2 Fig. D. Scale bars: lower magnification = 1 μm; insets = 500 nm. BHK, baby hamster kidney; CCP, clathrin-coated pits; ChBP, mCherry-binding peptide; cs, conditionally stable; GBP, GFP-nanobody/binding peptide; PM, plasma membrane; POI, protein of interest. https://doi.org/10.1371/journal.pbio.2005473.g002 To confirm efficient degradation of our new, conditionally stable csAPEX-ChBP via the proteasomal pathway, we used the well-established proteasome inhibitor MG132 [12]. Cells expressing csAPEX-ChBP alone showed negligible reaction product following the DAB reaction, whereas cells expressing both csAPEX-ChBP and cytoplasmic mCherry showed intense staining throughout (S2B Fig). However, following a 5-h supplementation with 10 μM MG132, cells expressing csAPEX-ChBP alone retained DAB staining in the cytoplasm, indicating that, under normal conditions, csAPEX-ChBP is degraded by the proteasome. Bimolecular fluorescence complementation (BiFC) is a technique for testing pairwise protein-protein interactions in fixed or living cells by genetically tagging candidates with different halves of a “split” fluorescent protein [13]. If these candidates attain sufficient proximity, the full length fluorescent protein is reconstituted, can fold and emit photons under excitation by a suitable wavelength of light. We hypothesised that by using the conditionally stable APEX nanobody system, we should be able to extend the resolution of bifluorescence complementation to the ultrastructural level. Indeed, the nanobody binding site in GFP (and its variants) straddles the split site in commonly used BiFC pairs [13,14]. Furthermore, folding is absolutely required for the GFP–nanobody interaction, such that recognition of the unfolded halves of the split protein by GBP is a theoretical impossibility. Using this technique, we were able to directly visualize interactions between Cavin1 and Cavin3 by EM using split mVenus, a YFP derivative recognised by the GBP. We transfected BHK cells with vectors encoding Cavin1 fused to the N-terminal fragment of mVenus, (Cavin1-mVenus1–155), Cavin3 fused to the C-terminal fragment (Cavin3-mVenus156–239), and our csAPEX-GBP construct, schematically represented in Fig 2K. Using this technique, we were able to delineate surface caveolae and putative endocytic caveolar carriers associated with intracellular compartments (Fig 2L, further examples in S2C and S2D Fig). The reciprocal experiment, in which the N- and C-terminal fragments of mVenus were exchanged, showed similar results (Fig 2M). Unusually high-expressing cells were occasionally visible, showing aggregation of intracellular Cavin recognised by csAPEX-GBP (Fig 2N). Transfection with just one-half of the split YFP most commonly showed no cytoplasmic staining (Fig 2O). However, inclusions of increased density were sometimes noted in these controls (Fig 2P, further example in S2E Fig) and were absent from untransfected samples. This staining was clearly distinguishable from the specific signal shown in Fig 2L and 2M, although the importance of such controls is emphasized, particularly since different cell types may contain different numbers of proteasomes. These results clearly demonstrate that protein–protein interactions can be effectively visualized using bimolecular fluorescence complementation at the ultrastructural level using csAPEX-GBP. In summary, we have utilised cell-free expression and single-molecule analysis to screen a number putative ChBP for association with mCherry-tagged Caveolin-1. The single nanobody we identify is a selective, high-affinity binder of the mCherry tag, lacks detectable self-aggregation or cross-reactivity with GFP, can be linked to APEX for high-resolution analysis of mCherry-tagged proteins in cell culture systems, and is compatible with correlative light and EM. We have also employed conditional stabilisation of both GFP and mCherry binding nanobodies fused to APEX2 which results in the generation of an APEX reaction product only when bound to their target fluorescent proteins. By degrading unbound APEX-BP protein, this modification facilitates an improved signal-to-noise ratio and circumvents any potential oversaturation of the APEX-BP vector. Finally, we have coupled the csAPEX-GBP system with bimolecular fluorescence complementation. This now allows direct visualisation of intracellular protein–protein interactions at the ultrastructural level, far beyond the resolution of light microscopy. This system is immediately applicable (without any new cloning steps) to any system in which the fluorescent split GFP system has been used. Unlike labelling on sections, APEX methods are compatible with 3D EM methods [4] such as focused ion beam-scanning EM, serial blockface-scanning EM, and electron tomography and can be used in whole animal systems [5]. As cellular function depends not on single proteins but on protein–protein interactions, these methods will be a vital complement to dynamic light microscopic methods. Materials and methods Single-molecule counting and coincidence detection Single-molecule spectroscopy was performed as previously described [9]. Briefly, samples (20 μl) were loaded into a custom-made silicone 192-well plate adhered to glass coverslips (ProSciTech Australia). Samples were analysed with two lasers (488 nm and 561 nm) using a Zeiss LSM710 microscope with a Conforcor3 module for single-molecule counting and a single 488-nm laser for aggregation analyses. The fluorescence emission was filtered with 505–540-nm band pass filter (GFP) and 580-nm long-pass filter (mCherry). Measurements were taken with photon counts in the approximate range of 750–2,000 which corresponds to a GFP concentration of around 1–2.5 μg/ml. Three replicates were carried out for each construct pair, and consistent results were obtained for each. Cell culture BHK cells were passaged in Dulbecco’s Modified Eagle Medium (Gibco) supplemented with 10% Fetal Bovine Serum and L-Glutamine. Cells were seeded onto 35-mm culture dishes (TPP), transfected with Lipofectamine 3000 as per the manufacturer’s instructions and processed for EM 24 h later. For bimolecular fluorescence complementation experiments, an 8-h incubation in 50 μM cyclohexamide prior to processing was used to reduce background staining. EM EM was performed exactly as described previously [5,6]. Briefly, cells were fixed with 2.5% glutaraldehyde in 0.1-M sodium cacodylate buffer for 1 h at room temperature. Cells were washed with cacodylate buffer to remove the fixative, then washed with DAB in cacodylate buffer for 1 min and subsequently treated with DAB in cacodylate buffer containing H2O2 for 30 min at room temperature. Cells were post-fixed with 1% OsO4 for 2 min to provide contrast. Cells were then washed in water and serially dehydrated in increasing percentages of ethanol before serial infiltration with LX112 resin in a BioWave microwave (Pelco). Resin was polymerised to hardness at 60°C overnight. Ultrathin sections were cut on an ultramicrotome (UC6: Leica) and imaged at 80 kV on a JEOL1011 transmission electron microscope. Sections were not post-stained. Correlative light and EM Cells were grown on 35-mm gridded MatTek dishes (with an in-plane alphanumeric code) and co-transfected with nls-mCherry and APEX-ChBP. Live cell imaging was performed on an EVOS FL epifluorescent microscope (ThermoFisher Scientific) at 10x and 20x magnification. Cells were processed as described above with the following exceptions. Post-polymerisation, the flat-embedded cells were removed from the dish and the region of interest was trimmed using the now-imprinted grid coordinates on the block face. Ultrathin sections were cut, placed on a slot grid, and imaged on a Tecnai 12 transmission electron microscope fitted with a 4K x 4K LC1100 camera (Direct Electron) at 120 kV under the control of SerialEM. Low-magnification (4,400 XMag) montages were acquired at a binning of 1 and stitched together using the Blendmont program in IMOD. Correlation of light and EM images was performed using Photoshop (Adobe Inc.). Constructs and cloning Split mVenus constructs were made by first removing the Fos and Jun inserts from pcs_kmVenus1-155_FosLZ135-171 and pcs_kmVenus156-239_JunLZ253-289 using EcoRV/SpeI. Human Cavin1 and Cavin3 open reading frames were amplified by PCR using the primer tags forward 5′-AGCGGCGGCGGCTCTGATATC-3′ and reverse 5′-ACAAGAAAGCTGGGTACTAGT-3′ and subcloned using infusion (BD). The series of ChBP-GFP expression vectors for L. tarentolae expression were constructed by PCR subcloning from the original templates [7] into the cell-free gateway cloning vector ‘N-term 8xHis eGFP pCellFree_G03’ [8] (Genbank KJ541667) using the following primer tags: forward 5′-GGGGACAAGTTTGTACAAAAAAGCAGGCTC-3′, reverse 5′-GGGGACCACTTTGTACAAGAAAGCTGGGTT-3′. Previously described vectors used for expression or subcloning were pmCherry-N1 (Clontech PT3974-5), pEGFP-N1 (Clontech PT3027-5), GFP-CaaX(Kras) [15], GFP-2xFYVEhrs [16], mCherry-2xFYVEhrs [17], Cavin1-mCherry [10], Cavin2-GFP and Cavin3-GFP [18], pCSDEST2 [19], pDEST-Tol2-pA2, p5E-CMV/SP6, pME-mCherry-CaaX (Hras) and p3E-pA [20], APEX2-GBP, mKate2-P2A-APEX2-GBP, and pME-APEX2-NS [6]. All other constructs were made using the Multisite Gateway system (Invitrogen). These new vectors have been deposited in the Addgene repository with the following identifiers: APEX2-csGBP (108874), mKate2-P2A-APEX2-csGBP (108875), APEX2-csChBP (108876), EGFP-P2A-APEX2-csChBP (108877), APEX2-ChBP (108878), EGFP-P2A-APEX2-ChBP (108879), H2B-mCherry (108880), nls-mCherry (108881), pME-nls (108882), pME-H2B (108883), p3E-mCherry (108884), pME-mCherry-NS (108885), mCherry-CaaX(Hras) (108886), mVenusN-Cavin1 (108887), mVenusC-Cavin1 (108888), mVenusN-Cavin3 (108889), mVenusC-Cavin3 (108890), p3E-csGBP (108891), p3E-ChBP (108892), p3E-csChBP (108893), p3E-APEX2 (108894), pME-EGFP-P2A-APEX2-NS (108895), and p3E-APEX2-P2A-EGFP (108896). Single-molecule counting and coincidence detection Single-molecule spectroscopy was performed as previously described [9]. Briefly, samples (20 μl) were loaded into a custom-made silicone 192-well plate adhered to glass coverslips (ProSciTech Australia). Samples were analysed with two lasers (488 nm and 561 nm) using a Zeiss LSM710 microscope with a Conforcor3 module for single-molecule counting and a single 488-nm laser for aggregation analyses. The fluorescence emission was filtered with 505–540-nm band pass filter (GFP) and 580-nm long-pass filter (mCherry). Measurements were taken with photon counts in the approximate range of 750–2,000 which corresponds to a GFP concentration of around 1–2.5 μg/ml. Three replicates were carried out for each construct pair, and consistent results were obtained for each. Cell culture BHK cells were passaged in Dulbecco’s Modified Eagle Medium (Gibco) supplemented with 10% Fetal Bovine Serum and L-Glutamine. Cells were seeded onto 35-mm culture dishes (TPP), transfected with Lipofectamine 3000 as per the manufacturer’s instructions and processed for EM 24 h later. For bimolecular fluorescence complementation experiments, an 8-h incubation in 50 μM cyclohexamide prior to processing was used to reduce background staining. EM EM was performed exactly as described previously [5,6]. Briefly, cells were fixed with 2.5% glutaraldehyde in 0.1-M sodium cacodylate buffer for 1 h at room temperature. Cells were washed with cacodylate buffer to remove the fixative, then washed with DAB in cacodylate buffer for 1 min and subsequently treated with DAB in cacodylate buffer containing H2O2 for 30 min at room temperature. Cells were post-fixed with 1% OsO4 for 2 min to provide contrast. Cells were then washed in water and serially dehydrated in increasing percentages of ethanol before serial infiltration with LX112 resin in a BioWave microwave (Pelco). Resin was polymerised to hardness at 60°C overnight. Ultrathin sections were cut on an ultramicrotome (UC6: Leica) and imaged at 80 kV on a JEOL1011 transmission electron microscope. Sections were not post-stained. Correlative light and EM Cells were grown on 35-mm gridded MatTek dishes (with an in-plane alphanumeric code) and co-transfected with nls-mCherry and APEX-ChBP. Live cell imaging was performed on an EVOS FL epifluorescent microscope (ThermoFisher Scientific) at 10x and 20x magnification. Cells were processed as described above with the following exceptions. Post-polymerisation, the flat-embedded cells were removed from the dish and the region of interest was trimmed using the now-imprinted grid coordinates on the block face. Ultrathin sections were cut, placed on a slot grid, and imaged on a Tecnai 12 transmission electron microscope fitted with a 4K x 4K LC1100 camera (Direct Electron) at 120 kV under the control of SerialEM. Low-magnification (4,400 XMag) montages were acquired at a binning of 1 and stitched together using the Blendmont program in IMOD. Correlation of light and EM images was performed using Photoshop (Adobe Inc.). Constructs and cloning Split mVenus constructs were made by first removing the Fos and Jun inserts from pcs_kmVenus1-155_FosLZ135-171 and pcs_kmVenus156-239_JunLZ253-289 using EcoRV/SpeI. Human Cavin1 and Cavin3 open reading frames were amplified by PCR using the primer tags forward 5′-AGCGGCGGCGGCTCTGATATC-3′ and reverse 5′-ACAAGAAAGCTGGGTACTAGT-3′ and subcloned using infusion (BD). The series of ChBP-GFP expression vectors for L. tarentolae expression were constructed by PCR subcloning from the original templates [7] into the cell-free gateway cloning vector ‘N-term 8xHis eGFP pCellFree_G03’ [8] (Genbank KJ541667) using the following primer tags: forward 5′-GGGGACAAGTTTGTACAAAAAAGCAGGCTC-3′, reverse 5′-GGGGACCACTTTGTACAAGAAAGCTGGGTT-3′. Previously described vectors used for expression or subcloning were pmCherry-N1 (Clontech PT3974-5), pEGFP-N1 (Clontech PT3027-5), GFP-CaaX(Kras) [15], GFP-2xFYVEhrs [16], mCherry-2xFYVEhrs [17], Cavin1-mCherry [10], Cavin2-GFP and Cavin3-GFP [18], pCSDEST2 [19], pDEST-Tol2-pA2, p5E-CMV/SP6, pME-mCherry-CaaX (Hras) and p3E-pA [20], APEX2-GBP, mKate2-P2A-APEX2-GBP, and pME-APEX2-NS [6]. All other constructs were made using the Multisite Gateway system (Invitrogen). These new vectors have been deposited in the Addgene repository with the following identifiers: APEX2-csGBP (108874), mKate2-P2A-APEX2-csGBP (108875), APEX2-csChBP (108876), EGFP-P2A-APEX2-csChBP (108877), APEX2-ChBP (108878), EGFP-P2A-APEX2-ChBP (108879), H2B-mCherry (108880), nls-mCherry (108881), pME-nls (108882), pME-H2B (108883), p3E-mCherry (108884), pME-mCherry-NS (108885), mCherry-CaaX(Hras) (108886), mVenusN-Cavin1 (108887), mVenusC-Cavin1 (108888), mVenusN-Cavin3 (108889), mVenusC-Cavin3 (108890), p3E-csGBP (108891), p3E-ChBP (108892), p3E-csChBP (108893), p3E-APEX2 (108894), pME-EGFP-P2A-APEX2-NS (108895), and p3E-APEX2-P2A-EGFP (108896). Supporting information S1 Fig. APEX-ChBP for EM-based subcellular localisation of Cherry-tagged proteins. Six putative ChBPs were GFP-tagged and co-expressed in cell free Leishmania lysate with mCherry tagged Caveolin1. A) ChBP1, B) ChBP2, C) ChBP3, D) ChBP4, E) ChBP6, F) ChBP8. Left-hand panels show GFP intensity through the confocal volume determined by single-molecule counting over time. Middle panels show simultaneous detection of coincidence of ChBP-GFP and mCherry-Cav1 over time. Right-hand panels show plots of the coincidence ratio between red and green channels. Only ChBP2 demonstrated a lack of self-aggregation/cross-reactivity with GFP (B, left panel), equivalent detection of red and green signal intensity over time (B, middle panel), and a 1:1 coincidence ratio of GFP to mCherry (B, right panel). Data underlying all middle panels is available in S1 Data. ChBP, mCherry binding peptide. https://doi.org/10.1371/journal.pbio.2005473.s001 (TIF) S2 Fig. Detection of protein-protein interactions using bimolecular fluorescence complementation. A) Quantitation of the effect of GFP presence on stabilisation of the conditionally stable APEX-GBP. Co-transfection of GFP with the csAPEX-GBP construct results in greater than 40% of cells with cytoplasmic density, compared to approximately 5% with transfection of csAPEX-GBP alone. Chi squared, p < 0.0001. See also Fig 2C. B) Validation of proteasome-mediated degradation of conditionally stable ChBP. Cells expressing csAPEX-ChBP alone show negligible reaction product following the DAB reaction (top row), whereas cells expressing both csAPEX-ChBP and cytoplasmic mCherry show intense staining throughout (middle row). Follow a 5-hr supplementation with 10 uM MG132, cells expressing csAPEX-ChBP alone retain DAB staining in the cytoplasm indicating that under normal conditions csAPEX-ChBP is degraded by the proteasome (bottom row). C-E) Further examples of Cavin1-YFP-N and Cavin3-YFP-C co-expression giving specific labelling associated with PM pits and vesicular profiles characteristic of caveolae. See also Fig 2L and 2M. E) Further example of APEX positive inclusions are seen in a small percentage of control cells. See also Fig 2P). Scale bars: B = 20 μm, C–E = 1 μm. Data underlying panel A is available in S2 Data. GBP, GFP binding peptide; PM, plasma membrane. https://doi.org/10.1371/journal.pbio.2005473.s002 (TIF) S1 Data. Data used to generate the middle panels in S1A–S1F Fig. https://doi.org/10.1371/journal.pbio.2005473.s003 (XLSX) S2 Data. Data used to generate the middle panels in S2 Fig. https://doi.org/10.1371/journal.pbio.2005473.s004 (XLSX) Acknowledgments The authors acknowledge the facilities of the Australian Microscopy & Microanalysis Research Facility at the Centre for Microscopy and Microanalysis, The University of Queensland, and the Australian Cancer Research Foundation (ACRF)/Institute for Molecular Bioscience (IMB) Dynamic Imaging Facility for Cancer Biology, established with funding from the ACRF. Professor Michael P. Rout supplied the initial mCherry nanobody vector series. Dr Andy Badrock supplied the vector backbones for the split-mVenus expression vectors. Professor Fred Meunier provided intellectual input into experimental design. We are particularly grateful to Associate Professor Brett Collins for advice on the GBP/split-YFP interaction.

doi: 10.1371/journal.pbio.2002846pmid: 29659566

Data as an asset Research data continue to be produced at ever-growing rates due to both technological advances and decreasing costs for their generation [1]. Understanding what makes data assets distinct from other types of assets is fundamental in terms of their valuation and effective management [2]. To briefly summarise, from an economic perspective, its unique characteristics are these: Information is infinitely shareable without any decrease in its intrinsic value. For example, the same sequence retrieved from the National Center for Biotechnology Information (NCBI) can be shared by an unlimited number of people without any loss of value. Unlike physical assets—e.g., sequencing equipment, which depreciates with use—information sharing actually increases its value in a compound fashion; and reciprocally, unshared information is less valuable [3,4,5]. Further, the more accurate and complete the information is, the more valuable it is. In other words, quality is at least as important as quantity [6,7,8]. Since inferences are only as good as the information they are based upon, inaccuracies and omissions compel scientists to spend valuable research time winnowing out poor-quality or inaccurate information or, even worse, inadvertently ploughing research funds into dead ends. Moreover, with the increasing role of automatic inference systems for high-throughput data and data analytics, there is a growing dependency on the availability of robust, high-quality knowledge resources, and the gold-standard data sets they contain, for benchmarking. Lastly, when information is combined, its value increases. For example, genetic testing can reveal hundreds of thousands of variants per individual, yet for most variants, the clinical consequences are not yet known [9]. If our goal is to advance research, instantiation of known connections is essential to accelerate the process of relating genotypes to phenotypes in a way that is impossible when using individual data sets in isolation [10,11,12,13,14]. Managing a biological information resource relies on a range of intersecting skills: Bioinformaticians, application developers, system administrators, biocurators, journal editors, etc. are all involved in this collective effort. Within this context, biocurators focus on information content rather than technology. Their overarching goal is to maximise the value of the information assets researchers are generating by assuring their accuracy, comprehensiveness, integration, accessibility, and reuse. What is biocuration? Biocuration is the extraction of knowledge from unstructured biological data into a structured, computable form. In this context, knowledge is most commonly extracted from published manuscripts, as well as from other sources such as experimental data sets and unpublished results from data analysis. Biocurators are typically PhD-level biologists, often with lab bench experience coupled with specialised expertise in computational knowledge representation. Their work entails the synthesis and integration of information from multiple sources—including, for example, peer-reviewed papers; large-scale projects, such as the Encyclopedia of DNA Elements (ENCODE); or conference abstracts. They contact authors directly for clarification, digest supplemental information, and resolve identifiers in order to accurately capture a researcher’s conclusion and their evidence for that conclusion. Biocurators strive to distil the current ‘best view’ from conflicting sources and ensure that their resources provide data that are not only findable, accessible, interoperable, and reproducible (FAIR), but also traceable, appropriately licensed, and interconnected (collectively, the FAIR-TLC principles [15]). Biocuration motivation Scientific communication is shifting in this ‘information age’, with researchers increasingly relying on curated resources [16,17,18,19]. For example, when comparing an entry in the Worldwide Protein Data Bank (wwPDB; https://www.wwpdb.org)—a resource containing detailed reviewed information on macromolecular structures—with a portable document format (PDF) file containing a figure of the same structure, it is obvious that the latter, non-computer-readable representation is insufficient for downstream comparative use. The political processes in the scientific community that led to designating wwPDB [20], the International Nucleotide Sequence Database Collaboration [21], and others such as the International Molecular Exchange (IMEx) [22] and ProteomeXchange consortia [23] as official depositories have proven to be well worth the effort. These examples highlight the importance of collaboration and synergy between journal editors and databases. The definition of what it means to publish is expanding [24], since results only published as a PDF have limited accessibility. To promote impact and reuse, the full semantic spectrum must be employed, from human-readable language to fully computationally interpretable. Biocuration costs Although expert biocuration is clearly labour intensive, it scales surprisingly well with the growth of biomedical literature, as demonstrated by two recent studies [25,26]. Advanced tools are also increasing efficiency and accuracy, and biocurators are often actively engaged as team members in developing machine learning and natural-language processing techniques. Although these methods currently lack the necessary precision and recall required for a real-world setting [27,28,29], they are beginning to provide assistance [30,31,32,33,34] and will continue to incrementally improve. The costs for sustaining a useful research resource in which biocuration plays an essential role represent only a tiny fraction of the original research funding [35]. An independent survey assessing the value of biological database services concluded that the benefits to users and their funders are equivalent to more than 20 times the direct operational cost of the institute [36]. Additionally, the hidden cost of an individual researcher’s time spent trawling the literature to find the information pertinent to their own specialist field is impossible to estimate, but having the required data easily accessible in a structured format represents a considerable saving in person-hours and, therefore, money for every funder, academic institute, and biomedical enterprise. Actionable recommendations Everyone can be a biocurator—Data reporting fit for knowledge synthesis Seriously addressing seemingly mundane issues—such as identifying gene symbols, isoforms, strains, antibodies, and cell lines—is essential if experimental results are to be correctly integrated within the existing body of knowledge. For example, a recent study found that almost 40% of the gene lists submitted to the Gene Expression Omnibus (GEO) and 20% of the gene lists in the supplementary material of published articles contain gene symbol errors introduced by the software used during data handling prior to publication [37]. This will continue to be a significant problem until infrastructure is in place at key junctions in the research life cycle. New tools and workflows are needed for connecting researchers, journals, reviewers, and repositories and easily conveying standards-compliant information. Progress is being made; notably, community guides for provisioning and referencing life science identifiers have recently been published [38,39], outlining best practices for facilitating large-scale data integration. Likewise, in the lab, software applications that support autocompletion within individual cells of spreadsheets, as well as more sophisticated standards-aware data collection tools, ensure that standard terminologies are applied as data are collected [40,41,42]. Through the use of such electronic laboratory notebook and manuscript submission software and the adoption of recommended formats and community-endorsed terminologies and ontologies, the goal of ‘born computable’ lab data generation will be realised. Initiatives have also started in scientific journals. A good example is provided by SourceData, a project initiated by the European Molecular Biology Organization (EMBO) press, which involves the biocuration of article figures prior to publication [43]. Support for standards—Development, usage, and sustainability Common standards for describing and classifying biology are indispensable for reproducible interactions, information exchange, interoperability, comparability, and discoverability [44]. Without standards, database search results will inevitably miss key information or include irrelevant material. Biocurators regularly lead efforts in standards development: engaging with experts, building consensus, fostering adoption, and maintaining biological fidelity. Yet apart from a very limited number of cases, funding for standards development is unavailable. Even in the case of the Gene Ontology Consortium [45], the funding for this indispensable standard is significantly aided through other projects. On the other side of the spectrum, the Human Phenotype Ontology [46,47,48] operates using donated time from a handful of dedicated individuals, despite its widespread adoption (e.g., the Unified Medical Language System [UMLS], United Kingdom 100,000 Genomes Project, and the Global Alliance for Genomics and Health [GA4GH]). While the lack of dedicated funding poses a risk, the harmful consequences of not using any standard are vastly greater. More can be done to inform and educate data producers and consumers on the importance of standards to ensure research data are not wasted or lost in the wrong format, with the wrong metadata descriptions, or described using a private or personal set of terms. Efforts such as FAIRsharing [30] (fairsharing.org), which maps the landscape of databases and standards and links them to the journal and funder data policies that endorse their use, go a long way to making sure that existing standards are adopted. However, more funding is needed for these infrastructure projects to aid data and knowledge sharing, to minimise the duplication of effort, and to ensure that researchers can easily employ appropriate standards. Expediting the collection and processing of data Recently, there has been considerable excitement about the strategy of crowdsourcing, putting biocuration tools into a researcher’s hands so that they may directly contribute and publish their results into knowledge resources [49,50,51,52]. There is a tremendous potential in this approach, but to ensure success, there are clear prerequisites that must be satisfied—(i) editorial oversight, (ii) automated integrity checks, and (iii) citation mechanisms. Successful community-sourced projects universally include editorial control, which is where biocurators can play a key role, to avoid collecting poor-quality data that would decrease the value of a resource overall. In addition, support for developing user interfaces, batch submission tools, and utilities to computationally validate content—such as simple checks for syntactical correctness, falling outside standard deviations, or using disallowed values—is needed for direct data submission. Here again, biocurators often play a role in defining validation standards. Machine-readable standards are critical in this step, as they enable validation to be carried out programmatically. Continuous integration and contextual analysis approaches may even suggest what a contributor might do to improve their data before making a final submission. Notably, biologists are already beginning to use community curation tools when they are available, such as Canto [53]—which is used by researchers working on Schizosaccharomyces pombe to directly submit their data to a resource—and Apollo [54], which is used for community-based curation of gene structures for improving automated gene sets. Lastly, citation mechanisms need to be built into the contribution process. This both acts as an incentive and fosters reproducibility, since information is traceable to the original experimental work that led to a conclusion. Currently, existing biological data resources associate every assertion they contain with its underlying experimental justification by linking it to a PubMed identifier, which is an indirect route to the actual researcher(s) who contributed this information. Literature citations are mere proxies for assessing productivity and impact. Embedding a traceable authorship facility directly into laboratory software or a resource’s submission software would provide a much more direct and accurate means of assessing a researcher’s impact. By associating a researcher (e.g., an Open Researcher and Contributor ID [ORCID] persistent identifier, https://orcid.org/) with an identified piece of information (e.g., a persistent identifier, such as a digital object identifier [DOI]), their contributions become citable objects [55,56,57], and the subsequent use of this information by other researchers can be tracked. If this is encouraged, one can envision a time when community curation tools become the first place for digitally publishing research conclusions, shared directly into digital community resources. Biocuration is a necessity for scientific progress Actively promoting innovations in fundamental data and information capture will yield enormous return on our research investment. The existing pain points—the time wasted by individual researchers discovering information, collecting it, manually verifying it, and integrating it in a piecemeal fashion—all impede scientific advancement. For researchers, biocuration means they can easily find extensive and interlinked information at well-documented, stable resources. It means they can access this information through multiple channels by browsing websites, downloading it from repositories, or retrieving it dynamically via web services. It likewise means the information will be as accurate and reliable as possible. And—because biocurators have integrated information by describing it using community semantic standards, applying authoritative identifiers, and transforming it into standard formats—disparate data sets collected from multiple research projects can be directly compared. Everyone can be a biocurator—Data reporting fit for knowledge synthesis Seriously addressing seemingly mundane issues—such as identifying gene symbols, isoforms, strains, antibodies, and cell lines—is essential if experimental results are to be correctly integrated within the existing body of knowledge. For example, a recent study found that almost 40% of the gene lists submitted to the Gene Expression Omnibus (GEO) and 20% of the gene lists in the supplementary material of published articles contain gene symbol errors introduced by the software used during data handling prior to publication [37]. This will continue to be a significant problem until infrastructure is in place at key junctions in the research life cycle. New tools and workflows are needed for connecting researchers, journals, reviewers, and repositories and easily conveying standards-compliant information. Progress is being made; notably, community guides for provisioning and referencing life science identifiers have recently been published [38,39], outlining best practices for facilitating large-scale data integration. Likewise, in the lab, software applications that support autocompletion within individual cells of spreadsheets, as well as more sophisticated standards-aware data collection tools, ensure that standard terminologies are applied as data are collected [40,41,42]. Through the use of such electronic laboratory notebook and manuscript submission software and the adoption of recommended formats and community-endorsed terminologies and ontologies, the goal of ‘born computable’ lab data generation will be realised. Initiatives have also started in scientific journals. A good example is provided by SourceData, a project initiated by the European Molecular Biology Organization (EMBO) press, which involves the biocuration of article figures prior to publication [43]. Support for standards—Development, usage, and sustainability Common standards for describing and classifying biology are indispensable for reproducible interactions, information exchange, interoperability, comparability, and discoverability [44]. Without standards, database search results will inevitably miss key information or include irrelevant material. Biocurators regularly lead efforts in standards development: engaging with experts, building consensus, fostering adoption, and maintaining biological fidelity. Yet apart from a very limited number of cases, funding for standards development is unavailable. Even in the case of the Gene Ontology Consortium [45], the funding for this indispensable standard is significantly aided through other projects. On the other side of the spectrum, the Human Phenotype Ontology [46,47,48] operates using donated time from a handful of dedicated individuals, despite its widespread adoption (e.g., the Unified Medical Language System [UMLS], United Kingdom 100,000 Genomes Project, and the Global Alliance for Genomics and Health [GA4GH]). While the lack of dedicated funding poses a risk, the harmful consequences of not using any standard are vastly greater. More can be done to inform and educate data producers and consumers on the importance of standards to ensure research data are not wasted or lost in the wrong format, with the wrong metadata descriptions, or described using a private or personal set of terms. Efforts such as FAIRsharing [30] (fairsharing.org), which maps the landscape of databases and standards and links them to the journal and funder data policies that endorse their use, go a long way to making sure that existing standards are adopted. However, more funding is needed for these infrastructure projects to aid data and knowledge sharing, to minimise the duplication of effort, and to ensure that researchers can easily employ appropriate standards. Expediting the collection and processing of data Recently, there has been considerable excitement about the strategy of crowdsourcing, putting biocuration tools into a researcher’s hands so that they may directly contribute and publish their results into knowledge resources [49,50,51,52]. There is a tremendous potential in this approach, but to ensure success, there are clear prerequisites that must be satisfied—(i) editorial oversight, (ii) automated integrity checks, and (iii) citation mechanisms. Successful community-sourced projects universally include editorial control, which is where biocurators can play a key role, to avoid collecting poor-quality data that would decrease the value of a resource overall. In addition, support for developing user interfaces, batch submission tools, and utilities to computationally validate content—such as simple checks for syntactical correctness, falling outside standard deviations, or using disallowed values—is needed for direct data submission. Here again, biocurators often play a role in defining validation standards. Machine-readable standards are critical in this step, as they enable validation to be carried out programmatically. Continuous integration and contextual analysis approaches may even suggest what a contributor might do to improve their data before making a final submission. Notably, biologists are already beginning to use community curation tools when they are available, such as Canto [53]—which is used by researchers working on Schizosaccharomyces pombe to directly submit their data to a resource—and Apollo [54], which is used for community-based curation of gene structures for improving automated gene sets. Lastly, citation mechanisms need to be built into the contribution process. This both acts as an incentive and fosters reproducibility, since information is traceable to the original experimental work that led to a conclusion. Currently, existing biological data resources associate every assertion they contain with its underlying experimental justification by linking it to a PubMed identifier, which is an indirect route to the actual researcher(s) who contributed this information. Literature citations are mere proxies for assessing productivity and impact. Embedding a traceable authorship facility directly into laboratory software or a resource’s submission software would provide a much more direct and accurate means of assessing a researcher’s impact. By associating a researcher (e.g., an Open Researcher and Contributor ID [ORCID] persistent identifier, https://orcid.org/) with an identified piece of information (e.g., a persistent identifier, such as a digital object identifier [DOI]), their contributions become citable objects [55,56,57], and the subsequent use of this information by other researchers can be tracked. If this is encouraged, one can envision a time when community curation tools become the first place for digitally publishing research conclusions, shared directly into digital community resources. Biocuration is a necessity for scientific progress Actively promoting innovations in fundamental data and information capture will yield enormous return on our research investment. The existing pain points—the time wasted by individual researchers discovering information, collecting it, manually verifying it, and integrating it in a piecemeal fashion—all impede scientific advancement. For researchers, biocuration means they can easily find extensive and interlinked information at well-documented, stable resources. It means they can access this information through multiple channels by browsing websites, downloading it from repositories, or retrieving it dynamically via web services. It likewise means the information will be as accurate and reliable as possible. And—because biocurators have integrated information by describing it using community semantic standards, applying authoritative identifiers, and transforming it into standard formats—disparate data sets collected from multiple research projects can be directly compared. Acknowledgments Most instrumental for the composition of this paper are the members of the International Society for Biocuration (ISB) themselves, whose valuable contributions, helpful comments, sound suggestions, and eye for details helped to shape this article. Special thanks go to these ISB members, in alphabetical order: Mais Ammari, Andrew Chatr Aryamontri, Helen Attrill, Amos Bairoch, Tanya Berardini, Judith Blake, Qingyu Chen, Julio Collado, Delphine Dauga, Joel T. Dudley, Stacia Engel, Ivan Erill, Petra Fey, Richard Gibson, Henning Hermjakob, Gemma Holliday, Doug Howe, Chris Hunter, David Landsman, Ruth Lovering, Deepa Manthravadi, Aron Marchler-Bauer, Beverley Matthews, Ellen M. McDonagh, Birgit Meldal, Gos Micklem, Daniel Mietchen, Christopher J. Mungall, Kim Pruitt, Vidhya Sagar Rajamanickam, James M. Reecy, Alix Rey, Khader Shameer, Aleksandra Shipitsyna, Ana Luisa Toribio, Mary Ann Tuli, Peter Uetz, Ulrike Wittig, Valerie Wood, and all the many other ISB members (biocuration.org). For their contribution, we recognise previous members of the ISB executive committee: Teresa Attwood, Alex Bateman, Tanya Berardini, Lydie Bougueleret, Pascale Gaudet, Jennifer Harrow, Tadashi Imanishi, Renate Kania, Lorna Richardson, Marc Robinson-Rechavi, Owen White, Ioannis Xenarios, and Chisato Yamasaki. Appreciation for steering the manuscript forward goes to the following members of current and recent ISB executive committees: Cecilia N. Arighi, Rama Balakrishnan, J. Michael Cherry, Melissa Haendel, Suzanna E. Lewis, Peter McQuilton, Monica Muñoz-Torres, Claire O’Donovan, Sandra Orchard, Sylvain Poux, Andrew Su, Nicole Vasilevsky, and Zhang Zhang.

doi: 10.1371/journal.pbio.2005377pmid: 29668686

Introduction Endocytosis has continued to evolve from a simple mode of ingestion and compartmentalization into a complex, multicomponent process that developed a bidirectional relationship with surface signaling [1,2]. In particular, evolutionary steps towards this complexity, which are associated with multicellularity, include the expansion to multiple isoforms of endocytic accessory proteins [3,4] and the introduction of dynamin [4,5]. Dynamin is the prototypical member of a family of large Guanosine Triphosphate hydrolases (GTPases) that catalyze membrane fission and fusion [6–8]. While encoded by single genes in Drosophila and Caenorhabditis elegans, further expansion of endocytic dynamins to three differentially expressed isoforms occurred in vertebrates [9]. Dynamin-1 (Dyn1), the first identified vertebrate isoform, has been extensively studied, and its mechanism of action as a fission GTPase is well understood [6,8,10]. The three dynamin isoforms are >70% identical in sequence, with most differences occurring in the C-terminal proline/arginine rich domain (PRD) that mediates interactions with numerous SRC Homology 3 (SH3) domain-containing binding partners. Dyn1 and Dyn3 appear to be functionally redundant [11]. However, Dyn2 is unable to substitute fully for Dyn1 or Dyn3 in supporting rapid synaptic vesicle recycling in neurons [12], and correspondingly, Dyn1 could not fully substitute for Dyn2 to support clathrin-mediated endocytosis (CME) in fibroblastic cells, even when overexpressed [13]. A direct comparison of the biochemical properties of Dyn1 and Dyn2 revealed differences in their in vitro curvature generating abilities: Dyn1 can potently induce membrane curvature and independently catalyze vesicle release from planar membrane surfaces, whereas Dyn2 requires the synergistic activity of curvature-generating Bin/Amphiphysin/Rvs (BAR) domain-containing proteins [14,15]. Less understood but still controversial [7,16–18] is dynamin’s suggested role in regulating early stages of CME [19–22]. Based on their differential biochemical properties, it was suggested that Dyn1 might be a more effective fission GTPase, while Dyn2 might be positioned to regulate early stages of CME [14]. However, whether dynamin isoforms play distinct roles in regulating CME has not been studied. Previously assumed to be neuron specific, recent studies have provided strong evidence that Dyn1 is indeed widely expressed but maintained in an inactive state in non-neuronal cells through phosphorylation at Serine 774 (S774) by the constitutively active kinase, glycogen synthase kinase-3 beta (GSK3β) [23]. Acute inhibition of GSK3β in retinal pigment epithelial (ARPE) cells accelerates CME due to increased rates of clathrin-coated pit (CCP) initiation and maturation [23]. The effects of GSK3β inhibition on CME depend on Dyn1 but not Dyn2, suggesting, unexpectedly, that Dyn1 might selectively function to regulate early stages of CME in non-neuronal cells. As the GSK3β phosphorylation site, S774, is located within the PRD, its phosphorylation is presumed to alter interactions with dynamin’s SH3 domain-containing binding partners, as has been shown for binding partners enriched in the synapse [24,25]. Which interactions are affected in non-neuronal cells and whether these might be dynamin isoform specific is not known. Immunoelectron microscopic studies using an antibody that recognizes both Dyn1 and Dyn2 have localized endogenous dynamin to both flat and deeply invaginated CCPs in A431 adenocarcinoma cells [26,27]. Live-cell imaging has shown that, when overexpressed, both Dyn1-eGFP and Dyn2-eGFP are recruited at low levels to nascent CCPs, that their association with CCPs fluctuates, and that they undergo a burst of recruitment prior to membrane scission and vesicle release [17,22,28–31]. Indeed, when compared directly, transiently overexpressed Dyn1-eGFP and Dyn2-eGFP had indistinguishable profiles for their recruitment to CCPs [30,31]. Analysis of the recruitment of genome-edited Dyn2-eGFP to CCPs has similarly revealed a burst of recruitment at late stages of CME, as well as more transient interactions of lower numbers of Dyn2 molecules during earlier stages of CCP maturation [17,32]. To date, direct and quantitative comparisons of the nature of Dyn1 and Dyn2 association with CCPs when they are expressed at endogenous levels do not exist, nor is it known how activation of Dyn1 affects its association with CCPs. Here, we explore the isoform-specific behaviors of genome-edited Dyn1 and Dyn2, both at steady state and in cells where Dyn1 is activated. We provide evidence for an early function of low levels of activated Dyn1 in regulating CCP initiation and maturation rates and that sorting nexin 9 (SNX9) serves as an isoform-selective and activity-dependent binding partner of Dyn1 to regulate CCP maturation. Finally, we show that Dyn1 can be activated, under physiological conditions, downstream of epidermal growth factor receptors (EGFRs) to alter CCP dynamics. Results Dynamin isoforms are differentially recruited to CCPs Recent studies have shown that Dyn1 is widely expressed in non-neuronal cells [2]; but, like at the neuronal synapse [33], it is mostly inactive at steady state due to phosphorylation by the constitutively active kinase GSK3β. Dyn1 function and its recruitment to CCPs have been studied in non-neuronal cells, albeit under conditions of overexpression and/or without an awareness of its phosphoregulation [14,21]. Therefore, to explore potential isoform-specific functions of Dyn1 and Dyn2, as well as the role of GSK3β in regulating Dyn1 activity, we generated genome-edited H1299 non-small cell lung cancer cells, which we previously showed partially utilize Dyn1 for CME [23]. Cells expressing endogenously tagged Dyn2-mRuby2 were generated using previously validated Zinc Finger Nucleases (ZFNs) [32,34] to introduce double-stranded breaks and insert the mRuby tag with complementary flanking regions by homology-driven repair (HDR) (Fig 1A). The resulting cells were single-cell sorted for mRuby2 fluorescence to obtain a heterozygous clone (clone 235, designated Dyn2-mRuby2end) expressing a single mRuby2-tagged allele of Dyn2 (Fig 1B). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Isoform-specific differences in recruitment of dynamin to CCPs. (A) Diagram of Zinc-finger and CRISPR/Cas9n knock-in strategies for endogenous labeling of Dyn2 and Dyn1 in H1299 cells with C-terminal mRuby2 and eGFP tags, respectively. For Dyn2, a short linker and mRuby2 (red) were placed at the stop codon in exon 22. For canonical Dyn1 splice isoform “a,” the 19 C-terminal amino acids (blue) were inserted in exon 21, followed by a short linker, eGFP (green), with stop codon and a polyadenylation sequence (yellow). In both constructs, flanking homology arms (HAs) of roughly 800 bp were used to promote recombination (dashed lines). See S1 Fig for details. (B) Western blot analysis of tagged isoforms. The low levels of Dyn1 in H1299 cells could not be directly detected by western blotting but can be detected after pulldown with GST-Amphiphysin II SH3 domains. Representative TIRF images (see S1 and S2 Movies) showing membrane recruitment of endogenous Dyn2-mRuby2end (C) or Dyn1a-eGFPend (E) and corresponding lentiviral transduced SNAP(647)-CLCa images. (D,F) Clathrin labeled puncta were identified and thresholded to define bona fide CCPs [22]. Shown are the averaged kinetics of recruitment of SNAP-CLCa and Dyn2-mRuby2end (D) or Dyn1a-eGFPend (F) for all tracks with lifetimes between 40 and 60 s (831 CCPs from 5 movies containing a total of 15 cells for Dyn2-mRuby2end and 13,346 CCPs from 10 movies containing a total of 29 cells for Dyn1a-eGFPend). CLCa, clathrin light chain a; CRISPR/Cas9n, Clustered Regularly Interspaced Short Palindromic Repeats-associated nucleases 9 nickase; Dyn1, dynamin-1; GST, Glutathione S-transferase; HA, homology arm; SH3, SRC Homology 3. https://doi.org/10.1371/journal.pbio.2005377.g001 Endogenously tagging Dyn1 was complicated by the fact that the DNM1 gene encodes C-terminal splice variants derived from differential splicing of exons 21 and 22 (S1A Fig), whose differential utilization could lead to partial loss of the fusion tag. Previous studies involving CRISPR/Cas9-mediated knockout and reconstitution with the Dyn1a C-terminal splice variant had confirmed that it fully reconstituted the GSK3β phosphoregulated activity of endogenous Dyn1 in H1299 cells [23], including its ability to be activated by calmodulin [35]. Therefore, using a Clustered Regularly Interspaced Short Palindromic Repeats-associated nucleases 9 nickase (CRISPR/Cas9n) strategy, we targeted the Dyn1 gene at the end of exon 21 and introduced sequences encoding the remaining 19 amino acids of the Dyn1a isoform, followed by a seven amino acid linker [32], monomeric eGFP fusion tag with stop codon, and finally, the SV40 polyadenylation signal to ensure unique expression of the “a” splice variant (Fig 1A and S1B Fig). Single-cell sorting by fluorescence-activated cell sorting (FACS) for eGFP fluorescence, followed by clonal amplification generated a heterozygous clone (clone 1B6, designated Dyn1a-eGFPend) expressing one eGFP-tagged allele of Dyn1a (Fig 1B). Note that although Dyn1 is expressed at very low levels in H1299 cells, it can be readily detected following enrichment by amphiphysin-II SH3 domain pulldown. As a robust fiduciary marker for CCPs, clathrin light chain a (CLCa) carrying an N-terminal SNAP-fusion tag was stably introduced in parallel into both cell lines via a lentiviral vector with puromycin selection of SNAP-CLCa expressing cells. As previously reported by several groups, mild overexpression of Fluorescent Probe (FP)-CLCa has no effect on CME as measured by transferrin endocytosis [22,31,32,36] and no effect on CCP dynamics compared to AP2 or other markers [20,29,31]. We then performed live-cell dual-channel total internal reflection fluorescence microscopy (TIRFM) and analyzed CCP dynamics and Dyn recruitment using the master–slave (CLCa–Dyn) approach introduced with the cmeAnalysis software [22,37,38]. As expected based on previous studies using either overexpressed [28,29,31] or endogenously tagged Dyn2 [17,22,32], Dyn2-mRuby2end was observed, on average, to gradually accumulate and then exhibit a burst of recruitment coincident with clathrin-coated vesicle (CCV) release. This can be seen in class-averaged tracks of bona fide CCPs with lifetimes ranging from 40–60 s (Fig 1C and 1D) and in all other CCP lifetime cohorts (S2A and S2B Fig, S1 movie). In contrast, Dyn1a-eGFPendo recruitment was barely detectable above background, and no burst was evident (Fig 1E and 1F, S2C and S2D Fig, S2 movie). This could reflect isoform-specific differences, very low levels of Dyn1 expression relative to Dyn2, and/or the inactivation of Dyn1 by GSK3β phosphorylation. Thus, we further explored these possibilities. Inhibition of constitutively active GSK3β kinase stimulates Dyn1 to accelerate CCP initiation and maturation We first tested whether activation of Dyn1 alters CCP dynamics and/or the recruitment of Dyn1a-eGFPend in H1299 cells. As expected based on earlier studies in ARPE cells [23], we confirmed that acute inhibition of GSK3β by incubation with the specific inhibitor, CHIR99021, leads to decreased phosphorylation of Dyn1 at S774 within 30 min (Fig 2A and 2B) and increased rates of CME, as measured by transferrin receptor (TfnR) internalization (Fig 2C). Importantly, the effects of GSK3β inhibition were dependent on Dyn1 expression, as treatment of Dyn1 knockout (Dyn1KO) H1299 cells [23] with CHIR99021 had no effect on CME (Fig 2C). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Activated Dyn1 regulates early stages of CME, even when recruited at low levels to CCPs. (A) Schematic representation of Dyn1 regulation by phosphorylation/dephosphorylation and activation upon GSK3β kinase inhibition. (B) Dephosphorylation of Dyn1 S774 upon GSK3β inhibition by 20 μM CHIR99021 observed by immunoblotting using a Dyn1 phosphospecific antibody and the quantification of pDyn1/Dyn1 intensity ratios (mean ± SD, n = 3). Lysates were loaded at two different dilutions (1x and 0.3x). (C) Transferrin receptor (TfnR) internalization efficiency of parental H1299 cells and Dyn1KO cells and their sensitivity to GSK3β inhibition (mean ± SD, n = 3). (D) Initiation densities of bona fide CCPs and (E) their median lifetimes. Each dot represents the average value per movie, where each movie contained 1–5 cells (see Materials and methods). (F) The distribution of CCP lifetimes measured in the absence or presence of GSK3β inhibitor. Data are derived from 10 movies each; 13,346 CCPs of 40–60-s lifetimes were analyzed from 74,807 bona fide CCPs, and 13,494 CCPs of 40–60-s lifetimes were analyzed from 75,426 bona fide CCPs, respectively, for control and GSK3β inhibition. Similarly, the initiation densities (G), median lifetime (H), and the lifetime distribution of bona fide CCPs (I) were analyzed for H1299 Dyn1KO cells with or without GSK3β inhibition. Average recruitment of Dyn1a-eGFPend to CLCa-labeled CCPs with lifetimes of 40–60 s measured in the absence (J) or presence (K) of GSK3β inhibitor. (L) Maximum intensity of Dyn1a-eGFPend detected at any point throughout the lifetime of an individual CCPs measured in the absence or presence of GSK3β inhibitor. The underlying data of panels B–I and L can be found in S1 Data. (* p ≤ 0.05, ** p ≤ 0.01, *** p ≤ 0.001; see Materials and methods for description of statistical analysis used in this and other figures.) CCP, clathrin-coated pit; CLCa, clathrin light chain a; CME, clathrin-mediated endocytosis; Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta. https://doi.org/10.1371/journal.pbio.2005377.g002 To further probe the mechanism by which activated Dyn1 accelerates CME, we introduced mRuby2-labeled CLCa into H1299 parent Dyn1KO cells and measured CCP dynamics by TIRFM. Analysis of the rates of assembly and departure of CCPs revealed that GSK3β inhibition resulted in a significant increase in the rate of coated pit initiation per unit cell area (Fig 2D), as well as an increase in maturation rates (i.e., decrease in lifetimes) of CCPs (Fig 2E). The latter was evident in the change in lifetime distribution of all bona fide CCPs (Fig 2F), which displayed a more quasi-exponential profile than untreated cells, indicative of a less-regulated process during early stages of CCP maturation [22]. Importantly, similar effects were observed for H1299 Dyn1a-eGFPend (S3A–S3C Fig), confirming that the C-terminally eGFP-tagged splice variant, Dyn1a, was functional and activated by dephosphorylation. Again, GSK3β inhibition had no effect on CCP initiation rates or lifetimes in H1299 Dyn1KO cells (Fig 2G–2I), confirming that these changes in CCP dynamics are a result of activation of Dyn1. We then asked whether GSK3β inhibition and activation of Dyn1 altered its recruitment to CCPs. Surprisingly, there was no significant difference in the average recruitment intensity (Fig 2J) of Dyn1 at CCPs. Previous studies had shown that the appearance of dynamin fluctuates at CCPs [21,32]; thus, it was possible that GSK3β inhibition induces asynchronous and transient appearances of Dyn1 at CCPs that could be obscured by measuring average recruitment. Therefore, we also quantified the maximum intensity of Dyn1 recruited at any time along a CCP track. Using this orthogonal measurement, we again saw no effect of GSK3β inhibition on Dyn1 recruitment to CCPs (Fig 2K). Together, these data suggest that dephosphorylation and activation of Dyn1 can alter CCP dynamics and CME even when Dyn1 is present at low amounts and that the effects of activation of Dyn1 on CCP dynamics are not likely explained simply by its increased recruitment to CCPs. Substoichiometric levels of Dyn1 are sufficient to stimulate CCP dynamics It remained possible that the extremely low expression levels of Dyn1 in H1299 might limit our ability to detect GSK3β-dependent changes in its recruitment. To test this, we stably overexpressed Dyn1aWT-eGFP in H1299 Dyn1KO cells at approximately 20-fold levels higher than endogenous to generate Dyn1aWT-eGFPo/x cells (Fig 3A). Importantly, overexpression of Dyn1aWT-eGFP itself did not result in any additional increase in TfnR uptake compared to the normal low endogenous levels (Fig 3B, see also Fig 4G). However, as in parental and genome-edited H1299 cells, acute GSK3β inhibition in the Dyn1aWT-eGFPO/X cells resulted in increased rates of TfnR uptake (Fig 3B) and alterations in CCP dynamics, including increased rates of CCP initiation and maturation (Fig 3C–3E). Yet similar to the Dyn1a-eGFPend-cells, GSK3β inhibition did not result in significantly enhanced recruitment of Dyn1aWT-eGFP to the membrane, either on average (Fig 3F and 3G) or when measured as maximum peak intensity (Fig 3H). Moreover, there was no evidence of a burst of Dyn1 recruitment prior to CCV formation (Fig 3G). Together, these results suggest that the observed changes in CCP dynamics are the result of a scission-independent early role for low levels of Dyn1 in regulating CME. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Dyn1 is inefficiently recruited to CCPs, even when overexpressed and activated in Dyn1KO H1299 cells. (A) Western blot showing overexpression of Dyn1WT-eGFP or Dyn1S774/8A-eGFP in Dyn1KO H1299 cells. Note that endogenous Dyn1 is not detected in parental cells at this loading level (see Fig 1B). (B) Effect of siRNA knockdown of Dyn2 on TfnR internalization in Dyn1KO cells reconstituted with Dyn1aWT-eGFP and treated or not with GSK3β inhibitor. Results are normalized to rates of endocytosis in parental H1299 cells. The data represents mean ± SEM of n = 3 experiments containing four replicates each (*p ≤ 0.05, **p ≤ 0.01 and ****p ≤ 0.0001). Initiation densities (C), median lifetimes (D), and the lifetime distribution (E) of bona fide CCPs analyzed in H1299 Dyn1KO cells reconstituted with Dyn1WT-eGFP with or without GSK3β inhibition, determined as in Fig 2. (F) Representative TIRFM images of overexpressed Dyn1WT-eGFP and mRuby2-CLCa and (G) quantification of the average recruitment of Dyn1WT-eGFP to CCPs, identified by mRuby2-CLCa, with lifetimes between 40 and 60 s (14,495 CCPs from a pool of 100,050 bona fide Dyn1-positive CCPs from 18 movies and 9,651 CCPs from a pool of 68,909 bona fide CCPs from 12 movies were analyzed from control and GSK3β, respectively). (H) Maximum Dyn1aWT-eGFP intensity averaged among individual bona fide CCP tracks in the absence or presence of GSK3β inhibitor. The underlying data of panels B–E and H can be found in S1 Data. CCP, clathrin-coated pit; CLCa, clathrin light chain a; Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta; siRNA, small interfering RNA; TfnR, transferrin receptor; TIRFM, total internal reflection fluorescence microscopy. https://doi.org/10.1371/journal.pbio.2005377.g003 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Nonphosphorylatable Dyn1 mutant mimics GSK3β effects and can partially substitute for Dyn2. CCP initiation densities (A), median lifetimes (B), and the lifetime distribution (C) of bona fide CCPs analyzed in H1299 Dyn1KO cells reconstituted with Dyn1WT- or Dyn1S774/8A-eGFP, determined as described in Fig 2. (D) Representative TIRFM images of overexpressed Dyn1WT-eGFP or Dyn1S774/8A-eGFP and mRuby2-CLCa and (E) quantification of their average recruitment to CCPs with lifetimes between 40 and 60 s. (F) Maximum intensities of Dyn1WT-eGFP or Dyn1S774/8A-eGFP averaged among individual bona fide CCP tracks. (G) Effect of siRNA knockdown of Dyn2 on TfnR endocytosis in parental and Dyn1KO H1299 cells and Dyn1KO cells reconstituted with either Dyn1aWT-eGFP or Dyn1aS774/8A-eGFP. (H) Representative TIRFM images of Dyn2 siRNA-treated Dyn1KO cells overexpressing Dyn1aWT-eGFP and mRuby2-CLCa treated or not with GSK3β inhibitor and (I) quantification of the average recruitment of Dyn1WT-eGFP to CCPs with lifetimes between 40 and 60 s in Dyn2 knockdown cells treated or not with GSK3β inhibitor. The underlying data of panels A–C, F, and G can be found in S1 Data. CCP, clathrin-coated pit; CLCa, clathrin light chain a; Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta; siRNA, small interfering RNA; TfnR, transferrin receptor; TIRFM, total internal reflection fluorescence microscopy. https://doi.org/10.1371/journal.pbio.2005377.g004 Dephosphorylated Dyn1 regulates early stages of CME Based on our finding that Dyn1 expression is required for the inhibitory effects of GSK3β on CME, we hypothesized that dephosphorylation of residues in Dyn1’s PRD should be sufficient to enhance CME efficiency. To test this, we introduced point mutations in Dyn1 at the serine residue phosphorylated by GSK3β (S774) and at the priming serine site that is responsible for recruiting GSK3β (S778). We expressed this mutant as an eGFP fusion in H1299 cells, Dyn1aS774/8A-eGFP, at comparable levels to Dyn1aWT-eGFP (Fig 3A). As predicted, Dyn1S774/8A-eGFP cells exhibited increased rates of CCP initiation (Fig 4A), decreased CCP lifetimes (i.e., increased rates of CCP maturation, Fig 4B), and changed the lifetime distribution to a quasi-exponential profile (Fig 4C). From these data, we conclude that dephosphorylated Dyn1 is sufficient to account for the effects of GSK3β inhibition on CCP dynamics. Surprisingly, even the nonphosphorylatable Dyn1aS774/8A-eGFP mutant was not efficiently recruited to CCPs and failed to display a pronounced late burst of recruitment accompanying membrane scission (Fig 4D–4F). Interestingly, the changes in CCP dynamics in Dyn1aS774/8A-eGFP-expressing cells were not reflected in significantly increased rates of TfnR uptake, presumably due to compensatory changes that occur upon prolonged expression of activated Dyn1 versus acute activation (Fig 4G). However, unlike parental H1299 cells or Dyn1aWT-eGFP cells, Dyn1KO cells reconstituted with Dyn1aS774/8A-eGFP exhibited significant residual levels of TfnR uptake upon siRNA knockdown of Dyn2 (Fig 4G), consistent with functional activation of Dyn1. Moreover, upon siRNA knockdown of Dyn2, we detected an increase in Dyn1aWT-eGFP recruitment to CCPs (Fig 4H and 4I), suggesting its activation as part of a compensatory mechanism to restore CME [23]. From these data, we conclude that Dyn1 is negatively regulated in non-neuronal cells through GSK3β-dependent phosphorylation of S774 and that dephosphorylated, active Dyn1 regulates early stages of CME even when present at low (nearly undetectable, in the case of parental H1299 cells) levels on CCPs. Importantly, overexpressed Dyn1, even when activated by mutation or GSK3β inhibition (Fig 3B), does not fully compensate for loss of Dyn2 function in CME, hence the two isoforms have partially divergent functions. A549 cells express high levels of Dyn1 that can partially substitute for Dyn2 We previously reported that several lung cancer cell lines express high levels of Dyn1 [35,39]. For example, A549 non-small cell lung cancer cells express approximately 5-fold higher levels of Dyn1 than Dyn2 [39], corresponding to approximately 20-fold higher levels of Dyn1 than in H1299 cells (S4A Fig). Reflective of these high levels of Dyn1 expression, siRNA knockdown of both Dyn1 and Dyn2 is necessary for potent inhibition of TfnR uptake in A549 cells (S4B Fig). Therefore, we reasoned that it might be possible to individually knockout Dyn1 and the otherwise essential Dyn2 in A549 cell lines for reconstitution experiments. Thus, we used CRISPR/Cas9n to generate a complete knockout of Dyn1 (Dyn1KO) or Dyn2 (Dyn2KO) in A549 cells (Fig 5A, S4C Fig) and then introduced mRuby2-CLCa to track CCP dynamics. Acute inhibition of GSK3β had no effect on the rates of CCP initiation or maturation in Dyn1KO A549 cells but significantly stimulated the rate of CCP initiation and decreased the lifetimes of CCPs in Dyn2KO A549 cells (Fig 5B and 5C). These data show that the two isoforms differentially regulate early stages of CME and confirm that the effects of GSK3β inhibition on CME depend on Dyn1 but not Dyn2. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Dyn1 and Dyn2 are differentially recruited to CCPs and differentially required for GSK3β-regulated CME. (A) Immunoblot validation of Dyn1 and Dyn2 KO A549 cells and their corresponding reconstitution at near endogenous levels with eGFP-labelled Dyn1 or Dyn2. GFP blot shows that in A549 cells Dyn1 is expressed at approximately 5-fold higher levels than Dyn2. CCP initiation densities (B), and median lifetimes (C) in Dyn1 or Dyn2 knockout cells with or without GSK3β inhibition, determined as described in Fig 1. (D) Representative TIRFM and epi images of co-cultured Dyn1KO:Dyn1a-eGFP:SNAP-CLCa and Dyn2KO:Dyn2-eGFP:mRuby2-CLCa cells allowing direct comparison of Dyn1a-eGFP versus Dyn2-eGFP recruitment to CCPs in A549 cells. (E) Quantification of the average recruitment of Dyn1a-eGFP or Dyn2-eGFP to CCPs with lifetimes between 40 and 60 s (4,420 CCPs from a pool of 12,555 Dyn1a-eGFP-positive CCPs and 3,961 CCPs from a pool of 12,766 Dyn2-eGFP-positive CCPs from a total of 11 movies were identified to have a lifetime between 40 and 60 s). Data are obtained from cells co-imaged either for SNAP(647)-CLCa (and Dyn1a-eGFP) or mRuby2-CLCa (and Dyn2-eGFP). (F) Maximum intensities of Dyn1a-eGFP or Dyn2-eGFP averaged among individual bona fide CCP tracks. (G) Subcellular fractionation of parental A549 cells into membrane (P) versus cytosolic (S) fractions and western blotted for the indicated proteins. Cytosolic MEK1/2 and membrane-associated TfnR serve as controls for fractionation. Quantification is shown in red above each band as the fraction of total protein in the P versus S fraction. Results are representative of 3 independent experiments. The underlying data of panels B, C, and F can be found in S1 Data. CCP, clathrin-coated pit; CLCa, clathrin light chain a; CME, clathrin-mediated endocytosis; Dyn1, dynamin-1; epi, epifluorescent; GSK3β, glycogen synthase kinase-3 beta; KO, knockout; TfnR, transferrin receptor; TIRFM, total internal reflection fluorescence microscopy. https://doi.org/10.1371/journal.pbio.2005377.g005 To directly and quantitatively compare the relative recruitment efficiencies of the two isoforms to CCPs, we reconstituted these knockout cells with their respective eGFP-tagged isoforms and sorted for expression comparable to their endogenous levels (i.e., in these A549 cells we chose cells in which Dyn1a-eGFP levels were approximately 5-fold higher than Dyn2-eGFP) (Fig 5A). Additionally, we introduced SNAP-CLCa and mRuby2-CLCa in Dyn1a-eGFP and Dyn2-eGFP cells, respectively, so that we could distinguish the two A549 cell lines (i.e., Dyn1KO:Dyn1a-eGFP:SNAP-CLCa from Dyn2KO:Dyn2-eGFP:mRuby2-CLCa) while imaging them in the same TIRFM field of view under the same conditions (Fig 5D). These data directly show the differential recruitment efficiencies of Dyn1 and Dyn2 to CCPs. Live-cell imaging revealed the typical gradual accumulation and burst of Dyn2-eGFP recruitment to CCPs when averaged over the cohort of 40–60-s lifetime CCPs (Fig 5E). However, under identical imaging conditions of the same fluorophore, Dyn1a-eGFP was recruited, on average, at least 10-fold less efficiently, even though it is expressed at higher abundance. The maximum intensity of tagged Dyn2 versus Dyn1 recruitment was also higher, albeit showing only an approximately 3-fold differential (Fig 5F). A likely explanation for the differences in average and peak measurements is that in A459 cells, Dyn1a-eGFP does display a slight burst of recruitment at late stages of CCV formation that is visible when the Dyn1 signal is rescaled (S4D Fig). Finally, to verify our results using an independent method, we performed Western blotting after subcellular fractionation and isolation of membrane versus cytosolic fractions, as confirmed using membrane-associated TfnR and cytosolic MEK1/2 as markers (Fig 5G). Under these fractionation conditions, approximately 90% of Dyn2 is membrane associated, whereas only 50% of Dyn1 sediments with the membrane fraction (Fig 5G). We observed a consistent, approximately 20% increase of membrane-associated Dyn1 upon GSK3β inhibition that was not detected by TIRFM. These biochemical data indicate a greater extent of membrane association of both active and inactive Dyn1 than detected at CCPs by TIRFM. The differences could reflect recruitment of Dyn1 to sites on the plasma membrane other than CCPs, as has been previously reported [40]. The approximately 20% increase in recruitment of activated Dyn1 likely reflects the increase in number of CCPs that occurs upon GSK3β inhibition, rather than an increase in Dyn1 per CCP. Consistent with TIRFM data, the distribution of phosphorylated Dyn1 (detected with an S774 phosphospecific antibody) was indistinguishable from total Dyn1 (i.e., there was no de-enrichment of phosphorylated Dyn1 in the membrane-bound fractions). These data confirmed that dephosphorylation of Dyn1 on S774 by GSK3β inhibition does not enhance its recruitment to CCPs. Thus, the effects of activated Dyn1 on CCP initiation and maturation occur either independently of its direct association with CCPs or, more likely, are manifested by very low levels of CCP-associated dephosphorylated Dyn1. Dyn1 and Dyn2 do not efficiently co-assemble Dynamin exists as a tetramer in solution [41,42] and assembles into higher-order helical oligomers on the membrane. Exploiting Dyn1KO and Dyn2KO A549 cells reconstituted with Dyn1a- or Dyn2-eGFP, respectively, we next assessed the degree to which Dyn1 and Dyn2 form hetero-tetramers in solution. Dyn1- or Dyn2-eGFP were efficiently immunoprecipitated with anti-eGFP nanobodies and the immunobeads were washed with 300 mM salt to disrupt any potential higher-order dynamin assemblies before measuring the fraction of Dyn2 or Dyn1 that coprecipitated. Under these conditions, we pulled down nearly 100% of the eGFP-tagged dynamins but only approximately 30% of Dyn2 with Dyn1-eGFP and <5% of Dyn1 with Dyn2-eGFP (S5A Fig). The difference in the extent of hetero-tetramerization is consistent with the approximately 5-fold higher levels of expression of Dyn1 versus Dyn2 in these cells. Thus, the two isoforms predominantly exist as homo-tetramers in solution. We also examined the relative abilities of Dyn1 and Dyn2 to co-assemble into higher-order structures in vitro. For this, we used a dominant-negative Dyn1 mutant (Dyn1S45N) defective in GTPase activity, which, when co-assembled with wild-type dynamin into higher-order oligomers on lipid nanotubes, will inhibit total assembly-stimulated GTPase activity through the intercalation of GTPase-defective subunits adjacent to wild-type subunits [43,44]. As expected, Dyn1S45N efficiently co-assembles with Dyn1WT such that, when present at equimolar levels, the total assembly-stimulated GTPase activity is inhibited by 50%. In contrast, at the same concentrations of Dyn1S45N, Dyn2 GTPase activity was significantly less affected (S5B Fig), indicating that Dyn2 less efficiently co-assembles into higher-order oligomers with the mutant Dyn1. Thus, consistent with their differential recruitment to CCPs, even when present at comparable levels of expression in the same cell type, the two isoforms only weakly interact. Genome-edited cells reveal that Dyn1 and Dyn2 are recruited to most CCPs in A549 cells Our results establish that Dyn1 and Dyn2 are differentially recruited to CCPs in non-neuronal cells and that, on average, Dyn1 is recruited at much lower levels than Dyn2. Despite this, acute activation of Dyn1 globally alters CCP dynamics. Thus, we next directly compared the recruitment of Dyn1 and Dyn2 to CCPs to determine whether Dyn1 is recruited at low levels to all CCPs or instead might be recruited at higher levels to a subpopulation of CCPs. Such heterogeneity would be lost by averaging. For this, we took advantage of the higher levels of Dyn1 expression in A549 cells and generated double genome-edited cells expressing Dyn1a-eGFP and Dyn2-mRuby2. We first used ZFNs to generate Dyn2 mRuby2-edited A549 cells and subsequently introduced a C-terminal eGFP to the Dyn1a splice variant using CRISPR/Cas9, as described earlier (Fig 1A, see Materials and methods). This yielded an A549 cell line homozygous for endogenously tagged Dyn2-mRuby2 and heterozygous for endogenously tagged Dyn1a-eGFP (2 of 3 Dyn1 alleles tagged in these triploid A549 cells) (Fig 6A). We confirmed that the double genome-edited cells exhibited comparable rates of TfnR uptake, as well as the degree of dependence on Dyn2 for CME, relative to the parent cells (Fig 6B). SNAP-CLCa was introduced into these cells by lentiviral transfection (Fig 6C), and we confirmed that GSK3β inhibition resulted in increased rates of CCP initiation, reduced CCP lifetimes, and altered the lifetime distributions of CCPs (Fig 6D–6F), as in the parental cells. Thus, the genome-edited Dyn isoforms were functionally active. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Generation and characterization of dual genome-edited Dyn1a-eGFP and Dyn2-mRuby2 A549 cells. (A) Immunoblot validation of Dyn1a-eGFP and Dyn2-mRuby2 single- and dual-genome-edited A549 cells. (B) TfnR endocytosis in dual genome edited A549 cells compared to parental A549 cells and their sensitivity to siRNA-mediated Dyn2 knockdown. (C) Representative TIRF images of Dyn1 and Dyn2 distribution relative to CLCa in dual genome-edited A549 cells. CCP initiation densities (D), median lifetimes (E), and the lifetime distribution (F) of bona fide CCPs in dual-genome-edited A549 cells with or without GSK3β inhibition, determined as in Fig 2. The underlying data of panels B and D–F can be found in S1 Data. Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta; TfnR, transferrin receptor. https://doi.org/10.1371/journal.pbio.2005377.g006 We next assessed the interplay between Dyn1a-eGFP and Dyn2-mRuby2 using three-color live-cell TIRFM imaging at 0.5 Hz (2 s per frame) (Fig 7A, S3 Movie). As reported previously, we detected fluctuations of both Dyn1 and Dyn2 at CCPs over their lifetimes (examples shown in Fig 7B) and frequently detected a burst of Dyn2 just prior to CCV formation. In many cases, we also detected a burst of Dyn1 recruitment, albeit to a lesser degree. For more quantitative analysis of these data, we applied the 3-channel functionality of our cmeAnalysis package to perform three-color master/slave analyses [22]. Using clathrin as the “master” channel and Dyn1 and 2 as “slave” channels, we determined whether the clathrin tracks contained either Dyn1, Dyn2, both, or neither. Individual CCP tracks were considered positive for Dyn1 and/or Dyn2 if the intensities of Dyn1/2 signals detected at the position of the clathrin tag were significantly higher than the local Dyn1/2 background signal around the clathrin tag position for a period of time exceeding random associations, as previously described [22]. This analysis revealed that in double genome-edited Dyn1a-eGFPend/Dyn2-mRuby2end A549 cells, both Dyn2 and Dyn1 could be robustly detected in approximately 75% of all bona fide CCPs (Fig 8A). Moreover, in this population of CCPs, a clear burst of recruitment of both Dyn1a-eGFP and Dyn2-mRuby2 could be detected prior to CCV formation. Importantly, the apparently higher levels of recruitment of Dyn1-eGFP versus Dyn2-mRuby2 in these genome-edited cells is not a reflection of protein levels but rather of imaging conditions and brightness for two different fluorophores (compare with Fig 5E). The remaining CCPs were roughly equally distributed as Dyn1 only, Dyn2 only, and both Dyn1- and Dyn2-negative subpopulations (Fig 8A). Note that the Dyn2 levels in the “Dyn1 only” CCPs were still on average higher than background (Dyn1/Dyn2 negative), reflecting the stringency of our master/slave detection and suggesting that Dyn2 is recruited to >90% of all CCPs, albeit to variable extents. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Tracking clathrin and dynamins in dual genome-edited Dyn1a-eGFP and Dyn2-mRuby2 A549 cells. (A) Representative TIRF images and corresponding kymographs of dynamic behavior of overexpressed SNAP-CLCa, Dyn2-mRuby2end, and Dyn1a-eGFPend in dual genome-edited A549 cells. See S3 Movie. (B) Examples of Dyn1 and Dyn2 dynamics at individual CCPs (i–iv) and (C) their corresponding quantitative traces. CLCa, clathrin light chain a; Dyn1, dynamin-1. https://doi.org/10.1371/journal.pbio.2005377.g007 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 8. Dyn1 and Dyn2 are recruited to the same CCPs and Dyn1 activation alters the dynamics of all CCP subpopulations. (A) Triple-color master/slave analyses of average dynamics of recruitment of Dyn2-mRuby2endo and/or Dyn1a-eGFPend to lifetime cohorts of SNAP-CLCa labeled CCPs identifies Dyn1/Dyn2 positive, Dyn1 only, Dyn2 only, and Dyn1/2-negative subpopulations of CCPs. The percentage of detected CCPs in each class is indicated. (B) Effect of GSK3β inhibition on the median lifetimes of compositionally distinct CCP subpopulations. The underlying data of panel B can be found in S1 Data. CCP, clathrin-coated pit; CLCa, clathrin light chain a; Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta. https://doi.org/10.1371/journal.pbio.2005377.g008 We next compared per cell median lifetimes of CCPs relative to their dynamin isoform composition and found that CCPs bearing higher levels of Dyn2 and Dyn1 exhibited longer lifetimes (median approximately 80 s) than single-positive CCPs (median approximately 38 s) (Fig 8B). CCPs that failed to detectably recruit either isoform were the shortest lived (median approximately 20 s) and likely represent abortive CCPs. These findings are consistent with previous data suggesting that a threshold level of Dyn2 recruitment is required for efficient CCP maturation [22,34]. All of these CCP subpopulations showed a significant decrease in CCP lifetimes upon inhibition of GSK3β, consistent with other data that only low levels of Dyn1 are required to alter CCP maturation. SNX9 is required for activated Dyn1-dependent effects on CCP maturation Our findings thus far point to isoform-specific functions of Dyn1 and Dyn2 and hence suggest the existence of isoform-specific binding partners. Dyn1 and Dyn2 are >80% identical except for their C-terminal PRDs, which are only 50% identical and likely determine isoform-specific interactions with SH3 domain-containing proteins. The Dyn1KO and Dyn2KO A549 cells provide an opportunity to measure Dyn2 and Dyn1-dependent CME, respectively, without the possibility of compensation. Thus, we measured, by TfnR uptake, the effects of siRNA knockdown of several known SH3 domain-containing binding partners on Dyn2-dependent CME in the Dyn1KO cells and on Dyn1-dependent CME in the Dyn2KO cells. Knockdown of these dynamin partners has only mild effects on TfnR uptake in parental A549 cells and in Dyn1KO cells, whose endocytosis is exclusively Dyn2 dependent (Fig 9A). Whether these mild effects reflect partial redundancy with other dynamin partners, activation of compensatory mechanisms [23], or that these factors, which were identified primarily as dynamin partners in brain lysates, play only minor roles in TfnR uptake in non-neuronal cells, cannot be discerned from these studies. Interestingly, siRNA knockdown of Grb2 appeared to inhibit TfnR uptake in Dyn1KO cells by approximately 20%, while not affecting TfnR uptake in either parental or Dyn2KO cells. This suggests that Grb2 might preferentially function together with Dyn2 in CME and that its depletion in parental cells can be compensated for by Grb2-independent Dyn1 activity. In contrast, siRNA knockdown of SNX9 only mildly inhibited TfnR uptake in parental A549 and had no significant effect on Dyn2-dependent TfnR uptake in Dyn1KO cells, but decreased TfnR uptake in Dyn2KO cells by >50% (Fig 9A). Thus Dyn1-dependent endocytosis appears to be particularly sensitive to SNX9 knockdown. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 9. SNX9 preferentially binds activated Dyn1 and is required for Dyn1-dependent changes in the lifetime distribution of CCPs. (A) Effect of siRNA knockdown of the indicated dynamin SH3 domain-containing endocytic accessory proteins on TfnR endocytosis in parental, Dyn1KO, and Dyn2KO A549 cells. siEndo refers to siRNA knockdown of endophilin A1, 2, and 3; siITSN refers to siRNA knockdown of intersectins 1 and 2; all others were single siRNAs. Knockdown efficiencies were determined to be >85% by western blotting. Data are normalized to the extent of TfnR uptake in control siRNA-treated parental, Dyn1KO, and Dyn2KO cells, which is set to 100, to allow direct comparison of the relative effects of siRNA knockdowns. (B) eGFP pulldown of Dyn1aWT-eGFP, Dyn1aS774/8A-eGFP, or Dyn2WT-eGFP expressed in Dyn1KO or Dyn2KO A549 cells, respectively, using anti-eGFP nAbs. Parental cells that do not express an eGFP-tagged protein (Dyn1end) are used as control. The pulldown fractions were analyzed by immunoblot. Effect of SNX9 siRNA-mediated knockdown on (C) CCP initiation densities, (D) median lifetimes, and (E) lifetime distribution of bona fide CCPs in Dyn1KO H1299 cells overexpressing either Dyn1aWT-eGFP or Dyn1aS774/8A-eGFP (data are derived from 7 movies for each condition, with each movie consisting of 1–3 cells). Each data point is the average value from a single movie. The underlying data of panels A, C–E, and L can be found in S1 Data. (* p ≤ 0.05, ** p ≤ 0.01, *** p ≤ 0.001) CCP, clathrin-coated pit; Dyn1, dynamin-1; SH3, SRC Homology 3; siRNA, small interfering RNA; SNX9, sorting nexin 9; TfnR, transferrin receptor. https://doi.org/10.1371/journal.pbio.2005377.g009 We next tested whether SNX9 preferentially interacts with Dyn1 versus Dyn2 by GFP pulldown assays using Dyn1KO cells reconstituted with either Dyn1aWT- or Dyn1S774/8A-eGFP and Dyn2KO A549 reconstituted with Dyn2-eGFP. Consistent with previous results [45,46], we confirmed that SNX9 binds both Dyn1 and Dyn2 (Fig 9B). However, the ratio of SNX9 binding to Dyn1 versus Dyn2 was 1.7 ± 0.6 (mean ± SEM, n = 3), indicative of a slight preference for Dyn1. Importantly, SNX9 showed a marked preference for binding to the nonphosphorylated and active Dyn1S774/8A-eGFP. The ratio of SNX9 binding to Dyn1S774/8A versus Dyn1WT was 3.6 ± 0.9 (mean ± SEM, n = 3). These data suggested that SNX9 might be a preferential functional partner of activated Dyn1. To test whether SNX9–Dyn1 interactions were required for the effects of activated Dyn1 on CCP initiation rates, CCP maturation, or both, we asked returned to the Dyn1KO H1299 cells reconstituted with Dyn1WT versus Dyn1S774/8A and tested whether the selective effects of Dyn1S774/8A on CCP dynamics (Fig 4A–4C) were dependent on SNX9. Knockdown of SNX9 decreased the rate of CCP initiation in Dyn1WT but was not required for the enhanced rate of CCP initiation triggered by Dyn1S774/8A expression (Fig 9C). Thus, other, yet-unidentified binding partners are responsible for the Dyn1-dependent effect on CCP initiation. SNX9 knockdown also led to an increase in the median CCP lifetimes in both Dyn1WT- and Dyn1S774/8A-expressing cells (Fig 9D). These data suggest that SNX9 functions in both Dyn1-dependent and independent stages of CCP maturation. Consistent with this, SNX9 knockdown also abrogated the effects of Dyn1S774/8A expression on the lifetime distribution of bona fide CCPs (Fig 9E), reverting the quasi-exponential distribution seen in Dyn1S774/8A to a distribution nearer to control. The strong effect of SNX9 knockdown is also seen in the rightward shift of the lifetime distribution of Dyn1WT cells treated with SNX9 siRNA. Together, these data suggest multiple roles of SNX9 at multiple stages of CME, including the support of Dyn1’s early functions in accelerating CCP maturation. Dyn1 is activated downstream of the EGFR We have shown that strong pharmacological inhibition of GSK3β activates Dyn1 in non-neuronal cells and results in increased rates of CCP initiation and maturation, leading to increased rates of TfnR uptake via CME. However, it is not clear whether this regulatory effect on Dyn1 function modulates CME under more physiologically relevant conditions. To test this, we treated serum-starved A549 cells with epidermal growth factor (EGF), which is known to activate Akt and in turn to phosphorylate and inactivate GSK3β [47]. We confirmed that GSK3β is phosphorylated in EGF-treated cells and that this resulted in reduced levels of phosphorylation of Dyn1 at S774 (Fig 10A, quantified in Fig 10B and 10C). As predicted by the results of inhibitor experiments, EGF treatment of serum-starved cells also increased the rate of CCP initiation (Fig 10D), decreased CCP lifetimes (Fig 10E), and, compared to control cells, resulted in a shift of the lifetime distributions of bona fide CCPs to a more quasi-exponential distribution (Fig 10F). Importantly, the effects of EGF treatment on CCP initiation rate and lifetimes were not seen in A549 Dyn1KO cells (Fig 10G and 10H). These data suggest that Dyn1 can be activated to alter CCP dynamics under physiological conditions through signaling downstream of EGFR. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 10. EGF stimulation alters CCP dynamics in a Dyn1-dependent manner. (A) Immunoblot analysis of changes in phosphorylation state of Dyn1 S774 and GSK3β upon EGF stimulation of parental A549 cells. (B,C) Quantification of the EGF-triggered changes in phosphorylation state (i.e., ratio of phosphorylated/ total protein) of GSK3β and Dyn1 (mean ± SD of n = 3 experiments, data are normalized to 0 min time point). CCP initiation densities (D), median lifetimes (E), and the lifetime distribution (F) of bona fide CCPs in serum starved A549 cells before (Control, Ctrl) or after incubation with EGF (20 ng/ml) for 10 min prior to imaging. CCP initiation densities (G) and median lifetimes (H) in serum-starved Dyn1KO A549 cells before (Ctrl) or 10 min after incubation with EGF (EGF). The underlying data of panels B–H can be found in S1 Data. CCP, clathrin-coated pit; Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta. https://doi.org/10.1371/journal.pbio.2005377.g010 Dynamin isoforms are differentially recruited to CCPs Recent studies have shown that Dyn1 is widely expressed in non-neuronal cells [2]; but, like at the neuronal synapse [33], it is mostly inactive at steady state due to phosphorylation by the constitutively active kinase GSK3β. Dyn1 function and its recruitment to CCPs have been studied in non-neuronal cells, albeit under conditions of overexpression and/or without an awareness of its phosphoregulation [14,21]. Therefore, to explore potential isoform-specific functions of Dyn1 and Dyn2, as well as the role of GSK3β in regulating Dyn1 activity, we generated genome-edited H1299 non-small cell lung cancer cells, which we previously showed partially utilize Dyn1 for CME [23]. Cells expressing endogenously tagged Dyn2-mRuby2 were generated using previously validated Zinc Finger Nucleases (ZFNs) [32,34] to introduce double-stranded breaks and insert the mRuby tag with complementary flanking regions by homology-driven repair (HDR) (Fig 1A). The resulting cells were single-cell sorted for mRuby2 fluorescence to obtain a heterozygous clone (clone 235, designated Dyn2-mRuby2end) expressing a single mRuby2-tagged allele of Dyn2 (Fig 1B). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Isoform-specific differences in recruitment of dynamin to CCPs. (A) Diagram of Zinc-finger and CRISPR/Cas9n knock-in strategies for endogenous labeling of Dyn2 and Dyn1 in H1299 cells with C-terminal mRuby2 and eGFP tags, respectively. For Dyn2, a short linker and mRuby2 (red) were placed at the stop codon in exon 22. For canonical Dyn1 splice isoform “a,” the 19 C-terminal amino acids (blue) were inserted in exon 21, followed by a short linker, eGFP (green), with stop codon and a polyadenylation sequence (yellow). In both constructs, flanking homology arms (HAs) of roughly 800 bp were used to promote recombination (dashed lines). See S1 Fig for details. (B) Western blot analysis of tagged isoforms. The low levels of Dyn1 in H1299 cells could not be directly detected by western blotting but can be detected after pulldown with GST-Amphiphysin II SH3 domains. Representative TIRF images (see S1 and S2 Movies) showing membrane recruitment of endogenous Dyn2-mRuby2end (C) or Dyn1a-eGFPend (E) and corresponding lentiviral transduced SNAP(647)-CLCa images. (D,F) Clathrin labeled puncta were identified and thresholded to define bona fide CCPs [22]. Shown are the averaged kinetics of recruitment of SNAP-CLCa and Dyn2-mRuby2end (D) or Dyn1a-eGFPend (F) for all tracks with lifetimes between 40 and 60 s (831 CCPs from 5 movies containing a total of 15 cells for Dyn2-mRuby2end and 13,346 CCPs from 10 movies containing a total of 29 cells for Dyn1a-eGFPend). CLCa, clathrin light chain a; CRISPR/Cas9n, Clustered Regularly Interspaced Short Palindromic Repeats-associated nucleases 9 nickase; Dyn1, dynamin-1; GST, Glutathione S-transferase; HA, homology arm; SH3, SRC Homology 3. https://doi.org/10.1371/journal.pbio.2005377.g001 Endogenously tagging Dyn1 was complicated by the fact that the DNM1 gene encodes C-terminal splice variants derived from differential splicing of exons 21 and 22 (S1A Fig), whose differential utilization could lead to partial loss of the fusion tag. Previous studies involving CRISPR/Cas9-mediated knockout and reconstitution with the Dyn1a C-terminal splice variant had confirmed that it fully reconstituted the GSK3β phosphoregulated activity of endogenous Dyn1 in H1299 cells [23], including its ability to be activated by calmodulin [35]. Therefore, using a Clustered Regularly Interspaced Short Palindromic Repeats-associated nucleases 9 nickase (CRISPR/Cas9n) strategy, we targeted the Dyn1 gene at the end of exon 21 and introduced sequences encoding the remaining 19 amino acids of the Dyn1a isoform, followed by a seven amino acid linker [32], monomeric eGFP fusion tag with stop codon, and finally, the SV40 polyadenylation signal to ensure unique expression of the “a” splice variant (Fig 1A and S1B Fig). Single-cell sorting by fluorescence-activated cell sorting (FACS) for eGFP fluorescence, followed by clonal amplification generated a heterozygous clone (clone 1B6, designated Dyn1a-eGFPend) expressing one eGFP-tagged allele of Dyn1a (Fig 1B). Note that although Dyn1 is expressed at very low levels in H1299 cells, it can be readily detected following enrichment by amphiphysin-II SH3 domain pulldown. As a robust fiduciary marker for CCPs, clathrin light chain a (CLCa) carrying an N-terminal SNAP-fusion tag was stably introduced in parallel into both cell lines via a lentiviral vector with puromycin selection of SNAP-CLCa expressing cells. As previously reported by several groups, mild overexpression of Fluorescent Probe (FP)-CLCa has no effect on CME as measured by transferrin endocytosis [22,31,32,36] and no effect on CCP dynamics compared to AP2 or other markers [20,29,31]. We then performed live-cell dual-channel total internal reflection fluorescence microscopy (TIRFM) and analyzed CCP dynamics and Dyn recruitment using the master–slave (CLCa–Dyn) approach introduced with the cmeAnalysis software [22,37,38]. As expected based on previous studies using either overexpressed [28,29,31] or endogenously tagged Dyn2 [17,22,32], Dyn2-mRuby2end was observed, on average, to gradually accumulate and then exhibit a burst of recruitment coincident with clathrin-coated vesicle (CCV) release. This can be seen in class-averaged tracks of bona fide CCPs with lifetimes ranging from 40–60 s (Fig 1C and 1D) and in all other CCP lifetime cohorts (S2A and S2B Fig, S1 movie). In contrast, Dyn1a-eGFPendo recruitment was barely detectable above background, and no burst was evident (Fig 1E and 1F, S2C and S2D Fig, S2 movie). This could reflect isoform-specific differences, very low levels of Dyn1 expression relative to Dyn2, and/or the inactivation of Dyn1 by GSK3β phosphorylation. Thus, we further explored these possibilities. Inhibition of constitutively active GSK3β kinase stimulates Dyn1 to accelerate CCP initiation and maturation We first tested whether activation of Dyn1 alters CCP dynamics and/or the recruitment of Dyn1a-eGFPend in H1299 cells. As expected based on earlier studies in ARPE cells [23], we confirmed that acute inhibition of GSK3β by incubation with the specific inhibitor, CHIR99021, leads to decreased phosphorylation of Dyn1 at S774 within 30 min (Fig 2A and 2B) and increased rates of CME, as measured by transferrin receptor (TfnR) internalization (Fig 2C). Importantly, the effects of GSK3β inhibition were dependent on Dyn1 expression, as treatment of Dyn1 knockout (Dyn1KO) H1299 cells [23] with CHIR99021 had no effect on CME (Fig 2C). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Activated Dyn1 regulates early stages of CME, even when recruited at low levels to CCPs. (A) Schematic representation of Dyn1 regulation by phosphorylation/dephosphorylation and activation upon GSK3β kinase inhibition. (B) Dephosphorylation of Dyn1 S774 upon GSK3β inhibition by 20 μM CHIR99021 observed by immunoblotting using a Dyn1 phosphospecific antibody and the quantification of pDyn1/Dyn1 intensity ratios (mean ± SD, n = 3). Lysates were loaded at two different dilutions (1x and 0.3x). (C) Transferrin receptor (TfnR) internalization efficiency of parental H1299 cells and Dyn1KO cells and their sensitivity to GSK3β inhibition (mean ± SD, n = 3). (D) Initiation densities of bona fide CCPs and (E) their median lifetimes. Each dot represents the average value per movie, where each movie contained 1–5 cells (see Materials and methods). (F) The distribution of CCP lifetimes measured in the absence or presence of GSK3β inhibitor. Data are derived from 10 movies each; 13,346 CCPs of 40–60-s lifetimes were analyzed from 74,807 bona fide CCPs, and 13,494 CCPs of 40–60-s lifetimes were analyzed from 75,426 bona fide CCPs, respectively, for control and GSK3β inhibition. Similarly, the initiation densities (G), median lifetime (H), and the lifetime distribution of bona fide CCPs (I) were analyzed for H1299 Dyn1KO cells with or without GSK3β inhibition. Average recruitment of Dyn1a-eGFPend to CLCa-labeled CCPs with lifetimes of 40–60 s measured in the absence (J) or presence (K) of GSK3β inhibitor. (L) Maximum intensity of Dyn1a-eGFPend detected at any point throughout the lifetime of an individual CCPs measured in the absence or presence of GSK3β inhibitor. The underlying data of panels B–I and L can be found in S1 Data. (* p ≤ 0.05, ** p ≤ 0.01, *** p ≤ 0.001; see Materials and methods for description of statistical analysis used in this and other figures.) CCP, clathrin-coated pit; CLCa, clathrin light chain a; CME, clathrin-mediated endocytosis; Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta. https://doi.org/10.1371/journal.pbio.2005377.g002 To further probe the mechanism by which activated Dyn1 accelerates CME, we introduced mRuby2-labeled CLCa into H1299 parent Dyn1KO cells and measured CCP dynamics by TIRFM. Analysis of the rates of assembly and departure of CCPs revealed that GSK3β inhibition resulted in a significant increase in the rate of coated pit initiation per unit cell area (Fig 2D), as well as an increase in maturation rates (i.e., decrease in lifetimes) of CCPs (Fig 2E). The latter was evident in the change in lifetime distribution of all bona fide CCPs (Fig 2F), which displayed a more quasi-exponential profile than untreated cells, indicative of a less-regulated process during early stages of CCP maturation [22]. Importantly, similar effects were observed for H1299 Dyn1a-eGFPend (S3A–S3C Fig), confirming that the C-terminally eGFP-tagged splice variant, Dyn1a, was functional and activated by dephosphorylation. Again, GSK3β inhibition had no effect on CCP initiation rates or lifetimes in H1299 Dyn1KO cells (Fig 2G–2I), confirming that these changes in CCP dynamics are a result of activation of Dyn1. We then asked whether GSK3β inhibition and activation of Dyn1 altered its recruitment to CCPs. Surprisingly, there was no significant difference in the average recruitment intensity (Fig 2J) of Dyn1 at CCPs. Previous studies had shown that the appearance of dynamin fluctuates at CCPs [21,32]; thus, it was possible that GSK3β inhibition induces asynchronous and transient appearances of Dyn1 at CCPs that could be obscured by measuring average recruitment. Therefore, we also quantified the maximum intensity of Dyn1 recruited at any time along a CCP track. Using this orthogonal measurement, we again saw no effect of GSK3β inhibition on Dyn1 recruitment to CCPs (Fig 2K). Together, these data suggest that dephosphorylation and activation of Dyn1 can alter CCP dynamics and CME even when Dyn1 is present at low amounts and that the effects of activation of Dyn1 on CCP dynamics are not likely explained simply by its increased recruitment to CCPs. Substoichiometric levels of Dyn1 are sufficient to stimulate CCP dynamics It remained possible that the extremely low expression levels of Dyn1 in H1299 might limit our ability to detect GSK3β-dependent changes in its recruitment. To test this, we stably overexpressed Dyn1aWT-eGFP in H1299 Dyn1KO cells at approximately 20-fold levels higher than endogenous to generate Dyn1aWT-eGFPo/x cells (Fig 3A). Importantly, overexpression of Dyn1aWT-eGFP itself did not result in any additional increase in TfnR uptake compared to the normal low endogenous levels (Fig 3B, see also Fig 4G). However, as in parental and genome-edited H1299 cells, acute GSK3β inhibition in the Dyn1aWT-eGFPO/X cells resulted in increased rates of TfnR uptake (Fig 3B) and alterations in CCP dynamics, including increased rates of CCP initiation and maturation (Fig 3C–3E). Yet similar to the Dyn1a-eGFPend-cells, GSK3β inhibition did not result in significantly enhanced recruitment of Dyn1aWT-eGFP to the membrane, either on average (Fig 3F and 3G) or when measured as maximum peak intensity (Fig 3H). Moreover, there was no evidence of a burst of Dyn1 recruitment prior to CCV formation (Fig 3G). Together, these results suggest that the observed changes in CCP dynamics are the result of a scission-independent early role for low levels of Dyn1 in regulating CME. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Dyn1 is inefficiently recruited to CCPs, even when overexpressed and activated in Dyn1KO H1299 cells. (A) Western blot showing overexpression of Dyn1WT-eGFP or Dyn1S774/8A-eGFP in Dyn1KO H1299 cells. Note that endogenous Dyn1 is not detected in parental cells at this loading level (see Fig 1B). (B) Effect of siRNA knockdown of Dyn2 on TfnR internalization in Dyn1KO cells reconstituted with Dyn1aWT-eGFP and treated or not with GSK3β inhibitor. Results are normalized to rates of endocytosis in parental H1299 cells. The data represents mean ± SEM of n = 3 experiments containing four replicates each (*p ≤ 0.05, **p ≤ 0.01 and ****p ≤ 0.0001). Initiation densities (C), median lifetimes (D), and the lifetime distribution (E) of bona fide CCPs analyzed in H1299 Dyn1KO cells reconstituted with Dyn1WT-eGFP with or without GSK3β inhibition, determined as in Fig 2. (F) Representative TIRFM images of overexpressed Dyn1WT-eGFP and mRuby2-CLCa and (G) quantification of the average recruitment of Dyn1WT-eGFP to CCPs, identified by mRuby2-CLCa, with lifetimes between 40 and 60 s (14,495 CCPs from a pool of 100,050 bona fide Dyn1-positive CCPs from 18 movies and 9,651 CCPs from a pool of 68,909 bona fide CCPs from 12 movies were analyzed from control and GSK3β, respectively). (H) Maximum Dyn1aWT-eGFP intensity averaged among individual bona fide CCP tracks in the absence or presence of GSK3β inhibitor. The underlying data of panels B–E and H can be found in S1 Data. CCP, clathrin-coated pit; CLCa, clathrin light chain a; Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta; siRNA, small interfering RNA; TfnR, transferrin receptor; TIRFM, total internal reflection fluorescence microscopy. https://doi.org/10.1371/journal.pbio.2005377.g003 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Nonphosphorylatable Dyn1 mutant mimics GSK3β effects and can partially substitute for Dyn2. CCP initiation densities (A), median lifetimes (B), and the lifetime distribution (C) of bona fide CCPs analyzed in H1299 Dyn1KO cells reconstituted with Dyn1WT- or Dyn1S774/8A-eGFP, determined as described in Fig 2. (D) Representative TIRFM images of overexpressed Dyn1WT-eGFP or Dyn1S774/8A-eGFP and mRuby2-CLCa and (E) quantification of their average recruitment to CCPs with lifetimes between 40 and 60 s. (F) Maximum intensities of Dyn1WT-eGFP or Dyn1S774/8A-eGFP averaged among individual bona fide CCP tracks. (G) Effect of siRNA knockdown of Dyn2 on TfnR endocytosis in parental and Dyn1KO H1299 cells and Dyn1KO cells reconstituted with either Dyn1aWT-eGFP or Dyn1aS774/8A-eGFP. (H) Representative TIRFM images of Dyn2 siRNA-treated Dyn1KO cells overexpressing Dyn1aWT-eGFP and mRuby2-CLCa treated or not with GSK3β inhibitor and (I) quantification of the average recruitment of Dyn1WT-eGFP to CCPs with lifetimes between 40 and 60 s in Dyn2 knockdown cells treated or not with GSK3β inhibitor. The underlying data of panels A–C, F, and G can be found in S1 Data. CCP, clathrin-coated pit; CLCa, clathrin light chain a; Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta; siRNA, small interfering RNA; TfnR, transferrin receptor; TIRFM, total internal reflection fluorescence microscopy. https://doi.org/10.1371/journal.pbio.2005377.g004 Dephosphorylated Dyn1 regulates early stages of CME Based on our finding that Dyn1 expression is required for the inhibitory effects of GSK3β on CME, we hypothesized that dephosphorylation of residues in Dyn1’s PRD should be sufficient to enhance CME efficiency. To test this, we introduced point mutations in Dyn1 at the serine residue phosphorylated by GSK3β (S774) and at the priming serine site that is responsible for recruiting GSK3β (S778). We expressed this mutant as an eGFP fusion in H1299 cells, Dyn1aS774/8A-eGFP, at comparable levels to Dyn1aWT-eGFP (Fig 3A). As predicted, Dyn1S774/8A-eGFP cells exhibited increased rates of CCP initiation (Fig 4A), decreased CCP lifetimes (i.e., increased rates of CCP maturation, Fig 4B), and changed the lifetime distribution to a quasi-exponential profile (Fig 4C). From these data, we conclude that dephosphorylated Dyn1 is sufficient to account for the effects of GSK3β inhibition on CCP dynamics. Surprisingly, even the nonphosphorylatable Dyn1aS774/8A-eGFP mutant was not efficiently recruited to CCPs and failed to display a pronounced late burst of recruitment accompanying membrane scission (Fig 4D–4F). Interestingly, the changes in CCP dynamics in Dyn1aS774/8A-eGFP-expressing cells were not reflected in significantly increased rates of TfnR uptake, presumably due to compensatory changes that occur upon prolonged expression of activated Dyn1 versus acute activation (Fig 4G). However, unlike parental H1299 cells or Dyn1aWT-eGFP cells, Dyn1KO cells reconstituted with Dyn1aS774/8A-eGFP exhibited significant residual levels of TfnR uptake upon siRNA knockdown of Dyn2 (Fig 4G), consistent with functional activation of Dyn1. Moreover, upon siRNA knockdown of Dyn2, we detected an increase in Dyn1aWT-eGFP recruitment to CCPs (Fig 4H and 4I), suggesting its activation as part of a compensatory mechanism to restore CME [23]. From these data, we conclude that Dyn1 is negatively regulated in non-neuronal cells through GSK3β-dependent phosphorylation of S774 and that dephosphorylated, active Dyn1 regulates early stages of CME even when present at low (nearly undetectable, in the case of parental H1299 cells) levels on CCPs. Importantly, overexpressed Dyn1, even when activated by mutation or GSK3β inhibition (Fig 3B), does not fully compensate for loss of Dyn2 function in CME, hence the two isoforms have partially divergent functions. A549 cells express high levels of Dyn1 that can partially substitute for Dyn2 We previously reported that several lung cancer cell lines express high levels of Dyn1 [35,39]. For example, A549 non-small cell lung cancer cells express approximately 5-fold higher levels of Dyn1 than Dyn2 [39], corresponding to approximately 20-fold higher levels of Dyn1 than in H1299 cells (S4A Fig). Reflective of these high levels of Dyn1 expression, siRNA knockdown of both Dyn1 and Dyn2 is necessary for potent inhibition of TfnR uptake in A549 cells (S4B Fig). Therefore, we reasoned that it might be possible to individually knockout Dyn1 and the otherwise essential Dyn2 in A549 cell lines for reconstitution experiments. Thus, we used CRISPR/Cas9n to generate a complete knockout of Dyn1 (Dyn1KO) or Dyn2 (Dyn2KO) in A549 cells (Fig 5A, S4C Fig) and then introduced mRuby2-CLCa to track CCP dynamics. Acute inhibition of GSK3β had no effect on the rates of CCP initiation or maturation in Dyn1KO A549 cells but significantly stimulated the rate of CCP initiation and decreased the lifetimes of CCPs in Dyn2KO A549 cells (Fig 5B and 5C). These data show that the two isoforms differentially regulate early stages of CME and confirm that the effects of GSK3β inhibition on CME depend on Dyn1 but not Dyn2. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Dyn1 and Dyn2 are differentially recruited to CCPs and differentially required for GSK3β-regulated CME. (A) Immunoblot validation of Dyn1 and Dyn2 KO A549 cells and their corresponding reconstitution at near endogenous levels with eGFP-labelled Dyn1 or Dyn2. GFP blot shows that in A549 cells Dyn1 is expressed at approximately 5-fold higher levels than Dyn2. CCP initiation densities (B), and median lifetimes (C) in Dyn1 or Dyn2 knockout cells with or without GSK3β inhibition, determined as described in Fig 1. (D) Representative TIRFM and epi images of co-cultured Dyn1KO:Dyn1a-eGFP:SNAP-CLCa and Dyn2KO:Dyn2-eGFP:mRuby2-CLCa cells allowing direct comparison of Dyn1a-eGFP versus Dyn2-eGFP recruitment to CCPs in A549 cells. (E) Quantification of the average recruitment of Dyn1a-eGFP or Dyn2-eGFP to CCPs with lifetimes between 40 and 60 s (4,420 CCPs from a pool of 12,555 Dyn1a-eGFP-positive CCPs and 3,961 CCPs from a pool of 12,766 Dyn2-eGFP-positive CCPs from a total of 11 movies were identified to have a lifetime between 40 and 60 s). Data are obtained from cells co-imaged either for SNAP(647)-CLCa (and Dyn1a-eGFP) or mRuby2-CLCa (and Dyn2-eGFP). (F) Maximum intensities of Dyn1a-eGFP or Dyn2-eGFP averaged among individual bona fide CCP tracks. (G) Subcellular fractionation of parental A549 cells into membrane (P) versus cytosolic (S) fractions and western blotted for the indicated proteins. Cytosolic MEK1/2 and membrane-associated TfnR serve as controls for fractionation. Quantification is shown in red above each band as the fraction of total protein in the P versus S fraction. Results are representative of 3 independent experiments. The underlying data of panels B, C, and F can be found in S1 Data. CCP, clathrin-coated pit; CLCa, clathrin light chain a; CME, clathrin-mediated endocytosis; Dyn1, dynamin-1; epi, epifluorescent; GSK3β, glycogen synthase kinase-3 beta; KO, knockout; TfnR, transferrin receptor; TIRFM, total internal reflection fluorescence microscopy. https://doi.org/10.1371/journal.pbio.2005377.g005 To directly and quantitatively compare the relative recruitment efficiencies of the two isoforms to CCPs, we reconstituted these knockout cells with their respective eGFP-tagged isoforms and sorted for expression comparable to their endogenous levels (i.e., in these A549 cells we chose cells in which Dyn1a-eGFP levels were approximately 5-fold higher than Dyn2-eGFP) (Fig 5A). Additionally, we introduced SNAP-CLCa and mRuby2-CLCa in Dyn1a-eGFP and Dyn2-eGFP cells, respectively, so that we could distinguish the two A549 cell lines (i.e., Dyn1KO:Dyn1a-eGFP:SNAP-CLCa from Dyn2KO:Dyn2-eGFP:mRuby2-CLCa) while imaging them in the same TIRFM field of view under the same conditions (Fig 5D). These data directly show the differential recruitment efficiencies of Dyn1 and Dyn2 to CCPs. Live-cell imaging revealed the typical gradual accumulation and burst of Dyn2-eGFP recruitment to CCPs when averaged over the cohort of 40–60-s lifetime CCPs (Fig 5E). However, under identical imaging conditions of the same fluorophore, Dyn1a-eGFP was recruited, on average, at least 10-fold less efficiently, even though it is expressed at higher abundance. The maximum intensity of tagged Dyn2 versus Dyn1 recruitment was also higher, albeit showing only an approximately 3-fold differential (Fig 5F). A likely explanation for the differences in average and peak measurements is that in A459 cells, Dyn1a-eGFP does display a slight burst of recruitment at late stages of CCV formation that is visible when the Dyn1 signal is rescaled (S4D Fig). Finally, to verify our results using an independent method, we performed Western blotting after subcellular fractionation and isolation of membrane versus cytosolic fractions, as confirmed using membrane-associated TfnR and cytosolic MEK1/2 as markers (Fig 5G). Under these fractionation conditions, approximately 90% of Dyn2 is membrane associated, whereas only 50% of Dyn1 sediments with the membrane fraction (Fig 5G). We observed a consistent, approximately 20% increase of membrane-associated Dyn1 upon GSK3β inhibition that was not detected by TIRFM. These biochemical data indicate a greater extent of membrane association of both active and inactive Dyn1 than detected at CCPs by TIRFM. The differences could reflect recruitment of Dyn1 to sites on the plasma membrane other than CCPs, as has been previously reported [40]. The approximately 20% increase in recruitment of activated Dyn1 likely reflects the increase in number of CCPs that occurs upon GSK3β inhibition, rather than an increase in Dyn1 per CCP. Consistent with TIRFM data, the distribution of phosphorylated Dyn1 (detected with an S774 phosphospecific antibody) was indistinguishable from total Dyn1 (i.e., there was no de-enrichment of phosphorylated Dyn1 in the membrane-bound fractions). These data confirmed that dephosphorylation of Dyn1 on S774 by GSK3β inhibition does not enhance its recruitment to CCPs. Thus, the effects of activated Dyn1 on CCP initiation and maturation occur either independently of its direct association with CCPs or, more likely, are manifested by very low levels of CCP-associated dephosphorylated Dyn1. Dyn1 and Dyn2 do not efficiently co-assemble Dynamin exists as a tetramer in solution [41,42] and assembles into higher-order helical oligomers on the membrane. Exploiting Dyn1KO and Dyn2KO A549 cells reconstituted with Dyn1a- or Dyn2-eGFP, respectively, we next assessed the degree to which Dyn1 and Dyn2 form hetero-tetramers in solution. Dyn1- or Dyn2-eGFP were efficiently immunoprecipitated with anti-eGFP nanobodies and the immunobeads were washed with 300 mM salt to disrupt any potential higher-order dynamin assemblies before measuring the fraction of Dyn2 or Dyn1 that coprecipitated. Under these conditions, we pulled down nearly 100% of the eGFP-tagged dynamins but only approximately 30% of Dyn2 with Dyn1-eGFP and <5% of Dyn1 with Dyn2-eGFP (S5A Fig). The difference in the extent of hetero-tetramerization is consistent with the approximately 5-fold higher levels of expression of Dyn1 versus Dyn2 in these cells. Thus, the two isoforms predominantly exist as homo-tetramers in solution. We also examined the relative abilities of Dyn1 and Dyn2 to co-assemble into higher-order structures in vitro. For this, we used a dominant-negative Dyn1 mutant (Dyn1S45N) defective in GTPase activity, which, when co-assembled with wild-type dynamin into higher-order oligomers on lipid nanotubes, will inhibit total assembly-stimulated GTPase activity through the intercalation of GTPase-defective subunits adjacent to wild-type subunits [43,44]. As expected, Dyn1S45N efficiently co-assembles with Dyn1WT such that, when present at equimolar levels, the total assembly-stimulated GTPase activity is inhibited by 50%. In contrast, at the same concentrations of Dyn1S45N, Dyn2 GTPase activity was significantly less affected (S5B Fig), indicating that Dyn2 less efficiently co-assembles into higher-order oligomers with the mutant Dyn1. Thus, consistent with their differential recruitment to CCPs, even when present at comparable levels of expression in the same cell type, the two isoforms only weakly interact. Genome-edited cells reveal that Dyn1 and Dyn2 are recruited to most CCPs in A549 cells Our results establish that Dyn1 and Dyn2 are differentially recruited to CCPs in non-neuronal cells and that, on average, Dyn1 is recruited at much lower levels than Dyn2. Despite this, acute activation of Dyn1 globally alters CCP dynamics. Thus, we next directly compared the recruitment of Dyn1 and Dyn2 to CCPs to determine whether Dyn1 is recruited at low levels to all CCPs or instead might be recruited at higher levels to a subpopulation of CCPs. Such heterogeneity would be lost by averaging. For this, we took advantage of the higher levels of Dyn1 expression in A549 cells and generated double genome-edited cells expressing Dyn1a-eGFP and Dyn2-mRuby2. We first used ZFNs to generate Dyn2 mRuby2-edited A549 cells and subsequently introduced a C-terminal eGFP to the Dyn1a splice variant using CRISPR/Cas9, as described earlier (Fig 1A, see Materials and methods). This yielded an A549 cell line homozygous for endogenously tagged Dyn2-mRuby2 and heterozygous for endogenously tagged Dyn1a-eGFP (2 of 3 Dyn1 alleles tagged in these triploid A549 cells) (Fig 6A). We confirmed that the double genome-edited cells exhibited comparable rates of TfnR uptake, as well as the degree of dependence on Dyn2 for CME, relative to the parent cells (Fig 6B). SNAP-CLCa was introduced into these cells by lentiviral transfection (Fig 6C), and we confirmed that GSK3β inhibition resulted in increased rates of CCP initiation, reduced CCP lifetimes, and altered the lifetime distributions of CCPs (Fig 6D–6F), as in the parental cells. Thus, the genome-edited Dyn isoforms were functionally active. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Generation and characterization of dual genome-edited Dyn1a-eGFP and Dyn2-mRuby2 A549 cells. (A) Immunoblot validation of Dyn1a-eGFP and Dyn2-mRuby2 single- and dual-genome-edited A549 cells. (B) TfnR endocytosis in dual genome edited A549 cells compared to parental A549 cells and their sensitivity to siRNA-mediated Dyn2 knockdown. (C) Representative TIRF images of Dyn1 and Dyn2 distribution relative to CLCa in dual genome-edited A549 cells. CCP initiation densities (D), median lifetimes (E), and the lifetime distribution (F) of bona fide CCPs in dual-genome-edited A549 cells with or without GSK3β inhibition, determined as in Fig 2. The underlying data of panels B and D–F can be found in S1 Data. Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta; TfnR, transferrin receptor. https://doi.org/10.1371/journal.pbio.2005377.g006 We next assessed the interplay between Dyn1a-eGFP and Dyn2-mRuby2 using three-color live-cell TIRFM imaging at 0.5 Hz (2 s per frame) (Fig 7A, S3 Movie). As reported previously, we detected fluctuations of both Dyn1 and Dyn2 at CCPs over their lifetimes (examples shown in Fig 7B) and frequently detected a burst of Dyn2 just prior to CCV formation. In many cases, we also detected a burst of Dyn1 recruitment, albeit to a lesser degree. For more quantitative analysis of these data, we applied the 3-channel functionality of our cmeAnalysis package to perform three-color master/slave analyses [22]. Using clathrin as the “master” channel and Dyn1 and 2 as “slave” channels, we determined whether the clathrin tracks contained either Dyn1, Dyn2, both, or neither. Individual CCP tracks were considered positive for Dyn1 and/or Dyn2 if the intensities of Dyn1/2 signals detected at the position of the clathrin tag were significantly higher than the local Dyn1/2 background signal around the clathrin tag position for a period of time exceeding random associations, as previously described [22]. This analysis revealed that in double genome-edited Dyn1a-eGFPend/Dyn2-mRuby2end A549 cells, both Dyn2 and Dyn1 could be robustly detected in approximately 75% of all bona fide CCPs (Fig 8A). Moreover, in this population of CCPs, a clear burst of recruitment of both Dyn1a-eGFP and Dyn2-mRuby2 could be detected prior to CCV formation. Importantly, the apparently higher levels of recruitment of Dyn1-eGFP versus Dyn2-mRuby2 in these genome-edited cells is not a reflection of protein levels but rather of imaging conditions and brightness for two different fluorophores (compare with Fig 5E). The remaining CCPs were roughly equally distributed as Dyn1 only, Dyn2 only, and both Dyn1- and Dyn2-negative subpopulations (Fig 8A). Note that the Dyn2 levels in the “Dyn1 only” CCPs were still on average higher than background (Dyn1/Dyn2 negative), reflecting the stringency of our master/slave detection and suggesting that Dyn2 is recruited to >90% of all CCPs, albeit to variable extents. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Tracking clathrin and dynamins in dual genome-edited Dyn1a-eGFP and Dyn2-mRuby2 A549 cells. (A) Representative TIRF images and corresponding kymographs of dynamic behavior of overexpressed SNAP-CLCa, Dyn2-mRuby2end, and Dyn1a-eGFPend in dual genome-edited A549 cells. See S3 Movie. (B) Examples of Dyn1 and Dyn2 dynamics at individual CCPs (i–iv) and (C) their corresponding quantitative traces. CLCa, clathrin light chain a; Dyn1, dynamin-1. https://doi.org/10.1371/journal.pbio.2005377.g007 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 8. Dyn1 and Dyn2 are recruited to the same CCPs and Dyn1 activation alters the dynamics of all CCP subpopulations. (A) Triple-color master/slave analyses of average dynamics of recruitment of Dyn2-mRuby2endo and/or Dyn1a-eGFPend to lifetime cohorts of SNAP-CLCa labeled CCPs identifies Dyn1/Dyn2 positive, Dyn1 only, Dyn2 only, and Dyn1/2-negative subpopulations of CCPs. The percentage of detected CCPs in each class is indicated. (B) Effect of GSK3β inhibition on the median lifetimes of compositionally distinct CCP subpopulations. The underlying data of panel B can be found in S1 Data. CCP, clathrin-coated pit; CLCa, clathrin light chain a; Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta. https://doi.org/10.1371/journal.pbio.2005377.g008 We next compared per cell median lifetimes of CCPs relative to their dynamin isoform composition and found that CCPs bearing higher levels of Dyn2 and Dyn1 exhibited longer lifetimes (median approximately 80 s) than single-positive CCPs (median approximately 38 s) (Fig 8B). CCPs that failed to detectably recruit either isoform were the shortest lived (median approximately 20 s) and likely represent abortive CCPs. These findings are consistent with previous data suggesting that a threshold level of Dyn2 recruitment is required for efficient CCP maturation [22,34]. All of these CCP subpopulations showed a significant decrease in CCP lifetimes upon inhibition of GSK3β, consistent with other data that only low levels of Dyn1 are required to alter CCP maturation. SNX9 is required for activated Dyn1-dependent effects on CCP maturation Our findings thus far point to isoform-specific functions of Dyn1 and Dyn2 and hence suggest the existence of isoform-specific binding partners. Dyn1 and Dyn2 are >80% identical except for their C-terminal PRDs, which are only 50% identical and likely determine isoform-specific interactions with SH3 domain-containing proteins. The Dyn1KO and Dyn2KO A549 cells provide an opportunity to measure Dyn2 and Dyn1-dependent CME, respectively, without the possibility of compensation. Thus, we measured, by TfnR uptake, the effects of siRNA knockdown of several known SH3 domain-containing binding partners on Dyn2-dependent CME in the Dyn1KO cells and on Dyn1-dependent CME in the Dyn2KO cells. Knockdown of these dynamin partners has only mild effects on TfnR uptake in parental A549 cells and in Dyn1KO cells, whose endocytosis is exclusively Dyn2 dependent (Fig 9A). Whether these mild effects reflect partial redundancy with other dynamin partners, activation of compensatory mechanisms [23], or that these factors, which were identified primarily as dynamin partners in brain lysates, play only minor roles in TfnR uptake in non-neuronal cells, cannot be discerned from these studies. Interestingly, siRNA knockdown of Grb2 appeared to inhibit TfnR uptake in Dyn1KO cells by approximately 20%, while not affecting TfnR uptake in either parental or Dyn2KO cells. This suggests that Grb2 might preferentially function together with Dyn2 in CME and that its depletion in parental cells can be compensated for by Grb2-independent Dyn1 activity. In contrast, siRNA knockdown of SNX9 only mildly inhibited TfnR uptake in parental A549 and had no significant effect on Dyn2-dependent TfnR uptake in Dyn1KO cells, but decreased TfnR uptake in Dyn2KO cells by >50% (Fig 9A). Thus Dyn1-dependent endocytosis appears to be particularly sensitive to SNX9 knockdown. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 9. SNX9 preferentially binds activated Dyn1 and is required for Dyn1-dependent changes in the lifetime distribution of CCPs. (A) Effect of siRNA knockdown of the indicated dynamin SH3 domain-containing endocytic accessory proteins on TfnR endocytosis in parental, Dyn1KO, and Dyn2KO A549 cells. siEndo refers to siRNA knockdown of endophilin A1, 2, and 3; siITSN refers to siRNA knockdown of intersectins 1 and 2; all others were single siRNAs. Knockdown efficiencies were determined to be >85% by western blotting. Data are normalized to the extent of TfnR uptake in control siRNA-treated parental, Dyn1KO, and Dyn2KO cells, which is set to 100, to allow direct comparison of the relative effects of siRNA knockdowns. (B) eGFP pulldown of Dyn1aWT-eGFP, Dyn1aS774/8A-eGFP, or Dyn2WT-eGFP expressed in Dyn1KO or Dyn2KO A549 cells, respectively, using anti-eGFP nAbs. Parental cells that do not express an eGFP-tagged protein (Dyn1end) are used as control. The pulldown fractions were analyzed by immunoblot. Effect of SNX9 siRNA-mediated knockdown on (C) CCP initiation densities, (D) median lifetimes, and (E) lifetime distribution of bona fide CCPs in Dyn1KO H1299 cells overexpressing either Dyn1aWT-eGFP or Dyn1aS774/8A-eGFP (data are derived from 7 movies for each condition, with each movie consisting of 1–3 cells). Each data point is the average value from a single movie. The underlying data of panels A, C–E, and L can be found in S1 Data. (* p ≤ 0.05, ** p ≤ 0.01, *** p ≤ 0.001) CCP, clathrin-coated pit; Dyn1, dynamin-1; SH3, SRC Homology 3; siRNA, small interfering RNA; SNX9, sorting nexin 9; TfnR, transferrin receptor. https://doi.org/10.1371/journal.pbio.2005377.g009 We next tested whether SNX9 preferentially interacts with Dyn1 versus Dyn2 by GFP pulldown assays using Dyn1KO cells reconstituted with either Dyn1aWT- or Dyn1S774/8A-eGFP and Dyn2KO A549 reconstituted with Dyn2-eGFP. Consistent with previous results [45,46], we confirmed that SNX9 binds both Dyn1 and Dyn2 (Fig 9B). However, the ratio of SNX9 binding to Dyn1 versus Dyn2 was 1.7 ± 0.6 (mean ± SEM, n = 3), indicative of a slight preference for Dyn1. Importantly, SNX9 showed a marked preference for binding to the nonphosphorylated and active Dyn1S774/8A-eGFP. The ratio of SNX9 binding to Dyn1S774/8A versus Dyn1WT was 3.6 ± 0.9 (mean ± SEM, n = 3). These data suggested that SNX9 might be a preferential functional partner of activated Dyn1. To test whether SNX9–Dyn1 interactions were required for the effects of activated Dyn1 on CCP initiation rates, CCP maturation, or both, we asked returned to the Dyn1KO H1299 cells reconstituted with Dyn1WT versus Dyn1S774/8A and tested whether the selective effects of Dyn1S774/8A on CCP dynamics (Fig 4A–4C) were dependent on SNX9. Knockdown of SNX9 decreased the rate of CCP initiation in Dyn1WT but was not required for the enhanced rate of CCP initiation triggered by Dyn1S774/8A expression (Fig 9C). Thus, other, yet-unidentified binding partners are responsible for the Dyn1-dependent effect on CCP initiation. SNX9 knockdown also led to an increase in the median CCP lifetimes in both Dyn1WT- and Dyn1S774/8A-expressing cells (Fig 9D). These data suggest that SNX9 functions in both Dyn1-dependent and independent stages of CCP maturation. Consistent with this, SNX9 knockdown also abrogated the effects of Dyn1S774/8A expression on the lifetime distribution of bona fide CCPs (Fig 9E), reverting the quasi-exponential distribution seen in Dyn1S774/8A to a distribution nearer to control. The strong effect of SNX9 knockdown is also seen in the rightward shift of the lifetime distribution of Dyn1WT cells treated with SNX9 siRNA. Together, these data suggest multiple roles of SNX9 at multiple stages of CME, including the support of Dyn1’s early functions in accelerating CCP maturation. Dyn1 is activated downstream of the EGFR We have shown that strong pharmacological inhibition of GSK3β activates Dyn1 in non-neuronal cells and results in increased rates of CCP initiation and maturation, leading to increased rates of TfnR uptake via CME. However, it is not clear whether this regulatory effect on Dyn1 function modulates CME under more physiologically relevant conditions. To test this, we treated serum-starved A549 cells with epidermal growth factor (EGF), which is known to activate Akt and in turn to phosphorylate and inactivate GSK3β [47]. We confirmed that GSK3β is phosphorylated in EGF-treated cells and that this resulted in reduced levels of phosphorylation of Dyn1 at S774 (Fig 10A, quantified in Fig 10B and 10C). As predicted by the results of inhibitor experiments, EGF treatment of serum-starved cells also increased the rate of CCP initiation (Fig 10D), decreased CCP lifetimes (Fig 10E), and, compared to control cells, resulted in a shift of the lifetime distributions of bona fide CCPs to a more quasi-exponential distribution (Fig 10F). Importantly, the effects of EGF treatment on CCP initiation rate and lifetimes were not seen in A549 Dyn1KO cells (Fig 10G and 10H). These data suggest that Dyn1 can be activated to alter CCP dynamics under physiological conditions through signaling downstream of EGFR. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 10. EGF stimulation alters CCP dynamics in a Dyn1-dependent manner. (A) Immunoblot analysis of changes in phosphorylation state of Dyn1 S774 and GSK3β upon EGF stimulation of parental A549 cells. (B,C) Quantification of the EGF-triggered changes in phosphorylation state (i.e., ratio of phosphorylated/ total protein) of GSK3β and Dyn1 (mean ± SD of n = 3 experiments, data are normalized to 0 min time point). CCP initiation densities (D), median lifetimes (E), and the lifetime distribution (F) of bona fide CCPs in serum starved A549 cells before (Control, Ctrl) or after incubation with EGF (20 ng/ml) for 10 min prior to imaging. CCP initiation densities (G) and median lifetimes (H) in serum-starved Dyn1KO A549 cells before (Ctrl) or 10 min after incubation with EGF (EGF). The underlying data of panels B–H can be found in S1 Data. CCP, clathrin-coated pit; Dyn1, dynamin-1; GSK3β, glycogen synthase kinase-3 beta. https://doi.org/10.1371/journal.pbio.2005377.g010 Discussion Our experiments provide further evidence that Dyn1, in addition to its well-studied roles in membrane fission during synaptic vesicle recycling, also has noncanonical functions as a regulator of the earliest stages of CME in non-neuronal cells. As in neurons, Dyn1 activity is negatively regulated by constitutive phosphorylation and activated by dephosphorylation. When studied at endogenous levels of expression, we show that Dyn1 and Dyn2 have distinct functions in CME, reflected in their quantitatively and qualitatively different recruitment to CCPs. While Dyn1 is expressed at very high levels in the brain, it is, like Dyn2, also widely expressed, albeit at lower levels, in all tissues and cells [2]. Importantly, we show that acute activation of even low, nearly undetectable levels of Dyn1 can increase the rates of CCP initiation and maturation to accelerate CME. These effects of Dyn1 activation occur well upstream of membrane fission and are not accompanied by a pronounced burst of recruitment prior to CCV formation. Hence, they reflect noncanonical activities of Dyn1, likely mediated by unassembled tetramers, that are distinct from its well-studied role in fission. As in neurons [33], Dyn1 is constitutively inactivated in non-neuronal cells by phosphorylation at S774 in the PRD by GSK3β. Acute chemical inhibition of GSK3β activates Dyn1 to alter CCP dynamics and increase the rate of CME. While GSK3β has numerous substrates, we show that Dyn1 is both necessary and sufficient to account for the effects of GSK3β inhibition on CCP dynamics and CME. Specifically, the effects of GSK3β inhibition on CME are dependent on Dyn1 but not Dyn2 expression, and Dyn1KO cells reconstituted with a nonphosphorylatable mutant of Dyn1 show increased rates of CCP initiation and maturation that phenocopy the effects of GSK3β inhibition. Even when Dyn1 is activated, either by GSK3β inhibition or by mutation of S774 and S778 to alanine, CME remains dependent on Dyn2 (data herein and [23]). Thus, the two isoforms play functionally distinct roles in CME. It is unlikely that Dyn1 activation merely releases an inhibitory effect of inactive Dyn1 on CME, for example by competing with Dyn2, because even high levels of overexpression of wild-type Dyn1 does not inhibit CME (see for example, [26]). Further studies are needed to elucidate the mechanisms underlying these distinct roles. A direct comparison of the in vitro properties of Dyn1 and Dyn2 established that they differ in their curvature generating/sensing properties [14]. While Dyn1 is an efficient curvature generator that is able to tubulate and catalyze fission from planar lipid templates, Dyn2 is a curvature sensor that is able to catalyze membrane fission of highly curved lipid templates but requires the synergistic activity of curvature-generating N-terminal Bin/Amphiphysin/Rvs (N-BAR) domain-containing accessory factors to drive curvature generation and fission from planar templates [14,15]. Strikingly, these biochemical differences could be ascribed to a single residue (Y600 in Dyn1, L600 in Dyn2) encoded within hydrophobic loops of the curvature-generating Pleckstrin homology (PH) domain of dynamin [14]. Based on these biochemical differences, it was suggested that the unique properties of Dyn2 might enable this isoform to monitor CCP maturation and to catalyze fission only after the development of a narrow membrane neck connecting deeply invaginated CCPs to the plasma membrane. Unexpectedly, the findings presented here and elsewhere [23,35,39] establish that Dyn1 uniquely functions to regulate the earliest stages of CME, including the rate of CCP initiation and maturation. The mechanisms underlying these Dyn1-specific activities remain to be elucidated. Activation of Dyn1 also altered the shape of the lifetime distribution curve for CCPs from a broad Rayleigh-like distribution with a distinct peak at approximately 30 s to a more exponential distribution. We have previously suggested that the Rayleigh-like shape reflects rate-limiting regulatory processes operating during the first 30 s of CCP progression [20,22,38]. It is possible that, due to its curvature-generating ability and/or through interactions with other partner proteins, Dyn1 activation accelerates these complex early processes of CCP maturation. Although Dyn1 and Dyn2 exhibit >80% sequence identity within their GTPase, middle, PH domains, and GTPase effector domain (GED), previous studies of the cellular activities Dyn1/Dyn2 chimeras have nonetheless revealed striking isoform-specific functional differences conferred by both the PH and GTPase domains [14,48]. Most divergent among mammalian dynamin isoforms is the PRD, which functions to mediate interactions with numerous SH3 domain-containing binding partners, and has been shown to target dynamin to CCPs [40]. Earlier comparative studies of Dyn1 and Dyn2 [13], as well as Dyn1/Dyn2 PRD chimeras expressed at near endogenous levels [14], have shown that Dyn2 is more efficiently recruited to CCPs in a PRD-dependent manner. However, these studies did not take into account the negative regulation of Dyn1 by GSK3β phosphorylation. Here, we reproduce and extend these findings by showing that the differential recruitment of Dyn1 is not due to phosphorylation of its PRD, at least on S774 or 778. Indeed, the recruitment of Dyn1 to CCPs was not significantly enhanced by GSK3β phosphorylation or when S774/S778 were mutated to nonphosphorylatable alanines. Thus, surprisingly, the effects of activated Dyn1 on CCP dynamics appear to occur independent of detectably enhanced recruitment to CCPs. It will be important to identify isoform-specific binding partners for Dyn1 and Dyn2 in non-neuronal cells. To date, most dynamin binding partners, including endophilin, amphiphysin, and intersectin have been identified in brain lysates in which Dyn1 is highly expressed and may play a specialized function in rapid synaptic vesicle recycling. Thus, it is perhaps not surprising that siRNA knockdown of these dynamin binding partners in non-neuronal cells has only mild effects on primarily Dyn2-dependent TfnR endocytosis. Further studies are needed to identify essential non-neuronal effectors of both Dyn2 and Dyn1 function in CME. Unexpectedly, our data suggests that SNX9, which was first identified as a major binding partner of Dyn2 in HeLa cells [45], interacts most strongly with dephosphorylated Dyn1 and that the effects of Dyn1 activation on early CCP maturation are dependent on SNX9. Published findings on SNX9 function in CME are enigmatic. Consistent with our findings in A549 NSCLC cells, siRNA-mediated knockdown of SNX9 has only mild effects on CME in several cell lines studied [46,49]. While it has been suggested that these mild effects are due to redundant functions of the distantly related (40% sequence identity) SNX18 [50], this is not the case in all cell types [49,50]. TIRFM studies on the recruitment of overexpressed SNX9-GFP to CCPs have also yielded differing results: It has been reported to be recruited coincident with [46], after [31], and before dynamin [50]. Interestingly, one study reported that SNX9 might linger at endocytic “hot-spots,” where it could function as an organizer of CCP nucleation [51]. Our results further suggest a more complex role for SNX9 at multiple stages of CME. SNX9 is not required for activated Dyn1-dependent increases in the rates of CCP initiation. However, it is required for the effects of Dyn1 activation on accelerating CCP maturation, as indicated by the marked switch from Rayleigh-like to quasi-exponential CCP lifetime distributions, which is reversed by SNX9 knockdown. That SNX9 knockdown alone decreases the rate of CCP initiation and slows CCP maturation in cells expressing Dyn1WT suggests other, potentially Dyn2-dependent and/or dynamin-independent functions in CME. More work is required to define both the multiple functions of SNX9 in CME and to identify Dyn1-specific binding partners required for CCP initiation. Recent studies have shown that Dyn1 is upregulated and/or activated in several cancer cell lines [35,39], leading to the suggestion that Dyn1 might function as a nexus between signaling and CME [2]. Here, we show that Dyn1 can be activated downstream of EGFR to alter CCP dynamics. Previous studies showed that tumor necrosis factor-related apoptosis-inducing ligand (TRAIL)-activated death receptors can activate Dyn1 to drive their selective uptake via CME [35]. Similarly, elegant studies on clathrin-mediated endocytosis of the G-protein coupled β-adrenergic receptors have shown that they alter the maturation kinetics of the CCPs in which they reside through delayed recruitment of Dyn2 [52]. These authors did not examine Dyn1 recruitment or function. Further studies will be needed to determine whether other signaling receptors can selectively alter the composition and/or maturation kinetics of CCPs in which they reside and, if so, whether these changes, as suggested by our present data, are at least in part Dyn1 dependent. Based on our results and other recent findings [23,35,39] regarding Dyn1 function downstream of signaling in non-neuronal cells, it is perhaps surprising that Dyn1 knockout mice develop normally, can live for up to 2 weeks after birth, and exhibit primarily neuronal defects [12]. This could, in part, be due to the redundant function of Dyn-3, as Dyn1/Dyn3-null mice exhibit a more severe phenotype and die within hours of birth [11]. However, we speculate that kinase-based activation of Dyn1 might function at the level of individual CCPs to foster their initiation and accelerate maturation, perhaps at a threshold of signaling not reached during normal development. Indeed, recent evidence has pointed to cargo-selective roles of Dyn1 in regulating CME and signaling in cancer cells [2,35,39]. Therefore, it might be interesting to probe Dyn1 or Dyn1/Dyn3 knockout mice for other, potentially more subtle, non-neuronal phenotypes related to signaling in health and disease. Materials and methods Cell culture, vector preparation, transfection, and culture perturbations Non-small cell lung cancer cell lines A549 and H1299 were kindly provided by Dr. John Minna (The Hamon Center for Therapeutic Oncology, Depts. of Internal Medicine and Pharmacology, UTSW) and were grown in RPMI 1640 (Life Technologies) with 5% FBS at 37°C and 5% CO2 and imaged in a temperature-controlled chamber mimicking similar culture conditions. The retroviral expression vector pMIEG was a modified pMIB (CMV-IRES-BFP) vector encoding Dyn1 cDNA with N-terminal HA-tag and C-terminal eGFP fusion tag. Point mutations to introduce S774/8A in Dyn1 cDNA were performed by site-directed mutagenesis. The lentiviral expression vector pLVX-puro (Clontech) encoded CLCa N-terminally tagged with mRuby 2 [53] or SNAP-tag [54] spaced with a 6 amino acid GGSGGS linker. The constructs were assembled from PCR fragments of mRuby2, SNAP, and CLCa (for primers, see S1 Table) in yeast as described below and subsequently cloned into pLVX-puro. Lentiviruses were generated in 293T packaging cells following standard transfection protocols [55] and were used for subsequent infections. They were prepared to transduce fluorescently tagged (mRuby2 or SNAP tag) CLCa fused to its N-terminus. Infected cells were selected using 10 μg/ml puromycin for 4 d, conditions under which uninfected cells perished. The cells were passaged for 2 w before imaging for CME analysis. Retroviruses were also generated in 293T cells and used to stably transduce eGFP tagged Dyn1WT, Dyn1S774/8A, and Dyn2WT proteins. Gene transduction was performed by exposing A549 or H1299 cells to retrovirus-containing cell culture supernatants through two rounds of viral transduction spread across 5 d. The recipient cells were further expanded to confluency in a 10-cm culture dish and FACS sorted for eGFP levels comparable to endogenous Dyn1-eGFP in A549 cells. Transfections for siRNA knockdown experiments were carried out using Lipofectamine 2000 or Lipofectamine RNAi-Max (Life Technologies), following manufacturer’s protocol. For siRNA mediated knockdown, approximately 2 × 105 cells (H1299) or 3 × 105 cells (A549) were plated in each well of a six-well plate. Twenty nmol siRNA was used per well, and two rounds of transfection across 5 days was sufficient to achieve over 90% knockdown. Perturbation of culture conditions by GSK3β inhibitor involved the addition of 10 μM CHIR99021 (Sigma) to prewarmed culture media and incubation of cells for 30 min before additional analysis. Growth factor stimulation was performed by adding 20 ng/ml EGF (Invitrogen) to prewarmed, serum-free culture media. Cells were analyzed after 10 min of incubation with EGF. Generation of genome-edited cell lines Genome-edited A549 and H1299 cells were generated by editing Dyn1 and Dyn2 to carry fusion tags. For the endogenous labeling of Dyn1 with fluorescent reporter proteins, we chose an approach based on site directed introduction of CRISPR/CAS9n-targeted DNA breaks and template assisted homology driven repair. eGFP fused to Dyn1 at its C-terminus was generated by CRISPR/Cas9n nickase strategy targeting the end of exon 21 of the DNM1 gene, inserting the last 19 amino acids of splice isoform “a,” a seven amino acid linker [32], monomeric eGFP with a stop codon, and the SV40 polyadenylation signal. In the donor plasmid, this inserted sequence was flanked by approximately 950 base pair homology arms for HDR. The +gRNA pair was designed using publicly available software (http://crispr.mit.edu/) and prepared as described [56] with oligos DNM1-Nuclease-A-f/ DNM1-Nuclease-A-r and DNM1-Nuclease-B-f/ DNM1-Nuclease-B-r, respectively (S1 Table). For assembly of the donor vector, the segments were amplified with oligonucleotides coding approximately 30 nucleotide overhangs. The bacterial artificial chromosome clone RP11-348G11 (BACPAC Resources Center, Children’s Hospital Oakland Research Institute, Oakland, California) covering the end of human DNM1 gene was used as template for the left and right homology arms (Fig 1A). The left and right homology arms were amplified using primer pairs DNM1-LH-f/DNM1-LH-r and DNM1-RH-f/DNM1-RH-r, respectively (see S1 Table). The 19 C-terminal amino acids of splice isoform “a” and the linker sequence DPPVATL [32] were covered with oligonucleotides DNM1-C-assembly-f and DNM1-C-assembly-r and amplified with short primers DNM1-Cterm-f and DNM1-Cterm-r. The sequence coding for monomeric eGFP and the SV40 polyadenylation signal were amplified from plasmid peGFP-N1 (Clontech), which carried the A206K mutation [57] with primers DNM1-eGFP-f/DNM1-eGFP-r and DNM1-pA-f/DNM1-pA-r, respectively. The first and last primers (DNM1-LH-f, DNM1-RH-r) also included overhangs for the E coli/yeast shuttle vector pRS424 [58]. Dyn2-mRuby genome-edited cells were generated using previously validated ZFNs [32,34]. For the DNM2-mRuby2 donor vector, the homology arms were amplified from the published [34] DNM2-eGFP construct (gift from D. Drubin, University of California, Berkeley) with primers DNM2-LH-f/DNM2-LH-r and DNM2-RH-f/DNM2-RH-r, respectively. The mRuby2 segment together with the linker sequence, DPPVATL [32], was amplified from pmRuby2-C1, a gift from Michael Lin (Addgene #40260) [53]. The PCR products were purified on 1% agarose gels and extracted using standard protocols before transformation into YPH500 yeast cells [59]. Yeast transformation, plasmid extraction, and plasmid validation were performed as described earlier [60]. The guide-RNA plasmids for DNM1-eGFP and donor vectors for both DNM1-eGFP and DNM2-mRuby2 are available from Addgene (IDs 107795, 107796, 107794, and 107793, respectively). For both the edits, the nCas9 nickase + gRNA pairs or the ZFN nuclease pairs were added at 1 μg DNA concentration, and 2 μg of the donor plasmid was added to this mixture in 150 ul OptiMEM (Life Technologies). This mixture was then added to 5.5 μl of Lipofectamine 2000 (Life Technologies) in 150 μl OptiMEM (Life Technologies), briefly vortexed, and incubated at room temperature for 15 min. The mixture was then added to cells plated 12 h earlier at 70% confluency (approximately 3 × 106 cells per well in six-well dish) with freshly replaced media. Transfect-containing media was replaced by prewarmed fresh media and the cells were allowed to grow for the next 48 h and then passaged for expansion in a 10-cm dish. The expanded cells were sorted as eGFP (or mRuby2) gene-edited single cells into 96 well plates 4 days after transfection using a FACSAria 2-SORP (BD Biosciences, San Jose, CA) instrument equipped with a 300-mW, 488-nm laser and a 100-μm nozzle. Clonal expansion ensued by incrementing the culture dish area and maintaining a minimum 50% cell confluency. Single clones were then assayed for edits by western blotting, and cells positive for genome edits were expanded. In order to generate double genome edited A549 cells, the A549 clone, 2C8, with homozygous Dyn2-mRuby2 knock-in was chosen and subsequently edited for Dyn1, and cell selection was performed as before using FACS preliminary screen followed by western blotting for validation. Dyn1 KO H1299 cells were generated as previously described [23] and the same strategy was employed to generate A549 Dyn1 KO cells. Briefly, cells plated in six-well plates were transfected with 1 μg each of single-guide RNAs (sgRNAs) and Cas9 nickase encoding plasmids and cotransfected with a 20th of peGFP plasmid. eGFP-positive cells were assumed to have harbored both the sgRNA guides and single-cell-sorted by FACS. In addition, Dyn2 KO cells were generated using a similar double nickase strategy with sgRNAs CGATCTGCGGCAGGTCCAGGTGG and CGCCGGCAAGAGCTCGGTGCTGG in the pX335 vector. Complete knockout of Dyn1 and 2 was validated by western blotting. Immunoprecipitation, pulldowns, and subcellular fractionation Glutathione S-transferase (GST)-SH3 pulldown. GST-Amph II SH3 pulldown involved lysing H1299 cells in lysis buffer (50 mM Tris, 150 mM NaCl, 1X Protease Inhibitor Cocktail [Roche]) containing 0.2% Triton X-100. Cells were dounced with 27.5-G syringe about 20 times or until most of the cells were ruptured to release intact nuclei. The post-nuclear fraction (PNF) was obtained by spinning the lysate at 10,000 xg at 4°C for 10 min and collecting the supernatant. About 3 mg of PNF in 1-ml volume was used for each pulldown. Addition of beads (approximately 20 μl) with bait protein (GST-Amph II SH3) in PNF followed by gentle rotation for 1 h allowed binding of target proteins. The bound fraction was washed twice with lysis buffer containing 0.2% Triton X-100 and the resulting beads were denatured using 2X Laemmli buffer (Bio-Rad) reduced with 5% β-mercaptethanol, boiled and run on SDS-PAGE gel of appropriate separation capacity (7.5% or 12%, based on the target protein size). The pulldowns were analyzed by western blotting. Subcellular fractionation. Confluent A549 cells in a 60-mm dish were detached with 1 ml of 10-mM EDTA at 37 °C for 10 min and washed with PBS by centrifugation and then resuspended in 0.5 ml buffer 2 (25 mM HEPES, 250 mM sucrose, 1 mM MgCl2, 2 mm EGTA, pH 7.4). The resuspended cells were lysed through 3 cycles of freeze-thaw (rapid freezing in liquid nitrogen and slow thawing in room temperature water). Cytoplasm and membrane portions were separated by 30 min ultracentrifugation at 110 kg in a Beckman Coulter rotor (TLA55). Pellets were resuspended with 0.5 ml buffer 2, and both supernatant and pellets were solubilized in 0.5% Triton X-100 for 10 min on ice and then precipitated with 10% TCA, followed by 2 rounds of 1 ml acetone wash. SDS-PAGE gel electrophoresis and western blot were applied as described above [61]. GFP-nAb immunoprecipitation. Confluent A549 cells in a 10-cm dish were detached with 2 ml of 10-mM EDTA at 37 °C for 10 min and washed with PBS by centrifugation and then resuspended and gently lysed for 15 min on ice with 2 ml buffer 3 (0.5% Triton X-100, 25 mM HEPES, 150 mM KCl, 1 mM MgCl2, 2 mM EGTA, 1X Protease Inhibitor Cocktail [Roche], 1X Phosphatase Inhibitor Cocktail, pH = 7.4). Lysates were centrifuged at 5000 xg, 4 °C for 5 min to remove nuclei, and protein concentration in the PNF was determined by Bradford assay. One-half mg of the PNF was added to 30 μl GFP-nAb agarose (Allele Biotech), rotated for 2 h at 4 °C, and then spun down at 2500 xg, 4 °C, 2 min. The agarose was washed twice (1 ml/each) with two different buffers to fulfill different experimental purposes: (1) to probe the dynamin interactors, the agarose was washed with buffer 3; (2) to probe dynamin self-assembly, salt concentration in buffer 3 was brought up to 300 mM to remove indirect dynamin–dynamin interactions. Ten percent cell lysate, which is used to determine immunoprecipitation efficiency, was precipitated with 10% TCA and washed twice with acetone from −20 °C freezer. The samples were subjected to SDS-PAGE and Western blotting for analysis. TIRFM Cells expressing appropriate fluorophores were cultured overnight on an acid-etched and gelatin-coated coverslip, placed in a well in a six-well plate. At the time of imaging, cells were checked for adherence and spreading. When imaging SNAP-tagged proteins, labeling was performed by incubating cells in 1 ml of fresh, prewarmed media containing 1 μl of predissolved SNAP-CELL 647-SiR dye (NEB). After 30 min incubation under standard incubator conditions, the media was aspirated, washed twice with sterile PBS, and reincubated in fresh culture media. The coverslips were mounted on glass slides with spacers and sealed with the same media. For experiments involving the addition of growth factor or inhibitor, cells were preincubated for the appropriate times and the coverslips were mounted as before with the treated media. The coverslips were then imaged using a 60x 1.49 NA Apo TIRF objective (Nikon) mounted on a Ti-Eclipse inverted microscope with Perfect Focus System (Nikon) equipped with an additional 1.8x tube lens, yielding at a total magnification of 108x. TIRF illumination was achieved using a Diskovery Platform (Andor Technology). During imaging, cells were maintained at 37°C in RPMI supplemented with 5% fetal calf serum. Time-lapse image sequences were acquired at a penetration depth of 80 nm and a frame rate of 2 Hz (three or two channels) or 1Hz (single channel) using a sCMOS camera with 6.5mm pixel size (pco.edge). Quantitative analysis of imaging The detection, tracking and analysis of all clathrin-labeled structures and thresholding to identify bona fide CCPs was done as previously described using the cmeAnalysis software package [22]. Briefly, diffraction-limited clathrin structures were detected using a Gaussian-based model method to approximate the point-spread function [22], and trajectories were determined from clathrin structure detections using the u-track software [37]. Subthreshold clathrin-labeled structures (sCLSs) were distinguished from bona fide CCPs, based on the quantitative and unbiased analysis of clathrin intensity progression in the early stages of structure formation [22,62]. Both sCLSs and CCPs represent nucleation events, but only bona fide CCPs represent structures that undergo stabilization, maturation, and, in some cases, scission to produce intracellular vesicles [22,62]). We report the rate of bona fide CCP formation, distribution of their lifetimes, and intensity cohorts, as described previously [22]. As these values will depend on day-to-day variations in the threshold, we image experimental and control conditions on the same day and apply the same threshold to both data sets to ensure that effects we detect are due to the specific experimental variable being assessed. We also report mean and maximum signal intensities in two or three channels for each individual CCP. These are average and maximum signal intensities for individual CCPs as they are extracted by the previously described analysis software [22]. The extraction of CCPs is achieved by a new function added to the cmeAnalysis software published in [22] that allows us to link the classification of events, CCPs, or sCLSs to more sophisticated analysis intensity time courses and lifetime. In this study, we focused merely on per-CCP mean and maximum intensity values, which were averaged per movie (1–5 cells) and finally presented as per-movie distributions covering 10–30 cells per experimental condition. Differences between conditions were assessed by comparison of the normal-distributed per-movie distributions using Student t test and a threshold of p < 0.01 to mark statistical significance. Statistical analysis Control and treatment datasets were statistically analyzed with two-tailed, unpaired Student t tests using Graphpad Prism 5.0 (Graphpad Software, La Jolla, CA), from which p values were derived (* p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001). Error bars representing standard error of the mean (SEM) for at least three independent experiments were calculated using Microsoft Excel. Receptor internalization (endocytosis) assay An in-cell ELISA approach was used to quantitate internalization of TfnR and EGFR, as previously described [23], using either anti-TfnR mAb (HTR-D65) [63] or biotinylated-EGF as ligands. Cells were grown overnight in 96-well plates at a density of 2 x 105 cells/well and incubated with 4 mg/ml of D65 or 20 ng/ml of biotinylated-EGF (Invitrogen) in assay buffer (PBS4+: PBS supplemented with 1 mM MgCl2, 1 mM CaCl2, 5 mM glucose, and 0.2% bovine serum albumin) at 37 °C for the indicated time points. Cells were then immediately cooled down (to 4 °C) to arrest internalization. The remaining surface-bound D65 or biotinylated-EGF was removed from the cells by an acid wash step (0.2 M acetic acid, 0.2 M NaCl, pH 2.5). Cells were then washed with cold PBS and then fixed in 4% paraformaldehyde (PFA) (Electron Microscopy Sciences) in PBS for 30 min and subsequently permeabilized with 0.1% Triton X-100/PBS for 10 min. Internalized D65 was assessed using a goat anti-mouse HRP-conjugated antibody (Life Technologies), and internalized biotinylated-EGF was assessed by streptavidin-POD (Roche). The reaction was developed by a colorimetric approach with OPD (Sigma-Aldrich), and color development was stopped by addition of 50 μl of 5M of H2SO4. The absorbance was read at 490 nm (Biotek Synergy H1 Hybrid Reader). Internalized ligand was expressed as the percentage of the total surface-bound ligand at 4 °C (i.e., without acid wash step), measured in parallel [23]. Well-to-well variability in cell number was accounted for by normalizing the reading at 490 nm with BCA readout at 560 nm. Cell culture, vector preparation, transfection, and culture perturbations Non-small cell lung cancer cell lines A549 and H1299 were kindly provided by Dr. John Minna (The Hamon Center for Therapeutic Oncology, Depts. of Internal Medicine and Pharmacology, UTSW) and were grown in RPMI 1640 (Life Technologies) with 5% FBS at 37°C and 5% CO2 and imaged in a temperature-controlled chamber mimicking similar culture conditions. The retroviral expression vector pMIEG was a modified pMIB (CMV-IRES-BFP) vector encoding Dyn1 cDNA with N-terminal HA-tag and C-terminal eGFP fusion tag. Point mutations to introduce S774/8A in Dyn1 cDNA were performed by site-directed mutagenesis. The lentiviral expression vector pLVX-puro (Clontech) encoded CLCa N-terminally tagged with mRuby 2 [53] or SNAP-tag [54] spaced with a 6 amino acid GGSGGS linker. The constructs were assembled from PCR fragments of mRuby2, SNAP, and CLCa (for primers, see S1 Table) in yeast as described below and subsequently cloned into pLVX-puro. Lentiviruses were generated in 293T packaging cells following standard transfection protocols [55] and were used for subsequent infections. They were prepared to transduce fluorescently tagged (mRuby2 or SNAP tag) CLCa fused to its N-terminus. Infected cells were selected using 10 μg/ml puromycin for 4 d, conditions under which uninfected cells perished. The cells were passaged for 2 w before imaging for CME analysis. Retroviruses were also generated in 293T cells and used to stably transduce eGFP tagged Dyn1WT, Dyn1S774/8A, and Dyn2WT proteins. Gene transduction was performed by exposing A549 or H1299 cells to retrovirus-containing cell culture supernatants through two rounds of viral transduction spread across 5 d. The recipient cells were further expanded to confluency in a 10-cm culture dish and FACS sorted for eGFP levels comparable to endogenous Dyn1-eGFP in A549 cells. Transfections for siRNA knockdown experiments were carried out using Lipofectamine 2000 or Lipofectamine RNAi-Max (Life Technologies), following manufacturer’s protocol. For siRNA mediated knockdown, approximately 2 × 105 cells (H1299) or 3 × 105 cells (A549) were plated in each well of a six-well plate. Twenty nmol siRNA was used per well, and two rounds of transfection across 5 days was sufficient to achieve over 90% knockdown. Perturbation of culture conditions by GSK3β inhibitor involved the addition of 10 μM CHIR99021 (Sigma) to prewarmed culture media and incubation of cells for 30 min before additional analysis. Growth factor stimulation was performed by adding 20 ng/ml EGF (Invitrogen) to prewarmed, serum-free culture media. Cells were analyzed after 10 min of incubation with EGF. Generation of genome-edited cell lines Genome-edited A549 and H1299 cells were generated by editing Dyn1 and Dyn2 to carry fusion tags. For the endogenous labeling of Dyn1 with fluorescent reporter proteins, we chose an approach based on site directed introduction of CRISPR/CAS9n-targeted DNA breaks and template assisted homology driven repair. eGFP fused to Dyn1 at its C-terminus was generated by CRISPR/Cas9n nickase strategy targeting the end of exon 21 of the DNM1 gene, inserting the last 19 amino acids of splice isoform “a,” a seven amino acid linker [32], monomeric eGFP with a stop codon, and the SV40 polyadenylation signal. In the donor plasmid, this inserted sequence was flanked by approximately 950 base pair homology arms for HDR. The +gRNA pair was designed using publicly available software (http://crispr.mit.edu/) and prepared as described [56] with oligos DNM1-Nuclease-A-f/ DNM1-Nuclease-A-r and DNM1-Nuclease-B-f/ DNM1-Nuclease-B-r, respectively (S1 Table). For assembly of the donor vector, the segments were amplified with oligonucleotides coding approximately 30 nucleotide overhangs. The bacterial artificial chromosome clone RP11-348G11 (BACPAC Resources Center, Children’s Hospital Oakland Research Institute, Oakland, California) covering the end of human DNM1 gene was used as template for the left and right homology arms (Fig 1A). The left and right homology arms were amplified using primer pairs DNM1-LH-f/DNM1-LH-r and DNM1-RH-f/DNM1-RH-r, respectively (see S1 Table). The 19 C-terminal amino acids of splice isoform “a” and the linker sequence DPPVATL [32] were covered with oligonucleotides DNM1-C-assembly-f and DNM1-C-assembly-r and amplified with short primers DNM1-Cterm-f and DNM1-Cterm-r. The sequence coding for monomeric eGFP and the SV40 polyadenylation signal were amplified from plasmid peGFP-N1 (Clontech), which carried the A206K mutation [57] with primers DNM1-eGFP-f/DNM1-eGFP-r and DNM1-pA-f/DNM1-pA-r, respectively. The first and last primers (DNM1-LH-f, DNM1-RH-r) also included overhangs for the E coli/yeast shuttle vector pRS424 [58]. Dyn2-mRuby genome-edited cells were generated using previously validated ZFNs [32,34]. For the DNM2-mRuby2 donor vector, the homology arms were amplified from the published [34] DNM2-eGFP construct (gift from D. Drubin, University of California, Berkeley) with primers DNM2-LH-f/DNM2-LH-r and DNM2-RH-f/DNM2-RH-r, respectively. The mRuby2 segment together with the linker sequence, DPPVATL [32], was amplified from pmRuby2-C1, a gift from Michael Lin (Addgene #40260) [53]. The PCR products were purified on 1% agarose gels and extracted using standard protocols before transformation into YPH500 yeast cells [59]. Yeast transformation, plasmid extraction, and plasmid validation were performed as described earlier [60]. The guide-RNA plasmids for DNM1-eGFP and donor vectors for both DNM1-eGFP and DNM2-mRuby2 are available from Addgene (IDs 107795, 107796, 107794, and 107793, respectively). For both the edits, the nCas9 nickase + gRNA pairs or the ZFN nuclease pairs were added at 1 μg DNA concentration, and 2 μg of the donor plasmid was added to this mixture in 150 ul OptiMEM (Life Technologies). This mixture was then added to 5.5 μl of Lipofectamine 2000 (Life Technologies) in 150 μl OptiMEM (Life Technologies), briefly vortexed, and incubated at room temperature for 15 min. The mixture was then added to cells plated 12 h earlier at 70% confluency (approximately 3 × 106 cells per well in six-well dish) with freshly replaced media. Transfect-containing media was replaced by prewarmed fresh media and the cells were allowed to grow for the next 48 h and then passaged for expansion in a 10-cm dish. The expanded cells were sorted as eGFP (or mRuby2) gene-edited single cells into 96 well plates 4 days after transfection using a FACSAria 2-SORP (BD Biosciences, San Jose, CA) instrument equipped with a 300-mW, 488-nm laser and a 100-μm nozzle. Clonal expansion ensued by incrementing the culture dish area and maintaining a minimum 50% cell confluency. Single clones were then assayed for edits by western blotting, and cells positive for genome edits were expanded. In order to generate double genome edited A549 cells, the A549 clone, 2C8, with homozygous Dyn2-mRuby2 knock-in was chosen and subsequently edited for Dyn1, and cell selection was performed as before using FACS preliminary screen followed by western blotting for validation. Dyn1 KO H1299 cells were generated as previously described [23] and the same strategy was employed to generate A549 Dyn1 KO cells. Briefly, cells plated in six-well plates were transfected with 1 μg each of single-guide RNAs (sgRNAs) and Cas9 nickase encoding plasmids and cotransfected with a 20th of peGFP plasmid. eGFP-positive cells were assumed to have harbored both the sgRNA guides and single-cell-sorted by FACS. In addition, Dyn2 KO cells were generated using a similar double nickase strategy with sgRNAs CGATCTGCGGCAGGTCCAGGTGG and CGCCGGCAAGAGCTCGGTGCTGG in the pX335 vector. Complete knockout of Dyn1 and 2 was validated by western blotting. Immunoprecipitation, pulldowns, and subcellular fractionation Glutathione S-transferase (GST)-SH3 pulldown. GST-Amph II SH3 pulldown involved lysing H1299 cells in lysis buffer (50 mM Tris, 150 mM NaCl, 1X Protease Inhibitor Cocktail [Roche]) containing 0.2% Triton X-100. Cells were dounced with 27.5-G syringe about 20 times or until most of the cells were ruptured to release intact nuclei. The post-nuclear fraction (PNF) was obtained by spinning the lysate at 10,000 xg at 4°C for 10 min and collecting the supernatant. About 3 mg of PNF in 1-ml volume was used for each pulldown. Addition of beads (approximately 20 μl) with bait protein (GST-Amph II SH3) in PNF followed by gentle rotation for 1 h allowed binding of target proteins. The bound fraction was washed twice with lysis buffer containing 0.2% Triton X-100 and the resulting beads were denatured using 2X Laemmli buffer (Bio-Rad) reduced with 5% β-mercaptethanol, boiled and run on SDS-PAGE gel of appropriate separation capacity (7.5% or 12%, based on the target protein size). The pulldowns were analyzed by western blotting. Subcellular fractionation. Confluent A549 cells in a 60-mm dish were detached with 1 ml of 10-mM EDTA at 37 °C for 10 min and washed with PBS by centrifugation and then resuspended in 0.5 ml buffer 2 (25 mM HEPES, 250 mM sucrose, 1 mM MgCl2, 2 mm EGTA, pH 7.4). The resuspended cells were lysed through 3 cycles of freeze-thaw (rapid freezing in liquid nitrogen and slow thawing in room temperature water). Cytoplasm and membrane portions were separated by 30 min ultracentrifugation at 110 kg in a Beckman Coulter rotor (TLA55). Pellets were resuspended with 0.5 ml buffer 2, and both supernatant and pellets were solubilized in 0.5% Triton X-100 for 10 min on ice and then precipitated with 10% TCA, followed by 2 rounds of 1 ml acetone wash. SDS-PAGE gel electrophoresis and western blot were applied as described above [61]. GFP-nAb immunoprecipitation. Confluent A549 cells in a 10-cm dish were detached with 2 ml of 10-mM EDTA at 37 °C for 10 min and washed with PBS by centrifugation and then resuspended and gently lysed for 15 min on ice with 2 ml buffer 3 (0.5% Triton X-100, 25 mM HEPES, 150 mM KCl, 1 mM MgCl2, 2 mM EGTA, 1X Protease Inhibitor Cocktail [Roche], 1X Phosphatase Inhibitor Cocktail, pH = 7.4). Lysates were centrifuged at 5000 xg, 4 °C for 5 min to remove nuclei, and protein concentration in the PNF was determined by Bradford assay. One-half mg of the PNF was added to 30 μl GFP-nAb agarose (Allele Biotech), rotated for 2 h at 4 °C, and then spun down at 2500 xg, 4 °C, 2 min. The agarose was washed twice (1 ml/each) with two different buffers to fulfill different experimental purposes: (1) to probe the dynamin interactors, the agarose was washed with buffer 3; (2) to probe dynamin self-assembly, salt concentration in buffer 3 was brought up to 300 mM to remove indirect dynamin–dynamin interactions. Ten percent cell lysate, which is used to determine immunoprecipitation efficiency, was precipitated with 10% TCA and washed twice with acetone from −20 °C freezer. The samples were subjected to SDS-PAGE and Western blotting for analysis. Glutathione S-transferase (GST)-SH3 pulldown. GST-Amph II SH3 pulldown involved lysing H1299 cells in lysis buffer (50 mM Tris, 150 mM NaCl, 1X Protease Inhibitor Cocktail [Roche]) containing 0.2% Triton X-100. Cells were dounced with 27.5-G syringe about 20 times or until most of the cells were ruptured to release intact nuclei. The post-nuclear fraction (PNF) was obtained by spinning the lysate at 10,000 xg at 4°C for 10 min and collecting the supernatant. About 3 mg of PNF in 1-ml volume was used for each pulldown. Addition of beads (approximately 20 μl) with bait protein (GST-Amph II SH3) in PNF followed by gentle rotation for 1 h allowed binding of target proteins. The bound fraction was washed twice with lysis buffer containing 0.2% Triton X-100 and the resulting beads were denatured using 2X Laemmli buffer (Bio-Rad) reduced with 5% β-mercaptethanol, boiled and run on SDS-PAGE gel of appropriate separation capacity (7.5% or 12%, based on the target protein size). The pulldowns were analyzed by western blotting. Subcellular fractionation. Confluent A549 cells in a 60-mm dish were detached with 1 ml of 10-mM EDTA at 37 °C for 10 min and washed with PBS by centrifugation and then resuspended in 0.5 ml buffer 2 (25 mM HEPES, 250 mM sucrose, 1 mM MgCl2, 2 mm EGTA, pH 7.4). The resuspended cells were lysed through 3 cycles of freeze-thaw (rapid freezing in liquid nitrogen and slow thawing in room temperature water). Cytoplasm and membrane portions were separated by 30 min ultracentrifugation at 110 kg in a Beckman Coulter rotor (TLA55). Pellets were resuspended with 0.5 ml buffer 2, and both supernatant and pellets were solubilized in 0.5% Triton X-100 for 10 min on ice and then precipitated with 10% TCA, followed by 2 rounds of 1 ml acetone wash. SDS-PAGE gel electrophoresis and western blot were applied as described above [61]. GFP-nAb immunoprecipitation. Confluent A549 cells in a 10-cm dish were detached with 2 ml of 10-mM EDTA at 37 °C for 10 min and washed with PBS by centrifugation and then resuspended and gently lysed for 15 min on ice with 2 ml buffer 3 (0.5% Triton X-100, 25 mM HEPES, 150 mM KCl, 1 mM MgCl2, 2 mM EGTA, 1X Protease Inhibitor Cocktail [Roche], 1X Phosphatase Inhibitor Cocktail, pH = 7.4). Lysates were centrifuged at 5000 xg, 4 °C for 5 min to remove nuclei, and protein concentration in the PNF was determined by Bradford assay. One-half mg of the PNF was added to 30 μl GFP-nAb agarose (Allele Biotech), rotated for 2 h at 4 °C, and then spun down at 2500 xg, 4 °C, 2 min. The agarose was washed twice (1 ml/each) with two different buffers to fulfill different experimental purposes: (1) to probe the dynamin interactors, the agarose was washed with buffer 3; (2) to probe dynamin self-assembly, salt concentration in buffer 3 was brought up to 300 mM to remove indirect dynamin–dynamin interactions. Ten percent cell lysate, which is used to determine immunoprecipitation efficiency, was precipitated with 10% TCA and washed twice with acetone from −20 °C freezer. The samples were subjected to SDS-PAGE and Western blotting for analysis. TIRFM Cells expressing appropriate fluorophores were cultured overnight on an acid-etched and gelatin-coated coverslip, placed in a well in a six-well plate. At the time of imaging, cells were checked for adherence and spreading. When imaging SNAP-tagged proteins, labeling was performed by incubating cells in 1 ml of fresh, prewarmed media containing 1 μl of predissolved SNAP-CELL 647-SiR dye (NEB). After 30 min incubation under standard incubator conditions, the media was aspirated, washed twice with sterile PBS, and reincubated in fresh culture media. The coverslips were mounted on glass slides with spacers and sealed with the same media. For experiments involving the addition of growth factor or inhibitor, cells were preincubated for the appropriate times and the coverslips were mounted as before with the treated media. The coverslips were then imaged using a 60x 1.49 NA Apo TIRF objective (Nikon) mounted on a Ti-Eclipse inverted microscope with Perfect Focus System (Nikon) equipped with an additional 1.8x tube lens, yielding at a total magnification of 108x. TIRF illumination was achieved using a Diskovery Platform (Andor Technology). During imaging, cells were maintained at 37°C in RPMI supplemented with 5% fetal calf serum. Time-lapse image sequences were acquired at a penetration depth of 80 nm and a frame rate of 2 Hz (three or two channels) or 1Hz (single channel) using a sCMOS camera with 6.5mm pixel size (pco.edge). Quantitative analysis of imaging The detection, tracking and analysis of all clathrin-labeled structures and thresholding to identify bona fide CCPs was done as previously described using the cmeAnalysis software package [22]. Briefly, diffraction-limited clathrin structures were detected using a Gaussian-based model method to approximate the point-spread function [22], and trajectories were determined from clathrin structure detections using the u-track software [37]. Subthreshold clathrin-labeled structures (sCLSs) were distinguished from bona fide CCPs, based on the quantitative and unbiased analysis of clathrin intensity progression in the early stages of structure formation [22,62]. Both sCLSs and CCPs represent nucleation events, but only bona fide CCPs represent structures that undergo stabilization, maturation, and, in some cases, scission to produce intracellular vesicles [22,62]). We report the rate of bona fide CCP formation, distribution of their lifetimes, and intensity cohorts, as described previously [22]. As these values will depend on day-to-day variations in the threshold, we image experimental and control conditions on the same day and apply the same threshold to both data sets to ensure that effects we detect are due to the specific experimental variable being assessed. We also report mean and maximum signal intensities in two or three channels for each individual CCP. These are average and maximum signal intensities for individual CCPs as they are extracted by the previously described analysis software [22]. The extraction of CCPs is achieved by a new function added to the cmeAnalysis software published in [22] that allows us to link the classification of events, CCPs, or sCLSs to more sophisticated analysis intensity time courses and lifetime. In this study, we focused merely on per-CCP mean and maximum intensity values, which were averaged per movie (1–5 cells) and finally presented as per-movie distributions covering 10–30 cells per experimental condition. Differences between conditions were assessed by comparison of the normal-distributed per-movie distributions using Student t test and a threshold of p < 0.01 to mark statistical significance. Statistical analysis Control and treatment datasets were statistically analyzed with two-tailed, unpaired Student t tests using Graphpad Prism 5.0 (Graphpad Software, La Jolla, CA), from which p values were derived (* p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001). Error bars representing standard error of the mean (SEM) for at least three independent experiments were calculated using Microsoft Excel. Receptor internalization (endocytosis) assay An in-cell ELISA approach was used to quantitate internalization of TfnR and EGFR, as previously described [23], using either anti-TfnR mAb (HTR-D65) [63] or biotinylated-EGF as ligands. Cells were grown overnight in 96-well plates at a density of 2 x 105 cells/well and incubated with 4 mg/ml of D65 or 20 ng/ml of biotinylated-EGF (Invitrogen) in assay buffer (PBS4+: PBS supplemented with 1 mM MgCl2, 1 mM CaCl2, 5 mM glucose, and 0.2% bovine serum albumin) at 37 °C for the indicated time points. Cells were then immediately cooled down (to 4 °C) to arrest internalization. The remaining surface-bound D65 or biotinylated-EGF was removed from the cells by an acid wash step (0.2 M acetic acid, 0.2 M NaCl, pH 2.5). Cells were then washed with cold PBS and then fixed in 4% paraformaldehyde (PFA) (Electron Microscopy Sciences) in PBS for 30 min and subsequently permeabilized with 0.1% Triton X-100/PBS for 10 min. Internalized D65 was assessed using a goat anti-mouse HRP-conjugated antibody (Life Technologies), and internalized biotinylated-EGF was assessed by streptavidin-POD (Roche). The reaction was developed by a colorimetric approach with OPD (Sigma-Aldrich), and color development was stopped by addition of 50 μl of 5M of H2SO4. The absorbance was read at 490 nm (Biotek Synergy H1 Hybrid Reader). Internalized ligand was expressed as the percentage of the total surface-bound ligand at 4 °C (i.e., without acid wash step), measured in parallel [23]. Well-to-well variability in cell number was accounted for by normalizing the reading at 490 nm with BCA readout at 560 nm. Supporting information S1 Fig. Design strategies for genome-edited H1299 and A549 cells. (A) Domain and genomic structure of DNM1 illustrating the C-terminal splice variants to illustrate splice variants A and B derived from exons 21 and 22 (filled red box) and 3′ UTR (open red box). To express the Dyn-1-EGFP fusion protein, the last 19 amino acids from splice variant A from exon 22 were introduced in frame in exon 21 (dark grey), followed by a 7-amino acid linker (blue), EGFP (green), and SV40 poly adenylation signal (orange). (B) Design of sgRNA guide A and B which targets the splice region in exon 21. The guide targeting sequences (underlined) and PAM sequences (red) are shown. For the donor vector, the DNA and amino acid sequences are shown for the junctions between exon 21, the inserted Dyn-1 C-term, the linker, EGFP and the poly adenylation signal. The color code is as in panel A. (C) Approach used for ZFN-mediated genome-editing of DNM2, as previously described [32,34] and the expected amino acid sequence for the Dyn2-mRuby2 fusion protein. Dyn1, dynamin-1; sgRNA, single-guide RNA; ZFN, Zinc Finger Nuclease. https://doi.org/10.1371/journal.pbio.2005377.s001 (TIF) S2 Fig. Isoform-specific differences in recruitment of Dyn1 and Dyn2 to CCPs. (A, C) Representative TIRF image and corresponding kymograph of dynamic SNAP(647)-CLCa-labeled CCPs and Dyn2-mRubyend (A) or Dyn1a-eGFPend (B) in genome-edited H1299 cells. (B, D) Corresponding quantification of the averaged intensities of CLCa and Dyn2-mRubyend (B) or Dyn1a-eGFPend (D) recruitment for the indicated lifetime cohorts. Data from 6,647 CCPs from 5 independent movies, containing a total of 15 cells (B) and data from 74,805 CCPs from 10 independent movies, containing a total of 29 cells (D). CCP, clathrin-coated pit; CLCa, clathrin light chain a; Dyn1, dynamin-1. https://doi.org/10.1371/journal.pbio.2005377.s002 (TIF) S3 Fig. CCP dynamics in genome-edited Dyn1a-eGFP H1299 cells. (A) CCP initiation rates, (B) CCP lifetimes, and (C) lifetime distributions of all CCPs in H1299 cells genome edited to express endogenously tagged Dyn1a-eGFP. Each point represents the value derived from a single movie, with 2–4 cells/movie. (** p ≤ 0.01, **** p ≤ 0.0001). The underlying data of panels A and B can be found in S1 Data. CCP, clathrin-coated pit; Dyn1, dynamin-1. https://doi.org/10.1371/journal.pbio.2005377.s003 (TIF) S4 Fig. Characterization of TfnR endocytosis and dynamin-isoform recruitment in A549 cells. (A) Differential expression of Dyn1 versus Dyn2 in H1299 versus A549 cells. In H1299 cells, Dyn2 is expressed at approximately 6-fold higher levels than Dyn1. In A549 cells, Dyn1 is expressed at approximately 5-fold higher levels than Dyn2 [39]. (B) TfnR endocytosis in parental A549 cells treated with the indicated siRNAs. (C) TfnR uptake at 10 min in parental, Dyn1KO, and Dyn2KO A549 cells. (D) Quantification of the average recruitment of Dyn1a-eGFP or Dyn2-eGFP to CCPs with lifetimes between 40 and 60 s (4,420 CCPs positive for Dyn1 and 3,961 CCPs positive for Dyn2 were identified and analyzed from 11 movies containing 2–4 cells per movie), as in Fig 5E; however, the Dyn1a-eGFP data is rescaled to illustrate that Dyn1, like Dyn2, peaks at late stages of CME in these cells. The underlying data of panels B and C can be found in S1 Data. CCP, clathrin-coated pit; CME, clathrin-mediated endocytosis; Dyn1, dynamin-1; siRNA, small interfering RNA; TfnR, transferrin receptor. https://doi.org/10.1371/journal.pbio.2005377.s004 (TIF) S5 Fig. Dynamin isoforms only weakly co-assemble. (A) Western blots and quantification (red) of bands showing extent of pulldown of Dyn1-eGFP or Dyn2-eGFP using anti-eGFP nAb-beads and coimmunoprecipitation of the other isoform. Data are representative of 3 independent experiments. (B) The inhibition of assembly stimulated GTPase activity of Dyn1 (blue) or Dyn2 (red) in the presence of increasing concentrations of GTPase-defective Dyn1S45N, which will inhibit assembly-stimulated GTPase activity by co-assembling with WT-dynamin on lipid nanotube templates. The underlying data of panel B can be found in S1 Data. Dyn1, dynamin-1; GTPase, Guanosine Triphosphate hydrolase; WT, wild-type. https://doi.org/10.1371/journal.pbio.2005377.s005 (TIF) S1 Movie. TIRFM movie of Dyn2-mRuby2end and SNAP647-CLCa in genome-edited H1299 cells. CLCa, clathrin light chain a; Dyn2, dynamin-2; TIRFM, total internal reflection fluorescence microscopy. https://doi.org/10.1371/journal.pbio.2005377.s006 (AVI) S2 Movie. TIRFM movie of Dyn1a-eGFPend and SNAP647-CLCa in genome-edited H1299 cells. CLCa, clathrin light chain a; Dyn1, dynamin-1; TIRFM, total internal reflection fluorescence microscopy. https://doi.org/10.1371/journal.pbio.2005377.s007 (AVI) S3 Movie. TIRFM movie of Dyn2-mRuby2end, Dyn1a-eGFPend, and SNAP647-CLCa in double genome-edited A549 cells. CLCa, clathrin light chain a; Dyn1, dynamin-1; TIRFM, total internal reflection fluorescence microscopy. https://doi.org/10.1371/journal.pbio.2005377.s008 (AVI) S1 Table. List of oligonucleotides used for genome editing, mutagenesis, and fusion constructs. https://doi.org/10.1371/journal.pbio.2005377.s009 (DOCX) S1 Data. Raw data for Fig 2 Panels C-L; Fig 3 Panels B-D, H; Fig 4 Panels A-C,F,G; Fig 5 Panels B.C.F; Fig 6 Panels B, D-F; Fig 8; Fig 9 Panels A, C-E; Fig 10 Panels B-E, G,H; S3 Panels A-C; S4 Panels A,C; S5 Panel B. https://doi.org/10.1371/journal.pbio.2005377.s010 (XLSX) Acknowledgments We thank members of the Schmid lab for helpful discussions and critically reading the manuscript, Phillipe Roudot for help with data analysis, Heather Grossman, Wesley Burford and Joseph Chi for technical assistance, and Marcel Mettlen for help with microscopy. FACS sorting was performed in The Moody Foundation Flow Cytometry Facility, Children’s Research Institute, UT Southwestern Medical Center. Carlos Reis generated the Dyn2KO A549 cells and David Drubin generously provided the ZFN constructs.

doi: 10.1371/journal.pbio.2003998pmid: 29649211

Introduction The centrosome is a major microtubule organising centre, with critical roles in cell migration, division, shape maintenance, and cilia function. Animal interphase cells are generally thought to have 1 centrosome, consisting of 2 mature microtubule-based structures, called centrioles. Centriole pairs are proteomically and functionally distinct [1] yet apparently remain physically associated, a phenomenon called centrosome cohesion [2–7]. Experimental changes to intercentriolar distance during interphase result in defects in cell migration, ciliary function, and mitosis [8–12], underscoring the functional importance of centrosome cohesion. How 2 centrioles coordinately assemble into a single centrosome yet maintain distinct functions is largely unexplored. Centrioles are not bounded by a lipid membrane but instead by 2 distinct structures, termed the pericentriolar material (PCM) and pericentriolar fibres [2,13]. Current models of PCM assembly emphasise high dynamics of constituent proteins, potentially as a liquid-like, toroidal structure [14,15]. In contrast, comparatively little is known about either the structure or assembly of pericentriolar fibres. Rootletin, or ciliary rootlet coiled-coil protein (gene symbol CROCC), localises to pericentriolar filaments, and rootletin knockout or knockdown results in both loss of filaments and centrosome cohesion [2,8,16,17]. CNAP1, a paralogous gene, interacts with rootletin, likely at centriole proximal ends [2,18]. One model posits that rootletin pericentriolar fibres directly connect centriole pairs to keep them spatially restricted [2,5,16,18]. Consistent with this proposal, rootletin is not found on mitotic centrosomes [5,18–20]. The kinetics of pericentriolar fibre dissolution, when they reform, and the principles governing their replication are poorly understood, however. To address these questions, this study uses fluorescence imaging, genome editing, and cell fusion to obtain unprecedented spatiotemporal information about the morphology, dynamics, and assembly properties of rootletin fibres, which are referred to as roots. Roots are bifurcating adhesive structures that are licensed to form on centrioles by polo-like kinase 1 (PLK1) enzymatic activity. Both mother and daughter centrioles form independent roots that do not remain connected in response to organelle movement in vivo. Thus, they adopt a structure and function that allow centriole pairs to independently position during interphase, providing new insight into centrosome self-organisation. Results Centrosomal roots are large bifurcating fibres licensed to form on procentrioles by PLK1 activity Pericentriolar filaments near centrosomes have been described for many decades [21], but their ubiquity in different cell types is unknown. The prevalence of rootletin fibres was systematically documented by immunofluorescent staining and enhanced confocal airyscan imaging in a range of human cell types, whether cancerous, immortalised, or primary. Thorough antibody validation, obtained by multiple independent lines of evidence, ensured specific recognition of rootletin (S1 Fig and summarised in Materials and methods). Endogenous rootletin almost ubiquitously formed bifurcating fibres at the centrosome, henceforth referred to as roots (Fig 1A). Costaining and segmentation of a range of markers of either centrioles or the PCM showed limited overlap with roots by either three-dimensional (3D) structured illumination microscopy (SIM) super-resolved imaging (Fig 1B) or confocal airyscan (S1E Fig), indicating that roots occupy a different locale, adjacent to the PCM and centrioles. Segmentation of both roots and centrioles, as marked by a stable green fluorescent protein–Centrin1 fusion (GFP-Centrin1), showed that roots are large relative to centrioles, at approximately 10-fold the size of a centriole on average in retinal pigment epithelium (RPE) cells (Fig 1C). Roots were much shorter than ciliary rootlets—the prominent rootletin fibres found in specialised ciliated cell types, including photoreceptor cells—however ([16,17]; S1F and S1G Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Centrosomal roots are large bifurcating fibres licensed to form on procentrioles by PLK1 activity. (A) Systematic immunofluorescent airyscan imaging of rootletin (green) and the PCM marker NEDD1 (red), using the same conditions throughout all cell types. Confocal slices are shown. Scale bar 1 μm. (B) 3D SIM imaging of rootletin (green) and various centrosomal components (red). Rootletin is stained either by anti-rootletin antibody or by anti-GFP nanobody. Z-projections and single z-slices with segmentation are shown on the top and bottom rows, respectively. Scale bar 1 μm. (C) Quantification of the ratio of rootletin immunostaining area relative to GFP-Centrin 1 area from maximum-intensity projected airyscan images. (D) Rootletin immunofluorescent staining is equal in unreplicated centrosomes and diplosomes. Centrosomes were classified based on GFP-Centrin1 foci number, and anti-rootletin staining was segmented. Scale bar 1 μm. The mean is shown as + and the median as a horizontal bar. n.s., t test. N = 21 cells. Note that rootletin is shown in red in this panel. (E) Cells were arrested in prometaphase with either STLC (Eg5 inhibition) or BI2536 (PLK1 kinase inhibition) before being forced into interphase with RO-3306 (CDK1 inhibition). (F) Cells expressing GFP-Centrin1 (green) were treated as depicted in panel E before staining with anti-rootletin antibody (red). Maximum-intensity projections are shown. Scale bar 1 μm. (G) Root area per cell was quantified by direct segmentation of rootletin staining from images obtained as described in panel F. The horizontal bar shows the median. *P = 0.0006, t test. See S1 Data for source data for the charts. 3D, three-dimensional; CDK1, cyclin-dependent kinase 1; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; PCM, pericentriolar material; PLK1, polo-like kinase 1; SIM, structured illumination microscopy; STLC, S-trityl-L-cysteine. https://doi.org/10.1371/journal.pbio.2003998.g001 Centriole replication normally proceeds through the appearance of a nascent procentriole from the base of an existing centriole during S and G2 phase [22,23]. To examine whether procentriole formation influences root structure, centrosomes containing either 2 or 4 GFP-Centrin1 marked centrioles were classified, corresponding to either unreplicated centrioles or diplosomes, respectively. No difference in rootletin intensity or size was detected (Fig 1D), suggesting that procentriole growth does not influence root structure. Procentrioles mature into centrioles during mitosis, dependent on PLK1 activity, becoming replication competent after physically moving away from a centriole (a process termed disengagement) [24]. Therefore, the effects of PLK1 kinase inhibition on root formation were investigated (Fig 1E). Cells arrested in mitosis through PLK1 blockade contained monopolar spindles [25], which were devoid of roots (Fig 1F), consistent with previous work [5,16,18–20]. Because the inhibition of PLK1 results in cell cycle arrest, mitotic exit was forced into an ensuing interphase without cell division—by addition of the cyclin-dependent kinase 1 (CDK1) inhibitor RO-3306 [26]—to understand subsequent effects on root structure in interphase. Control cells were also arrested in mitosis, but instead using the Eg5 kinesin motor inhibitor S-trityl-L-cysteine (STLC) followed by RO-3306. Cells forced into interphase in this manner reformed roots despite unsuccessful mitotic genome segregation (Fig 1F; right-hand panel). However, forced mitotic exit after PLK1 blockade resulted in partial root reformation relative to STLC control (Fig 1G). These results suggest that centrioles are capable of root reformation in G1 regardless of PLK1 activity in the previous mitosis. In contrast, procentrioles must be modified by PLK1-dependent processes before they are competent to form roots in the next cell cycle. Furthermore, because PLK1 promotes centrosomal PCM expansion during mitosis [14], mitotic centrosomes disassemble roots even in the absence of centrosome maturation. Taken together, roots are large bifurcating fibres, found commonly in a range of cell types on mature PLK1-modified centrioles during the interphase. Diffusionally stable roots are progressively formed from anaphase The dynamics and biophysical properties of enhanced GFP (eGFP)-tagged rootletin were examined in living cells by utilising both cDNA transgene overexpression and tagging of endogenous alleles. Consistent with previous work [2,16], overexpression of eGFP-rootletin resulted in fibres and bifurcating fork structures that were longer than endogenous rootletin (e.g., compare S2A Fig with Fig 1). Time-lapse imaging of eGFP-rootletin fibre formation following transfection showed that eGFP-rootletin first appeared focally in a single location, prior to the emergence of a larger network over many hours, eventually filling the cytoplasm (Fig 2A and S2A Fig). Fibres increased in size not only by extension in length outwards from a single point but additionally by the coalescence of multiple fibres to form larger aggregates, frequently through end-on fusions (compare arrows in Fig 2A; S1 Video). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Diffusionally stable roots are progressively formed from anaphase. (A) eGFP-rootletin fibres progressively assemble following transfection. The images are timepoints from a single cell, taken by live-cell 3D confocal time-lapse imaging. The arrows point to a fusion event of 2 preexisting fibres. Scale bar 3 μm. See also S1 Video for the full time course. (B) Representative images from single-cell 3-colour 3D confocal time-lapse imaging of rootletin-meGFP (green), NEDD1-mRuby3 (red; marking the PCM), and DNA (blue; marked by SiR-hoechst), showing root disassembly during mitosis. Images were smoothed for display purposes here using a 2-pixel median filter, but not for analysis. Scale bar 1 μm. See also S2 Video. (C) Cell cycle–dependent changes in rootletin-meGFP centrosomal fluorescence intensity. Centrosomes were automatically tracked as described in Materials and methods. Individual cell traces were manually aligned relative to anaphase onset based on SiR-hoechst staining of DNA (time 0). Mean +/- SD; N = 17 cells. (D) Root splitting during centrosome separation in early mitosis, showing rootletin-meGFP (green) and NEDD1-mRuby3 (red). Scale bar 2 μm. (E) Cell cycle–dependent changes in rootletin levels by western blot. Cells were synchronised in nocodazole, released, and harvested at different cell cycle stages, and western blotted with anti-rootletin antibody. (F) FRAP recovery curve over 15 hours, plotting the mean ± SD centrosomal intensity of rootletin-meGFP from 3D confocal imaging after bleaching the fluorescence of the whole centrosome, in thymidine-arrested cells. Centrosome position was tracked independently of rootletin-meGFP fluorescence through simultaneous NEDD1-mRuby3 imaging in a spectrally distinct channel. N = 11 cells. FRAP, fluorescence recovery after photobleaching; meGFP, monomeric enhanced green fluorescent protein; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; PCM, pericentriolar material. https://doi.org/10.1371/journal.pbio.2003998.g002 Cell cycle–dependent changes in the centrosomal intensity of meGFP-tagged rootletin were followed by 3D confocal time-lapse imaging. Because overexpressed rootletin fibres were larger than endogenous antibody stained roots, and overexpression can influence quantitative measures of protein function in vivo [27], CRISPR Cas9 was used to insert an in-frame fusion of meGFP into the endogenous rootletin (CROCC) locus and therefore study rootletin behaviour with live-cell microscopy at endogenous levels for the first time (S3 Fig). Homozygous tagging in the diploid breast cancer cell line Cal51 resulted in fluorescent signal closely resembling antibody staining (S3E Fig). Rootletin-meGFP was barely detectable at the centrosome during mitosis, consistent with immunofluorescent staining (Fig 1), and consequently, a stably coexpressed NEDD1-mRuby3 fluorescent fusion was used to mark the PCM and allow tracking of centrosomes throughout the cell cycle and independently of rootletin levels, in a spectrally distinct fluorescent channel (Fig 2B; S2 Video). Additionally, fluorescently labelled chromatin was monitored to visualise mitotic substages. Rootletin began to be released from the centrosome >2 hours prior to anaphase (Fig 2C). By anaphase, centrosomal rootletin could not be detected above cytoplasmic levels, suggesting disassembly of all centrosomal roots. Rootletin centrosomal levels increased from anaphase but, unexpectedly, continued to increase at a slow rate for approximately 9 hours and thus significantly into G1 phase. Staging of rootletin intensity relative to centrosome separation revealed that its release from the centrosome began prior to centrosome separation and continued after it, with low levels of rootletin still present during centrosome separation, which could be ripped apart during poleward centrosome migration (Fig 2D). Because roots were disassembled in mitosis, it was investigated whether cytoplasmic mitotic rootletin levels were decreased, by western blotting of synchronised cells (Fig 2E). Rootletin cytoplasmic levels were high in mitosis by western blot despite the absence of roots, indicating that cytoplasmic and centrosomal rootletin levels do not always correlate. Together, these results suggest that the removal of rootletin from centrosomes begins early in mitosis or in late G2 phase of the cell cycle, prior to both chromatin condensation and centrosome separation, and then continues during these processes. Rootletin assembly at the centrosome begins from anaphase and occurs slowly for approximately 9 hours into G1 phase. Previous work has implicated the centriole proximal factor CNAP1 in rootletin centrosomal localisation [2,12,18], and in agreement, small interfering RNA (siRNA)-mediated knockdown of CNAP1 resulted in reduced centrosomal rootletin localisation (S4A Fig). It was investigated whether ectopic CNAP1 plasma membrane localisation via a C-terminal CAAX domain fusion [28] was sufficient to induce root polymerisation outside of the centrosome (S4 Fig). However, neither plasma membrane–localised mScarlet-CNAP1-CAAX, nor the CNAP1-binding partner CEP135 (membrane localised as CEP135-mScarlet-CAAX [29]) induced the formation of ectopic roots. Some PCM components show dynamic exchange of subunits on the seconds timescale—a property that is thought to be important for centrosome assembly [13]. Fluorescence recovery after photobleaching (FRAP) was therefore used to ask whether rootletin forms steady state polymers. FRAP of extended eGFP-rootletin fibres showed almost no movement of eGFP-rootletin over a time period of 10 minutes, however—even after a relatively rapid bleach (S2B Fig; approximately 1 second). Lack of recovery was not due to image bleaching or fibre movement out of the field of view, because adjacent unbleached ends of the fibre remained unchanged. To investigate very slow dynamic exchange of endogenous centrosomal rootletin-meGFP, on the hours timescale, cells were arrested at the G1/S phase boundary of the cell cycle using thymidine, to circumvent the effects of cell cycle progression on root morphology (Fig 2C). Total centrosomal rootletin-meGFP fluorescent signal was then bleached, and recovery followed by tracking of NEDD1-mRuby3 marked centrosomes during time-lapse imaging (Fig 2F). Recovery of rootletin-meGFP fluorescence on this long timescale was limited to approximately 30%. Together, it can be surmised that eGFP-rootletin fibres are predominantly diffusionally stable structures that are progressively assembled slowly over hours following anaphase. Roots disentangle during transient centriole splitting in interphase How a single interphase cell coordinately organises 2 disengaged centrioles is unclear. The prevalence of centrosomal cohesion was systematically documented in a range of human tissue culture cell types by automated fluorescence imaging and analysis of centrosome position (Fig 3A). Quantification of the percentage of cells with split centrosomes—defined as 2 PCM foci >1.5 μm apart—showed that it was low at approximately 10%, dependent on cell type (see Materials and methods for further discussion of the definition of split centrosome). Thus, in most cell types, centrioles remain cohered in close proximity during interphase, consistent with previous work [6,30–34]. It was investigated whether the minority of split centrioles remain stably separated over time, perhaps due to a permanent failure of centrosome cohesion. However, single-cell 3D confocal live imaging of centriole pairs marked by GFP-Centrin1 showed transient splitting. Therefore, a single mother–daughter centriole pair would split into 2 and then rejoin, often repeatedly (Fig 3B; S3 Video). Transient centriole splitting was manifest in live-cell imaging of several different cell types, including Cal51, HeLa, RPE, and U2OS cells (Fig 3B–3E; S4 Video and S5 Video). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Roots disentangle during transient centriole splitting in interphase. (A) Quantification of centrosome cohesion in the interphase of various cell types through systematic immunofluorescent staining and analysis. The images show representative staining of PCNT (red; marking centrosomal PCM) and DNA (blue; hoechst 44432). The right panel shows representative segmentation of centrosomes (red), nuclei (blue), and cytoplasm (white) in Cal51 cells. The yellow asterisk denotes a cell containing 2 centrosome foci, separated by >1.5 μm. Scale bars 20 μm and 5 μm. The bar graph shows the mean percentage of cells with PCNT centroids separated by >1.5 μm, from a minimum of 500 cells. Error bars show SEM from 2 experiments. (B–E) Selected frames showing centriole splitting in live 3D confocal time-lapse imaging. Centrosomes are marked by either GFP-Centrin1 or NEDD1-mRuby3. Arrows denote centriole splitting events. The time intervals between frames are 12 minutes (panel B and C), 24 minutes (panel D), or 8 minutes (panel E). Scale bar 5 μm. See also S3–S5 Videos. (F) Centrosome cohesion in HeLa cells ± overexpression of eGFP-rootletin, measured by automated imaging and analysis. Horizontal bars show the mean of 2 experiments ± SD. *P < 0.001 by Fischer’s exact test. (G) Opposing models of root behaviour during centriole splitting, termed ‘Stable contact’ or ‘Disentangle’. (H) Representative 3D SIM images of roots (green) after centriole splitting, with the indicated costaining marking either the PCM or centrioles (red). Scale bar 1 μm. (I) Representative airyscan image of roots after centriole splitting. Scale bar 1μm. (J) Root linkage plotted as a function of centriole spacing distance. (K, L) Live-cell airyscan time-lapse imaging of endogenous rootletin-meGFP and NEDD1-mRuby3 during a centriole split (panel K) and when remaining stably cohered (panel L) in Cal51 cells. Scale bar 2 μm. See also S6 Video and S7 Video. See S1 Data for source data for the charts.; meGFP, monomeric enhanced green fluorescent protein; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; PCM, pericentriolar material; PCNT, Pericentrin; SIM, structured illumination microscopy. https://doi.org/10.1371/journal.pbio.2003998.g003 In agreement with a published report [31], HeLa Kyoto cells had high levels of centrosome separation, with approximately 50% of cells showing split centrioles in a fixed asynchronous population (Fig 3A). Because low levels of rootletin expression accompanied short roots in HeLa (Fig 1A) and previous work has shown that rootletin knockdown results in the loss of centrosome cohesion [2,35], the effect of increasing root length by rootletin overexpression on centrosome position was investigated in HeLa cells. This increase in fibre length significantly increased centrosome cohesion in interphase HeLa cells, as measured by automated imaging and analysis of immunofluorescently stained samples (Fig 3F, P < 0.001, Fischer’s exact test). Together, these results show that, although mother and daughter centrioles generally remain cohered into a single focal location, they are able to transiently split apart in interphase in a manner that is antagonised by eGFP-rootletin overexpression. How might rootletin fibres respond to transient centriole splitting? Two opposing models for root behaviour after centriole splitting were postulated (Fig 3G). The first was maintenance of a stable root contact between centrioles as they move apart, for example, due to stretching. The second was loss of physical connection and disentanglement (‘Stable contact’ versus ‘Disentangle’, respectively). Surprisingly and in contrast to cohered centrosomes, rootletin fibres were not detected between split centrioles (Fig 3H, Fig 3I and S5 Fig). Instead, roots from each centriole were generally only detected as linked together at a distance of less than approximately 1.5 μm (Fig 3J), thus supporting the disentanglement model. Simultaneous 2-colour airyscan microscopy of root disentanglement in living cells revealed that roots occupy markedly heterogenous orientations that change in response to in vivo centriole movement (Fig 3K; S6 Video). The centrosome distal ends of roots have the capacity to pivot relative to centrosome proximal ends, suggesting a common more stable attachment point at the proximal end. Pivoting of centrosome distal tips was not just observed in centrosomes with split centrioles but also in cohered centrosomes, with roots maintained stably at the centriole–centriole interface (Fig 3L; S7 Video). As centrosomes remerged after a split, roots did not necessarily join but could alternatively contact the PCM of the opposing centriole. Together, these observations indicate that although roots can be maintained stably at the interface between mother–daughter centrioles, their orientation is heterogeneous, and notably, in response to centriole movement, a continuous direct rootletin linkage is not detected. Independence of mother and daughter centrioles during interphase Because disengaged centrioles can transiently split (Fig 3), the comparative structure of roots and PCM on mother and daughter centrioles during splitting was investigated further. Root area was approximately halved in split versus cohered centrioles (Fig 4A), suggestive of equal partitioning of 2 independent roots. Indeed, discrimination of the mother and daughter using CEP164 immunostaining showed that roots are nucleated symmetrically on both mother and daughter (Fig 4B). A similar comparison of PCM structure with the PCM resident Pericentrin (PCNT) showed similarly that both mother and daughter centrioles individually nucleate PCM when split (Fig 4C), something also evident in the live-cell imaging of NEDD1-mRuby3 (Fig 3H–3K) and previous work [24,32,34]. These observations imply that both mature centrioles independently maintain roots and PCM during centrosome splitting in interphase. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Independence of mother and daughter centrioles during interphase. (A) Root fibre area is significantly lower (P < 0.0001, t test) in split versus cohered centrioles. Anti-rootletin immunofluorescent staining was imaged and segmented, N = 36 cells from 2 experiments. (B) Rootletin immunofluorescent staining (green) is the same at both the mother or daughter centriole (n.s., t test). “M” and “D” denote mother and daughter, respectively, on the basis of CEP164 positivity. N = 21 cells per sample. Scale bar 1 μm. (C) PCNT immunofluorescent staining (of the PCM) is the same (n.s., t test) on either mother or daughter centrioles. Cells were imaged and analysed as described in panel B, except segmenting PCNT. N = 21 cells. See S1 Data for source data for the charts. (D) Cells with 4 centrioles might either maintain them as separate pairs or cohere them together (“Grouped”). (E) The pie chart shows the proportion of each GFP-Centrin1 centriole configuration in cells with 4 centrioles, produced as depicted in Fig 1E. The images are representative of each configuration. N = 196 cells. (F) Cells expressing endogenously tagged rootletin-meGFP were fused with cells expressing endogenously tagged rootletin-mScarlet. (G) Representative SIM images of single-colour cells and a fused cell, created as depicted in Fig 4F and described in Materials and methods. Scale bars 1 μm throughout. (H) Interphase centriole pairs contain large bifurcating fibres that disentangle when centrioles move apart >1.5 μm relative to each other. Root dissolution begins prior to mitotic centrosome separation and chromosome condensation. At the time of centrosome separation, roots are diminished in quantity and ripped apart during poleward movement of centrosomes. Roots form slowly over many hours from anaphase, as diffusionally stable fibres. PLK1-dependent modification of procentrioles allows root formation in the ensuing cell cycle. meGFP, monomeric enhanced green fluorescent protein; PCNT, Pericentrin; PLK1, polo-like kinase 1; SIM, structured illumination microscopy. https://doi.org/10.1371/journal.pbio.2003998.g004 Given the dynamic nature of centrosome cohesion (Fig 3) and root disentanglement, it was of interest to investigate whether mother–daughter centriole pairs would be maintained in cells with 4 centrioles (Fig 4D). Centriole position in cells forced into interphase after a failed mitosis by STLC treatment (Fig 1E) showed all possible centrosome cohesion configurations. Most commonly, all 4 centrioles grouped as 1 (Fig 4E), but notably, other spatial arrangements were equally as likely as 2 pairs. Thus, 2 mother–daughter centriole pairs are not maintained separately but will cohere together, even in a grouping such as a single centriole and 3 cohered. Overexpression of eGFP-rootletin promoted centrosome cohesion in interphase cells with 4 mature centrioles created by sequential STLC/RO-3306 treatment (S6A–S6C Fig), a similar effect to that seen in cells with normal centrosome numbers (Fig 3F). Some cells with supernumerary centrosomes are able to cluster them to form a bipolar spindle during mitosis [36]. It was therefore investigated whether, in contrast to cells with 2 centrosomes (Fig 2), cells with supernumerary centrosomes retained centrosomal rootletin during mitosis. Cells clustering supernumerary centrosomes at spindle poles during mitosis did not contain roots, however (S7 Fig), consistent with a model wherein rootletin does not promote mitotic centriole clustering. Centrosome cohesion was further examined using polyethylene glycol–mediated cell fusion of 2 different cell lines, 1 expressing endogenously tagged rootletin-meGFP and the other expressing endogenously tagged rootletin-mScarlet (Fig 4F, S3F Fig and see Materials and methods for details of fusion). Fused cells contained centrosomes of 2 fluorescent colours, 1 from each different cell line, as well as 2 nuclei. Because rootletin shows very slow diffusional exchange (Fig 2), this allowed the origin of centriole pairs in fused cells to be distinguished based on the emitted fluorescence. As per after mitotic failure (Fig 4E), mother–daughter centriole pairs were not exclusively maintained after cell fusion, but instead, fluorescent roots of different colours engaged each other (Fig 4G and S6D Fig). Fusion of cell lines expressing rootletin-meGFP or NEDD1-mRuby3 similarly showed that fluorescent roots from 1 cell could embrace all 4 centrioles once fused (S6E Fig). Therefore, by 2 independent methods, mother–daughter centriole pairs are not stably maintained in cells with 4 centrioles. Centrosomal roots are large bifurcating fibres licensed to form on procentrioles by PLK1 activity Pericentriolar filaments near centrosomes have been described for many decades [21], but their ubiquity in different cell types is unknown. The prevalence of rootletin fibres was systematically documented by immunofluorescent staining and enhanced confocal airyscan imaging in a range of human cell types, whether cancerous, immortalised, or primary. Thorough antibody validation, obtained by multiple independent lines of evidence, ensured specific recognition of rootletin (S1 Fig and summarised in Materials and methods). Endogenous rootletin almost ubiquitously formed bifurcating fibres at the centrosome, henceforth referred to as roots (Fig 1A). Costaining and segmentation of a range of markers of either centrioles or the PCM showed limited overlap with roots by either three-dimensional (3D) structured illumination microscopy (SIM) super-resolved imaging (Fig 1B) or confocal airyscan (S1E Fig), indicating that roots occupy a different locale, adjacent to the PCM and centrioles. Segmentation of both roots and centrioles, as marked by a stable green fluorescent protein–Centrin1 fusion (GFP-Centrin1), showed that roots are large relative to centrioles, at approximately 10-fold the size of a centriole on average in retinal pigment epithelium (RPE) cells (Fig 1C). Roots were much shorter than ciliary rootlets—the prominent rootletin fibres found in specialised ciliated cell types, including photoreceptor cells—however ([16,17]; S1F and S1G Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Centrosomal roots are large bifurcating fibres licensed to form on procentrioles by PLK1 activity. (A) Systematic immunofluorescent airyscan imaging of rootletin (green) and the PCM marker NEDD1 (red), using the same conditions throughout all cell types. Confocal slices are shown. Scale bar 1 μm. (B) 3D SIM imaging of rootletin (green) and various centrosomal components (red). Rootletin is stained either by anti-rootletin antibody or by anti-GFP nanobody. Z-projections and single z-slices with segmentation are shown on the top and bottom rows, respectively. Scale bar 1 μm. (C) Quantification of the ratio of rootletin immunostaining area relative to GFP-Centrin 1 area from maximum-intensity projected airyscan images. (D) Rootletin immunofluorescent staining is equal in unreplicated centrosomes and diplosomes. Centrosomes were classified based on GFP-Centrin1 foci number, and anti-rootletin staining was segmented. Scale bar 1 μm. The mean is shown as + and the median as a horizontal bar. n.s., t test. N = 21 cells. Note that rootletin is shown in red in this panel. (E) Cells were arrested in prometaphase with either STLC (Eg5 inhibition) or BI2536 (PLK1 kinase inhibition) before being forced into interphase with RO-3306 (CDK1 inhibition). (F) Cells expressing GFP-Centrin1 (green) were treated as depicted in panel E before staining with anti-rootletin antibody (red). Maximum-intensity projections are shown. Scale bar 1 μm. (G) Root area per cell was quantified by direct segmentation of rootletin staining from images obtained as described in panel F. The horizontal bar shows the median. *P = 0.0006, t test. See S1 Data for source data for the charts. 3D, three-dimensional; CDK1, cyclin-dependent kinase 1; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; PCM, pericentriolar material; PLK1, polo-like kinase 1; SIM, structured illumination microscopy; STLC, S-trityl-L-cysteine. https://doi.org/10.1371/journal.pbio.2003998.g001 Centriole replication normally proceeds through the appearance of a nascent procentriole from the base of an existing centriole during S and G2 phase [22,23]. To examine whether procentriole formation influences root structure, centrosomes containing either 2 or 4 GFP-Centrin1 marked centrioles were classified, corresponding to either unreplicated centrioles or diplosomes, respectively. No difference in rootletin intensity or size was detected (Fig 1D), suggesting that procentriole growth does not influence root structure. Procentrioles mature into centrioles during mitosis, dependent on PLK1 activity, becoming replication competent after physically moving away from a centriole (a process termed disengagement) [24]. Therefore, the effects of PLK1 kinase inhibition on root formation were investigated (Fig 1E). Cells arrested in mitosis through PLK1 blockade contained monopolar spindles [25], which were devoid of roots (Fig 1F), consistent with previous work [5,16,18–20]. Because the inhibition of PLK1 results in cell cycle arrest, mitotic exit was forced into an ensuing interphase without cell division—by addition of the cyclin-dependent kinase 1 (CDK1) inhibitor RO-3306 [26]—to understand subsequent effects on root structure in interphase. Control cells were also arrested in mitosis, but instead using the Eg5 kinesin motor inhibitor S-trityl-L-cysteine (STLC) followed by RO-3306. Cells forced into interphase in this manner reformed roots despite unsuccessful mitotic genome segregation (Fig 1F; right-hand panel). However, forced mitotic exit after PLK1 blockade resulted in partial root reformation relative to STLC control (Fig 1G). These results suggest that centrioles are capable of root reformation in G1 regardless of PLK1 activity in the previous mitosis. In contrast, procentrioles must be modified by PLK1-dependent processes before they are competent to form roots in the next cell cycle. Furthermore, because PLK1 promotes centrosomal PCM expansion during mitosis [14], mitotic centrosomes disassemble roots even in the absence of centrosome maturation. Taken together, roots are large bifurcating fibres, found commonly in a range of cell types on mature PLK1-modified centrioles during the interphase. Diffusionally stable roots are progressively formed from anaphase The dynamics and biophysical properties of enhanced GFP (eGFP)-tagged rootletin were examined in living cells by utilising both cDNA transgene overexpression and tagging of endogenous alleles. Consistent with previous work [2,16], overexpression of eGFP-rootletin resulted in fibres and bifurcating fork structures that were longer than endogenous rootletin (e.g., compare S2A Fig with Fig 1). Time-lapse imaging of eGFP-rootletin fibre formation following transfection showed that eGFP-rootletin first appeared focally in a single location, prior to the emergence of a larger network over many hours, eventually filling the cytoplasm (Fig 2A and S2A Fig). Fibres increased in size not only by extension in length outwards from a single point but additionally by the coalescence of multiple fibres to form larger aggregates, frequently through end-on fusions (compare arrows in Fig 2A; S1 Video). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Diffusionally stable roots are progressively formed from anaphase. (A) eGFP-rootletin fibres progressively assemble following transfection. The images are timepoints from a single cell, taken by live-cell 3D confocal time-lapse imaging. The arrows point to a fusion event of 2 preexisting fibres. Scale bar 3 μm. See also S1 Video for the full time course. (B) Representative images from single-cell 3-colour 3D confocal time-lapse imaging of rootletin-meGFP (green), NEDD1-mRuby3 (red; marking the PCM), and DNA (blue; marked by SiR-hoechst), showing root disassembly during mitosis. Images were smoothed for display purposes here using a 2-pixel median filter, but not for analysis. Scale bar 1 μm. See also S2 Video. (C) Cell cycle–dependent changes in rootletin-meGFP centrosomal fluorescence intensity. Centrosomes were automatically tracked as described in Materials and methods. Individual cell traces were manually aligned relative to anaphase onset based on SiR-hoechst staining of DNA (time 0). Mean +/- SD; N = 17 cells. (D) Root splitting during centrosome separation in early mitosis, showing rootletin-meGFP (green) and NEDD1-mRuby3 (red). Scale bar 2 μm. (E) Cell cycle–dependent changes in rootletin levels by western blot. Cells were synchronised in nocodazole, released, and harvested at different cell cycle stages, and western blotted with anti-rootletin antibody. (F) FRAP recovery curve over 15 hours, plotting the mean ± SD centrosomal intensity of rootletin-meGFP from 3D confocal imaging after bleaching the fluorescence of the whole centrosome, in thymidine-arrested cells. Centrosome position was tracked independently of rootletin-meGFP fluorescence through simultaneous NEDD1-mRuby3 imaging in a spectrally distinct channel. N = 11 cells. FRAP, fluorescence recovery after photobleaching; meGFP, monomeric enhanced green fluorescent protein; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; PCM, pericentriolar material. https://doi.org/10.1371/journal.pbio.2003998.g002 Cell cycle–dependent changes in the centrosomal intensity of meGFP-tagged rootletin were followed by 3D confocal time-lapse imaging. Because overexpressed rootletin fibres were larger than endogenous antibody stained roots, and overexpression can influence quantitative measures of protein function in vivo [27], CRISPR Cas9 was used to insert an in-frame fusion of meGFP into the endogenous rootletin (CROCC) locus and therefore study rootletin behaviour with live-cell microscopy at endogenous levels for the first time (S3 Fig). Homozygous tagging in the diploid breast cancer cell line Cal51 resulted in fluorescent signal closely resembling antibody staining (S3E Fig). Rootletin-meGFP was barely detectable at the centrosome during mitosis, consistent with immunofluorescent staining (Fig 1), and consequently, a stably coexpressed NEDD1-mRuby3 fluorescent fusion was used to mark the PCM and allow tracking of centrosomes throughout the cell cycle and independently of rootletin levels, in a spectrally distinct fluorescent channel (Fig 2B; S2 Video). Additionally, fluorescently labelled chromatin was monitored to visualise mitotic substages. Rootletin began to be released from the centrosome >2 hours prior to anaphase (Fig 2C). By anaphase, centrosomal rootletin could not be detected above cytoplasmic levels, suggesting disassembly of all centrosomal roots. Rootletin centrosomal levels increased from anaphase but, unexpectedly, continued to increase at a slow rate for approximately 9 hours and thus significantly into G1 phase. Staging of rootletin intensity relative to centrosome separation revealed that its release from the centrosome began prior to centrosome separation and continued after it, with low levels of rootletin still present during centrosome separation, which could be ripped apart during poleward centrosome migration (Fig 2D). Because roots were disassembled in mitosis, it was investigated whether cytoplasmic mitotic rootletin levels were decreased, by western blotting of synchronised cells (Fig 2E). Rootletin cytoplasmic levels were high in mitosis by western blot despite the absence of roots, indicating that cytoplasmic and centrosomal rootletin levels do not always correlate. Together, these results suggest that the removal of rootletin from centrosomes begins early in mitosis or in late G2 phase of the cell cycle, prior to both chromatin condensation and centrosome separation, and then continues during these processes. Rootletin assembly at the centrosome begins from anaphase and occurs slowly for approximately 9 hours into G1 phase. Previous work has implicated the centriole proximal factor CNAP1 in rootletin centrosomal localisation [2,12,18], and in agreement, small interfering RNA (siRNA)-mediated knockdown of CNAP1 resulted in reduced centrosomal rootletin localisation (S4A Fig). It was investigated whether ectopic CNAP1 plasma membrane localisation via a C-terminal CAAX domain fusion [28] was sufficient to induce root polymerisation outside of the centrosome (S4 Fig). However, neither plasma membrane–localised mScarlet-CNAP1-CAAX, nor the CNAP1-binding partner CEP135 (membrane localised as CEP135-mScarlet-CAAX [29]) induced the formation of ectopic roots. Some PCM components show dynamic exchange of subunits on the seconds timescale—a property that is thought to be important for centrosome assembly [13]. Fluorescence recovery after photobleaching (FRAP) was therefore used to ask whether rootletin forms steady state polymers. FRAP of extended eGFP-rootletin fibres showed almost no movement of eGFP-rootletin over a time period of 10 minutes, however—even after a relatively rapid bleach (S2B Fig; approximately 1 second). Lack of recovery was not due to image bleaching or fibre movement out of the field of view, because adjacent unbleached ends of the fibre remained unchanged. To investigate very slow dynamic exchange of endogenous centrosomal rootletin-meGFP, on the hours timescale, cells were arrested at the G1/S phase boundary of the cell cycle using thymidine, to circumvent the effects of cell cycle progression on root morphology (Fig 2C). Total centrosomal rootletin-meGFP fluorescent signal was then bleached, and recovery followed by tracking of NEDD1-mRuby3 marked centrosomes during time-lapse imaging (Fig 2F). Recovery of rootletin-meGFP fluorescence on this long timescale was limited to approximately 30%. Together, it can be surmised that eGFP-rootletin fibres are predominantly diffusionally stable structures that are progressively assembled slowly over hours following anaphase. Roots disentangle during transient centriole splitting in interphase How a single interphase cell coordinately organises 2 disengaged centrioles is unclear. The prevalence of centrosomal cohesion was systematically documented in a range of human tissue culture cell types by automated fluorescence imaging and analysis of centrosome position (Fig 3A). Quantification of the percentage of cells with split centrosomes—defined as 2 PCM foci >1.5 μm apart—showed that it was low at approximately 10%, dependent on cell type (see Materials and methods for further discussion of the definition of split centrosome). Thus, in most cell types, centrioles remain cohered in close proximity during interphase, consistent with previous work [6,30–34]. It was investigated whether the minority of split centrioles remain stably separated over time, perhaps due to a permanent failure of centrosome cohesion. However, single-cell 3D confocal live imaging of centriole pairs marked by GFP-Centrin1 showed transient splitting. Therefore, a single mother–daughter centriole pair would split into 2 and then rejoin, often repeatedly (Fig 3B; S3 Video). Transient centriole splitting was manifest in live-cell imaging of several different cell types, including Cal51, HeLa, RPE, and U2OS cells (Fig 3B–3E; S4 Video and S5 Video). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Roots disentangle during transient centriole splitting in interphase. (A) Quantification of centrosome cohesion in the interphase of various cell types through systematic immunofluorescent staining and analysis. The images show representative staining of PCNT (red; marking centrosomal PCM) and DNA (blue; hoechst 44432). The right panel shows representative segmentation of centrosomes (red), nuclei (blue), and cytoplasm (white) in Cal51 cells. The yellow asterisk denotes a cell containing 2 centrosome foci, separated by >1.5 μm. Scale bars 20 μm and 5 μm. The bar graph shows the mean percentage of cells with PCNT centroids separated by >1.5 μm, from a minimum of 500 cells. Error bars show SEM from 2 experiments. (B–E) Selected frames showing centriole splitting in live 3D confocal time-lapse imaging. Centrosomes are marked by either GFP-Centrin1 or NEDD1-mRuby3. Arrows denote centriole splitting events. The time intervals between frames are 12 minutes (panel B and C), 24 minutes (panel D), or 8 minutes (panel E). Scale bar 5 μm. See also S3–S5 Videos. (F) Centrosome cohesion in HeLa cells ± overexpression of eGFP-rootletin, measured by automated imaging and analysis. Horizontal bars show the mean of 2 experiments ± SD. *P < 0.001 by Fischer’s exact test. (G) Opposing models of root behaviour during centriole splitting, termed ‘Stable contact’ or ‘Disentangle’. (H) Representative 3D SIM images of roots (green) after centriole splitting, with the indicated costaining marking either the PCM or centrioles (red). Scale bar 1 μm. (I) Representative airyscan image of roots after centriole splitting. Scale bar 1μm. (J) Root linkage plotted as a function of centriole spacing distance. (K, L) Live-cell airyscan time-lapse imaging of endogenous rootletin-meGFP and NEDD1-mRuby3 during a centriole split (panel K) and when remaining stably cohered (panel L) in Cal51 cells. Scale bar 2 μm. See also S6 Video and S7 Video. See S1 Data for source data for the charts.; meGFP, monomeric enhanced green fluorescent protein; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; PCM, pericentriolar material; PCNT, Pericentrin; SIM, structured illumination microscopy. https://doi.org/10.1371/journal.pbio.2003998.g003 In agreement with a published report [31], HeLa Kyoto cells had high levels of centrosome separation, with approximately 50% of cells showing split centrioles in a fixed asynchronous population (Fig 3A). Because low levels of rootletin expression accompanied short roots in HeLa (Fig 1A) and previous work has shown that rootletin knockdown results in the loss of centrosome cohesion [2,35], the effect of increasing root length by rootletin overexpression on centrosome position was investigated in HeLa cells. This increase in fibre length significantly increased centrosome cohesion in interphase HeLa cells, as measured by automated imaging and analysis of immunofluorescently stained samples (Fig 3F, P < 0.001, Fischer’s exact test). Together, these results show that, although mother and daughter centrioles generally remain cohered into a single focal location, they are able to transiently split apart in interphase in a manner that is antagonised by eGFP-rootletin overexpression. How might rootletin fibres respond to transient centriole splitting? Two opposing models for root behaviour after centriole splitting were postulated (Fig 3G). The first was maintenance of a stable root contact between centrioles as they move apart, for example, due to stretching. The second was loss of physical connection and disentanglement (‘Stable contact’ versus ‘Disentangle’, respectively). Surprisingly and in contrast to cohered centrosomes, rootletin fibres were not detected between split centrioles (Fig 3H, Fig 3I and S5 Fig). Instead, roots from each centriole were generally only detected as linked together at a distance of less than approximately 1.5 μm (Fig 3J), thus supporting the disentanglement model. Simultaneous 2-colour airyscan microscopy of root disentanglement in living cells revealed that roots occupy markedly heterogenous orientations that change in response to in vivo centriole movement (Fig 3K; S6 Video). The centrosome distal ends of roots have the capacity to pivot relative to centrosome proximal ends, suggesting a common more stable attachment point at the proximal end. Pivoting of centrosome distal tips was not just observed in centrosomes with split centrioles but also in cohered centrosomes, with roots maintained stably at the centriole–centriole interface (Fig 3L; S7 Video). As centrosomes remerged after a split, roots did not necessarily join but could alternatively contact the PCM of the opposing centriole. Together, these observations indicate that although roots can be maintained stably at the interface between mother–daughter centrioles, their orientation is heterogeneous, and notably, in response to centriole movement, a continuous direct rootletin linkage is not detected. Independence of mother and daughter centrioles during interphase Because disengaged centrioles can transiently split (Fig 3), the comparative structure of roots and PCM on mother and daughter centrioles during splitting was investigated further. Root area was approximately halved in split versus cohered centrioles (Fig 4A), suggestive of equal partitioning of 2 independent roots. Indeed, discrimination of the mother and daughter using CEP164 immunostaining showed that roots are nucleated symmetrically on both mother and daughter (Fig 4B). A similar comparison of PCM structure with the PCM resident Pericentrin (PCNT) showed similarly that both mother and daughter centrioles individually nucleate PCM when split (Fig 4C), something also evident in the live-cell imaging of NEDD1-mRuby3 (Fig 3H–3K) and previous work [24,32,34]. These observations imply that both mature centrioles independently maintain roots and PCM during centrosome splitting in interphase. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Independence of mother and daughter centrioles during interphase. (A) Root fibre area is significantly lower (P < 0.0001, t test) in split versus cohered centrioles. Anti-rootletin immunofluorescent staining was imaged and segmented, N = 36 cells from 2 experiments. (B) Rootletin immunofluorescent staining (green) is the same at both the mother or daughter centriole (n.s., t test). “M” and “D” denote mother and daughter, respectively, on the basis of CEP164 positivity. N = 21 cells per sample. Scale bar 1 μm. (C) PCNT immunofluorescent staining (of the PCM) is the same (n.s., t test) on either mother or daughter centrioles. Cells were imaged and analysed as described in panel B, except segmenting PCNT. N = 21 cells. See S1 Data for source data for the charts. (D) Cells with 4 centrioles might either maintain them as separate pairs or cohere them together (“Grouped”). (E) The pie chart shows the proportion of each GFP-Centrin1 centriole configuration in cells with 4 centrioles, produced as depicted in Fig 1E. The images are representative of each configuration. N = 196 cells. (F) Cells expressing endogenously tagged rootletin-meGFP were fused with cells expressing endogenously tagged rootletin-mScarlet. (G) Representative SIM images of single-colour cells and a fused cell, created as depicted in Fig 4F and described in Materials and methods. Scale bars 1 μm throughout. (H) Interphase centriole pairs contain large bifurcating fibres that disentangle when centrioles move apart >1.5 μm relative to each other. Root dissolution begins prior to mitotic centrosome separation and chromosome condensation. At the time of centrosome separation, roots are diminished in quantity and ripped apart during poleward movement of centrosomes. Roots form slowly over many hours from anaphase, as diffusionally stable fibres. PLK1-dependent modification of procentrioles allows root formation in the ensuing cell cycle. meGFP, monomeric enhanced green fluorescent protein; PCNT, Pericentrin; PLK1, polo-like kinase 1; SIM, structured illumination microscopy. https://doi.org/10.1371/journal.pbio.2003998.g004 Given the dynamic nature of centrosome cohesion (Fig 3) and root disentanglement, it was of interest to investigate whether mother–daughter centriole pairs would be maintained in cells with 4 centrioles (Fig 4D). Centriole position in cells forced into interphase after a failed mitosis by STLC treatment (Fig 1E) showed all possible centrosome cohesion configurations. Most commonly, all 4 centrioles grouped as 1 (Fig 4E), but notably, other spatial arrangements were equally as likely as 2 pairs. Thus, 2 mother–daughter centriole pairs are not maintained separately but will cohere together, even in a grouping such as a single centriole and 3 cohered. Overexpression of eGFP-rootletin promoted centrosome cohesion in interphase cells with 4 mature centrioles created by sequential STLC/RO-3306 treatment (S6A–S6C Fig), a similar effect to that seen in cells with normal centrosome numbers (Fig 3F). Some cells with supernumerary centrosomes are able to cluster them to form a bipolar spindle during mitosis [36]. It was therefore investigated whether, in contrast to cells with 2 centrosomes (Fig 2), cells with supernumerary centrosomes retained centrosomal rootletin during mitosis. Cells clustering supernumerary centrosomes at spindle poles during mitosis did not contain roots, however (S7 Fig), consistent with a model wherein rootletin does not promote mitotic centriole clustering. Centrosome cohesion was further examined using polyethylene glycol–mediated cell fusion of 2 different cell lines, 1 expressing endogenously tagged rootletin-meGFP and the other expressing endogenously tagged rootletin-mScarlet (Fig 4F, S3F Fig and see Materials and methods for details of fusion). Fused cells contained centrosomes of 2 fluorescent colours, 1 from each different cell line, as well as 2 nuclei. Because rootletin shows very slow diffusional exchange (Fig 2), this allowed the origin of centriole pairs in fused cells to be distinguished based on the emitted fluorescence. As per after mitotic failure (Fig 4E), mother–daughter centriole pairs were not exclusively maintained after cell fusion, but instead, fluorescent roots of different colours engaged each other (Fig 4G and S6D Fig). Fusion of cell lines expressing rootletin-meGFP or NEDD1-mRuby3 similarly showed that fluorescent roots from 1 cell could embrace all 4 centrioles once fused (S6E Fig). Therefore, by 2 independent methods, mother–daughter centriole pairs are not stably maintained in cells with 4 centrioles. Discussion Cells must carefully regulate centrosome number and position, coordinating 2 centrioles that are capable of distinct functions [23,32,37]. The data here provoke an interesting hypothesis: that interphase cells always have 2 centrosomes that are generally held together by stable fibres that reach outward into the cytoplasm. Three key pieces of evidence are provided. Firstly, both mature interphase centrioles in a pair independently nucleate roots, as well as PCM. Secondly, these units—consisting of a centriole/root/PCM—have the capability to transiently spatially separate during interphase, accompanied by root disentanglement. Thirdly, cells engineered with 2 centriole pairs do not maintain them separately but instead dynamically make new groupings. Thus, there is remarkable pliability in the maintenance of centrosome cohesion, with individual centrioles able to rearrange between pairs, through dynamic splitting of roots. These conclusions are consistent with previous observations of split centrioles [6,30–34]. It is possible that centriole independence may aid plasticity of centrosome function such that the 2 centrioles can either act as one or separately. Thus, the data explain previous observations that centrioles may have either different or coordinated functions [23,32,37]. It cannot be totally excluded that fine rootletin fibres exist, below the detection limit of imaging, which thus keeps centrioles continuously linked. However, there is no evidence for this, either in the motion of roots in living cells or in fixed-cell analyses (Fig 3 and Fig 4), nor is it consistent with previous electron microscopy [2]. How non–membrane-bound organelles regulate their position, size, and number within the cellular interior is still not understood. Recent work has postulated that organelles such as centrosomes and P granules phase-separate as liquid-like compartments [38]. This model is characterised by high internal turnover of components parts, spherical shape, and the ability of multiple organelles to fuse [15]. In contrast, roots are diffusionally stable, remain separate through multiple cycles of merging and splitting, and are not spherical, instead potentially engendering polarity to the centrosome as a branched organelle. Therefore, roots have surprisingly different organisational principles in comparison to the PCM. Further work will be needed to understand whether this has implications for how centrosomal position is regulated. Rootletin loss in mice results in mechanical fragility in ciliated tissues such as photoreceptors, apparently due to the loss of ciliary rootlets [17]. Whether roots contribute to cellular mechanics, in either specialised cell types or nonciliated cycling cells through the maintenance of diffusionally stable contacts, will be an interesting future topic. In conclusion, root-mediated splitting of 2 centrosomes might allow plasticity of cytoskeletal function, thus explaining how 2 non–membrane-bound organelles coordinately function in either 1 or 2 locations during the interphase [30,34,39]. It is tempting to suggest that progressively nucleated, diffusionally stable polymers might also regulate the subcellular position and number of other organelles. Materials and methods Antibody validation Multiple lines of evidence were obtained to indicate that a commercially available anti-rootletin antibody (Novus Biologicals NBP1-80820) specifically recognises the product of the CROCC gene. siRNA depletion of rootletin (CROCC) using RNA interference removed signal by both immunofluorescence and western blot in multiple cell types (S1A, S1B and S1D Fig). An antibody-independent method—GFP tagging—showed similar protein abundances to measurements made by immunofluorescence, both in time and space (this is apparent throughout Figs 1–3). For example, centrosomal rootletin signal was virtually undetectable in metaphase by either antibody or GFP tagging. Anti-rootletin antibody also stained eGFP-rootletin when overexpressed as a transgene (S1C Fig), and ciliary rootlets in mouse photoreceptor cells (S1F and S1G Fig). Cell culture, DNA constructs, and siRNA Cal51 (German Collection of Microorganisms and Cell Cultures ACC303), U2OS (American Type Culture Collection ATCC HTB-96), HeLa Kyoto, PANC-1, and IMR-90 cell lines were grown in Dulbecco's modified Eagle's Medium (DMEM) supplemented with 10% fetal calf serum, Glutamax, and 100 μg/ml penicillin/streptomycin. hTERT RPE1 cells were cultured in DMEM/F12 with 10% Fetal Bovine Serum (FBS), penicillin/streptomycin, and 4.2% sodium bicarbonate. h-TERT BJ-5ta (ATCC CRL-4001) were grown in a 4:1 mixture of DMEM to M199. h-TERT HPNE (ATCC CRL-4023) were grown in a 3:1 mixture of DMEM to M3:BaseF medium, with 5% fetal calf serum, 10 ng/ml EGF, 2 mM glutamine, and 750 ng/ml puromycin. All tissue culture reagents were purchased from Sigma-Aldrich. DNA transfection was with lipofectamine 3000 (Invitrogen) according to the manufacturer's instructions. SiR-Hoechst (Tebu Bio) was incubated for 30 minutes at 200 nM before replacing with fresh medium for imaging. siRNA transfection was with RNAiMax transfection reagent (ThermoFisher Scientific). siRNAs used against rootletin (CROCC), CEP250/CNAP1, and nontargeting were Dharmacon ON-TARGET plus SMARTpools. siRNA against ODF2 was silencer select from ThermoFisher Scientific (#4427037). NEDD1-mRuby3 contained 5 glycine residues as a linker between the gene and fluorescent protein and was expressed from the vector pcDNA 3.1(+). Plasma membrane targeting was with a C-terminal fusion of the CAAX motif of KRAS4b, consisting of the amino acid sequence KMSKDGKKKKKKSKTKCVIM. mScarlet-cNAP1-CAAX was constructed with HD In-fusion cloning (Clontech), according to the manufacturer’s instructions. CEP135-mScarlet-CAAX was synthesised by GeneArt (ThermoFisher Scientific). Five glycine residues were often, but not always, used as a linker between fusion proteins. SIM microscopy A Zeiss Elyra S.1 equipped with a 63x NA 1.4 lens was used to acquire 16-bit 3D SIM images with 3 rotations and 5 phases. Double-colour labelling was with various combinations of either Alexa 488 or ATTO 488, and either Alexa 594, Alexa 568, or ATTO 565, on cells seeded on high-precision 170-nm glass coverslips (Ibidi μ-slide, 80827). Reconstruction was using Zen Blue software, using automatic parameters. The median (± median absolute deviation) lateral and axial resolution of the system using these settings was measured at 114 ± 4 nm and 352 ± 15 nm, respectively (full-width at half-maximum). Channel alignment was performed in Zen Blue, using a double-colour bead calibration standard. cDNA- and CRISPR Cas9–mediated meGFP cell line production Stable cell lines expressing cDNA constructs were produced by transfection followed by culture for at least 4 days, either with or without antibiotic selection, followed by fluorescence-activated cell sorting. CRISPR clones were produced essentially as described in [27], with some modifications. Guide RNA was expressed from pSpCas9(BB)-2A-GFP (PX458) (Addgene plasmid #48138). Guide RNA sequences all overlapped the CROCC STOP codon and against the +ve strand were as follows (5'—3'): CCAGCAGGAGCTCATTTCTC, CCAGAGAAATGAGCTCCTGC, and CAGGAGCTCATTTCTCTGGG. Donor plasmids were constructed in the vector pUC19 by HD In-fusion cloning (Clontech). They consisted of 800 base pair homology arms from the C-terminus of the CROCC genomic reference sequence, surrounding the meGFP or mScarlet coding sequence. Five glycine residues linked the gene and fluorescent protein. This insert was cloned into the BamH1 site of the vector pUC19. Insertion of meGFP into the endogenous CROCC locus was detected by extraction of genomic DNA using QuickExtract DNA extraction solution (epicentre) according to the manufacturer’s instructions, followed by junction PCR with the following primers: forward: GGCTGGCCTTACCTTCCCTT; reverse: CTGGAAGGCCTGTCACTGTC. Immunofluorescence Tissue culture cells were fixed in 4% paraformaldehyde or ice-cold 100% methanol for 10 minutes, permeabilised in 0.1% Triton, and blocked in 3% bovine serum albumen (ThermoFisher Scientific). Mouse photoreceptor cells were isolated from retina by gentle dissection before fixation and staining as for tissue culture cells. Antibodies used were as follows: rabbit anti-CROCC (1:250–1:750, Novus Biologicals NBP1-80820), mouse anti-NEDD1 (1:500, Abcam ab57336), rabbit anti-PCNT (1:1000; Abcam ab4448), mouse anti-SAS6 (1:300, Santa Cruz Biotechnology sc-81431), mouse anti-CENPJ (1:100, Santa Cruz Biotechnology sc-81432), mouse anti-gamma Tubulin (1:1000, ice-cold methanol; GTU-88), mouse anti-CETN1 (1:4000, EMD Millipore 20H5), mouse anti-CEP164 (1:200, Santa Cruz Biotechnology sc-515403), rabbit anti-CEP350 (1:500, Atlas Antibodies HPA030845), rabbit anti-CNAP1 (Proteintech, 14498-1-AP), rabbit anti-CDK5RAP2 (Atlas Antibodies, HPA046529), alpaca anti-GFP nanobody (1:400, Chromotek gba-488), and FluoTag-X2 anti-mScarlet (1:500, NanoTag Biotechnologies, N1302-At565). Image analysis Images are presented as maximum-intensity projections from 3D data unless otherwise stated. Image brightness and contrast settings were changed linearly and consistently between samples for display purposes of representative images, but not for quantitation. The intensity of centrosomal rootletin-meGFP in cycling cells was determined by automated centrosome tracking after movie acquisition. Centrosomes were segmented and tracked using the Trackmate plugin in ImageJ/Fiji [40], using LAP Tracker, and confirmed as successful by manual analysis of tracking. NEDD1-mRuby3 was tracked, a marker of the PCM that was present throughout the cell cycle. Individual cell tracks were aligned in time relative to anaphase, or the nearest frame to anaphase, based on both bright-field and SiR-hoechst fluorescent DNA labelling. Segmentation from fixed images was in Cell Profiler software, with data analysis in Knime software. For calculation of per-cell centriole splitting, nuclei were detected based on hoechst staining and cytoplasm by using a watershed algorithm outwards from nuclei based on gamma-tubulin staining. Mitotic cells were excluded based on hoechst staining. Centrosomes were detected with PCNT staining and defined as split if a cell contained 2 PCNT foci centroids >1.5 μm apart by Euclidean straight-line distance. A length of 1.5 μm was chosen as the definition of split centrioles because this distance was the threshold above which roots rarely linked centrioles in imaging (Fig 3J), thus providing an unbiased definition of split centrioles. For segmentation of roots, various thresholding strategies were used in CellProfiler, including propagation outwards from a GFP-Centrin1 seed region, or direct thresholding. Spacing of PCM staining was measured by adaptive thresholding followed by calculation of 2D Euclidean distance between centroids. Roots were segmented using propagation from PCM and then manually classified as linked if 1 pixel overlap occurred between a root from each PCM. Segmentation of centriole and PCM markers in Fig 1B was in cell profiler using automatic threshold and declumping of adjacent objects based on shape. Western blotting Antibodies used were rabbit anti-CROCC (1:250–1:750 overnight; Novus Biologicals, 80820), rabbit anti-CROCC (1:250–1:750 overnight; Novus Biologicals, 80821), and mouse monoclonal beta-Actin (1:10000 1 hour at room temperature; Sigma-Aldrich). Cells were lysed for 20 minutes on ice in RIPA buffer (50 mM Tris HCl, pH 8, 150 mM NaCl, 1% NP40, 0.5 M sodium deoxycholate, 0.1% SDS, complete protease inhibitor cocktail, PhosSTOP [Roche]). Protein concentration was quantitated using the bicinchoninic acid method (Sigma-Aldrich). Whole-cell extracts were separated by electrophoresis on a 3% to 8% Tris-Acetate gel and transferred to PVDF membrane using the iBlot2 system (ThermoFisher Scientific) according to the manufacturer’s instructions. Membranes were blocked in 5% milk dissolved in 0.1% Tween/TBS. Live-cell time-lapse imaging and FRAP Cells were imaged without phenol red in either L15 CO2-indepdendent medium or in Fluorobrite Imaging medium with 5% CO2 at 37°C, in Ibidi u-slide 8-well dishes. Imaging was with a Carl Zeiss 880 airyscan, either in airyscan or standard confocal mode, using either a 63x NA 1.4 or 100x NA 1.4 oil immersion lens. Airyscan processing was performed with automatic settings in Zen Black. The median (± median absolute deviation) lateral and axial resolution of the system was measured at 198 ± 7.5 nm and 913 ± 50 nm (full-width at half-maximum), respectively. FRAP was performed essentially as described in [14], bleaching using a 488 argon laser at 100% for the minimum time required to cause approximately 50% fluorescence loss (keeping the same duration in all samples). Cell fusion Cells were fused using Hybri-Max 50% 1450 polyethylene glycol solution (Sigma-Aldrich). Briefly, cells were trypsinised, resuspended in PBS, and mixed at a 1:1 ratio. After spinning, the PBS was aspirated, and PEG was added dropwise over 30 seconds to the cell pellet and left for an additional 3.5 minutes at room temperature. Serum-free medium was then added dropwise for 1 minute before 10-minute incubation at 37°C with normal medium, followed by exchange for fresh medium. Fused cells constituted approximately 1% of the population and consequently were enriched by fluorescence-activated cell sorting. The fluorescence intensity from endogenously tagged rootletin (either meGFP or mScarlet) was dim as detected by flow cytometry, and so cells were labelled with either CellTrace Violet or CellTrace Far Red dye (ThermoFisher Scientific) to enable efficient sorting. Labelling was for 1 minute at room temperature in PBS, at 500 nM or 20 nM for CellTrace Violet or CellTrace Far Red, respectively. Cells were FACS sorted by gating for either CellTrace Violet, CellTrace Far Red, or NEDD1-mRuby3 positivity relative to negative controls, directly into imaging dishes. The majority of these cells were aneuploid relative to the single-colour lines as expected. Cells with centrosomes marked by NEDD1-mRuby3 fluorescence contained up to 4 foci, due to turnover of this marker at the centrosome. Mitotic arrest and release Cells were arrested for 12 hours in either 200 nM BI2536 (Sigma-Aldrich), 10 μM STLC, or 50 ng/ml Nocodazole. Only mitotically arrested cells were analysed further, by mitotic shake-off. Mitotic exit was forced with RO-3306 (10 μM) for 6 hours, or cells were released from mitotic blockade using 2 washes in warm medium. Dihydrocytochalasin B (DCB) treatment was at 4 μM for 18 hours, followed by 3 washes in fresh medium. eGFP-rootletin overexpression in cells with supernumerary centrosomes Cells were transfected with eGFP-rootletin for 24 hours before overnight arrest in STLC. Mitotic shake-off was performed into RO-3306, allowing a 6.5-hour release. Imaging was by tile-scanning confocal z-stacks. Transfected cells were identified in CellProfiler through segmentation of eGFP-rootletin filaments by global Robust Background threshold. Centrosome cohesion was measured by segmentation of PCNT foci without declumping, thus grouping cohered centrosomes as 1 focus. Split centrosomes were identified in this case as cells with 2, 3, or 4 PCNT foci by Robust Threshold segmentation. Cells either without any detected centrosomes or with greater than 4 foci constituted around 10% of cells, and these were discarded from further analysis. Antibody validation Multiple lines of evidence were obtained to indicate that a commercially available anti-rootletin antibody (Novus Biologicals NBP1-80820) specifically recognises the product of the CROCC gene. siRNA depletion of rootletin (CROCC) using RNA interference removed signal by both immunofluorescence and western blot in multiple cell types (S1A, S1B and S1D Fig). An antibody-independent method—GFP tagging—showed similar protein abundances to measurements made by immunofluorescence, both in time and space (this is apparent throughout Figs 1–3). For example, centrosomal rootletin signal was virtually undetectable in metaphase by either antibody or GFP tagging. Anti-rootletin antibody also stained eGFP-rootletin when overexpressed as a transgene (S1C Fig), and ciliary rootlets in mouse photoreceptor cells (S1F and S1G Fig). Cell culture, DNA constructs, and siRNA Cal51 (German Collection of Microorganisms and Cell Cultures ACC303), U2OS (American Type Culture Collection ATCC HTB-96), HeLa Kyoto, PANC-1, and IMR-90 cell lines were grown in Dulbecco's modified Eagle's Medium (DMEM) supplemented with 10% fetal calf serum, Glutamax, and 100 μg/ml penicillin/streptomycin. hTERT RPE1 cells were cultured in DMEM/F12 with 10% Fetal Bovine Serum (FBS), penicillin/streptomycin, and 4.2% sodium bicarbonate. h-TERT BJ-5ta (ATCC CRL-4001) were grown in a 4:1 mixture of DMEM to M199. h-TERT HPNE (ATCC CRL-4023) were grown in a 3:1 mixture of DMEM to M3:BaseF medium, with 5% fetal calf serum, 10 ng/ml EGF, 2 mM glutamine, and 750 ng/ml puromycin. All tissue culture reagents were purchased from Sigma-Aldrich. DNA transfection was with lipofectamine 3000 (Invitrogen) according to the manufacturer's instructions. SiR-Hoechst (Tebu Bio) was incubated for 30 minutes at 200 nM before replacing with fresh medium for imaging. siRNA transfection was with RNAiMax transfection reagent (ThermoFisher Scientific). siRNAs used against rootletin (CROCC), CEP250/CNAP1, and nontargeting were Dharmacon ON-TARGET plus SMARTpools. siRNA against ODF2 was silencer select from ThermoFisher Scientific (#4427037). NEDD1-mRuby3 contained 5 glycine residues as a linker between the gene and fluorescent protein and was expressed from the vector pcDNA 3.1(+). Plasma membrane targeting was with a C-terminal fusion of the CAAX motif of KRAS4b, consisting of the amino acid sequence KMSKDGKKKKKKSKTKCVIM. mScarlet-cNAP1-CAAX was constructed with HD In-fusion cloning (Clontech), according to the manufacturer’s instructions. CEP135-mScarlet-CAAX was synthesised by GeneArt (ThermoFisher Scientific). Five glycine residues were often, but not always, used as a linker between fusion proteins. SIM microscopy A Zeiss Elyra S.1 equipped with a 63x NA 1.4 lens was used to acquire 16-bit 3D SIM images with 3 rotations and 5 phases. Double-colour labelling was with various combinations of either Alexa 488 or ATTO 488, and either Alexa 594, Alexa 568, or ATTO 565, on cells seeded on high-precision 170-nm glass coverslips (Ibidi μ-slide, 80827). Reconstruction was using Zen Blue software, using automatic parameters. The median (± median absolute deviation) lateral and axial resolution of the system using these settings was measured at 114 ± 4 nm and 352 ± 15 nm, respectively (full-width at half-maximum). Channel alignment was performed in Zen Blue, using a double-colour bead calibration standard. cDNA- and CRISPR Cas9–mediated meGFP cell line production Stable cell lines expressing cDNA constructs were produced by transfection followed by culture for at least 4 days, either with or without antibiotic selection, followed by fluorescence-activated cell sorting. CRISPR clones were produced essentially as described in [27], with some modifications. Guide RNA was expressed from pSpCas9(BB)-2A-GFP (PX458) (Addgene plasmid #48138). Guide RNA sequences all overlapped the CROCC STOP codon and against the +ve strand were as follows (5'—3'): CCAGCAGGAGCTCATTTCTC, CCAGAGAAATGAGCTCCTGC, and CAGGAGCTCATTTCTCTGGG. Donor plasmids were constructed in the vector pUC19 by HD In-fusion cloning (Clontech). They consisted of 800 base pair homology arms from the C-terminus of the CROCC genomic reference sequence, surrounding the meGFP or mScarlet coding sequence. Five glycine residues linked the gene and fluorescent protein. This insert was cloned into the BamH1 site of the vector pUC19. Insertion of meGFP into the endogenous CROCC locus was detected by extraction of genomic DNA using QuickExtract DNA extraction solution (epicentre) according to the manufacturer’s instructions, followed by junction PCR with the following primers: forward: GGCTGGCCTTACCTTCCCTT; reverse: CTGGAAGGCCTGTCACTGTC. Immunofluorescence Tissue culture cells were fixed in 4% paraformaldehyde or ice-cold 100% methanol for 10 minutes, permeabilised in 0.1% Triton, and blocked in 3% bovine serum albumen (ThermoFisher Scientific). Mouse photoreceptor cells were isolated from retina by gentle dissection before fixation and staining as for tissue culture cells. Antibodies used were as follows: rabbit anti-CROCC (1:250–1:750, Novus Biologicals NBP1-80820), mouse anti-NEDD1 (1:500, Abcam ab57336), rabbit anti-PCNT (1:1000; Abcam ab4448), mouse anti-SAS6 (1:300, Santa Cruz Biotechnology sc-81431), mouse anti-CENPJ (1:100, Santa Cruz Biotechnology sc-81432), mouse anti-gamma Tubulin (1:1000, ice-cold methanol; GTU-88), mouse anti-CETN1 (1:4000, EMD Millipore 20H5), mouse anti-CEP164 (1:200, Santa Cruz Biotechnology sc-515403), rabbit anti-CEP350 (1:500, Atlas Antibodies HPA030845), rabbit anti-CNAP1 (Proteintech, 14498-1-AP), rabbit anti-CDK5RAP2 (Atlas Antibodies, HPA046529), alpaca anti-GFP nanobody (1:400, Chromotek gba-488), and FluoTag-X2 anti-mScarlet (1:500, NanoTag Biotechnologies, N1302-At565). Image analysis Images are presented as maximum-intensity projections from 3D data unless otherwise stated. Image brightness and contrast settings were changed linearly and consistently between samples for display purposes of representative images, but not for quantitation. The intensity of centrosomal rootletin-meGFP in cycling cells was determined by automated centrosome tracking after movie acquisition. Centrosomes were segmented and tracked using the Trackmate plugin in ImageJ/Fiji [40], using LAP Tracker, and confirmed as successful by manual analysis of tracking. NEDD1-mRuby3 was tracked, a marker of the PCM that was present throughout the cell cycle. Individual cell tracks were aligned in time relative to anaphase, or the nearest frame to anaphase, based on both bright-field and SiR-hoechst fluorescent DNA labelling. Segmentation from fixed images was in Cell Profiler software, with data analysis in Knime software. For calculation of per-cell centriole splitting, nuclei were detected based on hoechst staining and cytoplasm by using a watershed algorithm outwards from nuclei based on gamma-tubulin staining. Mitotic cells were excluded based on hoechst staining. Centrosomes were detected with PCNT staining and defined as split if a cell contained 2 PCNT foci centroids >1.5 μm apart by Euclidean straight-line distance. A length of 1.5 μm was chosen as the definition of split centrioles because this distance was the threshold above which roots rarely linked centrioles in imaging (Fig 3J), thus providing an unbiased definition of split centrioles. For segmentation of roots, various thresholding strategies were used in CellProfiler, including propagation outwards from a GFP-Centrin1 seed region, or direct thresholding. Spacing of PCM staining was measured by adaptive thresholding followed by calculation of 2D Euclidean distance between centroids. Roots were segmented using propagation from PCM and then manually classified as linked if 1 pixel overlap occurred between a root from each PCM. Segmentation of centriole and PCM markers in Fig 1B was in cell profiler using automatic threshold and declumping of adjacent objects based on shape. Western blotting Antibodies used were rabbit anti-CROCC (1:250–1:750 overnight; Novus Biologicals, 80820), rabbit anti-CROCC (1:250–1:750 overnight; Novus Biologicals, 80821), and mouse monoclonal beta-Actin (1:10000 1 hour at room temperature; Sigma-Aldrich). Cells were lysed for 20 minutes on ice in RIPA buffer (50 mM Tris HCl, pH 8, 150 mM NaCl, 1% NP40, 0.5 M sodium deoxycholate, 0.1% SDS, complete protease inhibitor cocktail, PhosSTOP [Roche]). Protein concentration was quantitated using the bicinchoninic acid method (Sigma-Aldrich). Whole-cell extracts were separated by electrophoresis on a 3% to 8% Tris-Acetate gel and transferred to PVDF membrane using the iBlot2 system (ThermoFisher Scientific) according to the manufacturer’s instructions. Membranes were blocked in 5% milk dissolved in 0.1% Tween/TBS. Live-cell time-lapse imaging and FRAP Cells were imaged without phenol red in either L15 CO2-indepdendent medium or in Fluorobrite Imaging medium with 5% CO2 at 37°C, in Ibidi u-slide 8-well dishes. Imaging was with a Carl Zeiss 880 airyscan, either in airyscan or standard confocal mode, using either a 63x NA 1.4 or 100x NA 1.4 oil immersion lens. Airyscan processing was performed with automatic settings in Zen Black. The median (± median absolute deviation) lateral and axial resolution of the system was measured at 198 ± 7.5 nm and 913 ± 50 nm (full-width at half-maximum), respectively. FRAP was performed essentially as described in [14], bleaching using a 488 argon laser at 100% for the minimum time required to cause approximately 50% fluorescence loss (keeping the same duration in all samples). Cell fusion Cells were fused using Hybri-Max 50% 1450 polyethylene glycol solution (Sigma-Aldrich). Briefly, cells were trypsinised, resuspended in PBS, and mixed at a 1:1 ratio. After spinning, the PBS was aspirated, and PEG was added dropwise over 30 seconds to the cell pellet and left for an additional 3.5 minutes at room temperature. Serum-free medium was then added dropwise for 1 minute before 10-minute incubation at 37°C with normal medium, followed by exchange for fresh medium. Fused cells constituted approximately 1% of the population and consequently were enriched by fluorescence-activated cell sorting. The fluorescence intensity from endogenously tagged rootletin (either meGFP or mScarlet) was dim as detected by flow cytometry, and so cells were labelled with either CellTrace Violet or CellTrace Far Red dye (ThermoFisher Scientific) to enable efficient sorting. Labelling was for 1 minute at room temperature in PBS, at 500 nM or 20 nM for CellTrace Violet or CellTrace Far Red, respectively. Cells were FACS sorted by gating for either CellTrace Violet, CellTrace Far Red, or NEDD1-mRuby3 positivity relative to negative controls, directly into imaging dishes. The majority of these cells were aneuploid relative to the single-colour lines as expected. Cells with centrosomes marked by NEDD1-mRuby3 fluorescence contained up to 4 foci, due to turnover of this marker at the centrosome. Mitotic arrest and release Cells were arrested for 12 hours in either 200 nM BI2536 (Sigma-Aldrich), 10 μM STLC, or 50 ng/ml Nocodazole. Only mitotically arrested cells were analysed further, by mitotic shake-off. Mitotic exit was forced with RO-3306 (10 μM) for 6 hours, or cells were released from mitotic blockade using 2 washes in warm medium. Dihydrocytochalasin B (DCB) treatment was at 4 μM for 18 hours, followed by 3 washes in fresh medium. eGFP-rootletin overexpression in cells with supernumerary centrosomes Cells were transfected with eGFP-rootletin for 24 hours before overnight arrest in STLC. Mitotic shake-off was performed into RO-3306, allowing a 6.5-hour release. Imaging was by tile-scanning confocal z-stacks. Transfected cells were identified in CellProfiler through segmentation of eGFP-rootletin filaments by global Robust Background threshold. Centrosome cohesion was measured by segmentation of PCNT foci without declumping, thus grouping cohered centrosomes as 1 focus. Split centrosomes were identified in this case as cells with 2, 3, or 4 PCNT foci by Robust Threshold segmentation. Cells either without any detected centrosomes or with greater than 4 foci constituted around 10% of cells, and these were discarded from further analysis. Supporting information S1 Video. Growth of cDNA eGFP-rootletin fibres in a single Cal51 cell after transfection. Each frame is taken at a 6-minute interval and shows a maximum-intensity z-projection from a 3D confocal stack. eGFP, enhanced green fluorescent protein. https://doi.org/10.1371/journal.pbio.2003998.s001 (AVI) S2 Video. Cell cycle–dependent changes in centrosomal rootletin-meGFP intensity (green; roots) in Cal51cells coexpressing NEDD1-mRuby3 (red; marking the PCM) and stained with SiR-Hoechst (blue; DNA). Each frame is taken at a 12-minute interval and shows a maximum-intensity z-projection from a 3D confocal stack. meGFP, monomeric enhanced green fluorescent protein; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; PCM, pericentriolar material. https://doi.org/10.1371/journal.pbio.2003998.s002 (AVI) S3 Video. Centriole splitting and cohesion visualised by 3D confocal time-lapse imaging of GFP-Centrin1 (centrioles) in Cal51 cells. Each frame is taken at a 12-minute interval and shows a maximum-intensity z-projection. Note that this cell divides after 25 frames. Cal51,; GFP, green fluorescent protein. https://doi.org/10.1371/journal.pbio.2003998.s003 (AVI) S4 Video. Centriole splitting and cohesion visualised by 3D confocal time-lapse imaging of GFP-Centrin1 (centrioles) in HeLa cells. Each frame is taken at a 12-minute interval and shows a maximum-intensity z-projection. GFP, green fluorescent protein; HeLa. https://doi.org/10.1371/journal.pbio.2003998.s004 (AVI) S5 Video. Centriole splitting and cohesion, visualised by 3D confocal time-lapse imaging of GFP-Centrin1 (centrioles) in RPE cells. Each frame is taken at a 24-minute interval and shows a maximum-intensity z-projection. GFP, green fluorescent protein; RPE, retinal pigment epithelium. https://doi.org/10.1371/journal.pbio.2003998.s005 (AVI) S6 Video. Root disentanglement during centriole splitting and remerging, visualised by 3D confocal airyscan time-lapse imaging of rootletin-meGFP (green; roots) and NEDD1-mRuby3 (red; PCM). Each frame is taken at a 10-minute interval and shows a maximum-intensity z-projection. meGFP, monomeric enhanced green fluorescent protein; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; PCM, pericentriolar material. https://doi.org/10.1371/journal.pbio.2003998.s006 (AVI) S7 Video. Root behaviour in a stably cohered centrosome, visualised by 3D confocal airyscan time-lapse imaging of rootletin-meGFP (green; roots) and NEDD1-mRuby3 (red; PCM). Each frame is taken at a 10-minute interval and shows a maximum-intensity z-projection. meGFP, monomeric enhanced green fluorescent protein; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; PCM, pericentriolar material. https://doi.org/10.1371/journal.pbio.2003998.s007 (AVI) S1 Fig. Validation of anti-rootletin antibody (related to Fig 1). (A, B) Anti-rootletin immunofluorescent staining (green) is not evident at centrosomes costained with anti-NEDD1 antibody (red) after rootletin (CROCC) siRNA, in multiple cell types. Anti-rootletin staining (green) is present after nontargeting siRNA negative control (nt) treatment. Arrows have been annotated manually to indicate centrosomes. Imaging conditions and brightness and contrast settings are consistent between control and siRNA treated samples. Panel A shows U2OS cells and panel B shows RPE cells. (C) Anti-rootletin antibody immunofluorescently stains (red) eGFP-rootletin overexpressed from a cDNA transgene (green). (D) Anti-rootletin bands are not detected by western blot of whole-cell lysate after rootletin (CROCC) siRNA, demonstrating antibody specificity in multiple cell types. (E) Pairwise anti-rootletin antibody costaining of rootletin (green) and other centrosomal genes (red), either in the PCM or centrioles. Maximum-intensity projections of confocal airyscan images; scale bar 1 μm. (F) Anti-rootletin staining (green) in mouse photoreceptor cells. Rootlets extend as part of the inner segment, from nuclei (blue) in the outer nuclear layer. Scale bars 40 μm (left and centre panels) and 20 μm (right panel). (G) Ciliary rootlets (green) costained with anti-NEDD1 antibody in mouse photoreceptor cells. Scale bar 5 μm. eGFP, enhanced green fluorescent protein; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; RPE, retinal pigment epithelium; siRNA, small interfering RNA; nt, nontargeting. https://doi.org/10.1371/journal.pbio.2003998.s008 (PDF) S2 Fig. Overexpression of eGFP-rootletin progressively assembles fibres that are diffusionally stable over minutes (related to Fig 2). (A) Representative maximum-intensity z-projection airyscan (i) and SIM (ii) images of overexpressed eGFP-rootletin fibres (green), costained with the PCM marker PCNT (red). Scale bars 5 μm and 1 μm, respectively. (B) FRAP of eGFP-rootletin in the location denoted by the arrow. The graphs show the fluorescence intensity of a line profile along the fibre at each timepoint. Scale bar 1μm. See S1 Data for source data. eGFP, enhanced green fluorescent protein; FRAP, fluorescence recovery after photobleaching; PCM, pericentriolar material; PCNT, Pericentrin; SIM, structured illumination microscopy. https://doi.org/10.1371/journal.pbio.2003998.s009 (PDF) S3 Fig. CRISPR Cas9–mediated tagging of endogenous rootletin/CROCC (related to Figs 2 and 3). (A) Schematic of guide RNAs targeting the STOP codon of CROCC as well as donor plasmid containing fluorescent protein and homology arms. (B) Clones were screened sequentially by FACS sorting, fluorescence microscopy, and junction PCR. (C) Example overlapping genomic PCR screen of clones expressing rootletin-meGFP. Clone 4_1 was used in this study because it has homozygous tagging of rootletin. Clones 4_7 and 20 are examples of heterozygous and negative clones, respectively. (D) Representative fluorescence microscopy screening of clones expressing endogenous rootletin-meGFP. The bottom panel shows centrosomal fluorescence in positive clones. Scale bar 5 μm. (E) Rootletin-meGFP centrosomal fluorescent signal closely resembles anti-rootletin antibody staining. The image shows clone 4_1 stained with anti-rootletin antibody and imaged by airyscan imaging. Scale bar 1 μm. (F) Overlapping genomic PCR screen of clones expressing rootletin-mScarlet. FACS, fluorescence-activated cell sorting; PCR, polymerase chain reaction. https://doi.org/10.1371/journal.pbio.2003998.s010 (PDF) S4 Fig. Ectopic CNAP1/CEP135 localisation to the plasma membrane with a CAAX motif is not sufficient for root formation. (A) siRNA-mediated knockdown of CNAP1 reduces the mean intensity of rootletin immunofluorescent staining at the centrosome. Cells were treated with the indicated siRNA for 18 hours, before immunofluorescent staining with anti-rootletin antibody. Horizontal bars show the mean of the distribution, dots show single cells. nt denotes nontargeting siRNA, -ve denotes untransfected. See S1 Data for source data. (B) Representative 3D SIM image of mScarlet-CNAP1-CAAX (red), costained with anti-rootletin (green) and DNA (Hoechst 44432). The right panel shows a zoomed region of the left panel image. Scale bar 5 μm. Arrows denote plasma membrane. (C) Representative 3D SIM image of CEP135-mScarlet-CAAX (red), costained with anti-rootletin (green) and DNA (Hoechst 44432), as described in panel A. AU, arbitrary unit; nt, nontargeting; SIM, structured illumination microscopy; siRNA, small interfering RNA. https://doi.org/10.1371/journal.pbio.2003998.s011 (PDF) S5 Fig. Rootletin links between centriole pairs are not detected using high brightness and contrast settings (related to Fig 3). Rootletin was stained with either anti-rootletin antibody (A) or rootletin-meGFP was stained with anti-GFP nanobody (B) and imaged with 3D SIM. Centriolar PCM was costained with either anti-gamma TUB or anti-PCNT (red). Scale bar 1 μm. meGFP, monomeric enhanced green fluorescent protein; PCM, pericentriolar material; PCNT, Pericentrin; SIM, structured illumination microscopy; g-TUB, tubulin gamma 1 gene. https://doi.org/10.1371/journal.pbio.2003998.s012 (PDF) S6 Fig. Centrosome cohesion in cells with supernumerary centrosomes (related to Fig 4). (A) Following transfection with eGFP-rootletin, supernumerary centrosomes were induced as described in Fig 1E, and cells were imaged by confocal tile scanning. Representative immunofluorescent staining (left panel) and segmentation (right panel) of nuclei, PCM, and eGFP-rootletin. Scale bar 5 μm. (B) eGFP-rootletin expressing cells had significantly higher centrosome cohesion relative to untransfected cells. The graph plots the proportion of cells with unsplit (cohered) centrosomes, from the data described in (A). n = 778 and 374 cells in -ve and eGFP-rootletin categories respectively. The values are significantly different, p<0.0001, Fischer’s exact test. See S1 Data for source data. (C) The proportion of each configuration of PCNT foci detected from the data described in (A) and (B). (D) Cells expressing endogenously tagged rootletin-meGFP were fused with cells expressing endogenously tagged rootletin-mScarlet and imaged by SIM. Each panel shows a maximum-intensity z-projection of centrosomes in 1 fused cell. Scale bar 1 μm. (E) Cells expressing rootletin-meGFP were fused with cells stably expressing NEDD1-mRuby3. Root arrangement in 3 different fused cells. Scale bar 1 μm. meGFP, monomeric enhanced green fluorescent protein; NEDD1, neural precursor cell expressed, developmentally down-regulated 1; PCM, pericentriolar material; PCNT, Pericentrin. https://doi.org/10.1371/journal.pbio.2003998.s013 (PDF) S7 Fig. Cells with supernumerary centrosomes do not retain mitotic roots (related to Fig 4). (A) Supernumerary centrosomes were induced by DCB, which causes cytokinesis failure and tetraploidy. Roots (red) were not detected at centrioles (marked in green by GFP-Centrin1) in mitotic cells, regardless of whether spindles were multipolar (A) or bipolar (B). Image acquisition and presentation settings are constant throughout all 3 panels. Scale bar 5 μm. DCB, Dihydrocytochalasin B; GFP, green fluorescent protein. https://doi.org/10.1371/journal.pbio.2003998.s014 (PDF) S1 Data. Raw values from the microscopy data shown in Figs 1, 3, 4, S2, S4 and S6. https://doi.org/10.1371/journal.pbio.2003998.s015 (XLSX) Acknowledgments pEGFP rootletin (Nigg pFL2 [CW499]) was a gift from Erich Nigg (Addgene plasmid #41166). pSpCas9(BB)-2A-GFP (PX458) was a gift from Feng Zhang (Addgene plasmid #48138). pIRES GFP Centrin1 Hygro was a gift from Matthieu Piel (Addgene plasmid #64339). IMR-90 cells were a gift from Liam Cassidy. CNAP1 cDNA was provided by Tara Hardy and Andrew Fry. CAAX domain cDNA was a gift from Anchal Chandra. The Cambridge Institute for Medical Research flow cytometry facility and Susan Grant provided excellent support. Paul M. W. French and Ashok R. Venkitaraman provided guidance and feedback throughout. Amy Emery, Petr Strnad, and Tom Miller provided critical comments on the manuscript.

doi: 10.1371/journal.pbio.2003993pmid: 29624578

Introduction One of the most challenging questions in cognitive neuroscience is how abstract knowledge emerges from more basic dimensions of information, such as visual shapes and patterns of motor action. How do we proceed from the visual shape of a pair of scissors to the knowledge that they can be used to cut things and that they are semantically related to an axe, which looks different and is manipulated differently from scissors? Research on the neural basis of semantic memory—the storage of general knowledge about the world—has revealed widely distributed brain regions supporting modality-specific attributes of objects, such as shape, color, and motion (e.g., [1,2]; see review in [3]). Nonetheless, such attribute-specific knowledge and its simple pairings are not adequate to explain the actual semantic space of objects that have quite different sensory/motor attributes but that may nonetheless be considered to be semantically similar (e.g., [4–7]). To achieve such a semantic space, various steps of binding and abstraction are assumed to occur at specific gray matter (GM) regions [6,8–11]. Although past research on semantic representation has focused on the roles of cortical regions, specific white matter (WM) tracts have been found to be necessary for semantic processing, including the left inferior fronto-occipital fasciculus (IFOF), the left uncinate fasciculus (UF), and the left anterior thalamic radiation. Damage to these tracts is associated with semantic deficits in patients [12–17]. WM is classically assumed to relay information [18–20]. In accord with this general notion, these WM tracts that are necessary for semantic processing are assumed to relay distributed information to particular GM regions (e.g., the anterior temporal lobe or angular gyrus) for binding, where concepts are represented and the “deep structures” of semantic space are formed [6,7,21]. The nature of the potential information carried by WM has never been discussed or examined. Herein, we present results for a new notion that the WM connections, being natural binding structures, provide an alternative basis to achieve semantic representation. Distributed GM regions that represent different attribute dimensions (e.g., shape, color, manner of interaction) of the same object are connected by WM. The WM linking pattern itself would then contain multiple dimensions of information in these GM regions and, importantly, additional information about the manner of mapping among various attributes. The incorporation of these elements has been argued to be necessary for the “higher-order” semantic similarity relationships, which are not explained by attribute-specific spaces, to emerge (e.g., [7, 22]). To investigate the information coded in WM connections, we extended representational similarity analysis (RSA) [23], a highly productive method that tests the nature of representation in functional magnetic resonance imaging (fMRI) studies of cortical regions [24–26], to lesion data and WM connections. RSA examines the relationship between the representational dissimilarity matrix (RDM) derived from neural patterns and RDMs based on various types of stimulus information as a measure of information representation. The conventional neural RDM is measured by the dissimilarity of brain activity patterns induced by stimulus conditions. Here, we compute the neural RDMs with a machine-learning model using the voxel-wise lesion patterns as features to predict behavioral performance in patients with brain damage (see Fig 1). The performance in picture naming of 100 object items and the structural MRI data of 80 patients were collected. For each WM connection, a training model was built for each item (e.g., scissors) using the support vector machine (SVM) classifier with patients’ voxel-wise lesion patterns as predictive features and the naming performances of that item as labels (0, incorrect; 1, correct). The correlation between the predicted score using the classifier from that item and the actual scores of another item (e.g., axe) was taken as the neural similarity basis of these two items, based on the assumption that if this connection pattern contains certain aspects of information shared by the naming process of these two items, models trained with one item (useful features relevant for such information) should also predict naming accuracy of the other item. Once the neural RDMs are obtained from various WM connections or GM regions using this method, they can be correlated with behavioral RDMs of various object property dimensions, including the semantic RDM and four modality-specific attribute RDMs (shape, manipulation, color, and motion). Neural RDMs that are correlated with the semantic RDM even after controlling for the attribute RDMs are considered to contain “higher-order” semantic information. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. A flowchart for constructing a neural RDM in a WM connection. (A) The neuropsychological test. We asked patients to complete a picture-naming task containing 100 items. The response for each item was scored as 1 if correct or 0 if wrong. (B) The lesion mask (manually traced in T1 image) in a given patient was converted to MNI space and was then overlapped with the WM connection template constructed from a healthy population [27] to extract the voxel-wise lesion pattern on each WM connection. (C) The SVM classifier was trained on the naming accuracy of one item i (e.g., scissors) and lesion patterns on a WM connection in some patients (see Materials and methods) and then used to generate the predicted naming score in the testing patients (1 or 0). The correspondence (simple matching coefficient) between the predicted score and the actual naming score of each of the other items (e.g., axe) across patients was calculated. This correlation reflects to what degree the lesion features that were useful to predict naming accuracy of item i could also be useful to predict item j, and thus was taken as the neural similarity between the naming process of the training item i and this other item j (scissors–axe similarity) on this connection. All cross-item and within-item similarity could be obtained this way, resulting in a 100 × 100 similarity matrix for this connection. (D) A sample neural RDM of a WM connection (between MTG and superior ATL). The values of dissimilarity were 1-similarity (obtained in C); red indicates low dissimilarity (high similarity) and blue high dissimilarity (low similarity). The object line drawings were done by the first author Y.F.; The brain figure was generated using BrainNet Viewer [28]. ATL, anterior temporal lobe; MTG, middle temporal gyrus; RDM, representational dissimilarity matrix; SVM, support vector machine; WM, white matter. https://doi.org/10.1371/journal.pbio.2003993.g001 Results Behavioral RDMs: Semantic and modality-specific attributes Behavioral RDMs for the semantic, shape, manipulation, color, and motion features of 100 objects (20 animals, 20 fruits and vegetables, 20 tools, 20 non-tool small objects, and 20 large non-manipulable objects) were generated using a multi-arrangement method [29]. In this task, 20 college students were instructed to arrange the items by a particular dimension of interest on a computer screen, and the distance among items was derived, resulting in an RDM (see Fig 2A). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. The construction and result of behavioral RDMs. (A) The multi-arrangement method. Twenty college students were asked to arrange object pictures according to their semantic (or modality-specific attribute) relatedness by dragging the items on a screen with a mouse. The distances between items on the screen would transform into an RDM. If two items, e.g., scissors and axe, showed a close distance, then they were assigned a low value in the RDM. (B) The results of the behavioral RDMs. Three broad types of distances were measured: semantic similarity, modality-specific attributes (shape, manipulation, color, and motion), and control models that are also potentially relevant to object naming (early visual, phonological, and object category matrix). The values of dissimilarity were transformed to percentile for display. Red indicates low dissimilarity (high similarity) and blue high dissimilarity (low similarity). (C) Visualization of the semantic RDM using multidimensional scaling. (D) The correlations among various behavioral RDMs. The object line drawings were done by the first author Y.F. The underlying data for this figure can be found at https://osf.io/h7upk/?view_only=52b8f86cffa14ed4844e4a1b9cd429cb. F&V, fruit and vegetable; MDS, multidimensional scaling; RDM, representational dissimilarity matrix. https://doi.org/10.1371/journal.pbio.2003993.g002 The semantic RDM was visually clustered into three domains: animals, fruits and vegetables, and man-made objects (tools, small non-tool objects, large non-manipulable objects; see Fig 2B & 2C). Visualization of the semantic RDM using multidimensional scaling (Fig 2C) further revealed that within each category, words with closer semantics tended to share similar function (e.g., scissors and knife), share certain distinct features, or belong to finer subordinate categories (e.g., peanut and potato). The semantic RDM and the four modality-specific attribute RDMs were intercorrelated to various degrees (Fig 2D; semantic with shape: r = 0.35; with manipulation: r = 0.47; with color: r = 0.23; with motion: r = 0.27; p < 10−9). WM neural RDMs: Lesion-naming model decoding Neural RDMs were generated for each of the 688 WM connections (S1A Fig) that were identified through deterministic tractography across 90 automated anatomical labeling (AAL) regions based on the diffusion tensor imaging (DTI) data of 48 healthy controls [27]. To generate the neural RDM for each WM connection, we performed lesion-naming model decoding using voxel-wise lesion patterns and item-level naming responses. For 80 patients with brain damage, lesion patterns in each WM connection (with each voxel in the WM connections labeled as “lesion” or “intact”) for each patient were obtained by overlapping the manually traced lesion mask (converted to the MNI) space) with the WM mask (see Fig 1). A total of 680 out of 688 WM connections with adequate lesion coverage (see Materials and methods; see also S1E Fig for the lesion distribution map) were included in the following analyses. The patients’ naming performances for each of the 100 pictures were collected (performance distribution in S1B Fig). WM neural RDMs were generated using item-based lesion-naming prediction models. For 197 connections, the lesion-naming models had successful within-item prediction averaged across all items (Bonferroni p < 0.05; diagonal in Fig 1D). That is, they yielded successful naming prediction models and were the connections that we considered in the following analyses. Of these connections, 185 were located in the left hemisphere and 12 in the right hemisphere (S1C Fig). For each of these WM connections, we computed the correspondence between the predicted scores using SVM classifiers built using the training patients’ lesion patterns and the naming scores of one item and the actual naming score of another item in the testing samples across testing iterations. This between-item correlation was taken as the similarity value for this item pair in the neural RDM, based on the assumption that if this connection pattern contains certain aspects of information shared by the naming process of these two items being captured by the SVM model, models trained with one item should also predict naming accuracy of the other item. Worth clarifying is that this procedure does not depend fully on the correlation between the actual naming accuracies across item pairs but also to what degree the potentially shared underlying properties for their naming process are supported by each WM connection (as captured by the SVM models). For example, for connections supporting phonological processing, the SVM models may pick up phonological properties and result in higher correlation between phonologically related pairs; those supporting semantic processing may pick up semantic properties and result in correlation between semantically related pairs. The resulting 100 × 100 (-item) lesion-naming prediction similarity matrix was transformed to be the neural RDM of this connection (1-prediction similarity, Fig 1D). RSA results: Semantic representation in WM connections Using RSA, the correlations between the WM neural RDMs and the semantic RDM were assessed. Significantly positive correlations were obtained in 60 WM connections (r = 0.03–0.11, false discovery rate [FDR] q < 0.05; see S1D Fig). These WM connections connected widely distributed regions across the left hemisphere, and approximately half (31/60) of the connections had at least one of the connected nodes located in the temporal lobe. The most densely connected regions (degree z-score > 1) were the middle temporal gyrus (MTG), superior temporal gyrus (STG), orbital part of middle frontal gyrus, inferior parietal lobule (IPL), and precentral gyrus. What about semantic effects that could not be explained by modality-specific attributes, peripheral factors, or broad semantic categorical effects? We controlled for the effects of all four modality-specific attributes, two peripheral variables (the early visual and phonological) and semantic category matrix (labeling within-category pairs 1 and between-category pair 0) using partial correlation. The semantic effect was consistently significant in eight WM connections (r = 0.03–0.07, FDR q < 0.05; Fig 3A–3C). Table 1 presents the detailed statistical results before and after, including these variables as covariates. These eight connections were considered to represent (relatively) higher-order semantic space. Five of them were located in the left ventral visual pathway and connected occipital regions (middle occipital gyrus, calcarine sulcus, and lingual gyrus) and temporal regions (STG, MTG, superior anterior temporal lobe [ATL], and middle ATL). The three remaining WM connections were located in the right hemisphere, connecting the postcentral gyrus with the thalamus, lingual gyrus, and parahippocampal gyrus. These reconstructed connections are shown in Fig 3B and S2 Fig. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. WM connections representing higher-order semantic space. (A) Eight WM connections representing higher-order semantic space, with 11 GM regions being connected. The regions that fail to show successful within-item prediction or in right hemisphere are rendered gray. The four colored regions represent raw semantic effects or modality-specific attributes (red for manipulation, shape, and semantic; orange for manipulation and shape; and purple for shape and color). (B) The WM connections reconstructed using the HCP database. The blue streamlines are the WM connections between two GM regions (rendered in red and green). The masks of WM connections reconstructed with current data are shown in S2 Fig. The RSA results of the eight WM connections, with bars showing the correlation strength (r value) between neural and semantic RDMs and error bars indicating ±1 standard error based on 1,000 times bootstrap resampling (see [23] for details) of the neural and behavioral RDM sets. The three WM connections did not survive all validation tests were shown in the dashed box. (C) The GM nodes representing semantic and modality-specific knowledge. The bar figure shows the RSA correlation strength (r value) of the semantic and modality-specific attributes in the colored GM regions in (A); the error bars indicate ±1 standard error; only positive values are shown. Note that for the superior ATL, in which RSA with semantic space was significant, its semantic effects diminished when controlling for modality-specific attribute RDMs. Asterisks indicate FDR q < 0.05. The object line drawings were done by first author Y.F.; the brain illustrations were generated using BrainNet Viewer [28] and DSI Studio (http://dsi-studio.labsolver.org/). The underlying data can be found at https://osf.io/h7upk/?view_only=52b8f86cffa14ed4844e4a1b9cd429cb. ATL, anterior temporal lobe; CAL, calcarine sulcus; FDR, false discovery rate; GM, gray matter; HCP, Human Connectome Project; LING, lingual gyrus; midATL, middle anterior temporal lobe; MOG, middle occipital gyrus; MTG, middle temporal gyrus; PHG, parahippocampal gyrus; PoCG, postcentral gyrus; RDM, representational dissimilarity matrix; RSA, representational similarity analysis; STG, superior temporal gyrus; supATL, superior anterior temporal lobe; WM, white matter. https://doi.org/10.1371/journal.pbio.2003993.g003 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. The RSA results of the WM connections showing significant effects of higher-order semantic space. For each connection, results (r(p)) are shown for the higher-order semantic space, raw semantic space, and broad object category, before or after controlling for various types of stimulus properties and in various subsets of patients. R values are the Spearman r between the neural RDM in the corresponding connection and the semantic RDM with various other properties controlled for. https://doi.org/10.1371/journal.pbio.2003993.t001 To examine the degree to which the semantic effects we observed on these WM connections reflect effects of broad semantic category, we also checked the RSA effect of the category matrix (correlating the neural RDM and the category RDM) and found that none of these connections had significant effects of the semantic category (p > 0.05, Table 1). Validation analyses To consolidate the main results above, we further performed validation analyses to test the following concerns: (1) The WM mask we adopted was constructed using DTI data acquired on a scanner with a low magnetic field (1.5 T) and 32 directions. Was the WM connection construction accurate and unaffected by crossing-fiber problems? (2) To maximize power, we included patients with multiple etiologies (84% stroke and 16% traumatic brain injury [TBI]) and lesion distributions (37.5% lesion in the left hemisphere only, 43.8% lesion in bilateral hemispheres, and 18.8% in the right hemisphere only). Were the results systematically affected by disease type or hemispheric differences? The quality of the WM fiber tracking. We reconstructed WM connections using diffusion data from a public state-of-the-art connectome database—the Human Connectome Project (HCP)—to ensure that the WM connections we adopted [27] were not false connections. The HCP data were acquired with a high angular resolution diffusion imaging (HARDI) sequence and therefore allowed for complex diffusion models to handle the cross-fiber issues. For the eight WM connections with higher-order semantic representation that we observed in the main results above, the WM connections reconstructed from the HCP data and our DTI data were visually very similar when projected onto the 3D brain (see S2 Fig). Controlling for the effects of patient disease type and lesion hemisphere. There was no systematic difference in the naming scores between stroke and TBI patients (t78 = −1.30, p = 0.20). We computed the neural RDMs in WM connections using only data from the 67 stroke patients, and the RSA results across the five left higher-order semantic WM connections remained highly consistent with those in the results using all patients, but the effects in right hemisphere diminished (r = −0.004–0.017, FDR q > 0.05). Using only patients with unilateral left hemispheric lesions (30 patients), we also obtained results that were generally consistent with the main results: all but one higher-order semantic WM connections retained significance (FDR q < 0.05, except for the one connecting the MTG and calcarine sulcus, r = 0.03, uncorrected p < 0.05, see Table 1). RSA results: The representation content of GM nodes that are connected by semantic WM connections What types of representations are linked by the WM connections that represent the semantic space? Do the WM connections simply relay semantic information that has already been encoded in the GM nodes, or do they contain information that cannot be accounted for by representation in the GM nodes? We tested the representational contents of the seven GM nodes that were connected by the five higher-order semantic WM connections whose effects remained robust in the validation tests (see Table 1). Four GM regions had successful within-item naming prediction and were considered in the RSA analysis: superior ATL, middle ATL, MTG, and STG. The neural RDM for each GM node was constructed using the same method as with the neural RDMs of the WM connections. We found that the higher-order semantic representation in the five semantic WM connections cannot be simply explained by GM information (Fig 3D; S1 Table): when correlating the GM neural RDMs with the semantic RDM (controlling for peripheral and categorical matrices), only the superior ATL reached significance (r = 0.04, FDR q < 0.05). However, this effect could be explained by modality-specific attribute representations. After controlling for the four modality-specific attribute matrices, none of the four GM nodes significantly correlated with the semantic RDM at either the conventional threshold (FDR q < 0.05) or a less stringent threshold (uncorrected p < 0.05, see S1 Table). Additionally, when testing the higher-order semantic representation in the five WM connections by further adding the neural RDMs of the two GM nodes being connected as additional confounding variables, the results remained unchanged (see Table 1). We further constrained our WM connection mask with a WM mask constructed by T1 segmentation (conducted using SPM8 in MNI T1 template, default parameters) to offer a clear WM boundary, i.e., containing only WM voxels. We then recomputed the higher-order semantic RSA in these WM connections using only the voxels within the WM mask and found that the effects in all five WM connections remained significant (FDR q < 0.05, r = 0.03–0.07, SD = 0.01). If not semantic, do these GM nodes code modality-specific attributes? We correlated the neural RDM of each GM node with each of the four modality-specific attribute RDMs (shape, manipulation, color, and motion; Fig 3D & S1 Table; the three control matrices—low-level visual, phonological, category—were controlled for). The superior ATL, MTG, and STG were significantly correlated with the shape and manipulation RDMs (shape: r = 0.04–0.08, manipulation: r = 0.12–0.16, FDR q < 0.05). The middle ATL was significantly correlated with the shape and color RDMs (shape: r = 0.04, color: r = 0.06, FDR q < 0.05). Finally, we conducted a whole-brain analysis across all 90 AAL GM nodes. In addition to superior ATL, the neural RDMs of the left IPL, precentral gyrus, and postcentral gyrus were significantly correlated with the semantic RDM (r = 0.04–0.05, FDR q < 0.05), but none of these or any other GM regions retained significance after controlling for the four modality-specific attribute matrices (FDR q < 0.05). Behavioral RDMs: Semantic and modality-specific attributes Behavioral RDMs for the semantic, shape, manipulation, color, and motion features of 100 objects (20 animals, 20 fruits and vegetables, 20 tools, 20 non-tool small objects, and 20 large non-manipulable objects) were generated using a multi-arrangement method [29]. In this task, 20 college students were instructed to arrange the items by a particular dimension of interest on a computer screen, and the distance among items was derived, resulting in an RDM (see Fig 2A). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. The construction and result of behavioral RDMs. (A) The multi-arrangement method. Twenty college students were asked to arrange object pictures according to their semantic (or modality-specific attribute) relatedness by dragging the items on a screen with a mouse. The distances between items on the screen would transform into an RDM. If two items, e.g., scissors and axe, showed a close distance, then they were assigned a low value in the RDM. (B) The results of the behavioral RDMs. Three broad types of distances were measured: semantic similarity, modality-specific attributes (shape, manipulation, color, and motion), and control models that are also potentially relevant to object naming (early visual, phonological, and object category matrix). The values of dissimilarity were transformed to percentile for display. Red indicates low dissimilarity (high similarity) and blue high dissimilarity (low similarity). (C) Visualization of the semantic RDM using multidimensional scaling. (D) The correlations among various behavioral RDMs. The object line drawings were done by the first author Y.F. The underlying data for this figure can be found at https://osf.io/h7upk/?view_only=52b8f86cffa14ed4844e4a1b9cd429cb. F&V, fruit and vegetable; MDS, multidimensional scaling; RDM, representational dissimilarity matrix. https://doi.org/10.1371/journal.pbio.2003993.g002 The semantic RDM was visually clustered into three domains: animals, fruits and vegetables, and man-made objects (tools, small non-tool objects, large non-manipulable objects; see Fig 2B & 2C). Visualization of the semantic RDM using multidimensional scaling (Fig 2C) further revealed that within each category, words with closer semantics tended to share similar function (e.g., scissors and knife), share certain distinct features, or belong to finer subordinate categories (e.g., peanut and potato). The semantic RDM and the four modality-specific attribute RDMs were intercorrelated to various degrees (Fig 2D; semantic with shape: r = 0.35; with manipulation: r = 0.47; with color: r = 0.23; with motion: r = 0.27; p < 10−9). WM neural RDMs: Lesion-naming model decoding Neural RDMs were generated for each of the 688 WM connections (S1A Fig) that were identified through deterministic tractography across 90 automated anatomical labeling (AAL) regions based on the diffusion tensor imaging (DTI) data of 48 healthy controls [27]. To generate the neural RDM for each WM connection, we performed lesion-naming model decoding using voxel-wise lesion patterns and item-level naming responses. For 80 patients with brain damage, lesion patterns in each WM connection (with each voxel in the WM connections labeled as “lesion” or “intact”) for each patient were obtained by overlapping the manually traced lesion mask (converted to the MNI) space) with the WM mask (see Fig 1). A total of 680 out of 688 WM connections with adequate lesion coverage (see Materials and methods; see also S1E Fig for the lesion distribution map) were included in the following analyses. The patients’ naming performances for each of the 100 pictures were collected (performance distribution in S1B Fig). WM neural RDMs were generated using item-based lesion-naming prediction models. For 197 connections, the lesion-naming models had successful within-item prediction averaged across all items (Bonferroni p < 0.05; diagonal in Fig 1D). That is, they yielded successful naming prediction models and were the connections that we considered in the following analyses. Of these connections, 185 were located in the left hemisphere and 12 in the right hemisphere (S1C Fig). For each of these WM connections, we computed the correspondence between the predicted scores using SVM classifiers built using the training patients’ lesion patterns and the naming scores of one item and the actual naming score of another item in the testing samples across testing iterations. This between-item correlation was taken as the similarity value for this item pair in the neural RDM, based on the assumption that if this connection pattern contains certain aspects of information shared by the naming process of these two items being captured by the SVM model, models trained with one item should also predict naming accuracy of the other item. Worth clarifying is that this procedure does not depend fully on the correlation between the actual naming accuracies across item pairs but also to what degree the potentially shared underlying properties for their naming process are supported by each WM connection (as captured by the SVM models). For example, for connections supporting phonological processing, the SVM models may pick up phonological properties and result in higher correlation between phonologically related pairs; those supporting semantic processing may pick up semantic properties and result in correlation between semantically related pairs. The resulting 100 × 100 (-item) lesion-naming prediction similarity matrix was transformed to be the neural RDM of this connection (1-prediction similarity, Fig 1D). RSA results: Semantic representation in WM connections Using RSA, the correlations between the WM neural RDMs and the semantic RDM were assessed. Significantly positive correlations were obtained in 60 WM connections (r = 0.03–0.11, false discovery rate [FDR] q < 0.05; see S1D Fig). These WM connections connected widely distributed regions across the left hemisphere, and approximately half (31/60) of the connections had at least one of the connected nodes located in the temporal lobe. The most densely connected regions (degree z-score > 1) were the middle temporal gyrus (MTG), superior temporal gyrus (STG), orbital part of middle frontal gyrus, inferior parietal lobule (IPL), and precentral gyrus. What about semantic effects that could not be explained by modality-specific attributes, peripheral factors, or broad semantic categorical effects? We controlled for the effects of all four modality-specific attributes, two peripheral variables (the early visual and phonological) and semantic category matrix (labeling within-category pairs 1 and between-category pair 0) using partial correlation. The semantic effect was consistently significant in eight WM connections (r = 0.03–0.07, FDR q < 0.05; Fig 3A–3C). Table 1 presents the detailed statistical results before and after, including these variables as covariates. These eight connections were considered to represent (relatively) higher-order semantic space. Five of them were located in the left ventral visual pathway and connected occipital regions (middle occipital gyrus, calcarine sulcus, and lingual gyrus) and temporal regions (STG, MTG, superior anterior temporal lobe [ATL], and middle ATL). The three remaining WM connections were located in the right hemisphere, connecting the postcentral gyrus with the thalamus, lingual gyrus, and parahippocampal gyrus. These reconstructed connections are shown in Fig 3B and S2 Fig. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. WM connections representing higher-order semantic space. (A) Eight WM connections representing higher-order semantic space, with 11 GM regions being connected. The regions that fail to show successful within-item prediction or in right hemisphere are rendered gray. The four colored regions represent raw semantic effects or modality-specific attributes (red for manipulation, shape, and semantic; orange for manipulation and shape; and purple for shape and color). (B) The WM connections reconstructed using the HCP database. The blue streamlines are the WM connections between two GM regions (rendered in red and green). The masks of WM connections reconstructed with current data are shown in S2 Fig. The RSA results of the eight WM connections, with bars showing the correlation strength (r value) between neural and semantic RDMs and error bars indicating ±1 standard error based on 1,000 times bootstrap resampling (see [23] for details) of the neural and behavioral RDM sets. The three WM connections did not survive all validation tests were shown in the dashed box. (C) The GM nodes representing semantic and modality-specific knowledge. The bar figure shows the RSA correlation strength (r value) of the semantic and modality-specific attributes in the colored GM regions in (A); the error bars indicate ±1 standard error; only positive values are shown. Note that for the superior ATL, in which RSA with semantic space was significant, its semantic effects diminished when controlling for modality-specific attribute RDMs. Asterisks indicate FDR q < 0.05. The object line drawings were done by first author Y.F.; the brain illustrations were generated using BrainNet Viewer [28] and DSI Studio (http://dsi-studio.labsolver.org/). The underlying data can be found at https://osf.io/h7upk/?view_only=52b8f86cffa14ed4844e4a1b9cd429cb. ATL, anterior temporal lobe; CAL, calcarine sulcus; FDR, false discovery rate; GM, gray matter; HCP, Human Connectome Project; LING, lingual gyrus; midATL, middle anterior temporal lobe; MOG, middle occipital gyrus; MTG, middle temporal gyrus; PHG, parahippocampal gyrus; PoCG, postcentral gyrus; RDM, representational dissimilarity matrix; RSA, representational similarity analysis; STG, superior temporal gyrus; supATL, superior anterior temporal lobe; WM, white matter. https://doi.org/10.1371/journal.pbio.2003993.g003 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. The RSA results of the WM connections showing significant effects of higher-order semantic space. For each connection, results (r(p)) are shown for the higher-order semantic space, raw semantic space, and broad object category, before or after controlling for various types of stimulus properties and in various subsets of patients. R values are the Spearman r between the neural RDM in the corresponding connection and the semantic RDM with various other properties controlled for. https://doi.org/10.1371/journal.pbio.2003993.t001 To examine the degree to which the semantic effects we observed on these WM connections reflect effects of broad semantic category, we also checked the RSA effect of the category matrix (correlating the neural RDM and the category RDM) and found that none of these connections had significant effects of the semantic category (p > 0.05, Table 1). Validation analyses To consolidate the main results above, we further performed validation analyses to test the following concerns: (1) The WM mask we adopted was constructed using DTI data acquired on a scanner with a low magnetic field (1.5 T) and 32 directions. Was the WM connection construction accurate and unaffected by crossing-fiber problems? (2) To maximize power, we included patients with multiple etiologies (84% stroke and 16% traumatic brain injury [TBI]) and lesion distributions (37.5% lesion in the left hemisphere only, 43.8% lesion in bilateral hemispheres, and 18.8% in the right hemisphere only). Were the results systematically affected by disease type or hemispheric differences? The quality of the WM fiber tracking. We reconstructed WM connections using diffusion data from a public state-of-the-art connectome database—the Human Connectome Project (HCP)—to ensure that the WM connections we adopted [27] were not false connections. The HCP data were acquired with a high angular resolution diffusion imaging (HARDI) sequence and therefore allowed for complex diffusion models to handle the cross-fiber issues. For the eight WM connections with higher-order semantic representation that we observed in the main results above, the WM connections reconstructed from the HCP data and our DTI data were visually very similar when projected onto the 3D brain (see S2 Fig). Controlling for the effects of patient disease type and lesion hemisphere. There was no systematic difference in the naming scores between stroke and TBI patients (t78 = −1.30, p = 0.20). We computed the neural RDMs in WM connections using only data from the 67 stroke patients, and the RSA results across the five left higher-order semantic WM connections remained highly consistent with those in the results using all patients, but the effects in right hemisphere diminished (r = −0.004–0.017, FDR q > 0.05). Using only patients with unilateral left hemispheric lesions (30 patients), we also obtained results that were generally consistent with the main results: all but one higher-order semantic WM connections retained significance (FDR q < 0.05, except for the one connecting the MTG and calcarine sulcus, r = 0.03, uncorrected p < 0.05, see Table 1). The quality of the WM fiber tracking. We reconstructed WM connections using diffusion data from a public state-of-the-art connectome database—the Human Connectome Project (HCP)—to ensure that the WM connections we adopted [27] were not false connections. The HCP data were acquired with a high angular resolution diffusion imaging (HARDI) sequence and therefore allowed for complex diffusion models to handle the cross-fiber issues. For the eight WM connections with higher-order semantic representation that we observed in the main results above, the WM connections reconstructed from the HCP data and our DTI data were visually very similar when projected onto the 3D brain (see S2 Fig). Controlling for the effects of patient disease type and lesion hemisphere. There was no systematic difference in the naming scores between stroke and TBI patients (t78 = −1.30, p = 0.20). We computed the neural RDMs in WM connections using only data from the 67 stroke patients, and the RSA results across the five left higher-order semantic WM connections remained highly consistent with those in the results using all patients, but the effects in right hemisphere diminished (r = −0.004–0.017, FDR q > 0.05). Using only patients with unilateral left hemispheric lesions (30 patients), we also obtained results that were generally consistent with the main results: all but one higher-order semantic WM connections retained significance (FDR q < 0.05, except for the one connecting the MTG and calcarine sulcus, r = 0.03, uncorrected p < 0.05, see Table 1). RSA results: The representation content of GM nodes that are connected by semantic WM connections What types of representations are linked by the WM connections that represent the semantic space? Do the WM connections simply relay semantic information that has already been encoded in the GM nodes, or do they contain information that cannot be accounted for by representation in the GM nodes? We tested the representational contents of the seven GM nodes that were connected by the five higher-order semantic WM connections whose effects remained robust in the validation tests (see Table 1). Four GM regions had successful within-item naming prediction and were considered in the RSA analysis: superior ATL, middle ATL, MTG, and STG. The neural RDM for each GM node was constructed using the same method as with the neural RDMs of the WM connections. We found that the higher-order semantic representation in the five semantic WM connections cannot be simply explained by GM information (Fig 3D; S1 Table): when correlating the GM neural RDMs with the semantic RDM (controlling for peripheral and categorical matrices), only the superior ATL reached significance (r = 0.04, FDR q < 0.05). However, this effect could be explained by modality-specific attribute representations. After controlling for the four modality-specific attribute matrices, none of the four GM nodes significantly correlated with the semantic RDM at either the conventional threshold (FDR q < 0.05) or a less stringent threshold (uncorrected p < 0.05, see S1 Table). Additionally, when testing the higher-order semantic representation in the five WM connections by further adding the neural RDMs of the two GM nodes being connected as additional confounding variables, the results remained unchanged (see Table 1). We further constrained our WM connection mask with a WM mask constructed by T1 segmentation (conducted using SPM8 in MNI T1 template, default parameters) to offer a clear WM boundary, i.e., containing only WM voxels. We then recomputed the higher-order semantic RSA in these WM connections using only the voxels within the WM mask and found that the effects in all five WM connections remained significant (FDR q < 0.05, r = 0.03–0.07, SD = 0.01). If not semantic, do these GM nodes code modality-specific attributes? We correlated the neural RDM of each GM node with each of the four modality-specific attribute RDMs (shape, manipulation, color, and motion; Fig 3D & S1 Table; the three control matrices—low-level visual, phonological, category—were controlled for). The superior ATL, MTG, and STG were significantly correlated with the shape and manipulation RDMs (shape: r = 0.04–0.08, manipulation: r = 0.12–0.16, FDR q < 0.05). The middle ATL was significantly correlated with the shape and color RDMs (shape: r = 0.04, color: r = 0.06, FDR q < 0.05). Finally, we conducted a whole-brain analysis across all 90 AAL GM nodes. In addition to superior ATL, the neural RDMs of the left IPL, precentral gyrus, and postcentral gyrus were significantly correlated with the semantic RDM (r = 0.04–0.05, FDR q < 0.05), but none of these or any other GM regions retained significance after controlling for the four modality-specific attribute matrices (FDR q < 0.05). Discussion To test the potential WM basis of semantic representation, we developed a structural-property-pattern-based RSA approach by applying machine learning to lesion and behavioral data in patients to derive item-based neural RDMs for WM connections. We found that a set of WM connections connecting occipital/middle temporal regions and anterior temporal regions represented a semantic space that was not explained by broad semantic categories or the effects of modality-specific attributes and, hence, was addressed as higher-order semantic representation. Such semantic effects were not fully explained by the properties of the GM nodes that were connected. Although the neural RDM of a connecting node—the superior ATL—correlated with the semantic RDM, such effect diminished after controlling for modality-specific attributes. Instead, these GM nodes tended to represent modality-specific attributes, including shape and manipulation in the superior ATL, MTG, and STG and shape and color in the middle ATL. First, it should be noted that we inferred semantic effects to be higher-order when they were not explained by linear combinations of the classical modality-specific attributes for objects. The potential effects of some untested modalities or certain nonlinear combinations across various modalities could not be fully excluded. Also, subjectively judged semantic distance might be a rather composite measure that is driven by multiple semantic dimensions, which may have different neural bases (e.g., [30]). Under the current (conventional) operation, these WM connections that represent higher-order semantics tend to lie in several major pathways that have been associated with semantic processing using univariate lesion-behavior correlation or intraoperative stimulation [12,16,21,27,31]. These connections partly belong to IFOF, and the inferior longitudinal fasciculus (ILF) (the overlapped voxels with the Johns Hopkins University WM template: IFOF [32%], ILF [71%], and minimally on the minor forceps [6%] and superior longitudinal fasciculus [8%]). Lesion or atrophy in IFOF is associated with semantic deficit severity in patients with stroke and in patients with semantic dementia [12,27,32]. A similar result was also found with ILF in semantic dementia [16,33]. Additionally, direct intraoperative stimulation of IFOF induces semantic errors [34,35]. Our current findings based on multivariate RSA demonstrate that the organization of specific connections among these large WM tract bundles represent the fine-grained semantic space. Items closer in semantic space are represented by more similar WM patterns in these specific connections. Note that it is well known that patients’ specific naming errors may vary from session to session [36]. The WM lesion pattern observed here is likely associated with some aspects of semantic space rather than with specific items. The damage of such specific aspects of semantic space would result in noisy/impaired representation for a range of items sharing that space, resulting in potentially different outputs at different time points. Such semantic space was nonetheless much finer than broad semantic categories, however, as the RSA results were robust after controlling for the categorical matrix. It is also well known that patients may make different types of errors, such as phonological and semantic paraphasias, which may be originated from different cognitive stages. Our approach here pulled all types of naming errors together, and the RSA results of correlating the neural RDM with different RDMs (semantic versus phonological/visual) presumably reflect the neural basis of different error types, which should be directly examined in future research. What is the relationship between the WM representations and the nature of the GM regions that they connect? First, we indeed observed that one of the seven linked GM regions was related to semantic space—superior ATL. The finding that lesion-pattern-behavior (neural) RDM in the superior ATL correlated with semantic space before regressing out the effects of modality-specific attributes converges nicely with the accumulated evidence about the cortical representation of semantics from fMRI and neuropsychological studies. ATL is the region with the strongest atrophy in patients with semantic dementia, which is marked by semantic deficits [6,7,31,37,38] and is sensitive to multiple modalities of object attributes [39,40]. Unlike the WM connections related to higher-order semantic space, the semantic effect in the superior ATL could be explained by the effects of modality-specific attributes. Worth noting is that ventral ATL was not scrutinized because it was not a node in the AAL parcellation we used but was included in the fusiform and inferior temporal nodes. What should be highlighted, however, is that the positive effects of higher-order semantic representation in the WM connections are significant and are not simply inheriting the properties of the connected GM nodes. Several higher-order semantic WM connections observed here connected ATL with other regions, inviting further questions about whether it is the integrity of ATL or of the ATL-related WM connections that make stronger contributions to the semantic deficits in semantic dementia patients. While our results certainly do not argue against the possibility that there are specific GM regions supporting semantic representation, we found that the GM nodes being connected by the WM connections obtained here tended to represent multiple modality-specific object properties. Of the four GM regions we could test, the MTG, STG, and superior ATL represented shape and manipulation properties, and the middle ATL represented shape and color properties. These results converge nicely with the fMRI literature studying the sensitivity of these regions for object attributes. For instance, the effects of various attributes were recently tested using parametric modulation analyses [2], which found that the posterior MTG was sensitive to both shape and manipulation knowledge. Coutanche and Thompson-Schill [39] found that the ATL codes the integration of color and shape, and Peelen and Caramazza [40] found that the ATL codes both manipulation and location. The STG was sensitive to motion properties in Fernandino et al. [2] but not in our study, perhaps due to different parcellation scales regarding the finer structure within this region. Note that many studies about the attribute-specific property representations have revealed results in sensory and motor cortices (e.g., shape in the lateral occipital/temporal cortex: [26,41]; color in the ventromedial occipital cortex such as lingual gyrus: [42–44]). However, these regions could not be tested in our data given their chance-level lesion-naming prediction performance, which could either be due to low lesion distribution in these regions (see S1E Fig for lesion distributions) or because the specific dimensions they represent are unnecessary for object picture-naming behavior. It may also be the case that higher-order semantic space is formed by binding multiple, rather than single, pairs of attributes. Consistent with this speculation, it has been shown that computation simulation models with a convergent architecture, in which intermediate units code multiple types of dimension pairings, were better at capturing the “deep” structure of conceptual space and promoting generalizations across semantically related items that were not apparently similar along single dimensions [22]. What is the mechanism of coding higher-order semantic information in WM that connects multiple modality-specific attributes? One potential mechanism could be through synchronized firing of specific sensory and motor patterns for objects. Consider when people use a pair of scissors: the neurons that represent the attributes across various modalities—e.g., shape, haptics, ways of grasping and manipulating it, seeing the consequence of using it (things being cut)—fire together. Such functional co-activation across a wide range of attributes occurs often when we see or use scissors, which enhances the structural connection between neurons within and across dimensions of the same object. WM provides a basis for such synchronization between distant cortical regions [45]. These synchronizations also lead to the building and tuning of WM connections, because neuronal activity traveling through axons can affect the properties of myelin sheaths in the active circuit; for example, electrical activity in the axon induces myelination [46,47]. This interactive process results in the WM basis of a multidimensional representation of “scissors,” which is closer in the higher-order semantic space to concepts such as “axe” or “paper.” The formation and modulation of the WM microstructure underlying these representations can be affected by our experiences, which is the basis of acquiring new concepts and of the coloring of existing concepts. Ample evidence describes how WM is affected by experience. Early-life experiential deprivation in animals and humans leads to decreased myelin sheath thickness and WM volume [48,49], whereas these parameters increase when the organism is placed in a rich experiential environment [50]. Reading training [51] and music practice [52,53] during childhood lead to increased fractional anisotropy in WM. The acquisition of motor skills changes the WM microstructure [54,55]. The exact relationship between WM microstructure and the functional coupling between cortical regions for various representational dimensions warrants further studies. A final methodological note is that the approach we developed here—building neural RDMs using machine learning with structural lesion data and condition-specific performances—could be easily adapted to other cognitive issues and all kinds of brain structural integrity measurements, including DTI indices (e.g., fractional anisotropy, mean diffusivity) or voxel-based morphometry measures for both patient and healthy populations. For the current study, we chose to focus on manually traced lesion on the T1 image (with reference to T2) because it captures the structural damage in our specific patient group (mostly stroke) in a most straightforward fashion. RSA, an approach that connects major branches of systems neuroscience—brain-activity measurement, behavioral measurement, and computational modeling [23]—could now be extended to an additional branch, i.e., brain structural measurement. In conclusion, using a structural-property-pattern-based RSA approach, we found that the WM structures mainly connecting occipital/middle temporal regions and anterior temporal regions represent fine-grained higher-order semantic information. Such semantic relatedness effects were not attributable to modality-specific attributes (shape, manipulation, color, and motion) or to the representation contents of the cortical regions that they connected and were above and beyond the broad categorical distinctions. By connecting multiple modality-specific attributes, higher-order semantic space can be formed through patterns of these connections. Materials and methods Participants Eighty patients with brain damage participated in the present study. The patient group (60 males, 20 females) was recruited from the China Rehabilitation Research Center with at least 1 month post-onset (mean = 6.09; SD = 11.69; range: 1–86 months) and premorbidly right-handed. The majority suffered from stroke (n = 67) and others suffered from TBI (n = 13). The patients’ mean age was 45 years (SD = 13; range: 19–76 years) and mean years of formal education was 13 (SD = 3; range: 2–19). Twenty additional college students (10 males; mean age = 22.9, SD = 2.45, range = 19–27) participated in the multi-arrangement experiment for the behavioral RDMs. This study was approved by the Institutional Review Board of the State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University (IORG0004944), adhering to the Declaration of Helsinki for research involving human subjects. All participants gave informed written consent. MRI data collection and preprocessing Each subject was scanned using a 1.5T GE SIGNA EXCITE scanner with an 8-channel split head coil at the China Rehabilitation Research Center. We collected two types of images: (1) high-resolution 3D T1-weighted MPRAGE images in the sagittal plane with a matrix size = 512 × 512, voxel size = 0.49 × 0.49 × 0.70 mm3, repetition time (TR) = 12.26 ms, echo time (TE) = 4.2 ms, inversion time = 400 ms, field of view (FOV) = 250 × 250 mm2, flip angle = 15°, and slice number = 248; and (2) FLAIR T2-weighted images in the axial plane with a matrix size = 512 × 512, voxel size = 0.49 × 0.49 × 5 mm3, TR = 8,002 ms, TE = 127.57 ms, inversion time = 2 s, FOV = 250 × 250 mm2, flip angle = 90°, and slice number = 28. To improve the image quality, the T1 image was scanned twice. The two scans were then co-registered and averaged for the following analyses. All imaging data can be found at the Open Science Framework database (URL: https://osf.io/h7upk/?view_only=52b8f86cffa14ed4844e4a1b9cd429cb). Materials, neuropsychological testing, and behavioral RDM construction Materials. One hundred colored photographs of objects, with an equal number of items from five semantic categories (animals, fruits and vegetables, tools, small non-tool artifacts, and large non-manipulable objects), were used in the neuropsychological testing and behavioral RDM construction. Neuropsychological testing. Patients underwent an oral picture-naming test outside the scanner. They were asked to name each object on a computer screen. The first complete response was scored. Responses were scored as 1 if correct or 0 if wrong. Behavioral RDM construction. The semantic RDM was based on a multi-arrangement method [29]. Each subject judged the semantic distance among 100 objects in the oral picture-naming task by arranging them on a computer screen. The distance between any two objects on the screen reflected their semantic distance. The subjects were instructed to “arrange objects according to how similar they are in meaning; for instance, the meaning of ‘rock–cell phone’ has little in common so they should be dragged far apart; ‘rock–sand’ has high similarity in meaning so they should be dragged close together; please consider only the aspect of ‘semantic similarity’ and disregard other aspects such as object size, color, materials, or pure associations (e.g., dog–bone).” The instruction was adapted from classical behavioral studies using semantic similarity ratings (e.g., [56]). To optimally estimate the dissimilarity matrix, all 100 images were only shown together in the first trial, and a subset was selected in every subsequent trial (see [27] for details). Modality-specific attribute (shape, manipulation, color, and motion) RDMs were based on the same arrangement method using different instructions (e.g., “Please arrange these objects according to their color/shape/manipulation/motion similarity”). Because some attributes may not be salient for some categories (e.g., it is not sensible to ask for the manipulation of a tiger or the motion of a monument), only those categories with explicit and lucid attributes were selected for a given attribute (i.e., shape for all five categories, small non-tool artifacts and tools for manipulation, fruits/vegetables and animals for color, and tools and animals for motion). All modality-specific attribute RDMs were then mapped to a complete 100 × 100 matrix by setting missing values to 1 (i.e., items without certain type of salient properties were labeled as being most dissimilar with other items on this property type; we also carried out an analysis, setting such missing values to NaN and the result pattern remained largely unchanged, see Table 1). Confounding variable RDMs were constructed based on the visual, phonological, and category properties of the items. We computed the low-level visual RDMs based on image silhouettes, because this method offers an effective prediction of the activation patterns in the early visual cortex [25]. The image pixels were binarized according to whether the pixel belonged to the object (pixel value = 1) or to the background (pixel value = 0). The dissimilarity between images was computed by 1 minus Jaccard similarity. For the phonological RDM, the dissimilarity of two-item names was measured by 1 minus the proportion of shared sub-syllabic units (onset or rhyme), regardless of position (e.g., [57]). The sub-syllabic units for a given syllable were defined based on the phonetic transcript of Chinese characters (the “pinyin” system), which transcribes each syllable with an onset consonant (“shengmu”) and a rhyme vowel or vowel-consonant (“yunmu”). The categorical RDM was constructed based on five object categories, with item pairs within the same category labeled 0 and other cells labeled 1. Neural RDM construction We used structural-property-pattern (lesion)-based RSA to investigate semantic and modality-specific attribute representation in WM connections and GM regions. Similar to the conventional RSA, which is a highly fruitful method to research the neural representation in cortical regions using functional imaging data, the structural-property-pattern (lesion)-based RSA computes the relationship between the neural RDMs and behavioral or theoretical RDMs. The main difference is that the neural RDMs in this study were constructed by machine-learning models based on performances on neuropsychological tests and patients’ brain structural lesion patterns. The main rationale for this neural similarity measure is that if a WM connection pattern contains certain aspects of information shared by the naming process of two items (e.g., some semantic features), models trained with one item should also be able to predict naming accuracy of the other item to some degree. We first extracted the lesion features, balanced item labels by bootstrapping, input the lesion features and balanced labels into SVM training and testing to obtain the neural RDM, and used permutation to estimate the significance level of the neural RDM. The full pipeline is shown in Fig 1 and the details for each of these steps are described below in turn. The scripts of the full pipeline can be found at https://osf.io/h7upk/?view_only=52b8f86cffa14ed4844e4a1b9cd429cb. Extracting the lesion features. As shown in Fig 1B, we first obtained the lesion mask (manually traced in T1 image) for each patient, then converted to MNI space, which was then overlapped with a WM connection template constructed from a healthy population [27], to extract the voxel-wise lesion pattern for each patient on each WM connection. We here focused on the structural (lesion) imaging data instead of performing analyses directly on patients’ DTI data (e.g., analyzing fractional anisotropy [FA] values or performing tractography), mainly because lesions from structural imaging (T1 and T2) are most straightforward in capturing brain structure damage properties in our specific patient type (mostly chronic stroke). For lesion identification, in each patient, a lesion mask was constructed from manually traced lesion contours on averaged T1 images slice-by-slice with reference to T2 images (see [27] for details). Lesion mapping in patients with brain damage is a challenging task and various automatic methods have been developed, with supervised or nonsupervised algorithms [58–61], but manual drawing is considered the gold standard [58,62], even in very recent works [61]. We chose this highly labor-intensive method to ensure the validity of the lesion data and have gone through several procedures to ensure the reliability (inter-rater reliability values between our two investigators and an experienced radiologist were: mean percentage volume difference, 9% ± 8% and 4% ± 3%; mean percentage of discrepant voxels, 7% ± 4% and 6% ± 2%). For WM connection, we adopted a previously reported template of the whole-brain WM network [27] to have a common reference template for the WM lesion patterns in the individual patients. Building neural RDMs in the current approach can only be done in the common template space where voxels are lined up, so that lesion patterns for different patients can be compared (i.e., for a same voxel, whether patients have lesion or not) and to be used as features for machine-learning model computation. The template we adopted was constructed using deterministic fiber tracking based on diffusion imaging data of 48 healthy participants ([27]; S1A Fig). This template contains 688 WM connections across 90 GM nodes (parcellated by the AAL atlas [63]). Briefly, the WM reconstruction was first applied in each healthy subject using determinative tracking among every two AAL regions. The resulting tracking maps in the subjects’ native space were transformed to a binary map in the MNI space. The binary maps of the MNI space for all subjects were then overlaid to generate a count map. Finally, a group-level threshold was set at voxel value (>25% of subjects; cluster size > 300 voxels) to determine whether a pair of brain regions was anatomically connected. The details of template construction can be found in [27]. Deterministic tracking was used because it has determinate termination conditions (FA values and fiber angles). It tends to suffer more false negatives but offers a clear border of WM connection to avoid invading to GM. While probabilistic tracking is generally considered more sensitive than deterministic tracking and thus revealing of more WM structures [64–66], it also increases the probability of false connections, and the biological meanings of the probabilistic values are uncertain, while it is relatively clear for the measurements used in deterministic tracking [64]. Note that the DTI imaging acquisition was suboptimal, according to standards nowadays, because of pragmatic issues in collecting patient and healthy group data using the same scanner. Nonetheless, the existence and shape of the connections showed generally good correspondence with WM networks constructed from other datasets (e.g., [67]; see also below for validation analyses). Then, the patients’ lesion mask was converted to MNI space and overlapped with the WM connection template or the GM region masks (see [27] for details). In each WM connection/GM region, intact voxels (i.e., without lesion) were labeled 0 and lesioned voxels were labeled 1. This resulted in a binary V × N matrix in which V denoted the total number of voxels in the WM connection/GM region and N the number of patients, constituting a feature set for each machine-learning model. To ensure that the WM connections/GM regions had enough subjects with lesion coverage, we only tested the WM connections and GM regions with at least five subjects having damage and with more than 20 voxels lesioned per patient (see lesion distribution at S1E Fig). A total of 680 out of 688 connections and 80 out of 90 AAL regions were included in the following analyses. Because the input feature data only contained binary values and the range was consistent with the behavioral data, no normalization was applied in the feature set. Bootstrapping. An item’s naming accuracy across all patients was not always 50%. Unbalanced training labels (e.g., the numbers of 0’s and 1’s in the training data were not equal) would ruin the classification ability because the training model always classifies the test sample into the group whose labels are predominant in number. A bootstrapping method was used to address this issue. Before classification, the subjects were reallocated into two groups: one group with correct responses and the other group with incorrect responses. We selected all subjects with the less common response of the two groups (e.g., if the accuracy of one item was 60%, all subjects with incorrect responses were selected) along with the same number of subjects randomly chosen from the other group. Thus, a new dataset for each item was constructed, with an accuracy across patients of 50%. The sample sizes of the training data for the 100 items ranged from 12 to 78 subjects (mean = 41.8 ± 8.5). This procedure was repeated 100 times for each item in each WM connection/GM region. SVM training and testing. For each WM connection or GM region, a linear SVM with default parameters [68] was used. For each item (e.g., scissors), an SVM classifier was trained based on the balanced naming labels and voxel-wise lesion patterns. The resultant classifier was used to predict the naming score (1 or 0) of all patients who were not included in the training set using their lesion patterns; for patients who were included in the training set, a leave-two-out cross-validation scheme was used. This combined procedure ended up with a predicted score for each patient (each patient was a testing case once across testing iterations). The correspondence (simple matching coefficient) between this predicted score (based on training model of one item) and the actual naming score of each of the other items (e.g., axe) was calculated and was considered the neural similarity between the training item and this other item (i.e., scissors–axe similarity) on the particular WM connection being tested. All cross-item and within-item similarity could be obtained this way, resulting in a 100 × 100 similarity matrix. We averaged the symmetrical cells in the matrix according to the principal diagonal to obtain a symmetric matrix. Each cell in the matrix was then averaged across all 100 bootstrapped samples to produce the final 100 × 100 (1-similarity) neural RDM. Significance testing (permutation and FDR). The nonparametric permutation test (10,000 times) was used to estimate the significance of the classification model for each individual edge. For each permutation, the patient labels were randomly exchanged to shuffle the relationship between behavioral data and lesion data. The averaged accuracy of the principal diagonal cells (i.e., within-item prediction accuracies) was then computed. The p-value was calculated as the fraction of accuracies from all permutations that were greater than the actual accuracy using correct labels. For each WM connection, an independent classification model was built. To control for false positives caused by comparisons across multiple edges, we applied FDR as a multi-comparison correction method. The neural RDMs of WM connections/GM regions with significant within-item prediction accuracies at the threshold of FDR q < 0.05 were considered meaningful and were used for further analyses. Representational similarity analyses: Correlating neural RDMs with behavioral RDMs The neural RDMs were correlated with behavioral RDMs using Spearman correlation. Specifically, for each WM connection, its neural RDM (a 100 [-item] × 100 [-item] matrix) and the semantic RDM (a 100 [-item] × 100 [-item] matrix) were both converted to a 1 × 4,950 vector. Correlation was computed on these two vectors (4,950 pairs of values). The r values were used to determine the extent of specific information encoded in the WM connections/GM regions. The FDR (q < 0.05) was used for multiple comparison correction. To investigate the higher-order semantic effects beyond modality-specific attributes, partial correlation analyses were performed between the semantic RDM and neural RDMs, with the modality-specific attribute RDMs (and the peripheral and categorical matrices) as nuisance variables. As explained in the “Behavioral RDM Construction” session, we adopted two ways of treating missing values in the modality-specific attributes (e.g., animal items were not rated on “manipulation” property)—setting it to be 1 (most dissimilar with other items on this modality) or to “NaN” (missing value). The RSA mapping procedure was implemented using a custom MATLAB function. Validation analyses Quality of the WM fiber tracking. We used the HCP database to check the WM template that we used in the main analyses, because HCP contains high-quality diffusion MRI data with advanced acquisition and processing methods [69]. Diffusion scans were acquired in a Siemens 3T Skyra scanner using a 2D spin-echo single-shot multiband EPI sequence with a multiband factor of three and a monopolar gradient pulse. The spatial resolution was 1.25 mm isotropic, TR = 5,500 ms, TE = 89 ms. A multishell diffusion scheme was used. The b-values were 1,000, 2,000, and 3,000 s/mm2. The total number of diffusion sampling directions was 270. We used the dataset “unrelated 40” on the ConnectomeDB website (https://db.humanconnectome.org/) for fiber reconstruction. After excluding two subjects with technical problems in acquisition, the remaining 38 subjects were included in the analyses. The preprocessing, reconstruction, and fiber tracking were performed with DSI-studio software (dsi-studio.labsolver.org). To reduce the fiber-crossing problem, we reconstructed the diffusion data using the generalized q-sampling imaging (GQI) method [70]. Controlling for the effects of patient disease type and lesion hemisphere. We computed the neural RDMs in WM connections using data from the 67 stroke patients or from the 30 patients with unilateral left hemispheric lesions. Participants Eighty patients with brain damage participated in the present study. The patient group (60 males, 20 females) was recruited from the China Rehabilitation Research Center with at least 1 month post-onset (mean = 6.09; SD = 11.69; range: 1–86 months) and premorbidly right-handed. The majority suffered from stroke (n = 67) and others suffered from TBI (n = 13). The patients’ mean age was 45 years (SD = 13; range: 19–76 years) and mean years of formal education was 13 (SD = 3; range: 2–19). Twenty additional college students (10 males; mean age = 22.9, SD = 2.45, range = 19–27) participated in the multi-arrangement experiment for the behavioral RDMs. This study was approved by the Institutional Review Board of the State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University (IORG0004944), adhering to the Declaration of Helsinki for research involving human subjects. All participants gave informed written consent. MRI data collection and preprocessing Each subject was scanned using a 1.5T GE SIGNA EXCITE scanner with an 8-channel split head coil at the China Rehabilitation Research Center. We collected two types of images: (1) high-resolution 3D T1-weighted MPRAGE images in the sagittal plane with a matrix size = 512 × 512, voxel size = 0.49 × 0.49 × 0.70 mm3, repetition time (TR) = 12.26 ms, echo time (TE) = 4.2 ms, inversion time = 400 ms, field of view (FOV) = 250 × 250 mm2, flip angle = 15°, and slice number = 248; and (2) FLAIR T2-weighted images in the axial plane with a matrix size = 512 × 512, voxel size = 0.49 × 0.49 × 5 mm3, TR = 8,002 ms, TE = 127.57 ms, inversion time = 2 s, FOV = 250 × 250 mm2, flip angle = 90°, and slice number = 28. To improve the image quality, the T1 image was scanned twice. The two scans were then co-registered and averaged for the following analyses. All imaging data can be found at the Open Science Framework database (URL: https://osf.io/h7upk/?view_only=52b8f86cffa14ed4844e4a1b9cd429cb). Materials, neuropsychological testing, and behavioral RDM construction Materials. One hundred colored photographs of objects, with an equal number of items from five semantic categories (animals, fruits and vegetables, tools, small non-tool artifacts, and large non-manipulable objects), were used in the neuropsychological testing and behavioral RDM construction. Neuropsychological testing. Patients underwent an oral picture-naming test outside the scanner. They were asked to name each object on a computer screen. The first complete response was scored. Responses were scored as 1 if correct or 0 if wrong. Behavioral RDM construction. The semantic RDM was based on a multi-arrangement method [29]. Each subject judged the semantic distance among 100 objects in the oral picture-naming task by arranging them on a computer screen. The distance between any two objects on the screen reflected their semantic distance. The subjects were instructed to “arrange objects according to how similar they are in meaning; for instance, the meaning of ‘rock–cell phone’ has little in common so they should be dragged far apart; ‘rock–sand’ has high similarity in meaning so they should be dragged close together; please consider only the aspect of ‘semantic similarity’ and disregard other aspects such as object size, color, materials, or pure associations (e.g., dog–bone).” The instruction was adapted from classical behavioral studies using semantic similarity ratings (e.g., [56]). To optimally estimate the dissimilarity matrix, all 100 images were only shown together in the first trial, and a subset was selected in every subsequent trial (see [27] for details). Modality-specific attribute (shape, manipulation, color, and motion) RDMs were based on the same arrangement method using different instructions (e.g., “Please arrange these objects according to their color/shape/manipulation/motion similarity”). Because some attributes may not be salient for some categories (e.g., it is not sensible to ask for the manipulation of a tiger or the motion of a monument), only those categories with explicit and lucid attributes were selected for a given attribute (i.e., shape for all five categories, small non-tool artifacts and tools for manipulation, fruits/vegetables and animals for color, and tools and animals for motion). All modality-specific attribute RDMs were then mapped to a complete 100 × 100 matrix by setting missing values to 1 (i.e., items without certain type of salient properties were labeled as being most dissimilar with other items on this property type; we also carried out an analysis, setting such missing values to NaN and the result pattern remained largely unchanged, see Table 1). Confounding variable RDMs were constructed based on the visual, phonological, and category properties of the items. We computed the low-level visual RDMs based on image silhouettes, because this method offers an effective prediction of the activation patterns in the early visual cortex [25]. The image pixels were binarized according to whether the pixel belonged to the object (pixel value = 1) or to the background (pixel value = 0). The dissimilarity between images was computed by 1 minus Jaccard similarity. For the phonological RDM, the dissimilarity of two-item names was measured by 1 minus the proportion of shared sub-syllabic units (onset or rhyme), regardless of position (e.g., [57]). The sub-syllabic units for a given syllable were defined based on the phonetic transcript of Chinese characters (the “pinyin” system), which transcribes each syllable with an onset consonant (“shengmu”) and a rhyme vowel or vowel-consonant (“yunmu”). The categorical RDM was constructed based on five object categories, with item pairs within the same category labeled 0 and other cells labeled 1. Materials. One hundred colored photographs of objects, with an equal number of items from five semantic categories (animals, fruits and vegetables, tools, small non-tool artifacts, and large non-manipulable objects), were used in the neuropsychological testing and behavioral RDM construction. Neuropsychological testing. Patients underwent an oral picture-naming test outside the scanner. They were asked to name each object on a computer screen. The first complete response was scored. Responses were scored as 1 if correct or 0 if wrong. Behavioral RDM construction. The semantic RDM was based on a multi-arrangement method [29]. Each subject judged the semantic distance among 100 objects in the oral picture-naming task by arranging them on a computer screen. The distance between any two objects on the screen reflected their semantic distance. The subjects were instructed to “arrange objects according to how similar they are in meaning; for instance, the meaning of ‘rock–cell phone’ has little in common so they should be dragged far apart; ‘rock–sand’ has high similarity in meaning so they should be dragged close together; please consider only the aspect of ‘semantic similarity’ and disregard other aspects such as object size, color, materials, or pure associations (e.g., dog–bone).” The instruction was adapted from classical behavioral studies using semantic similarity ratings (e.g., [56]). To optimally estimate the dissimilarity matrix, all 100 images were only shown together in the first trial, and a subset was selected in every subsequent trial (see [27] for details). Modality-specific attribute (shape, manipulation, color, and motion) RDMs were based on the same arrangement method using different instructions (e.g., “Please arrange these objects according to their color/shape/manipulation/motion similarity”). Because some attributes may not be salient for some categories (e.g., it is not sensible to ask for the manipulation of a tiger or the motion of a monument), only those categories with explicit and lucid attributes were selected for a given attribute (i.e., shape for all five categories, small non-tool artifacts and tools for manipulation, fruits/vegetables and animals for color, and tools and animals for motion). All modality-specific attribute RDMs were then mapped to a complete 100 × 100 matrix by setting missing values to 1 (i.e., items without certain type of salient properties were labeled as being most dissimilar with other items on this property type; we also carried out an analysis, setting such missing values to NaN and the result pattern remained largely unchanged, see Table 1). Confounding variable RDMs were constructed based on the visual, phonological, and category properties of the items. We computed the low-level visual RDMs based on image silhouettes, because this method offers an effective prediction of the activation patterns in the early visual cortex [25]. The image pixels were binarized according to whether the pixel belonged to the object (pixel value = 1) or to the background (pixel value = 0). The dissimilarity between images was computed by 1 minus Jaccard similarity. For the phonological RDM, the dissimilarity of two-item names was measured by 1 minus the proportion of shared sub-syllabic units (onset or rhyme), regardless of position (e.g., [57]). The sub-syllabic units for a given syllable were defined based on the phonetic transcript of Chinese characters (the “pinyin” system), which transcribes each syllable with an onset consonant (“shengmu”) and a rhyme vowel or vowel-consonant (“yunmu”). The categorical RDM was constructed based on five object categories, with item pairs within the same category labeled 0 and other cells labeled 1. Neural RDM construction We used structural-property-pattern (lesion)-based RSA to investigate semantic and modality-specific attribute representation in WM connections and GM regions. Similar to the conventional RSA, which is a highly fruitful method to research the neural representation in cortical regions using functional imaging data, the structural-property-pattern (lesion)-based RSA computes the relationship between the neural RDMs and behavioral or theoretical RDMs. The main difference is that the neural RDMs in this study were constructed by machine-learning models based on performances on neuropsychological tests and patients’ brain structural lesion patterns. The main rationale for this neural similarity measure is that if a WM connection pattern contains certain aspects of information shared by the naming process of two items (e.g., some semantic features), models trained with one item should also be able to predict naming accuracy of the other item to some degree. We first extracted the lesion features, balanced item labels by bootstrapping, input the lesion features and balanced labels into SVM training and testing to obtain the neural RDM, and used permutation to estimate the significance level of the neural RDM. The full pipeline is shown in Fig 1 and the details for each of these steps are described below in turn. The scripts of the full pipeline can be found at https://osf.io/h7upk/?view_only=52b8f86cffa14ed4844e4a1b9cd429cb. Extracting the lesion features. As shown in Fig 1B, we first obtained the lesion mask (manually traced in T1 image) for each patient, then converted to MNI space, which was then overlapped with a WM connection template constructed from a healthy population [27], to extract the voxel-wise lesion pattern for each patient on each WM connection. We here focused on the structural (lesion) imaging data instead of performing analyses directly on patients’ DTI data (e.g., analyzing fractional anisotropy [FA] values or performing tractography), mainly because lesions from structural imaging (T1 and T2) are most straightforward in capturing brain structure damage properties in our specific patient type (mostly chronic stroke). For lesion identification, in each patient, a lesion mask was constructed from manually traced lesion contours on averaged T1 images slice-by-slice with reference to T2 images (see [27] for details). Lesion mapping in patients with brain damage is a challenging task and various automatic methods have been developed, with supervised or nonsupervised algorithms [58–61], but manual drawing is considered the gold standard [58,62], even in very recent works [61]. We chose this highly labor-intensive method to ensure the validity of the lesion data and have gone through several procedures to ensure the reliability (inter-rater reliability values between our two investigators and an experienced radiologist were: mean percentage volume difference, 9% ± 8% and 4% ± 3%; mean percentage of discrepant voxels, 7% ± 4% and 6% ± 2%). For WM connection, we adopted a previously reported template of the whole-brain WM network [27] to have a common reference template for the WM lesion patterns in the individual patients. Building neural RDMs in the current approach can only be done in the common template space where voxels are lined up, so that lesion patterns for different patients can be compared (i.e., for a same voxel, whether patients have lesion or not) and to be used as features for machine-learning model computation. The template we adopted was constructed using deterministic fiber tracking based on diffusion imaging data of 48 healthy participants ([27]; S1A Fig). This template contains 688 WM connections across 90 GM nodes (parcellated by the AAL atlas [63]). Briefly, the WM reconstruction was first applied in each healthy subject using determinative tracking among every two AAL regions. The resulting tracking maps in the subjects’ native space were transformed to a binary map in the MNI space. The binary maps of the MNI space for all subjects were then overlaid to generate a count map. Finally, a group-level threshold was set at voxel value (>25% of subjects; cluster size > 300 voxels) to determine whether a pair of brain regions was anatomically connected. The details of template construction can be found in [27]. Deterministic tracking was used because it has determinate termination conditions (FA values and fiber angles). It tends to suffer more false negatives but offers a clear border of WM connection to avoid invading to GM. While probabilistic tracking is generally considered more sensitive than deterministic tracking and thus revealing of more WM structures [64–66], it also increases the probability of false connections, and the biological meanings of the probabilistic values are uncertain, while it is relatively clear for the measurements used in deterministic tracking [64]. Note that the DTI imaging acquisition was suboptimal, according to standards nowadays, because of pragmatic issues in collecting patient and healthy group data using the same scanner. Nonetheless, the existence and shape of the connections showed generally good correspondence with WM networks constructed from other datasets (e.g., [67]; see also below for validation analyses). Then, the patients’ lesion mask was converted to MNI space and overlapped with the WM connection template or the GM region masks (see [27] for details). In each WM connection/GM region, intact voxels (i.e., without lesion) were labeled 0 and lesioned voxels were labeled 1. This resulted in a binary V × N matrix in which V denoted the total number of voxels in the WM connection/GM region and N the number of patients, constituting a feature set for each machine-learning model. To ensure that the WM connections/GM regions had enough subjects with lesion coverage, we only tested the WM connections and GM regions with at least five subjects having damage and with more than 20 voxels lesioned per patient (see lesion distribution at S1E Fig). A total of 680 out of 688 connections and 80 out of 90 AAL regions were included in the following analyses. Because the input feature data only contained binary values and the range was consistent with the behavioral data, no normalization was applied in the feature set. Bootstrapping. An item’s naming accuracy across all patients was not always 50%. Unbalanced training labels (e.g., the numbers of 0’s and 1’s in the training data were not equal) would ruin the classification ability because the training model always classifies the test sample into the group whose labels are predominant in number. A bootstrapping method was used to address this issue. Before classification, the subjects were reallocated into two groups: one group with correct responses and the other group with incorrect responses. We selected all subjects with the less common response of the two groups (e.g., if the accuracy of one item was 60%, all subjects with incorrect responses were selected) along with the same number of subjects randomly chosen from the other group. Thus, a new dataset for each item was constructed, with an accuracy across patients of 50%. The sample sizes of the training data for the 100 items ranged from 12 to 78 subjects (mean = 41.8 ± 8.5). This procedure was repeated 100 times for each item in each WM connection/GM region. SVM training and testing. For each WM connection or GM region, a linear SVM with default parameters [68] was used. For each item (e.g., scissors), an SVM classifier was trained based on the balanced naming labels and voxel-wise lesion patterns. The resultant classifier was used to predict the naming score (1 or 0) of all patients who were not included in the training set using their lesion patterns; for patients who were included in the training set, a leave-two-out cross-validation scheme was used. This combined procedure ended up with a predicted score for each patient (each patient was a testing case once across testing iterations). The correspondence (simple matching coefficient) between this predicted score (based on training model of one item) and the actual naming score of each of the other items (e.g., axe) was calculated and was considered the neural similarity between the training item and this other item (i.e., scissors–axe similarity) on the particular WM connection being tested. All cross-item and within-item similarity could be obtained this way, resulting in a 100 × 100 similarity matrix. We averaged the symmetrical cells in the matrix according to the principal diagonal to obtain a symmetric matrix. Each cell in the matrix was then averaged across all 100 bootstrapped samples to produce the final 100 × 100 (1-similarity) neural RDM. Significance testing (permutation and FDR). The nonparametric permutation test (10,000 times) was used to estimate the significance of the classification model for each individual edge. For each permutation, the patient labels were randomly exchanged to shuffle the relationship between behavioral data and lesion data. The averaged accuracy of the principal diagonal cells (i.e., within-item prediction accuracies) was then computed. The p-value was calculated as the fraction of accuracies from all permutations that were greater than the actual accuracy using correct labels. For each WM connection, an independent classification model was built. To control for false positives caused by comparisons across multiple edges, we applied FDR as a multi-comparison correction method. The neural RDMs of WM connections/GM regions with significant within-item prediction accuracies at the threshold of FDR q < 0.05 were considered meaningful and were used for further analyses. Extracting the lesion features. As shown in Fig 1B, we first obtained the lesion mask (manually traced in T1 image) for each patient, then converted to MNI space, which was then overlapped with a WM connection template constructed from a healthy population [27], to extract the voxel-wise lesion pattern for each patient on each WM connection. We here focused on the structural (lesion) imaging data instead of performing analyses directly on patients’ DTI data (e.g., analyzing fractional anisotropy [FA] values or performing tractography), mainly because lesions from structural imaging (T1 and T2) are most straightforward in capturing brain structure damage properties in our specific patient type (mostly chronic stroke). For lesion identification, in each patient, a lesion mask was constructed from manually traced lesion contours on averaged T1 images slice-by-slice with reference to T2 images (see [27] for details). Lesion mapping in patients with brain damage is a challenging task and various automatic methods have been developed, with supervised or nonsupervised algorithms [58–61], but manual drawing is considered the gold standard [58,62], even in very recent works [61]. We chose this highly labor-intensive method to ensure the validity of the lesion data and have gone through several procedures to ensure the reliability (inter-rater reliability values between our two investigators and an experienced radiologist were: mean percentage volume difference, 9% ± 8% and 4% ± 3%; mean percentage of discrepant voxels, 7% ± 4% and 6% ± 2%). For WM connection, we adopted a previously reported template of the whole-brain WM network [27] to have a common reference template for the WM lesion patterns in the individual patients. Building neural RDMs in the current approach can only be done in the common template space where voxels are lined up, so that lesion patterns for different patients can be compared (i.e., for a same voxel, whether patients have lesion or not) and to be used as features for machine-learning model computation. The template we adopted was constructed using deterministic fiber tracking based on diffusion imaging data of 48 healthy participants ([27]; S1A Fig). This template contains 688 WM connections across 90 GM nodes (parcellated by the AAL atlas [63]). Briefly, the WM reconstruction was first applied in each healthy subject using determinative tracking among every two AAL regions. The resulting tracking maps in the subjects’ native space were transformed to a binary map in the MNI space. The binary maps of the MNI space for all subjects were then overlaid to generate a count map. Finally, a group-level threshold was set at voxel value (>25% of subjects; cluster size > 300 voxels) to determine whether a pair of brain regions was anatomically connected. The details of template construction can be found in [27]. Deterministic tracking was used because it has determinate termination conditions (FA values and fiber angles). It tends to suffer more false negatives but offers a clear border of WM connection to avoid invading to GM. While probabilistic tracking is generally considered more sensitive than deterministic tracking and thus revealing of more WM structures [64–66], it also increases the probability of false connections, and the biological meanings of the probabilistic values are uncertain, while it is relatively clear for the measurements used in deterministic tracking [64]. Note that the DTI imaging acquisition was suboptimal, according to standards nowadays, because of pragmatic issues in collecting patient and healthy group data using the same scanner. Nonetheless, the existence and shape of the connections showed generally good correspondence with WM networks constructed from other datasets (e.g., [67]; see also below for validation analyses). Then, the patients’ lesion mask was converted to MNI space and overlapped with the WM connection template or the GM region masks (see [27] for details). In each WM connection/GM region, intact voxels (i.e., without lesion) were labeled 0 and lesioned voxels were labeled 1. This resulted in a binary V × N matrix in which V denoted the total number of voxels in the WM connection/GM region and N the number of patients, constituting a feature set for each machine-learning model. To ensure that the WM connections/GM regions had enough subjects with lesion coverage, we only tested the WM connections and GM regions with at least five subjects having damage and with more than 20 voxels lesioned per patient (see lesion distribution at S1E Fig). A total of 680 out of 688 connections and 80 out of 90 AAL regions were included in the following analyses. Because the input feature data only contained binary values and the range was consistent with the behavioral data, no normalization was applied in the feature set. Bootstrapping. An item’s naming accuracy across all patients was not always 50%. Unbalanced training labels (e.g., the numbers of 0’s and 1’s in the training data were not equal) would ruin the classification ability because the training model always classifies the test sample into the group whose labels are predominant in number. A bootstrapping method was used to address this issue. Before classification, the subjects were reallocated into two groups: one group with correct responses and the other group with incorrect responses. We selected all subjects with the less common response of the two groups (e.g., if the accuracy of one item was 60%, all subjects with incorrect responses were selected) along with the same number of subjects randomly chosen from the other group. Thus, a new dataset for each item was constructed, with an accuracy across patients of 50%. The sample sizes of the training data for the 100 items ranged from 12 to 78 subjects (mean = 41.8 ± 8.5). This procedure was repeated 100 times for each item in each WM connection/GM region. SVM training and testing. For each WM connection or GM region, a linear SVM with default parameters [68] was used. For each item (e.g., scissors), an SVM classifier was trained based on the balanced naming labels and voxel-wise lesion patterns. The resultant classifier was used to predict the naming score (1 or 0) of all patients who were not included in the training set using their lesion patterns; for patients who were included in the training set, a leave-two-out cross-validation scheme was used. This combined procedure ended up with a predicted score for each patient (each patient was a testing case once across testing iterations). The correspondence (simple matching coefficient) between this predicted score (based on training model of one item) and the actual naming score of each of the other items (e.g., axe) was calculated and was considered the neural similarity between the training item and this other item (i.e., scissors–axe similarity) on the particular WM connection being tested. All cross-item and within-item similarity could be obtained this way, resulting in a 100 × 100 similarity matrix. We averaged the symmetrical cells in the matrix according to the principal diagonal to obtain a symmetric matrix. Each cell in the matrix was then averaged across all 100 bootstrapped samples to produce the final 100 × 100 (1-similarity) neural RDM. Significance testing (permutation and FDR). The nonparametric permutation test (10,000 times) was used to estimate the significance of the classification model for each individual edge. For each permutation, the patient labels were randomly exchanged to shuffle the relationship between behavioral data and lesion data. The averaged accuracy of the principal diagonal cells (i.e., within-item prediction accuracies) was then computed. The p-value was calculated as the fraction of accuracies from all permutations that were greater than the actual accuracy using correct labels. For each WM connection, an independent classification model was built. To control for false positives caused by comparisons across multiple edges, we applied FDR as a multi-comparison correction method. The neural RDMs of WM connections/GM regions with significant within-item prediction accuracies at the threshold of FDR q < 0.05 were considered meaningful and were used for further analyses. Representational similarity analyses: Correlating neural RDMs with behavioral RDMs The neural RDMs were correlated with behavioral RDMs using Spearman correlation. Specifically, for each WM connection, its neural RDM (a 100 [-item] × 100 [-item] matrix) and the semantic RDM (a 100 [-item] × 100 [-item] matrix) were both converted to a 1 × 4,950 vector. Correlation was computed on these two vectors (4,950 pairs of values). The r values were used to determine the extent of specific information encoded in the WM connections/GM regions. The FDR (q < 0.05) was used for multiple comparison correction. To investigate the higher-order semantic effects beyond modality-specific attributes, partial correlation analyses were performed between the semantic RDM and neural RDMs, with the modality-specific attribute RDMs (and the peripheral and categorical matrices) as nuisance variables. As explained in the “Behavioral RDM Construction” session, we adopted two ways of treating missing values in the modality-specific attributes (e.g., animal items were not rated on “manipulation” property)—setting it to be 1 (most dissimilar with other items on this modality) or to “NaN” (missing value). The RSA mapping procedure was implemented using a custom MATLAB function. Validation analyses Quality of the WM fiber tracking. We used the HCP database to check the WM template that we used in the main analyses, because HCP contains high-quality diffusion MRI data with advanced acquisition and processing methods [69]. Diffusion scans were acquired in a Siemens 3T Skyra scanner using a 2D spin-echo single-shot multiband EPI sequence with a multiband factor of three and a monopolar gradient pulse. The spatial resolution was 1.25 mm isotropic, TR = 5,500 ms, TE = 89 ms. A multishell diffusion scheme was used. The b-values were 1,000, 2,000, and 3,000 s/mm2. The total number of diffusion sampling directions was 270. We used the dataset “unrelated 40” on the ConnectomeDB website (https://db.humanconnectome.org/) for fiber reconstruction. After excluding two subjects with technical problems in acquisition, the remaining 38 subjects were included in the analyses. The preprocessing, reconstruction, and fiber tracking were performed with DSI-studio software (dsi-studio.labsolver.org). To reduce the fiber-crossing problem, we reconstructed the diffusion data using the generalized q-sampling imaging (GQI) method [70]. Controlling for the effects of patient disease type and lesion hemisphere. We computed the neural RDMs in WM connections using data from the 67 stroke patients or from the 30 patients with unilateral left hemispheric lesions. Quality of the WM fiber tracking. We used the HCP database to check the WM template that we used in the main analyses, because HCP contains high-quality diffusion MRI data with advanced acquisition and processing methods [69]. Diffusion scans were acquired in a Siemens 3T Skyra scanner using a 2D spin-echo single-shot multiband EPI sequence with a multiband factor of three and a monopolar gradient pulse. The spatial resolution was 1.25 mm isotropic, TR = 5,500 ms, TE = 89 ms. A multishell diffusion scheme was used. The b-values were 1,000, 2,000, and 3,000 s/mm2. The total number of diffusion sampling directions was 270. We used the dataset “unrelated 40” on the ConnectomeDB website (https://db.humanconnectome.org/) for fiber reconstruction. After excluding two subjects with technical problems in acquisition, the remaining 38 subjects were included in the analyses. The preprocessing, reconstruction, and fiber tracking were performed with DSI-studio software (dsi-studio.labsolver.org). To reduce the fiber-crossing problem, we reconstructed the diffusion data using the generalized q-sampling imaging (GQI) method [70]. Controlling for the effects of patient disease type and lesion hemisphere. We computed the neural RDMs in WM connections using data from the 67 stroke patients or from the 30 patients with unilateral left hemispheric lesions. Supporting information S1 Fig. The WM template and the results of lesion-naming predictions. (A) The WM template used in the current study was adopted from Fang et al. (2015), in which deterministic tractography was performed across 90 AAL regions using the DTI data of 48 healthy adults acquired in the same scanner as our patient imaging data. The resulting whole-brain anatomical network contained 688 WM connections. (B) Patients’ naming performance distribution for the 100 objects. (C) The 197 WM connections in which the lesion pattern predicted the naming performance for the same items with greater-than-chance accuracy in the SVM model. The RSA analyses were conducted on these connections. (D) The 60 WM connections in which the neural RDMs significantly positively correlated with the semantic RDM before controlling for the attribute RDMs (FDR q < 0.05). (E) Patient’s lesion distribution in the WM connections and GM nodes. The N value of each WM connection and each GM node was denoted by the number of patients with lesion in more than 20 voxels. The brain figures were generated using Brainnet Viewer (Xia et al. 2013). AAL, automated anatomical labeling; DTI, diffusion tensor imaging; FDR, false discovery rate; GM, gray matter; RDM, representational dissimilarity matrix; RSA, representational similarity analysis; SVM, support vector machine; WM, white matter. https://doi.org/10.1371/journal.pbio.2003993.s001 (DOCX) S2 Fig. Reconstruction of the eight WM connections that represent higher-order semantic space. The masks of the WM connections that were used in our main analyses (adopted from Fang et al. 2015) are shown in the two left columns, and the WM connections reconstructed using Human Connectome Project data are shown in the two right columns. The brain figures were generated using BrainNet Viewer (Xia et al. 2013). WM, white matter. https://doi.org/10.1371/journal.pbio.2003993.s002 (DOCX) S1 Table. The RSA results in the GM nodes that are connected by the WM connections that showed robust higher-order semantic effect. *Positive correlation values that survived FDR correction (q < 0.05). The missing values were set as “1” (most dissimilar) in modality-specific attributes matrix. #Low-level visual, phonological, category. FDR, false discovery rate; GM, gray matter; MTG, middle temporal gyrus; MidATL, middle anterior temporal lobe; RSA, representational similarity analysis; SupATL, superior anterior temporal lobe; STG, superior temporal gyrus; WM, white matter. https://doi.org/10.1371/journal.pbio.2003993.s003 (DOCX) S2 Table. Background information of the 80 patients. https://doi.org/10.1371/journal.pbio.2003993.s004 (DOCX)

doi: 10.1371/journal.pbio.2004356pmid: 29708964

Introduction The rise of antibiotic resistance is reducing the arsenal of available drugs to treat bacterial infections [1–3]. Some infections are already nearly untreatable because the infecting pathogens are resistant to virtually all available drugs [4,5]. The identification and establishment of new antibiotics has become a major focus of national and international health programs, and substantial investments have been directed towards drug discovery, for example, by the United States and the European Union [6–10]. Yet even if these attempts succeeded and dozens of novel compounds became available tomorrow, the antibiotic crisis would not subside. The evolution of resistance is inevitable, and new drugs will be incapacitated within short time periods [2,3]. So how can we hamper this evolutionary march towards resistance? To some extent, we cannot escape the open-ended arms race between compound discovery and resistance evolution. Nevertheless, we may still use evolutionary thinking to enhance treatment efficacy and sustainability [11]. Combination therapy, the simultaneous deployment of 2 or more drugs, is commonly proposed [12]. Indeed, WHO has endorsed it as the first-line strategy to treat diseases such as tuberculosis, malaria, or HIV [13–15]. However, the nature of the drug combination is crucial for treatment success because initially effective combinations may maximize selection for antibiotic resistance [16,17]. The approach of experimental evolution has proven highly informative on exploring the dynamics that shape the emergence and spread of drug resistance [11,18]. Using this approach, drug pairs were previously suggested to be most effective at limiting bacterial adaptation if (i) antimicrobials display collateral sensitivity, such that bacteria that evolve resistance to one of the compounds immediately suffer exacerbated suppression by the other [19–22], or (ii) antibiotics interact antagonistically, such that they inhibit each other’s effect [16,23,24]. A mathematical model indicated that the latter empirical findings may not be generally applicable but depend on the exact conditions during evolution [25]. In particular, synergistic drug pairs generally favor bacterial clearance but only sometimes low adaptation rates. The strong reduction in population size by synergistic drugs decreases the likelihood of resistance mutations emerging and increases the chances of population extinction. However, these effects only correlate with low adaptation rates when resource competition is weak. When resource competition is high, resistance mutations have a strong selective advantage and may spread rapidly through the population due to competitive release. Under these conditions, antagonistic rather than synergistic drugs are most efficient in reducing adaptation rates [25]. To date, few experimental data are available to explore these particular model predictions—and, moreover, test the role of evolutionary trade-offs, such as the evolved collateral sensitivities—on bacterial adaptation in multidrug environments. In the current study, we performed a systematic analysis using an experimental evolution approach and the gram-negative opportunistic human pathogen Pseudomonas aeruginosa as a model. We evaluated 38 drug pairs for their ability to effectively constrain bacterial adaptation in multidrug environments and calculated 2 antibiotic combination efficacy (ACE) networks based on either the rate of adaptation or bacterial clearance (i.e., frequency of population extinction). These measures provide complementary information on treatment efficacy. First, population extinction represents the ultimate aim of any antibiotic intervention; its frequency is a highly informative indicator of treatment efficacy under our specific experimental conditions, in which antibiotics are always applied at sublethal doses. Second, for the surviving populations, we further evaluated increases in growth rates as a measure of the bacteria’s adaptive potential in antibiotic environments [16]. We subsequently employed complementary statistical approaches, including an integrative Bayesian network (BN) analysis, to disentangle the relative impacts of drug interaction type and evolved collateral effects between individual drugs on the characteristics of the inferred ACE networks. For selected drug pairs, we additionally explored to what extent adaptation to the combinations is driven by the single-component drugs or by initial drug inhibitory levels. Results Most tested antibiotics interact synergistically in P. aeruginosa Antibiotic interactions are defined as synergistic, additive, or antagonistic when the drug pair has a stronger, equivalent, or weaker inhibitory effect on bacterial growth than the corresponding single drugs (i.e., monotherapies), respectively. Here, we determined this interaction quantitatively using an estimator denoted α [17]. This estimator is obtained from a quadratic regression applied to growth measurements as a function of different drug proportions of 2 drugs. The concentration of each of the single drugs is chosen to fall onto the line of equal dose, in our case defined to inhibit 75% of growth (i.e., inhibitory concentration [IC] 75; Fig 1A, S1 Fig and Table 1). The estimator α describes the shape of the resulting response in growth whereby positive values indicate synergism and negative values antagonism (Fig 1B). This approach has two advantages: first, it provides a statistical framework for testing the significance of positive or negative α; and second, its inference is less laborious than alternative procedures, thus facilitating characterization of a larger number of drug interactions. Even though the approach was carefully evaluated previously [17], we specifically validated its suitability for our model system. We compared the inferred α values for 8 selected combinations (S2A Fig and S3 Fig) to the corresponding results obtained with one of the commonly used alternative methods, based on Bliss independence and the checkerboard approach (S1 Data for a key to all datasets and S2 Data), as previously described for Escherichia coli [16,26]. This comparison demonstrated that α correlates significantly with the degree of synergy (S), irrespective of whether S is calculated from the average of all viable concentrations across a grid defined by the 2 drugs (ABij = Ai + Bj; S2B Fig) or from combinations for which the 2 individual drugs had the same level of inhibition (ABij for which IC50[Ai] = IC50[Bj], S2C Fig). We thus conclude that the α estimator provides an informative, quantitative indicator of a 2-drug interaction. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Drug interaction network for P. aeruginosa. (A) Schematic representation, adapted from [17], of the principle underlying the drug proportion parameter θ (line of equal dose; dashed lines), which is subsequently used to determine drug interactions, in comparison to different shapes of isobolograms (solid lines), as observed in synergistic (in red; top panel) or antagonistic (in blue; bottom panel) interactions. (B) Schematic illustration of the different interaction types as a function of the drug proportion parameter θ, ranging from synergism to antagonism. Drugs are combined in 9 different proportions (n = 9 for each combination), with each drug alone set to inhibit 75% of growth (S1 Fig). After a fixed time (12 h), bacterial growth is measured, and a quadratic model is used to fit the observed data. The α test [17] was used to determine significance of synergism or antagonism (S1 Table). (C) The α parameter was inferred from measured data to reconstruct a drug interaction network including 52 different antibiotic combinations. Combinations were formed from 12 different drugs, here represented as the nodes of the network, spanning 5 different antibiotic classes (see outer ring). The drug interaction profile is shown through the links (lines) formed between the nodes, and its strength is highlighted by the thickness of the lines and color. Red, black, and blue lines correspond to synergistic, additive, or antagonistic interactions, respectively (see also S3 Fig). The data for this panel are provided in S3 Data. AZL, azlocillin; CAR, carbenicillin; CEF, cefsulodin; CEZ, ceftazidime; CIP, ciprofloxacin; DOR, doripenem; GEN, gentamicin; IC75, concentration inhibiting 75% of bacterial growth; IMI, imipenem; PIT, piperacillin + tazobactam; STR, streptomycin; TIC, ticarcillin; TOB, tobramycin. https://doi.org/10.1371/journal.pbio.2004356.g001 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. List of antibiotics used in this study. https://doi.org/10.1371/journal.pbio.2004356.t001 We subsequently evaluated the interactions among 12 different antibiotics representing 5 classes (Table 1). We chose these drugs as representatives of the main classes of antibiotics, which are commonly used in combination to treat P. aeruginosa and to which most clinical P. aeruginosa strains are still susceptible [27–29]. Even though this choice could have introduced a bias in the overall pattern of inferred interaction types, these should nevertheless be representative of the clinically applied drug combinations. We characterized drug interactions for almost all of the possible combinations, resulting in a total of 52 measures that we summarized in an interaction network (Fig 1C, S3 Fig, S1 Table, S3 Data). Overall, synergistic combinations were more common than other interaction types (synergistic = 24/52; additive = 14/52; and antagonistic = 14/52). Combinations between cell wall inhibitors (β-lactams) and aminoglycosides most often produced synergisms, whereas those including ciprofloxacin (CIP) had exclusively antagonistic effects (Fig 1C). ACE networks demonstrate substantial variation in the effect of combinations on adaptation rates and population extinction We used evolution experiments to assess ACE, which is the ability of drug combinations to constrain bacterial adaptation either through population extinction or, in the case of surviving populations, reduced adaptation rates. Based on the inferred drug interactions and the previously obtained frequencies of collateral sensitivity between 8 of the considered antibiotics (Fig 2) [30], we selected 38 drug pairs covering all different types of drug interactions and collateral effects. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Collateral sensitivity network. The FCRs among 8 of the 12 drugs used in this study were obtained from our previous work [30]. FCR ranges from 0 to 1, such that 0 indicates that all populations (12–20 populations per combination) were sensitive to the corresponding other drug, thus having complete reciprocal sensitivity, whereas 1 highlights that none of the populations with resistance to one of the antibiotics in a pair suffered exacerbated sensitivity against the other. For the graphical illustration, we divided the combinations into 4 groups: complete collateral sensitivity (FCR ≤ 0.25; dark purple lines), partial collateral sensitivity (0.25 < FCR ≤ 0.5; light dashed pink lines), partial cross-resistance (0.5 < FCR < 0.75; light green dashed lines), and complete cross-resistance (FCR ≥ 0.75; dark green lines). CAR, carbenicillin; CEF, cefsulodin; CIP, ciprofloxacin; DOR, doripenem; FCR, frequency of collateral resistances; GEN, gentamicin; IMI, imipenem; PIT, piperacillin + tazobactam; STR, streptomycin. https://doi.org/10.1371/journal.pbio.2004356.g002 Based on this choice of drugs, we evolved a total of 1,672 populations through serial transfers into fresh media containing the respective antibiotics using a transfer period of 12 h and a total of 10 transfers (total duration of 120 h; Fig 3A and S4 Data). We assessed bacterial adaptive potential by integrating quantitative growth measurements taken in 15-min intervals from each evolving population (a total of 783,464 measurements for all treatments and populations; for a validation of our optical density (OD) measures as a proxy for bacterial growth, see Materials and methods and S4 Fig). For each population in a growth season, we then calculated the growth rate r during the exponential phase (Fig 3B). Following previous work [16], we defined the rate of adaptation as the change in growth rate over time for each evolving population (Fig 3C; for a validation of using growth characteristics as a proxy of evolutionary adaptation, see Materials and methods and S5 Fig). For subsequent analysis, we focused on the results of the 50:50 drug proportion (S6 Fig) and the single-drug treatments (S7 Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Experimental design and inference of adaptation rates. (A) Schematic representation of the evolution experiment with antibiotic combinations. Thirty-eight combinations were serially transferred every 12 h (season) into fresh medium containing antibiotics mixed in 5 different proportions (n = 8 per proportion and drug combination). An uninhibited control was also included, replicated 4 times, resulting in a total of 44 populations per combination and 1,672 for all combinations. Single-drug treatments of any drug A and B aimed at inhibiting 75% of growth relative to a drug-free environment (i.e., IC75). (B) An example of the quantitative growth measures obtained for a particular combination (CIP plus GEN) and the various drug proportions. Each panel shows 1 out of 5 seasons of growth (measured with OD as a proxy ± SD) over a 12-h period. Vertical grey lines denote the time window from which the slope was calculated to infer the growth rate r of each evolving population during exponential growth. All the drug proportions considered are highlighted in different colors (yellow to red), as well as the no-drug control (black). (C) Six exemplary populations from the CIP plus GEN combination experiments illustrating the change in growth rate r over 10 seasons of growth for each of the drug proportions. The rate of adaptation was calculated following previous work [16], and as indicated on the left of panel C, tadapt is defined as the time required to reach half of the change in growth rate, Δr. The data for this figure are provided in S4 Data. CFU, colony-forming unit; CIP, ciprofloxacin; GEN, gentamicin; OD, optical density; IC75, concentration inhibiting 75% of bacterial growth. https://doi.org/10.1371/journal.pbio.2004356.g003 We reconstructed the 2 ACE networks based either on adaptation rates of the surviving populations (Fig 4A) or on population extinctions (Fig 4B). Below, we first describe the patterns seen in the ACE networks, while their statistical analysis is explained in the next section. In all cases but one (for carbenicillin [CAR] plus gentamicin [GEN], all populations went extinct), adaptation to the combination treatment was possible. However, the rates of adaptation varied substantially across the different drug combinations, with lower rates of adaptation (below the 50th quantile) predominantly, but not exclusively, seen among antagonistic combinations that included CIP (Fig 4A; S8 Fig and S9 Fig show separate ACE networks for each drug interaction type and the 2 types of evolved collateral effects, respectively). Several synergistic drug pairs, combining an aminoglycoside with either a penicillin or carbapenem, led to similarly low rates of adaptation (below the 50th quantile, S8 Fig). Moreover, almost all cases of collateral sensitivity included in this study were associated with reduced adaptation rates (S9 Fig). This was not the case for combinations with cross-resistance. Furthermore, when estimating clearance efficacy, we found that extinctions almost exclusively occurred with the synergistic combinations (Fig 4B, S8 Fig). The synergistic combinations that did select for lower rates of adaptation did not necessarily have higher rates of extinction and vice versa (populations surviving synergistic combinations were not necessarily adapting more slowly; see azlocillin [AZL] plus streptomycin [STR], cefsulodin [CEF] plus CAR, or ticarcillin [TIC] plus GEN; S8 Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. The ACE networks. (A) ACE network built from the rates of adaptation of surviving populations in the combination environment. The color and thickness of the lines (links) formed between the drugs (nodes) reflect the quantiles within which the inferred adaptation rates are found relative to the entire distribution: orange thick lines denote the combinations with the slowest adaptation rates (one of the aims of treatment efficacy), and grey thin lines highlight those with fast adaptation. (B) ACE network on the number of extinction events observed in the combination treatments. Thickness and color of the links represent the number of extinct populations, ranging from 0 (grey) to 8 (dark orange). Adaptation rates and extinction frequencies are inferred from the growth characteristics provided in S4 Data. ACE, antibiotic combination efficacy; AZL, azlocillin; CAR, carbenicillin; CEF, cefsulodin; CEZ, ceftazidime; CIP, ciprofloxacin; DOR, doripenem; GEN, gentamicin; IMI, imipenem; PIT, piperacillin + tazobactam; STR, streptomycin; TIC, ticarcillin; TOB, tobramycin. https://doi.org/10.1371/journal.pbio.2004356.g004 Statistical ACE network analysis reveals complementary roles for synergism and collateral sensitivity in treatment efficacy We next performed 2 types of statistical analyses to assess to what extent the overall characteristics of the 2 ACE networks are determined by the 2 considered predictors of combination efficacy: interaction type inferred from α (Fig 1C) and collateral sensitivity profiles previously obtained from experimentally evolved resistant populations of P. aeruginosa (Fig 2, [30]). We first used a BN approach to assess the relationships among the considered variables (i.e., adaptation rate, extinction frequency, drug interaction, and frequency of collateral resistances [FCR]). The BN approach is based on a constraint-based interleaved incremental association algorithm [31–33] to dissect the relationships between our variables (see Materials and methods for details). The results are summarized in the BN (Fig 5A), in which nodes represent the different variables and arrows indicate the inferred dependencies. The BN analysis revealed that the type of antibiotic interaction strongly influenced the proportion of extinction, but not the rate of adaptation. Instead, the rate of adaptation was found to depend solely on the frequency of collateral sensitivities. No other dependency was inferred by the analysis. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. BN analysis of antibiotic resistance evolution under combination therapy. (A) BN obtained from a constraint-based interleaved incremental association algorithm including 4 different random variables: drug interaction types, FCR, proportion of extinctions, and rate of adaptation. (B) Based on the BN, we calculated the conditional probabilities of rate of adaptation for different types of collateral effects (top panel) and extinction frequencies for different antibiotic interaction characteristics (bottom panel). The Bayesian analysis is based on data for drug interaction characteristics (S2 Data), collateral effects [30], and extinction frequencies, and adaptation rates are inferred from growth characteristics during experimental evolution (S4 Data). Adap., rate of adaptation; BN, Bayesian network; Ext., proportion of extinctions; FCR, frequency of collateral resistances; Int., drug interaction types. https://doi.org/10.1371/journal.pbio.2004356.g005 Based on the BN structure, we calculated the conditional probabilities for the inferred dependencies between the frequencies of collateral sensitivity and the rates of adaptation as well as for the proportion of extinction and drug interaction type. In particular, we used the different types of evolved collateral effects (i.e., partial collateral sensitivity, partial cross-resistance, and cross-resistance; none of the combinations evaluated during evolution had complete collateral sensitivity between their components, as shown in Fig 2) and calculated the conditional probability of obtaining the distribution of observed adaptation rates across 5 equal quantile bins (Fig 5B, top panel). Similarly, given the different drug interaction types (synergism, additivity, and antagonism), we calculated the conditional probabilities of different extinction frequencies across 5 equal quantile bins (Fig 5B, bottom panel). These 2 additional analyses describe more clearly the inferred dependencies within the BN. Antibiotic combinations for which at least half of the populations had collateral sensitivity against one or both of the individual drug components (i.e., partial collateral sensitivity; purple bars in Fig 5B, top panel) have a higher probability of selecting for low but not high rates of adaptation. Conversely, combinations with partial or complete cross-resistance (green bars in Fig 5B, top panel) have a higher probability of producing the top scores of inferred adaptation rates. In addition, high probabilities of extinction are associated with synergistic and additive combinations, whereas the reverse is found for antagonistic drug pairs (Fig 5B, bottom panel). We further validated the inferred dependencies between variables using partial correlation analysis, following the approach previously established for a similar analysis of combination efficacy in E. coli [16]. This approach allowed us to control for drug pair membership using the average rate of adaptation towards the corresponding single drugs of a particular combination as a covariate (Materials and methods). Statistical significance was subsequently inferred using a permutation test [16]. This analysis revealed a significant correlation between the FCR and the rate of adaptation (ρs = 0.52, P = 0.038) and between the proportion of extinction and the drug interaction type α (ρs = 0.51, P = 0.043), but not between the FCR and the proportion of extinction (ρs = 0.39, P = 0.146) or the drug interaction α and the rate of adaptation (ρs = 0.3, P = 0.262). This analysis, based on a distinct statistical approach, thereby corroborated the findings of the BN analysis. We conclude that synergistic drug interactions enhance bacterial clearance, whereas collateral sensitivity limits the adaptive potential of the bacteria. Adaptation to the strongest component influences adaptation to multidrug environments We next assessed whether the ability of bacteria to adapt to the combination is mainly driven by adaptation to only one of the drugs rather than dependent on a unique property of the antibiotic pair. For our dataset, we related the inferred rates of adaptation in the combination treatments to those inferred for the corresponding single-drug environments (S10 Fig). We first compared the 2 corresponding monotherapies of a given drug pair and defined the drug leading to lower rates of adaptation as the stronger component (i.e., higher ability to minimize resistance evolution) and the other as the weaker component (i.e., lower ability to minimize resistance evolution). Thereafter, we calculated the relative rate of adaptation of the combination by standardizing it against either the stronger or the weaker component of the pair. The resulting ACE networks are shown in Fig 6A and 6B, respectively. Interestingly, the original ACE network for adaptation rates (Fig 4A) is more similar to that standardized by the weaker but not the stronger component drug (Fig 6; S2 Table). This suggests that the characteristics of the original ACE network (Fig 4A), and thus the efficacy of drug combinations to reduce adaptation rates, is primarily driven by adaptation to the stronger component, which—if accounted for by the standardizing scheme—removes important properties of the network (see as prominent examples the disappearance of the strong reduction in adaptation rate for doripenem [DOR] plus TIC, or DOR plus PIT [piperacillin + tazobactam]; Fig 4A and Fig 6A). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Weighted ACE networks and their Bayesian analysis. We assessed to what extent adaptation to one of the drugs of a pair determined the overall rate of adaptation to the combination treatment. The stronger component drug of each pair was identified as the one with lower adaptation rates in monotherapy. We subsequently standardized the adaptation rates towards the combination by those towards either (A) the stronger or (B) the weaker component drug, resulting in 2 weighted ACE networks. Orange thick lines indicate slower adaptation, while grey thin bands denote fast adaptation. (C) Results of the BN analysis on the original network versus the 2 standardized networks. The relationship between drug interaction type and extinction frequency was stable across all analyses, while the dependence of adaptation rate on evolved collateral effects disappeared when adaptation rates were standardized by the stronger component. Adaptation rates are inferred from the data on growth characteristics during experimental evolution, provided in S4 Data. ACE, antibiotic combination efficacy; AZL, azlocillin; BN, Bayesian network; CAR, carbenicillin; CEF, cefsulodin; CEZ, ceftazidime; CIP, ciprofloxacin; DOR, doripenem; GEN, gentamicin; IMI, imipenem; PIT, piperacillin + tazobactam; STR, streptomycin; TIC, ticarcillin; TOB, tobramycin. https://doi.org/10.1371/journal.pbio.2004356.g006 We further evaluated influence of the component drugs by repetition of the BN analysis. We found that the dependency observed between the FCR and the rates of adaptation of the combinations disappeared when the latter is weighted by the stronger but not the weaker component drug (Fig 6C). At the same time, the dependency between drug interaction and extinction frequency remained, while no additional relationship was revealed. Similar results were obtained when we repeated the correlation analysis with standardized adaptation rates. The originally identified correlation between the FCR and the rate of adaptation was no longer significant when the latter was standardized by adaptation to the stronger component drug (ρs = 0.33, P = 0.21), yet it still showed a statistical trend when we standardized by the weaker component drug (ρs = 0.45, P = 0.078). In these 2 analyses, drug interaction did not correlate significantly with the weighted adaptation rates (ρs < 0.47, P > 0.09). These results consistently indicate that adaptation to the stronger component drug influences adaptation to the combination and that this is dependent on the evolved collateral effects. Initial inhibition levels correlate with adaptation rates, while extinction events are almost exclusively restricted to synergistic combinations We next performed a separate evolution experiment with 4 selected combinations to assess to what extent the inherently different starting levels of inhibition—imposed by each type of interaction during the first season of growth (Fig 1B and S3 Fig)—influenced both the number of extinctions and adaptation rates. We performed this evolution experiment with 4 selected combinations with different interaction profiles: 2 interacting synergistically (GEN plus CAR and STR plus PIT) and 2 antagonistically (GEN plus CIP and Tobramycin [TOB] plus CIP). For these combinations, we varied the initial inhibition level of the combination across 8 steps, ranging from IC50 to >IC90. Populations were serially transferred into fresh media as explained before (S5 Data; and for the obtained changes in growth rate r, see S11 Fig). This separate evolution experiment revealed that initial inhibitory levels of the tested combinations are significantly related to the rates of adaptation, irrespective of combination identity or drug interaction type (GLM, F1,336 = 37.735, P < 0.001; Fig 7A and S3 Table). In particular, increasing levels of inhibition are generally associated with higher rates of adaptation, suggesting that strong inhibition increases selection for an adaptive response [34,35]. At higher levels of inhibition, the synergistic and antagonistic combinations produce clearly distinct responses, especially regarding population extinction. Here, the 2 synergistic pairs are associated with a significant increase in the number of extinct populations (logistic regression, F12,336 = 21.15, P < 0.001; Fig 7B and S4 Table), while antagonistic combinations produced almost no extinction at all. Moreover, at the very high initial inhibitory levels, antagonistic pairs showed a sudden drop in adaptation rates (Fig 7A), as expected from previous work [16,24]. A similarly strong reduction is not observed for the synergistic combinations, possibly owing to the fact that only few populations survived and could be used to infer adaptation rates. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Influence of the initial inhibition level on adaptation rate and population extinction. In a separate round of evolution experiments, we evaluated the consequences of the initial inhibition levels in synergistic and antagonistic combinations. The experiment was performed following the protocol of the main evolution experiment (Fig 3A), with the exception that the starting doses of the combinations were fixed at different levels of inhibition for the combinations. (A) Inferred rates of adaptation as a function of initial inhibitory levels. We determined initial inhibition by measuring in the first season of the evolution experiment the AUC of growth over time (measured as OD as a proxy) for each replicate population and then standardized it against the average AUC of the no-drug control (x-axis, AUCi). The rate of adaptation (y-axis) was inferred as for the main evolution experiment (Fig 3). (B) Extinction was significantly more often observed in synergistic rather than antagonistic combinations, even at the same level of inhibition (results for logistic regression analysis in S4 Table). Adaptation rates and extinction frequencies are inferred from growth characteristics during experimental evolution, provided in S5 Data. AUC, area under the curve; AUCi, area under the curve of relative inhibition of growth; CAR, carbenicillin; CIP, ciprofloxacin; GEN, gentamicin; OD, optical density; PIT, piperacillin + tazobactam; STR, streptomycin; TOB, tobramycin. https://doi.org/10.1371/journal.pbio.2004356.g007 Taken together, the results from this separate evolution experiment suggest that the generally higher inhibition levels of the synergistic pairs in our main evolution experiment could potentially have contributed to higher adaptation rates for this type of combination (even though these were not found to be significantly increased compared to those for other interaction types; see above). This seems less likely the case for extinction events, which are generally more frequent in treatments with synergistic rather than antagonistic combinations, irrespective of the initial inhibition level. Most tested antibiotics interact synergistically in P. aeruginosa Antibiotic interactions are defined as synergistic, additive, or antagonistic when the drug pair has a stronger, equivalent, or weaker inhibitory effect on bacterial growth than the corresponding single drugs (i.e., monotherapies), respectively. Here, we determined this interaction quantitatively using an estimator denoted α [17]. This estimator is obtained from a quadratic regression applied to growth measurements as a function of different drug proportions of 2 drugs. The concentration of each of the single drugs is chosen to fall onto the line of equal dose, in our case defined to inhibit 75% of growth (i.e., inhibitory concentration [IC] 75; Fig 1A, S1 Fig and Table 1). The estimator α describes the shape of the resulting response in growth whereby positive values indicate synergism and negative values antagonism (Fig 1B). This approach has two advantages: first, it provides a statistical framework for testing the significance of positive or negative α; and second, its inference is less laborious than alternative procedures, thus facilitating characterization of a larger number of drug interactions. Even though the approach was carefully evaluated previously [17], we specifically validated its suitability for our model system. We compared the inferred α values for 8 selected combinations (S2A Fig and S3 Fig) to the corresponding results obtained with one of the commonly used alternative methods, based on Bliss independence and the checkerboard approach (S1 Data for a key to all datasets and S2 Data), as previously described for Escherichia coli [16,26]. This comparison demonstrated that α correlates significantly with the degree of synergy (S), irrespective of whether S is calculated from the average of all viable concentrations across a grid defined by the 2 drugs (ABij = Ai + Bj; S2B Fig) or from combinations for which the 2 individual drugs had the same level of inhibition (ABij for which IC50[Ai] = IC50[Bj], S2C Fig). We thus conclude that the α estimator provides an informative, quantitative indicator of a 2-drug interaction. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Drug interaction network for P. aeruginosa. (A) Schematic representation, adapted from [17], of the principle underlying the drug proportion parameter θ (line of equal dose; dashed lines), which is subsequently used to determine drug interactions, in comparison to different shapes of isobolograms (solid lines), as observed in synergistic (in red; top panel) or antagonistic (in blue; bottom panel) interactions. (B) Schematic illustration of the different interaction types as a function of the drug proportion parameter θ, ranging from synergism to antagonism. Drugs are combined in 9 different proportions (n = 9 for each combination), with each drug alone set to inhibit 75% of growth (S1 Fig). After a fixed time (12 h), bacterial growth is measured, and a quadratic model is used to fit the observed data. The α test [17] was used to determine significance of synergism or antagonism (S1 Table). (C) The α parameter was inferred from measured data to reconstruct a drug interaction network including 52 different antibiotic combinations. Combinations were formed from 12 different drugs, here represented as the nodes of the network, spanning 5 different antibiotic classes (see outer ring). The drug interaction profile is shown through the links (lines) formed between the nodes, and its strength is highlighted by the thickness of the lines and color. Red, black, and blue lines correspond to synergistic, additive, or antagonistic interactions, respectively (see also S3 Fig). The data for this panel are provided in S3 Data. AZL, azlocillin; CAR, carbenicillin; CEF, cefsulodin; CEZ, ceftazidime; CIP, ciprofloxacin; DOR, doripenem; GEN, gentamicin; IC75, concentration inhibiting 75% of bacterial growth; IMI, imipenem; PIT, piperacillin + tazobactam; STR, streptomycin; TIC, ticarcillin; TOB, tobramycin. https://doi.org/10.1371/journal.pbio.2004356.g001 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. List of antibiotics used in this study. https://doi.org/10.1371/journal.pbio.2004356.t001 We subsequently evaluated the interactions among 12 different antibiotics representing 5 classes (Table 1). We chose these drugs as representatives of the main classes of antibiotics, which are commonly used in combination to treat P. aeruginosa and to which most clinical P. aeruginosa strains are still susceptible [27–29]. Even though this choice could have introduced a bias in the overall pattern of inferred interaction types, these should nevertheless be representative of the clinically applied drug combinations. We characterized drug interactions for almost all of the possible combinations, resulting in a total of 52 measures that we summarized in an interaction network (Fig 1C, S3 Fig, S1 Table, S3 Data). Overall, synergistic combinations were more common than other interaction types (synergistic = 24/52; additive = 14/52; and antagonistic = 14/52). Combinations between cell wall inhibitors (β-lactams) and aminoglycosides most often produced synergisms, whereas those including ciprofloxacin (CIP) had exclusively antagonistic effects (Fig 1C). ACE networks demonstrate substantial variation in the effect of combinations on adaptation rates and population extinction We used evolution experiments to assess ACE, which is the ability of drug combinations to constrain bacterial adaptation either through population extinction or, in the case of surviving populations, reduced adaptation rates. Based on the inferred drug interactions and the previously obtained frequencies of collateral sensitivity between 8 of the considered antibiotics (Fig 2) [30], we selected 38 drug pairs covering all different types of drug interactions and collateral effects. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Collateral sensitivity network. The FCRs among 8 of the 12 drugs used in this study were obtained from our previous work [30]. FCR ranges from 0 to 1, such that 0 indicates that all populations (12–20 populations per combination) were sensitive to the corresponding other drug, thus having complete reciprocal sensitivity, whereas 1 highlights that none of the populations with resistance to one of the antibiotics in a pair suffered exacerbated sensitivity against the other. For the graphical illustration, we divided the combinations into 4 groups: complete collateral sensitivity (FCR ≤ 0.25; dark purple lines), partial collateral sensitivity (0.25 < FCR ≤ 0.5; light dashed pink lines), partial cross-resistance (0.5 < FCR < 0.75; light green dashed lines), and complete cross-resistance (FCR ≥ 0.75; dark green lines). CAR, carbenicillin; CEF, cefsulodin; CIP, ciprofloxacin; DOR, doripenem; FCR, frequency of collateral resistances; GEN, gentamicin; IMI, imipenem; PIT, piperacillin + tazobactam; STR, streptomycin. https://doi.org/10.1371/journal.pbio.2004356.g002 Based on this choice of drugs, we evolved a total of 1,672 populations through serial transfers into fresh media containing the respective antibiotics using a transfer period of 12 h and a total of 10 transfers (total duration of 120 h; Fig 3A and S4 Data). We assessed bacterial adaptive potential by integrating quantitative growth measurements taken in 15-min intervals from each evolving population (a total of 783,464 measurements for all treatments and populations; for a validation of our optical density (OD) measures as a proxy for bacterial growth, see Materials and methods and S4 Fig). For each population in a growth season, we then calculated the growth rate r during the exponential phase (Fig 3B). Following previous work [16], we defined the rate of adaptation as the change in growth rate over time for each evolving population (Fig 3C; for a validation of using growth characteristics as a proxy of evolutionary adaptation, see Materials and methods and S5 Fig). For subsequent analysis, we focused on the results of the 50:50 drug proportion (S6 Fig) and the single-drug treatments (S7 Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Experimental design and inference of adaptation rates. (A) Schematic representation of the evolution experiment with antibiotic combinations. Thirty-eight combinations were serially transferred every 12 h (season) into fresh medium containing antibiotics mixed in 5 different proportions (n = 8 per proportion and drug combination). An uninhibited control was also included, replicated 4 times, resulting in a total of 44 populations per combination and 1,672 for all combinations. Single-drug treatments of any drug A and B aimed at inhibiting 75% of growth relative to a drug-free environment (i.e., IC75). (B) An example of the quantitative growth measures obtained for a particular combination (CIP plus GEN) and the various drug proportions. Each panel shows 1 out of 5 seasons of growth (measured with OD as a proxy ± SD) over a 12-h period. Vertical grey lines denote the time window from which the slope was calculated to infer the growth rate r of each evolving population during exponential growth. All the drug proportions considered are highlighted in different colors (yellow to red), as well as the no-drug control (black). (C) Six exemplary populations from the CIP plus GEN combination experiments illustrating the change in growth rate r over 10 seasons of growth for each of the drug proportions. The rate of adaptation was calculated following previous work [16], and as indicated on the left of panel C, tadapt is defined as the time required to reach half of the change in growth rate, Δr. The data for this figure are provided in S4 Data. CFU, colony-forming unit; CIP, ciprofloxacin; GEN, gentamicin; OD, optical density; IC75, concentration inhibiting 75% of bacterial growth. https://doi.org/10.1371/journal.pbio.2004356.g003 We reconstructed the 2 ACE networks based either on adaptation rates of the surviving populations (Fig 4A) or on population extinctions (Fig 4B). Below, we first describe the patterns seen in the ACE networks, while their statistical analysis is explained in the next section. In all cases but one (for carbenicillin [CAR] plus gentamicin [GEN], all populations went extinct), adaptation to the combination treatment was possible. However, the rates of adaptation varied substantially across the different drug combinations, with lower rates of adaptation (below the 50th quantile) predominantly, but not exclusively, seen among antagonistic combinations that included CIP (Fig 4A; S8 Fig and S9 Fig show separate ACE networks for each drug interaction type and the 2 types of evolved collateral effects, respectively). Several synergistic drug pairs, combining an aminoglycoside with either a penicillin or carbapenem, led to similarly low rates of adaptation (below the 50th quantile, S8 Fig). Moreover, almost all cases of collateral sensitivity included in this study were associated with reduced adaptation rates (S9 Fig). This was not the case for combinations with cross-resistance. Furthermore, when estimating clearance efficacy, we found that extinctions almost exclusively occurred with the synergistic combinations (Fig 4B, S8 Fig). The synergistic combinations that did select for lower rates of adaptation did not necessarily have higher rates of extinction and vice versa (populations surviving synergistic combinations were not necessarily adapting more slowly; see azlocillin [AZL] plus streptomycin [STR], cefsulodin [CEF] plus CAR, or ticarcillin [TIC] plus GEN; S8 Fig). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. The ACE networks. (A) ACE network built from the rates of adaptation of surviving populations in the combination environment. The color and thickness of the lines (links) formed between the drugs (nodes) reflect the quantiles within which the inferred adaptation rates are found relative to the entire distribution: orange thick lines denote the combinations with the slowest adaptation rates (one of the aims of treatment efficacy), and grey thin lines highlight those with fast adaptation. (B) ACE network on the number of extinction events observed in the combination treatments. Thickness and color of the links represent the number of extinct populations, ranging from 0 (grey) to 8 (dark orange). Adaptation rates and extinction frequencies are inferred from the growth characteristics provided in S4 Data. ACE, antibiotic combination efficacy; AZL, azlocillin; CAR, carbenicillin; CEF, cefsulodin; CEZ, ceftazidime; CIP, ciprofloxacin; DOR, doripenem; GEN, gentamicin; IMI, imipenem; PIT, piperacillin + tazobactam; STR, streptomycin; TIC, ticarcillin; TOB, tobramycin. https://doi.org/10.1371/journal.pbio.2004356.g004 Statistical ACE network analysis reveals complementary roles for synergism and collateral sensitivity in treatment efficacy We next performed 2 types of statistical analyses to assess to what extent the overall characteristics of the 2 ACE networks are determined by the 2 considered predictors of combination efficacy: interaction type inferred from α (Fig 1C) and collateral sensitivity profiles previously obtained from experimentally evolved resistant populations of P. aeruginosa (Fig 2, [30]). We first used a BN approach to assess the relationships among the considered variables (i.e., adaptation rate, extinction frequency, drug interaction, and frequency of collateral resistances [FCR]). The BN approach is based on a constraint-based interleaved incremental association algorithm [31–33] to dissect the relationships between our variables (see Materials and methods for details). The results are summarized in the BN (Fig 5A), in which nodes represent the different variables and arrows indicate the inferred dependencies. The BN analysis revealed that the type of antibiotic interaction strongly influenced the proportion of extinction, but not the rate of adaptation. Instead, the rate of adaptation was found to depend solely on the frequency of collateral sensitivities. No other dependency was inferred by the analysis. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. BN analysis of antibiotic resistance evolution under combination therapy. (A) BN obtained from a constraint-based interleaved incremental association algorithm including 4 different random variables: drug interaction types, FCR, proportion of extinctions, and rate of adaptation. (B) Based on the BN, we calculated the conditional probabilities of rate of adaptation for different types of collateral effects (top panel) and extinction frequencies for different antibiotic interaction characteristics (bottom panel). The Bayesian analysis is based on data for drug interaction characteristics (S2 Data), collateral effects [30], and extinction frequencies, and adaptation rates are inferred from growth characteristics during experimental evolution (S4 Data). Adap., rate of adaptation; BN, Bayesian network; Ext., proportion of extinctions; FCR, frequency of collateral resistances; Int., drug interaction types. https://doi.org/10.1371/journal.pbio.2004356.g005 Based on the BN structure, we calculated the conditional probabilities for the inferred dependencies between the frequencies of collateral sensitivity and the rates of adaptation as well as for the proportion of extinction and drug interaction type. In particular, we used the different types of evolved collateral effects (i.e., partial collateral sensitivity, partial cross-resistance, and cross-resistance; none of the combinations evaluated during evolution had complete collateral sensitivity between their components, as shown in Fig 2) and calculated the conditional probability of obtaining the distribution of observed adaptation rates across 5 equal quantile bins (Fig 5B, top panel). Similarly, given the different drug interaction types (synergism, additivity, and antagonism), we calculated the conditional probabilities of different extinction frequencies across 5 equal quantile bins (Fig 5B, bottom panel). These 2 additional analyses describe more clearly the inferred dependencies within the BN. Antibiotic combinations for which at least half of the populations had collateral sensitivity against one or both of the individual drug components (i.e., partial collateral sensitivity; purple bars in Fig 5B, top panel) have a higher probability of selecting for low but not high rates of adaptation. Conversely, combinations with partial or complete cross-resistance (green bars in Fig 5B, top panel) have a higher probability of producing the top scores of inferred adaptation rates. In addition, high probabilities of extinction are associated with synergistic and additive combinations, whereas the reverse is found for antagonistic drug pairs (Fig 5B, bottom panel). We further validated the inferred dependencies between variables using partial correlation analysis, following the approach previously established for a similar analysis of combination efficacy in E. coli [16]. This approach allowed us to control for drug pair membership using the average rate of adaptation towards the corresponding single drugs of a particular combination as a covariate (Materials and methods). Statistical significance was subsequently inferred using a permutation test [16]. This analysis revealed a significant correlation between the FCR and the rate of adaptation (ρs = 0.52, P = 0.038) and between the proportion of extinction and the drug interaction type α (ρs = 0.51, P = 0.043), but not between the FCR and the proportion of extinction (ρs = 0.39, P = 0.146) or the drug interaction α and the rate of adaptation (ρs = 0.3, P = 0.262). This analysis, based on a distinct statistical approach, thereby corroborated the findings of the BN analysis. We conclude that synergistic drug interactions enhance bacterial clearance, whereas collateral sensitivity limits the adaptive potential of the bacteria. Adaptation to the strongest component influences adaptation to multidrug environments We next assessed whether the ability of bacteria to adapt to the combination is mainly driven by adaptation to only one of the drugs rather than dependent on a unique property of the antibiotic pair. For our dataset, we related the inferred rates of adaptation in the combination treatments to those inferred for the corresponding single-drug environments (S10 Fig). We first compared the 2 corresponding monotherapies of a given drug pair and defined the drug leading to lower rates of adaptation as the stronger component (i.e., higher ability to minimize resistance evolution) and the other as the weaker component (i.e., lower ability to minimize resistance evolution). Thereafter, we calculated the relative rate of adaptation of the combination by standardizing it against either the stronger or the weaker component of the pair. The resulting ACE networks are shown in Fig 6A and 6B, respectively. Interestingly, the original ACE network for adaptation rates (Fig 4A) is more similar to that standardized by the weaker but not the stronger component drug (Fig 6; S2 Table). This suggests that the characteristics of the original ACE network (Fig 4A), and thus the efficacy of drug combinations to reduce adaptation rates, is primarily driven by adaptation to the stronger component, which—if accounted for by the standardizing scheme—removes important properties of the network (see as prominent examples the disappearance of the strong reduction in adaptation rate for doripenem [DOR] plus TIC, or DOR plus PIT [piperacillin + tazobactam]; Fig 4A and Fig 6A). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Weighted ACE networks and their Bayesian analysis. We assessed to what extent adaptation to one of the drugs of a pair determined the overall rate of adaptation to the combination treatment. The stronger component drug of each pair was identified as the one with lower adaptation rates in monotherapy. We subsequently standardized the adaptation rates towards the combination by those towards either (A) the stronger or (B) the weaker component drug, resulting in 2 weighted ACE networks. Orange thick lines indicate slower adaptation, while grey thin bands denote fast adaptation. (C) Results of the BN analysis on the original network versus the 2 standardized networks. The relationship between drug interaction type and extinction frequency was stable across all analyses, while the dependence of adaptation rate on evolved collateral effects disappeared when adaptation rates were standardized by the stronger component. Adaptation rates are inferred from the data on growth characteristics during experimental evolution, provided in S4 Data. ACE, antibiotic combination efficacy; AZL, azlocillin; BN, Bayesian network; CAR, carbenicillin; CEF, cefsulodin; CEZ, ceftazidime; CIP, ciprofloxacin; DOR, doripenem; GEN, gentamicin; IMI, imipenem; PIT, piperacillin + tazobactam; STR, streptomycin; TIC, ticarcillin; TOB, tobramycin. https://doi.org/10.1371/journal.pbio.2004356.g006 We further evaluated influence of the component drugs by repetition of the BN analysis. We found that the dependency observed between the FCR and the rates of adaptation of the combinations disappeared when the latter is weighted by the stronger but not the weaker component drug (Fig 6C). At the same time, the dependency between drug interaction and extinction frequency remained, while no additional relationship was revealed. Similar results were obtained when we repeated the correlation analysis with standardized adaptation rates. The originally identified correlation between the FCR and the rate of adaptation was no longer significant when the latter was standardized by adaptation to the stronger component drug (ρs = 0.33, P = 0.21), yet it still showed a statistical trend when we standardized by the weaker component drug (ρs = 0.45, P = 0.078). In these 2 analyses, drug interaction did not correlate significantly with the weighted adaptation rates (ρs < 0.47, P > 0.09). These results consistently indicate that adaptation to the stronger component drug influences adaptation to the combination and that this is dependent on the evolved collateral effects. Initial inhibition levels correlate with adaptation rates, while extinction events are almost exclusively restricted to synergistic combinations We next performed a separate evolution experiment with 4 selected combinations to assess to what extent the inherently different starting levels of inhibition—imposed by each type of interaction during the first season of growth (Fig 1B and S3 Fig)—influenced both the number of extinctions and adaptation rates. We performed this evolution experiment with 4 selected combinations with different interaction profiles: 2 interacting synergistically (GEN plus CAR and STR plus PIT) and 2 antagonistically (GEN plus CIP and Tobramycin [TOB] plus CIP). For these combinations, we varied the initial inhibition level of the combination across 8 steps, ranging from IC50 to >IC90. Populations were serially transferred into fresh media as explained before (S5 Data; and for the obtained changes in growth rate r, see S11 Fig). This separate evolution experiment revealed that initial inhibitory levels of the tested combinations are significantly related to the rates of adaptation, irrespective of combination identity or drug interaction type (GLM, F1,336 = 37.735, P < 0.001; Fig 7A and S3 Table). In particular, increasing levels of inhibition are generally associated with higher rates of adaptation, suggesting that strong inhibition increases selection for an adaptive response [34,35]. At higher levels of inhibition, the synergistic and antagonistic combinations produce clearly distinct responses, especially regarding population extinction. Here, the 2 synergistic pairs are associated with a significant increase in the number of extinct populations (logistic regression, F12,336 = 21.15, P < 0.001; Fig 7B and S4 Table), while antagonistic combinations produced almost no extinction at all. Moreover, at the very high initial inhibitory levels, antagonistic pairs showed a sudden drop in adaptation rates (Fig 7A), as expected from previous work [16,24]. A similarly strong reduction is not observed for the synergistic combinations, possibly owing to the fact that only few populations survived and could be used to infer adaptation rates. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Influence of the initial inhibition level on adaptation rate and population extinction. In a separate round of evolution experiments, we evaluated the consequences of the initial inhibition levels in synergistic and antagonistic combinations. The experiment was performed following the protocol of the main evolution experiment (Fig 3A), with the exception that the starting doses of the combinations were fixed at different levels of inhibition for the combinations. (A) Inferred rates of adaptation as a function of initial inhibitory levels. We determined initial inhibition by measuring in the first season of the evolution experiment the AUC of growth over time (measured as OD as a proxy) for each replicate population and then standardized it against the average AUC of the no-drug control (x-axis, AUCi). The rate of adaptation (y-axis) was inferred as for the main evolution experiment (Fig 3). (B) Extinction was significantly more often observed in synergistic rather than antagonistic combinations, even at the same level of inhibition (results for logistic regression analysis in S4 Table). Adaptation rates and extinction frequencies are inferred from growth characteristics during experimental evolution, provided in S5 Data. AUC, area under the curve; AUCi, area under the curve of relative inhibition of growth; CAR, carbenicillin; CIP, ciprofloxacin; GEN, gentamicin; OD, optical density; PIT, piperacillin + tazobactam; STR, streptomycin; TOB, tobramycin. https://doi.org/10.1371/journal.pbio.2004356.g007 Taken together, the results from this separate evolution experiment suggest that the generally higher inhibition levels of the synergistic pairs in our main evolution experiment could potentially have contributed to higher adaptation rates for this type of combination (even though these were not found to be significantly increased compared to those for other interaction types; see above). This seems less likely the case for extinction events, which are generally more frequent in treatments with synergistic rather than antagonistic combinations, irrespective of the initial inhibition level. Discussion Our study provides a systematic experimental analysis of the efficacy of antibiotic combination therapy in the opportunistic human pathogen P. aeruginosa. Based on evolution experiments with 38 distinct combinations, ACE networks were reconstructed for 2 complementary measures of treatment efficacy: the frequency of population extinctions and the reduction in adaptation rates. Subsequent statistical analyses identified the likely ACE determinants: Synergistic drug interactions enhanced the frequency of extinction, even at the same inhibitory level as antagonistic interactions, while reduced adaptation rates depended on the evolved collateral sensitivities among the drugs. The latter effect is likely driven by adaptation to the stronger component drug in a pair. Consequently, our findings suggest that treatment efficacy against P. aeruginosa can be optimized by drug combinations, which interact synergistically to increase bacterial clearance and which can evolve collateral sensitivity to each other to slow down the rate of adaptation. The use of BN analysis enhanced dissection of the determinants of ACE. The BN approach has been widely applied across different fields of biology in recent years but not yet in studies on antibiotic resistance evolution [33,36–39]. Its accessible graphical output and the underlying probabilistic theory facilitate the inference of causal relationships between different variables [31,32]. It further offers estimation of conditional probabilities that reflect the strength of the inferred dependencies; a strategy well suited for the stochastic nature of biological systems and their measurements [40]. The latter is important for the analysis of antibiotic resistance evolution, for which we are mainly interested in anticipating bacterial adaptation based on distinct drug properties or deployment strategies [11,12,41–43]. The suitability of the BN approach for analysis of drug resistance evolution was corroborated with a previously established statistical approach, based on partial correlation analysis [16], which identified a significant relationship for the same pairs of variables. Our analyses consistently revealed that synergistic drug interactions are an important ACE determinant, especially in terms of bacterial clearance (Fig 4A). The particular importance of bacterial elimination as a component of treatment efficacy was previously considered in a mathematical model [25] but has not yet been evaluated empirically. The previous model assessed the effect of antibiotic interactions on treatment efficacy [25] by modifying a previous infection model based on data from mice infected with P. aeruginosa [44]. The model is related to the design of our main evolution experiment in that the concentration of a particular drug in a combination is standardized by its inhibitory effect in monotherapy. The model predicted contrasting treatment outcomes for synergistic combinations: On the one hand, synergism enhances extinction, most likely because it strongly reduces population size, thereby decreasing the likelihood of new resistance mutations arising. On the other hand, if resistance emerges, synergism increases the selective advantage of the resistant mutants through competitive release, enhancing bacterial adaptation [25]. Our experimental results are consistent with both alternatives. Although synergism mainly favored bacterial extinction (Figs 5–7), it was in several cases associated with low adaptation rates (Fig 4A). However, in our study, the effect of drug interaction on adaptation rate was always insignificant, irrespective of the analytical approach. Interestingly, we found higher population extinction for synergistic rather than antagonistic combinations also at low initial inhibitory concentrations (Fig 7). This finding cannot have resulted from the stronger reduction in population size (i.e., inhibitory levels were the same for the 2 interaction types) but must have depended on other properties of the synergistic drug pairs. A likely explanation may be found in the mechanism underlying synergism, which can rely on increased membrane permeability induced by one of the drugs, subsequently enhancing cellular uptake of the second drug [45]. Such mechanisms may have a cumulative effect across time [45] and/or may generally be difficult to counter. This, in turn, limits the number of suitable resistance mutations and ultimately increases the likelihood of extinction. A detailed exploration of this effect clearly warrants further research. Our experiments further identified the potential to evolve collateral sensitivity as a key determinant of low adaptation rates. This result is generally consistent with previous work on E. coli and Staphylococcus aureus [46,47], although this is the first time it has been shown for P. aeruginosa. Adaptation rates are thus significantly influenced by evolutionary trade-offs, whereby adaptation to one of the drugs of a pair constrains adaptation to the other. Our findings and those of colleagues [46,47] thereby highlight that such trade-offs may not only improve treatment when drugs are applied sequentially, as originally proposed for evolved collateral sensitivities in E. coli (i.e., collateral sensitivity cycling; [20–22]). Instead, they can also optimize combination therapy. Our analysis further revealed that the involved dynamics are likely driven by adaptation to the stronger component drug of a pair (Fig 6). This suggests that, if adaptation to the stronger component comes with a higher likelihood of collateral sensitivity to the second drug, adaptation to the combination is systematically slowed down, as, for example, for CIP plus STR or CIP plus CAR (Fig 2, Fig 4A). In contrast, when adaptation to the stronger drug is more likely to cause cross-resistance, then this can enhance adaptation to the combination, as seen for GEN plus STR or CAR plus CEF (Fig 2, Fig 4A). The further exploration of these trade-offs represents a promising avenue to improve treatment efficacy. Our finding of the high clearance efficacy of synergistic combinations shows some consistency with clinical practice. For P. aeruginosa, we predominantly observed drug synergism between β-lactams and aminoglycosides (Fig 1C). These 2 antibiotic classes are also most commonly used in combination therapy against this pathogen [29,48,49]. Our results empirically confirm the potency of the β-lactam–aminoglycoside combinations, especially penicillin–aminoglycoside pairs, in causing higher numbers of extinct replicate populations (Fig 4B and S8 Fig). In some cases, the populations surviving these specific combinations also adapted more slowly (e.g., STR plus PIT or TIC plus TOB in Fig 4A and 4B, and S8 Fig). Furthermore, the effectiveness of these combinations may not only be caused by drug synergism but additionally by reciprocal collateral sensitivity that can evolve among these pairs [30]. Our systematic analysis performed under controlled laboratory conditions thus provides empirical support for the often experience-driven choice in clinical treatment. In the future, the clinical applicability of our results should be further explored. For example, we identified high clearance efficacy of certain combinations of penicillins and cephalosporins (Fig 4B) or low adaptation rates if fluoroquinolones (e.g., CIP) were combined with aminoglycosides or penicillins (Fig 4A). It would be of particular interest to corroborate these patterns for clinical isolates in laboratory experiments or under clinical conditions. In summary, our systematic analysis of antibiotic combinations identified the role of drug interactions and evolved collateral effects in determining 2 complementary properties of treatment efficacy. The comprehensive dataset collected in our study may serve as a useful reference for further exploration of effective therapy, including more detailed statistical analyses such as those that use the potency of pairwise interactions to estimate higher-order drug effects [50,51]. Our approach and the specific results obtained may, moreover, help to improve the design of medical treatment with the 2-fold aim of minimizing pathogen burden and reducing resistance evolution. A similar combined assessment of the efficacy of drug interaction and evolved collateral effects may not only be applicable to other pathogens and infectious diseases. It could similarly help to improve cancer therapy, as previously evaluated for selected cancer types and drug interactions [52–55]. Materials and methods Bacteria and media All experiments were conducted with P. aeruginosa PA14. Cells were grown at 37 °C in sterile M9 minimal medium supplemented with 0.2% glucose and 0.1% casamino acids. All antibiotics were prepared according to the manufacturer’s instructions and filter sterilized before each experiment (Table 1). All experiments were carried out in randomized 96-well plates shaken and incubated at 37 °C in BioTek Eon plate readers, which were also used for regular measurement of ODs in 15-min intervals. Randomization schemes of plates for each experiment were different from each other. All analyses were performed using the R platform (version 3.3.2) unless specified otherwise [56]. Dose-response curves and minimal inhibitory concentration We tested 14 different concentrations of each drug in order to establish dose-response relationships after 12 h of incubation. For all concentrations, a 1- to 2-ml 10× stock was prepared and then diluted in a randomized 96-well plate with 6 replicates per concentration, resulting in 90 replicates per antibiotic and 1,080 for all treatments. Ten microliters of an isogenic bacterial population of PA14 were added to a final volume of 100 μl, equivalent to 104 to 105 CFU/ml initial population size. In addition, 2 types of controls were included: one without antibiotic and a second one without both antibiotic and bacteria, each also replicated 6 times. We used a logistic regression to analyze the dose-response relationship of each drug using the package “drc” in R [57]. The obtained models (S1 Fig) allowed accurate calculation of different levels of inhibitory concentrations for each drug, including the minimum inhibitory concentration (MIC; here defined as the concentration inhibiting >90% of growth). Checkerboards and degree of synergy To measure the type of interaction using the checkerboard approach, we considered 9 concentrations of each antibiotic in a pair, including a no-drug control, and distributed them randomly across a 96-well plate. Each pair was evaluated twice. Plates were incubated at 37 °C for 12 h with constant shaking and regular OD measurements taken every 15 min. We then calculated the growth rate r for each individual well and combination by fitting a linear regression of growth over time during the exponential phase. Exponential phase was generally observed during 195 to 360 min of each season. We subsequently determined the degree of synergy of any drug pair AB using the Bliss independence method described previously [16]: such that rA0 represents the growth rate at a given concentration of drug A in the absence of B, and vice versa for r0B. r00 is the growth rate of the no-drug control, and rAB is the growth rate at any concentration in which drugs A and B are found together. The degree of synergy S was only calculated for drug combinations that had growth rates larger than 0. Positive values indicate synergism, whereas negative ones denote antagonism. Drug combinations and interaction profile To classify the interaction between 2 drugs, we considered an environment in which each drug separately inhibits 75% ± 10% of bacterial growth (IC75). For each combination, we evaluated 11 treatments: 9 different proportions of a given pair of antibiotics, a control of uninhibited growth, and a control with only M9 medium. Nine replicates for all treatments were considered, except for the M9 control that consisted of only 6 wells. This resulted in 81 replicates per drug combination and 4,212 for all 52 antibiotic pairs. OD measurements were taken every 15 min for 12 h, resulting in a total of 48 data points per individual replicate and 202,176 for all combinations and replicates. To determine whether interactions were antagonistic, synergistic, or additive, we used a t test on the second-order term (α) of a quadratic regression of our data, as established previously [17]. The α parameter expresses convexity or concavity of observed bacterial-density data in the model q(θ) = αθ2 + βθ + γ, such that θ represents any drug proportion between any drugs A and B (Fig 1B). Positive values of α indicate synergy and negative values antagonism. Collateral sensitivity network We considered our previously published data on the evolved collateral effects of highly resistant populations of P. aeruginosa PA14 [30] and used the frequency of cross-resistance in all possible pairwise combinations of 8 of the drugs considered in this study. Briefly, the FCR counts the number of populations resistant to drug A that show collateral resistance to drug B, and vice versa, relative to the total number of populations resistant to A and B. Values close to 0 indicate reciprocal collateral sensitivity, and those close to 1 denote cross-resistance. We categorized the obtained values into 4 different groups and built a collateral sensitivity network (Fig 2): complete collateral sensitivity (FCR ≤ 0.25), partial collateral sensitivity (0.25 < FCR ≤ 0.5), partial cross-resistance (0.5 < FCR < 0.75), and complete cross-resistance (FCR ≥ 0.75). Experimental evolution of antibiotic combinations Based on the interaction profile and the collateral sensitivity and/or resistance [30] scores, we selected a total of 38 different combinations for a series of evolution experiments (Fig 3A). For all combinations, we included 5 different proportions of the combined antibiotics, an uninhibited control, and an M9 control, resulting in 44 populations per combination, randomly distributed in a 96-well plate (2 combinations were included in a single plate), for a total of 1,672 populations. The concentration was set for each individual drug to inhibit bacterial growth by 75% (IC75). We considered 10 transfers (hereafter referred to as seasons) of 1% volume into fresh plates every 12 h (approximately 120 generations). For each season, OD600 measurements were taken every 15 min, resulting in 48 measurements per replicate and season and a total of 781,440 measurements across all replicate populations. All plates were frozen at −80 °C with 1:4 (v/v) of 86% glycerol. To validate our OD measurements as a proxy for bacterial growth during evolution, we replicated the conditions of the first season for 4 selected combinations (only the 1:1 proportion), 6 corresponding single-drug treatments, and a no-drug control. We focused on those combinations and the corresponding monotherapies for which we also evaluated the influence of initial drug inhibitory level (Fig 7) and the evolution of resistance (S5 Fig). Each treatment was replicated 8 times. After 12 h of evolution, we performed a dilution series and standard plating techniques to count viable colony-forming units (CFUs) for all replicates and treatments. The obtained CFUs were then correlated with the endpoint OD measurements (S4 Fig). We found a significant correlation between our OD measurements and the CFU counts at the end of season 1 (Spearman rank correlation test, ρs = 0.782, P < 0.001). To further validate the OD measurements, we performed a similar correlation analysis for the same combinations and corresponding monotherapies, using evolved bacteria from the final transfer of the separate, focused evolution experiment, in which the influence of initial drug inhibitory levels was assessed. The evolved material was thawed from the frozen stock cultures, then exposed to 1 full season of experimental evolution under the exact treatment conditions already experienced by populations during the evolution experiment. Thereafter, CFUs were counted using a dilution series on Agar plates, as outlined above, and then compared to the OD measures obtained during the above repetition of a full season. As before, CFUs were significantly correlated with the corresponding OD measurements (Spearman rank correlation test, ρs = 0.339, P = 0.002). We further validated the suitability of changes in growth characteristics as a proxy for evolutionary adaptation and therefore genetically fixed alterations by re-assessing cryo-preserved material from the last transfer of experimental evolution. This analysis was performed with material from the separate evolution experiment, which tested the influence of initial inhibitory levels, and further details are outlined below in the description of this experiment. Rates of adaptation We first calculated the growth rate r as described above for each evolving population, treatment, and season. Subsequently, we considered the rate of adaptation for each evolving line as defined previously [16]: such that Δr represents the change in growth rate over 10 seasons of growth, and the time of adaptation, tadapt, corresponds to the interpolated time at which a population reached half of its maximum growth rate. This measurement reflects how quickly resistance spreads in a population in a serial transfer experiment. To determine to what extent adaptation to the drug combinations was determined by adaptation to each of the individual drugs, we measured which of the individual components in a drug pair led to lower and higher rates of adaptation. The single antibiotic in a pair that alone led to lower rates of adaptation was considered as the stronger of the components and the other as the weaker one. The adaptation rate of each combination was then standardized by the adaptation rate of either its weaker or stronger component drug. The 2 types of standardized adaptation rates were visualized in ACE networks and statistically evaluated (see below). BN analysis We used BN analysis to assess the directional relationship between 4 variables, including the inferred drug interaction type, the frequency of collateral sensitivities, the adaptation rates, and the frequency of population extinctions. The entire BN analysis was repeated with the different types of inferred adaptation rates, including those obtained for the combinations in the main experiment and then those that we standardized by either the stronger or the weaker component drug. The BN analysis generally followed 2 steps. In the first step, the approach identifies variables that are related to each other and visualizes these as nodes in a network between variables. In this step, it further infers the direction of each relationship and represents these as arrows in the network, thereby implying a causality between the connected variables [31]. To achieve this first step, the model first infers the graphical structure of the network by analyzing the probabilistic relations between all nodes and thereafter constructs the network by setting directions for the identified connections while satisfying an acyclicity constraint [58]. We implemented BN analysis employing a constraint-based interleaved incremental association–optimized algorithm [59] to reduce the likelihood of obtaining false positives and to obtain possible probabilistic dependencies between our variables: drug interaction type (categorical: synergism, additivity, or antagonism), FCR (categorical: complete collateral sensitivity, partial collateral sensitivity, partial cross-resistance, and complete cross-resistance), proportion of extinction (numerical), and rates of adaptation (numerical). We only included combinations with complete sets of data and then followed the algorithm’s default parameters. From the obtained dependencies, we estimated the conditional probabilities associated with the linked variables over an array of different values. All tests were performed in R using the “bnlearn” package [60]. Additional correlation and partial correlation analysis To validate the inferred dependencies from the BN analysis, we additionally performed correlation analysis combined with permutation tests, following the approach previously established for a similar analysis of ACE in E. coli [16]. For each round of permutation, we calculated correlation coefficients, ρs, between any two given variables x and y by permuting the values of x while keeping y constant, as in [16]. For each test, we considered 10,000 permutations and estimated the P value as the proportion of the obtained distribution of correlation coefficients that had an absolute value larger than the absolute value obtained for the observed ρs [16]. This approach was used to correlate the measures of collateral effects and drug interaction to proportion of extinction and, later on, to the standardized adaptation rates. Furthermore, to account for the effect of adaptation to the single drugs (z) in the main analysis with nonstandardized adaptation rates, we performed a partial correlation analysis with z as a covariate, generally following the previously established approach [16]. For this, we first obtained the residuals from the linear regression of x on z and those of y on z, such that y corresponds to the adaptation rates of the combination. Then, to estimate the correlation coefficient between x and y, with z as a covariate, we employed the permutation test as explained above using the residuals of the corresponding regressions [16]. Experimental evolution with fixed inhibitory levels of antibiotic combinations To evaluate the effect of the starting inhibition level of the combinations, we considered a second round of evolution experiments as described above. This time, the level of inhibition of the combination was fixed instead of that of the individual drug treatments. Briefly, concentrations of each drug were mixed 1:1 so that each would inhibit between 50% and 75% of growth. These were then diluted to obtain a range of different inhibition levels and to evaluate their effect on growth in P. aeruginosa after 12 h of incubation at 37 °C. Evolution experiments were then initiated for 4 different combinations that included 11 different treatments: a no-drug control, the individual monotherapies, and 8 different inhibition levels ranging from approximately IC50 to >IC90 of each combination. Each treatment was replicated 8 times and distributed randomly in 96-well plates. Genetically fixed changes in growth characteristics We used the focused set-up of the above separate evolution experiment to validate the suitability of growth measurements as a proxy for evolutionary adaptation. Evolutionary adaptation assumes that changes are genetically fixed rather than due to phenotypic (i.e., physiological) responses. To assess this, we studied cryo-preserved material from the last drug-free season of the evolution experiment and regrew them under defined antibiotic conditions. Purely phenotypic adaptations to antibiotics are unlikely to have persisted for this material, which was grown under antibiotic-free conditions for 12 to 16 h (equivalent to a minimum of 6 generations) and additionally subjected to a cryo-preservation step. Therefore, any persistent changes in growth characteristics under antibiotic exposure are likely based on genetic changes and thus indicate evolutionary adaptation. For this analysis, we considered material evolved in the presence of 2 synergistic (i.e., GEN plus CAR and STR plus PIT) or 2 antagonistic combinations (i.e, CIP plus GEN and CIP plus TOB), in all cases set to either IC50 or >IC90, and also included material from the corresponding monotherapies. A total of 4 replicate populations was studied for each of the various evolution treatments and compared to the ancestral PA14. Changes in growth characteristics were inferred from dose-response curves in a 2-fold dilution series of each of the antibiotics included in the pair. The evolved relative changes in resistance were calculated as the area under the curve (AUC) of the dose-response curve for each of the populations and then divided by that of the ancestral PA14. The results are shown in S5 Fig. They highlight a general increase in growth characteristics and thus resistance across the various treatment groups even if not significant in all cases (based on a 1-sample Wilcoxon test with μ = 1). We conclude that, overall, the observed changes in growth characteristics have a genetic basis and are not exclusively due to phenotypic responses. Therefore, we consider the recorded changes in growth characteristics to provide a meaningful proxy for evolutionary adaptation. Bacteria and media All experiments were conducted with P. aeruginosa PA14. Cells were grown at 37 °C in sterile M9 minimal medium supplemented with 0.2% glucose and 0.1% casamino acids. All antibiotics were prepared according to the manufacturer’s instructions and filter sterilized before each experiment (Table 1). All experiments were carried out in randomized 96-well plates shaken and incubated at 37 °C in BioTek Eon plate readers, which were also used for regular measurement of ODs in 15-min intervals. Randomization schemes of plates for each experiment were different from each other. All analyses were performed using the R platform (version 3.3.2) unless specified otherwise [56]. Dose-response curves and minimal inhibitory concentration We tested 14 different concentrations of each drug in order to establish dose-response relationships after 12 h of incubation. For all concentrations, a 1- to 2-ml 10× stock was prepared and then diluted in a randomized 96-well plate with 6 replicates per concentration, resulting in 90 replicates per antibiotic and 1,080 for all treatments. Ten microliters of an isogenic bacterial population of PA14 were added to a final volume of 100 μl, equivalent to 104 to 105 CFU/ml initial population size. In addition, 2 types of controls were included: one without antibiotic and a second one without both antibiotic and bacteria, each also replicated 6 times. We used a logistic regression to analyze the dose-response relationship of each drug using the package “drc” in R [57]. The obtained models (S1 Fig) allowed accurate calculation of different levels of inhibitory concentrations for each drug, including the minimum inhibitory concentration (MIC; here defined as the concentration inhibiting >90% of growth). Checkerboards and degree of synergy To measure the type of interaction using the checkerboard approach, we considered 9 concentrations of each antibiotic in a pair, including a no-drug control, and distributed them randomly across a 96-well plate. Each pair was evaluated twice. Plates were incubated at 37 °C for 12 h with constant shaking and regular OD measurements taken every 15 min. We then calculated the growth rate r for each individual well and combination by fitting a linear regression of growth over time during the exponential phase. Exponential phase was generally observed during 195 to 360 min of each season. We subsequently determined the degree of synergy of any drug pair AB using the Bliss independence method described previously [16]: such that rA0 represents the growth rate at a given concentration of drug A in the absence of B, and vice versa for r0B. r00 is the growth rate of the no-drug control, and rAB is the growth rate at any concentration in which drugs A and B are found together. The degree of synergy S was only calculated for drug combinations that had growth rates larger than 0. Positive values indicate synergism, whereas negative ones denote antagonism. Drug combinations and interaction profile To classify the interaction between 2 drugs, we considered an environment in which each drug separately inhibits 75% ± 10% of bacterial growth (IC75). For each combination, we evaluated 11 treatments: 9 different proportions of a given pair of antibiotics, a control of uninhibited growth, and a control with only M9 medium. Nine replicates for all treatments were considered, except for the M9 control that consisted of only 6 wells. This resulted in 81 replicates per drug combination and 4,212 for all 52 antibiotic pairs. OD measurements were taken every 15 min for 12 h, resulting in a total of 48 data points per individual replicate and 202,176 for all combinations and replicates. To determine whether interactions were antagonistic, synergistic, or additive, we used a t test on the second-order term (α) of a quadratic regression of our data, as established previously [17]. The α parameter expresses convexity or concavity of observed bacterial-density data in the model q(θ) = αθ2 + βθ + γ, such that θ represents any drug proportion between any drugs A and B (Fig 1B). Positive values of α indicate synergy and negative values antagonism. Collateral sensitivity network We considered our previously published data on the evolved collateral effects of highly resistant populations of P. aeruginosa PA14 [30] and used the frequency of cross-resistance in all possible pairwise combinations of 8 of the drugs considered in this study. Briefly, the FCR counts the number of populations resistant to drug A that show collateral resistance to drug B, and vice versa, relative to the total number of populations resistant to A and B. Values close to 0 indicate reciprocal collateral sensitivity, and those close to 1 denote cross-resistance. We categorized the obtained values into 4 different groups and built a collateral sensitivity network (Fig 2): complete collateral sensitivity (FCR ≤ 0.25), partial collateral sensitivity (0.25 < FCR ≤ 0.5), partial cross-resistance (0.5 < FCR < 0.75), and complete cross-resistance (FCR ≥ 0.75). Experimental evolution of antibiotic combinations Based on the interaction profile and the collateral sensitivity and/or resistance [30] scores, we selected a total of 38 different combinations for a series of evolution experiments (Fig 3A). For all combinations, we included 5 different proportions of the combined antibiotics, an uninhibited control, and an M9 control, resulting in 44 populations per combination, randomly distributed in a 96-well plate (2 combinations were included in a single plate), for a total of 1,672 populations. The concentration was set for each individual drug to inhibit bacterial growth by 75% (IC75). We considered 10 transfers (hereafter referred to as seasons) of 1% volume into fresh plates every 12 h (approximately 120 generations). For each season, OD600 measurements were taken every 15 min, resulting in 48 measurements per replicate and season and a total of 781,440 measurements across all replicate populations. All plates were frozen at −80 °C with 1:4 (v/v) of 86% glycerol. To validate our OD measurements as a proxy for bacterial growth during evolution, we replicated the conditions of the first season for 4 selected combinations (only the 1:1 proportion), 6 corresponding single-drug treatments, and a no-drug control. We focused on those combinations and the corresponding monotherapies for which we also evaluated the influence of initial drug inhibitory level (Fig 7) and the evolution of resistance (S5 Fig). Each treatment was replicated 8 times. After 12 h of evolution, we performed a dilution series and standard plating techniques to count viable colony-forming units (CFUs) for all replicates and treatments. The obtained CFUs were then correlated with the endpoint OD measurements (S4 Fig). We found a significant correlation between our OD measurements and the CFU counts at the end of season 1 (Spearman rank correlation test, ρs = 0.782, P < 0.001). To further validate the OD measurements, we performed a similar correlation analysis for the same combinations and corresponding monotherapies, using evolved bacteria from the final transfer of the separate, focused evolution experiment, in which the influence of initial drug inhibitory levels was assessed. The evolved material was thawed from the frozen stock cultures, then exposed to 1 full season of experimental evolution under the exact treatment conditions already experienced by populations during the evolution experiment. Thereafter, CFUs were counted using a dilution series on Agar plates, as outlined above, and then compared to the OD measures obtained during the above repetition of a full season. As before, CFUs were significantly correlated with the corresponding OD measurements (Spearman rank correlation test, ρs = 0.339, P = 0.002). We further validated the suitability of changes in growth characteristics as a proxy for evolutionary adaptation and therefore genetically fixed alterations by re-assessing cryo-preserved material from the last transfer of experimental evolution. This analysis was performed with material from the separate evolution experiment, which tested the influence of initial inhibitory levels, and further details are outlined below in the description of this experiment. Rates of adaptation We first calculated the growth rate r as described above for each evolving population, treatment, and season. Subsequently, we considered the rate of adaptation for each evolving line as defined previously [16]: such that Δr represents the change in growth rate over 10 seasons of growth, and the time of adaptation, tadapt, corresponds to the interpolated time at which a population reached half of its maximum growth rate. This measurement reflects how quickly resistance spreads in a population in a serial transfer experiment. To determine to what extent adaptation to the drug combinations was determined by adaptation to each of the individual drugs, we measured which of the individual components in a drug pair led to lower and higher rates of adaptation. The single antibiotic in a pair that alone led to lower rates of adaptation was considered as the stronger of the components and the other as the weaker one. The adaptation rate of each combination was then standardized by the adaptation rate of either its weaker or stronger component drug. The 2 types of standardized adaptation rates were visualized in ACE networks and statistically evaluated (see below). BN analysis We used BN analysis to assess the directional relationship between 4 variables, including the inferred drug interaction type, the frequency of collateral sensitivities, the adaptation rates, and the frequency of population extinctions. The entire BN analysis was repeated with the different types of inferred adaptation rates, including those obtained for the combinations in the main experiment and then those that we standardized by either the stronger or the weaker component drug. The BN analysis generally followed 2 steps. In the first step, the approach identifies variables that are related to each other and visualizes these as nodes in a network between variables. In this step, it further infers the direction of each relationship and represents these as arrows in the network, thereby implying a causality between the connected variables [31]. To achieve this first step, the model first infers the graphical structure of the network by analyzing the probabilistic relations between all nodes and thereafter constructs the network by setting directions for the identified connections while satisfying an acyclicity constraint [58]. We implemented BN analysis employing a constraint-based interleaved incremental association–optimized algorithm [59] to reduce the likelihood of obtaining false positives and to obtain possible probabilistic dependencies between our variables: drug interaction type (categorical: synergism, additivity, or antagonism), FCR (categorical: complete collateral sensitivity, partial collateral sensitivity, partial cross-resistance, and complete cross-resistance), proportion of extinction (numerical), and rates of adaptation (numerical). We only included combinations with complete sets of data and then followed the algorithm’s default parameters. From the obtained dependencies, we estimated the conditional probabilities associated with the linked variables over an array of different values. All tests were performed in R using the “bnlearn” package [60]. Additional correlation and partial correlation analysis To validate the inferred dependencies from the BN analysis, we additionally performed correlation analysis combined with permutation tests, following the approach previously established for a similar analysis of ACE in E. coli [16]. For each round of permutation, we calculated correlation coefficients, ρs, between any two given variables x and y by permuting the values of x while keeping y constant, as in [16]. For each test, we considered 10,000 permutations and estimated the P value as the proportion of the obtained distribution of correlation coefficients that had an absolute value larger than the absolute value obtained for the observed ρs [16]. This approach was used to correlate the measures of collateral effects and drug interaction to proportion of extinction and, later on, to the standardized adaptation rates. Furthermore, to account for the effect of adaptation to the single drugs (z) in the main analysis with nonstandardized adaptation rates, we performed a partial correlation analysis with z as a covariate, generally following the previously established approach [16]. For this, we first obtained the residuals from the linear regression of x on z and those of y on z, such that y corresponds to the adaptation rates of the combination. Then, to estimate the correlation coefficient between x and y, with z as a covariate, we employed the permutation test as explained above using the residuals of the corresponding regressions [16]. Experimental evolution with fixed inhibitory levels of antibiotic combinations To evaluate the effect of the starting inhibition level of the combinations, we considered a second round of evolution experiments as described above. This time, the level of inhibition of the combination was fixed instead of that of the individual drug treatments. Briefly, concentrations of each drug were mixed 1:1 so that each would inhibit between 50% and 75% of growth. These were then diluted to obtain a range of different inhibition levels and to evaluate their effect on growth in P. aeruginosa after 12 h of incubation at 37 °C. Evolution experiments were then initiated for 4 different combinations that included 11 different treatments: a no-drug control, the individual monotherapies, and 8 different inhibition levels ranging from approximately IC50 to >IC90 of each combination. Each treatment was replicated 8 times and distributed randomly in 96-well plates. Genetically fixed changes in growth characteristics We used the focused set-up of the above separate evolution experiment to validate the suitability of growth measurements as a proxy for evolutionary adaptation. Evolutionary adaptation assumes that changes are genetically fixed rather than due to phenotypic (i.e., physiological) responses. To assess this, we studied cryo-preserved material from the last drug-free season of the evolution experiment and regrew them under defined antibiotic conditions. Purely phenotypic adaptations to antibiotics are unlikely to have persisted for this material, which was grown under antibiotic-free conditions for 12 to 16 h (equivalent to a minimum of 6 generations) and additionally subjected to a cryo-preservation step. Therefore, any persistent changes in growth characteristics under antibiotic exposure are likely based on genetic changes and thus indicate evolutionary adaptation. For this analysis, we considered material evolved in the presence of 2 synergistic (i.e., GEN plus CAR and STR plus PIT) or 2 antagonistic combinations (i.e, CIP plus GEN and CIP plus TOB), in all cases set to either IC50 or >IC90, and also included material from the corresponding monotherapies. A total of 4 replicate populations was studied for each of the various evolution treatments and compared to the ancestral PA14. Changes in growth characteristics were inferred from dose-response curves in a 2-fold dilution series of each of the antibiotics included in the pair. The evolved relative changes in resistance were calculated as the area under the curve (AUC) of the dose-response curve for each of the populations and then divided by that of the ancestral PA14. The results are shown in S5 Fig. They highlight a general increase in growth characteristics and thus resistance across the various treatment groups even if not significant in all cases (based on a 1-sample Wilcoxon test with μ = 1). We conclude that, overall, the observed changes in growth characteristics have a genetic basis and are not exclusively due to phenotypic responses. Therefore, we consider the recorded changes in growth characteristics to provide a meaningful proxy for evolutionary adaptation. Supporting information S1 Fig. Dose-response curves of the ancestral strain PA14 exposed to all different antibiotics used in the study. Each panel corresponds to a single antibiotic (see Table 1 for abbreviations). Boxplots show bacterial growth relative (n = 6 per concentration) to an antibiotic-free environment across different drug concentrations. The red dotted line indicates the 75% level of inhibition (IC75) used as a standard for subsequent experiments. https://doi.org/10.1371/journal.pbio.2004356.s001 (TIF) S2 Fig. Validation of the interaction strength measure α. (A) Checkerboards of 8 selected combinations. Each panel corresponds to an antibiotic combination, here from left to right and top to bottom: CAR plus GEN, CAR plus CEF, STR plus PIT, TIC plus TOB, CIP plus CAR, CIP plus CEF, CIP plus DOR, and PIT plus CAR. Growth relative to the drug-free environment is shown over a grid of concentrations of both drugs in different shades of grey: values close to 1 indicate normal growth (black), whereas those close to 0 correspond to no detectable growth after 12 h of incubation (white). Red, grey, and blue circles embedded within each panel highlight the different types of interactions determined using α, showing—respectively—synergism, additivity, and antagonism. We calculated the degree of synergy (S) using the Bliss independence method either by averaging all obtained values across the grid where the fitness effect was measurable (panel B), or by calculating S from the combination having the same level of inhibition for each drug (panel C). A significant correlation was obtained between the degree of synergy S obtained in panels B and C with our measurements of α (as in S2 Fig). The data used for these panels are provided in S2 Data. https://doi.org/10.1371/journal.pbio.2004356.s002 (TIF) S3 Fig. Interaction profile of 52 antibiotic combinations. Each panel shows the growth of the P. aeruginosa PA14 strain across 9 different drug proportions ranging from full dose of one drug (θ = 0) to a full dose of the second one (θ = 1), each set to inhibit 75% of normal growth. Points and error bars indicate the mean and 95% CI for bacterial growth, as inferred through OD (n = 9) after 12 h of incubation. Colored lines represent the quadratic fit of observed data, whereby the color itself indicates the interaction type, as deduced from the α parameter of the model. Synergy, additivity, and antagonism are shown as red, grey, or blue lines, respectively. The data used for these panels are provided in S3 Data. OD, optical density. https://doi.org/10.1371/journal.pbio.2004356.s003 (TIF) S4 Fig. Validation of OD as a proxy for bacterial growth in the presence of antibiotics. We used the ancestral PA14 strain to replicate the first season of evolution for 6 selected single-drug treatments, 4 corresponding antibiotic combinations, and a no-drug control (shown in the different colors) in a single 96-well plate. Antibiotic concentrations were set to IC75 for the single-drug treatments, and for the drug pairs, each antibiotic was set to IC75 and then combined in a 1:1 ratio, thereby following the same set-up used for the main evolution experiment. For each treatment, we included 8 replicates. The plate was incubated at 37 °C for 12 h under continuous shaking. At the end of the incubation period, a sample from each well was taken, plated on LB agar plates, and incubated for 16 to 20 h at 37 °C to count the number of viable cells as CFUs. We found a significant correlation between the obtained CFU counts and the endpoint OD measurements (Spearman rank test, ρs = 0.782, P < 0.001). CFU, colony-forming unit; IC75, inhibiting 75% of bacterial growth; LB, Luria-Bertani; OD, optical density. https://doi.org/10.1371/journal.pbio.2004356.s004 (TIF) S5 Fig. Changes in resistance upon experimental evolution in selected populations. We determined changes in resistance for selected populations from the separate evolution experiment with different initial drug inhibitory levels (Fig 7) that included 4 different combinations: (A) GEN plus CAR, (B) STR plus PIT, (C) CIP plus GEN, and (D) TOB plus CIP. For each of these combinations, we tested 4 populations adapted to each of the single drugs, 4 populations adapted to the combinations set to IC50, 4 populations adapted to those set to >IC90, and the ancestor PA14. All populations were from the final season with antibiotics. Antibiotic resistance was assessed with dose-response curves using 2-fold dilution series of each of the antibiotics included in the tested combination. Results are given as changes in resistance (for the changes in IC90, see S2 Table), calculated as the AUC for each evolved population relative to that of the ancestral PA14. Asterisks indicate significant differences obtained from a 1-sample Wilcoxon test (μ = 1, dotted red line). All P values were corrected for multiple comparison using FDR. AUC, area under the curve; FDR, false discovery rate; IC50, concentration inhibiting 50% of bacterial growth; IC90, concentration inhibiting 90% of bacterial growth. https://doi.org/10.1371/journal.pbio.2004356.s005 (TIF) S6 Fig. Growth rate of the 50:50 treatment of all combinations over 10 seasons of experimental evolution. Each panel corresponds to an antibiotic combination, as indicated by the abbreviations in the top of each panel. Grey lines and circles show the growth rate r of each replicate within the 50:50 proportion. Orange circles and lines highlight the mean of all surviving populations per combination (note the number of replicate populations varies between combinations because of extinction; GEN plus CAR has no surviving population, and therefore it is not shown). The data used for these panels are provided in S4 Data. https://doi.org/10.1371/journal.pbio.2004356.s006 (TIF) S7 Fig. Growth rates of the monotherapies over 10 seasons of experimental evolution. Each panel corresponds to a single-drug treatment. Grey lines and circles show the growth rate r of each replicate within a single-drug treatment (note that replicates among antibiotics differ; shown in brackets). Dark cyan circles and lines highlight the mean of all surviving populations per antibiotic. The data used for these panels are provided in S4 Data. https://doi.org/10.1371/journal.pbio.2004356.s007 (TIF) S8 Fig. ACE networks for each drug interaction type. ACE networks were calculated for only synergistic (top panels), antagonistic (middle panels), and additive combinations (bottom panels). As in the main text, they were built on 2 parameters: (A, C, E) rates of adaptation and (B, D, F) extinction rates. Rates of adaptation and extinction numbers were calculated from the data provided in S4 Data. Abbreviations indicate antibiotics. A, azlocillin; ACE, antibiotic combination efficacy; C, ciprofloxacin; D, doripenem; F, cefsulodin; G, gentamicin; I, imipenem; K, carbenicillin; P, piperacillin + tazobactam; Q, ticarcillin; S, streptomycin; T, tobramycin; Z, ceftazidime. https://doi.org/10.1371/journal.pbio.2004356.s008 (TIF) S9 Fig. ACE networks for the different collateral effects. ACE networks were calculated for only collaterally sensitive drug pairs (top panels) or collaterally resistant combinations (bottom panels). As in the main text, they were built on 2 parameters: (A, C) rates of adaptation and (B, D) extinction rates. Rates of adaptation and extinction numbers were calculated from the data provided in S4 Data. Abbreviations indicate antibiotics. A, azlocillin; ACE, antibiotic combination efficacy; C, ciprofloxacin; D, doripenem; F, cefsulodin; G, gentamicin; I, imipenem; K, carbenicillin; P, piperacillin + tazobactam; Q, ticarcillin; S, streptomycin; T, tobramycin; Z, ceftazidime. https://doi.org/10.1371/journal.pbio.2004356.s009 (TIF) S10 Fig. Rates of adaptation in single-drug treatments. Rates of adaptation are shown for the treatments with only 1 antibiotic. Colors indicate the different antibiotic classes: fluoroquinolones (red), carbapenems (green), cephalosporins (gold), penicillins (orange), and aminoglycosides (light blue). The number of populations in each treatment varies depending on the number of times a given drug is part of the tested combinations and the number of extinct populations (shown in brackets for each drug). The data used for these panels are provided in S4 Data. https://doi.org/10.1371/journal.pbio.2004356.s010 (TIF) S11 Fig. Changes in growth rate of 4 selected combinations with fixed initial inhibitory levels. Each column corresponds to a specific antibiotic combination, and the rows represent the different initial inhibitory levels considered: from top to bottom are shown the no-drug controls (black), the monotherapies (turquoise and purple; the numbers given in brackets after the antibiotic abbreviation indicates which monotherapy is shown first or second), and 8 different starting levels of inhibition of the combinations (from approximately IC50 in yellow to >IC95 in dark red). Grey points and lines indicate the replicate population, while the colored points and lines show the mean per treatment and combination. Rates of adaptation, extinction numbers, and inhibitory levels were calculated from the data provided in S5 Data. IC50, concentration inhibiting 50% of bacterial growth; IC90, concentration inhibiting 90% of bacterial growth. https://doi.org/10.1371/journal.pbio.2004356.s011 (TIF) S1 Table. The α test of 52 drug combinations used against P. aeruginosa. https://doi.org/10.1371/journal.pbio.2004356.s012 (DOCX) S2 Table. Rates of adaptation of all combinations relative to the weaker and stronger components in a drug pair. https://doi.org/10.1371/journal.pbio.2004356.s013 (DOCX) S3 Table. Effect test of the initial inhibitory level, interaction type, and combination on rates of adaptation. https://doi.org/10.1371/journal.pbio.2004356.s014 (DOCX) S4 Table. Effect test of the initial inhibitory level, interaction type, and combination on the number of extinctions. https://doi.org/10.1371/journal.pbio.2004356.s015 (DOCX) S1 Data. Key to datasets (Readme file). https://doi.org/10.1371/journal.pbio.2004356.s016 (RTF) S2 Data. OD measurements after 12 h of growth in different proportions of a given drug pair for all 52 antibiotic combinations. These data were used to infer the drug interaction type using the α estimator. https://doi.org/10.1371/journal.pbio.2004356.s017 (TXT) S3 Data. OD measurements after 12 h of growth in drug checkerboards evaluating the interaction of 8 selected antibiotic combinations. These data were used to calculate the degree of synergy S, to correlate it to the values obtained for the same drug pairs using the α estimator. https://doi.org/10.1371/journal.pbio.2004356.s018 (TXT) S4 Data. OD measurements taken every 15 min for 38 antibiotic pairs during a total of 120 h. These data were then used to infer adaptation rates in surviving replicate populations and the number of extinction events occurring per combination and treatment. https://doi.org/10.1371/journal.pbio.2004356.s019 (TXT) S5 Data. OD measurements taken every 15 min for 4 selected antibiotic pairs with varying levels of inhibition during a total of 120 h. These data were then used to infer adaptation rates in surviving replicate populations and the number of extinction events occurring per combination and treatment. https://doi.org/10.1371/journal.pbio.2004356.s020 (TXT) Acknowledgments We are grateful to Julia Bunk and Christopher Blake for support in the lab; and Lutz Becks, Anette Friedrichs, and the Schulenburg lab for valuable advice.

doi: 10.1371/journal.pbio.2004920pmid: 29689046

Introduction The acquisition of new protein functions through horizontal gene transfer (HGT) is known to confer selective advantages and enable the occupation of new ecological niches [1–6]. Examples include the acquisition of antibiotic resistance [7], virulence-promoting factors [8], expanded enzymatic capability [9–17], and tolerance of environmental extremes [18,19]. In well-understood cases of HGT, the transferred genes generally encode enzymes whose functions appear to be retained in the recipients. The ability to sense gravity allows plants and fungi to orient the growth of shoots and roots, and fruiting bodies, respectively. This response, known as gravitropism, depends on sedimentation of dense cytoplasmic bodies [20–22], which generate cell elongation-promoting signals at the cell cortex. Plant gravity sensing is mediated by starch bodies that form within specialized plastids [20]. In the fungi, gravitropism has been demonstrated in the multicellular Basidiomycota [21] and the Mucorales [22]. However, gravity-sensing organelles have only been examined in the Mucoralean Phycomyces blakesleeanus [23] where giant single-celled sporangiophores exhibit gravitropism through a combination of buoyant lipid globules and sedimenting protein crystals that form within vacuoles [24]. A crystal-less mutant grows normally, but displays defective gravitropism, indicating that the crystals indeed serve as gravity sensors [24–26]. Similar structures have been observed in other members of the Mucorales [22], suggesting that this function arose early in this lineage. However, its basis and evolutionary origin remain unknown. Here, we identify the octahedral crystal matrix protein (OCTIN). Phylogenetic analyses indicate that octin was acquired from a gram-negative bacterium. Both Phycomyces crystals and bacterial OCTIN form disulfide-bonded high-order oligomers, suggesting that they share elements of a conserved assembly mechanism. Given the size of bacterial cells, thermal fluctuations are expected to dominate the movement of OCTIN oligomers. This precludes any speculated role in bacterial gravity sensing. We conclude that HGT of a bacterial octin into the common ancestor of the Mucorales is likely to have relieved constraints on OCTIN oligomer size, allowing evolution of the gravity-sensing function. The data exemplify a general mechanism for the evolution of adaptations based on HGT and high-order protein assembly. Results and discussion To determine the molecular basis of gravity sensing, we purified vacuolar crystals from P. blakesleeanus sporangiophores using the method of Ootaki and Wolken (Fig 1) [27]. As previously observed, a highly purified crystal fraction contains two major proteins, p55 and p14 (Fig 1C) [28]. Mass spectrometry indicates that peptides from these bands are derived from the N- (p14) and C-terminus (p55) of a single predicted protein, which we named OCTIN (Fig 1D). Edman degradation defines the N’-termini of p14 and p55 and full-length octin transcript is detected exclusively in sporangiophores (Figs 1D and S1A). These data indicate that p14 and p55 are derived through proteolytic processing of an OCTIN precursor. Furthermore, sequencing the octin gene from the crystal-less mutant reveals a stop codon at W326 (Figs 1D and S1B). Together, these observations identify two OCTIN-derived proteins as primary components of Phycomyces gravity-sensing crystals. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. P. blakesleeanus OCTIN crystals. (A) The upper panel shows Phycomyces asexual fruiting body development tracked over the course of 10 hours. The stalk is a single-celled sporangiophore and the sphere at its tip contains nuclei that develop into asexual spores. The asterisk indicates the approximate region where protein crystals occur. The lower panel shows a close-up view of protein crystals within the sporangiophore central vacuole. (B) Three focal planes reveal the octahedral structure of a purified crystal. The bottom panel shows a cartoon of the crystal geometry. The lightest triangular face corresponds to the first panel. The darkest triangular face corresponds to the third panel. (C) The crystal-enriched fraction analyzed by SDS-PAGE. Two prominent proteins, p55 and p14 are indicated. The asterisk identifies a 46-kDa band whose peptides are mapped to the same region as p55 by mass spectrometry. (D) The cartoon depicts the full-length OCTIN sequence. Peptides identified from p14 and p55 are shown in blue and yellow, respectively. The N-termini of the mature proteins are indicated (arrowheads). The dashed line identifies the predicted region removed through proteolytic processing based on the molecular weight of p14. An asterisk marks the position of a stop-codon in the crystal-less mutant. (E) An organismal phylogeny showing the distribution of taxa where full-length OCTIN homologs are found. Names of these taxa are in colored or black labels. Gray colored groups do not contain OCTIN. OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.g001 Full-length OCTIN is sporadically present in eukaryotes and bacteria (Figs 1E and S2). In the fungi, OCTIN is found exclusively in members of the Mucoromycotina, suggesting that it was acquired early on in this lineage. Homologs are also found in the protozoan Stramenopiles, including all sequenced Oomycetes, the Pelagophyceae diatom Aureococcus anophagefferens and both sequenced Haptophytes (the brown alga Emiliania huxleyi and the phytoplankton Chrysochromulina). OCTIN also occurs sporadically in diverse bacterial clades, where it is found in Proteobacteria, Acidobacteria, Actinobacteria, and Bacteroidetes (Fig 1E). Mucorales octin sequences do not encode a predicted signal sequence, suggesting localization through the cytoplasm-to-vacuole targeting pathway, which has been associated with the import of oligomeric vacuolar resident proteins [29]. Predicted signal sequences are found in OCTIN homologs from gram-negative bacteria and the Oomycetes, suggesting that these proteins are directed to the periplasm and secretory pathway, respectively. The sporadic distribution of OCTIN in eukaryotes (Fig 1E) could be explained by an early origin followed by extensive gene loss. However, both maximum likelihood (ML) and Bayesian analyses provide strong support for independent acquisition of OCTIN by the Mucoromycotina and Oomycetes through HGT from bacteria. In the ML tree, the Mucorales and Oomycetes each have a distinct sister bacterial group (Figs 2 and S3 and S4), while in the Bayesian tree, the Mucorales are nested within a clade of acido- and proteobacteria (S5 Fig). Enforcing eukaryote monophyly on the ML OCTIN phylogeny results in a topology significantly less likely than the unconstrained phylogeny as judged by the Shimodaira’s Approximately Unbiased (AU) test (p-value = 0.021, S1 Table). The trees further suggest HGT among bacteria: acidobacteria and proteobacteria, as well as actinobacteria and proteobacteria, are interspersed to form distinct well-supported monophyletic groups (Figs 2 and S3, S4 and S5), while the constrained topology consistent with vertical transmission is significantly less likely (AU test p-value = 0.009, S2 Table). OCTIN is found in a large number of species in deep branching clades in the proteobacteria and acidobacteria (S6 and S7 Figs), suggesting an ancient origin in bacteria. Together, the phylogenetic analyses support an origin for the gravity-sensing protein crystal through HGT from a gram-negative bacterium. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. The OCTIN phylogeny indicates multiple HGT events. The OCTIN ML phylogenetic tree supports bacteria-Mucorales, bacteria-Oomycetes, and bacteria-bacteria HGTs. Support values greater than 50 are shown as node labels. Values of 100 are represented by thick horizontal lines. The tree is rooted with a distant homolog, human FGE. Sequences with predicted N-terminal signal sequences, and gram-negative and gram-positive bacteria are marked with the indicated symbols. The various taxa are color-coded according to the legend. The full ML tree constructed from 127 OCTIN sequences is shown in S3A Fig. FGE, formylglycine-generating enzyme; HGT, horizontal gene transfer; ML, maximum likelihood; OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.g002 The OCTIN C-terminus contains a full-length formylglycine-generating enzyme (FGE) domain (Fig 3A). In metazoans, FGE catalyzes the oxidation of cysteine to Cα-formylglycine to activate sulfatase enzymes in the endoplasmic reticulum (ER). In humans, its loss-of-function causes the fatal genetic disorder multiple sulfatase deficiency (MSD) [30]. Alignment between OCTIN from diverse species and human FGE reveals high overall sequence conservation, with many residues mutated in MSD being conserved in the OCTIN FGE domain. However, key FGE catalytic cysteines are absent in OCTIN sequences, suggesting that OCTIN does not function in sulfatase activation (S8 Fig). Interestingly, many other bacterial FGE domain-containing proteins lack FGE catalytic residues, and like OCTIN, have N-terminal sequence extensions (S9 Fig). In some cases, these extensions show similarity to known domains, which include Kinase, Caspase, DinB, NATCH, and PEGA domains. DinB-FGE has been shown to function as a sulfoxide synthase. This activity depends on DinB catalytic residues that form contacts with the FGE domain [31,32]. Together, these data identify a bacterial superfamily of OCTIN-related proteins. The extent to which these function through structural or enzymatic mechanisms remains to be determined. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. The Phycomyces OCTIN crystal lattice is stabilized by intermolecular disulphide bonds. (A) Conservation of cysteine residue position shown by sequence alignment. Positions exhibiting clade-specific conservation are shown in the color of the species to which they correspond. The positions of other cysteine residues are shown in black. Connected lines above the human FGE represent cysteine pairs that form disulphide bonds. The arrowhead indicates the catalytic cysteine pair. The FGE domain is indicated by the horizontal black bar. (B) Crystal-enriched fraction separated by SDS-PAGE in the presence (+) and absence (−) of the reducing agent 2-ME. P55, but not p14, migrates as a high-molecular–weight smear in the absence of 2-ME. Note that p46 also shifts in the absence of 2-ME, suggesting that it is a processing variant of p55. (C) Stills taken from a video recording the disassembly of Phycomyces OCTIN crystals by the reducing agent DTT (S1 Movie). (D) Synergistic disassembly of Phycomyces OCTIN crystals by SDS and DTT. While SDS alone is sufficient to completely shift p14 to the supernatant after centrifugation at 100,000 x g, only the combination of SDS and DTT has the same effect on p55. 2-ME, 2-Mercaptoethanol; DTT, dithiothreitol; FGE, formylglycine-generating enzyme; OCTIN, octahedral crystal matrix protein; P, pellet; S, supernatant; SDS, sodium dodecyl sulfate; T, total. https://doi.org/10.1371/journal.pbio.2004920.g003 The position and number of OCTIN cysteine residues show significant variation between the diverse OCTIN-containing clades. However, within clades, cysteine residues can be well conserved (Fig 3A), suggesting that they tailor OCTIN to its taxa-specific functions. When Phycomyces crystals are analyzed by SDS-PAGE under non-reducing conditions, p55 shifts to a high-molecular–weight species that migrates as smear around 250 kDa. By contrast, the migration of p14 is unchanged. These data indicate that p55 forms a disulphide-bonded sub-assembly (Fig 3B). Rapid swelling and disintegration of crystals upon treatment with DTT (dithiothreitol) reveal the importance of disulphide bonds for crystal lattice stability. (Fig 3C and S1 Movie). Centrifugation confirms this effect—p55 and p14 are pelleted by centrifugation at 100,000 x g, whereas DTT treatment shifts both into the supernatant fraction. Together, these data further show that p14 associates with p55 through non-covalent interactions. Crystals also swell and dissolve upon addition of the protein denaturant sodium dodecyle sulfate (SDS) (S2 Movie). Neither DTT nor SDS fully solubilizes p55. However, when combined, they synergize to promote disassembly (Fig 3D). Together, these data show that disulphide-bonded p55 sub-assemblies form a crystal lattice through additional non-covalent interactions. p14 is physically associated with the p55 lattice. However, its role in stabilizing this structure is unclear. The origin of a gravity-sensing crystal through HGT from a gram-negative bacterium raises the important question of how bacterial OCTIN might be predisposed to this function. Bacteria descended from the likely octin donor are not currently genetically manipulable. To investigate this question, we expressed OCTIN from the gram-negative acidobacterium Terriglobus saanensis (OCTINT) in Escherichia coli. OCTINT encodes a predicted signal sequence (SST) and a SST-mCherry fusion protein is targeted to the periplasm as indicated by a fluorescent ring around the cell periphery. By contrast, a full-length OCTINT-mCherry fusion protein produces punctate fluorescence at the cell periphery (Fig 4A). Both proteins are released upon lysis of the outer membrane, indicating that they are indeed periplasmic (Fig 4B). However, only OCTINT can be pelleted by centrifugation, suggesting that patches seen by fluorescence represent stable high-order oligomers (Fig 4C). Non-reducing SDS-PAGE shows that like Phycomyces OCTIN, OCTINT forms intermolecular disulphide bonds (Fig 4D). Furthermore, as with Phycomyces OCTIN, SDS and DTT synergize to promote OCTINT oligomer disassembly (Fig 4E). Compared with Phycomyces OCTIN, DTT alone has little effect, suggesting that these assemblies rely more on non-covalent interactions. Nevertheless, these data support a related underlying mechanism of self-assembly for Phycomyces and bacterial OCTIN. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Localization and assembly of bacterial and Phycomyces OCTIN upon ectopic expression. (A) Terriglobus OCTIN assembles into patches in the periplasm. The Terriglobus OCTIN signal sequence fused to the N′-terminus of mCherry (SST-mCherry) is localized in a ring around the cell periphery, while a full-length OCTINT-mCherry fusion protein is localized in patches. (B) Western blotting shows the enrichment of SST-mCherry and OCTINT-HA in an isolated periplasmic fraction. Cytoplasmic mCherry serves as a control for contamination of the periplasmic fraction through cell lysis. (C) OCTINT-HA assembles into high-order oligomers. Periplasmic OCTINT-HA, but not SST-mCherry is pelleted by centrifugation at 100,000 x g. (D) OCTINT forms intermolecular disulphide bonds. OCTINT-HA migrates as a high-molecular–weight smear in the absence (−) but not presence (+) of 2-ME. (E) As with Phycomyces OCTIN crystals (Fig 3D), SDS and DTT synergize to promote the disassembly of OCTINT oligomers. (F) OCTINT-mCherry and SST-OCTINP-mCherry are targeted to the ER upon expression in mammalian cells. The ER is defined by sfGFP with an N-terminal SS and C′-terminal ER retention signal (KDEL). (G) Western blotting for OCTINT-HA and SST-OCTINP-HA expressed in mammalian cells shows that SST-OCTINP does not undergo proteolytic processing in the ER. Arrowhead indicates the expected position of p55. 2-ME, 2-Mercaptoethanole; BF, brightfield; DTT, dithiothreitol; ER, endoplasmic reticulum; FL; Fluorescence; OCTIN, octahedral crystal matrix protein; P, pellet; S, supernatant; SDS, sodium dodecyl sulfate; sfGFP, superfolder GFP, SS, signal sequence; T, total. https://doi.org/10.1371/journal.pbio.2004920.g004 Phycomyces sporangiophores are approximately 100 μm in diameter [33] and OCTIN crystals have an average edge length of 5 μm [27]. By contrast, octin-containing bacteria whose sizes are known have diameters ranging from 0.3 to 0.8 μm [34–39]. To the best of our knowledge, bacterial gravitropism has not been observed. Moreover, assuming an OCTIN assembly size of 1 μm or less, and taking into account cytoplasmic viscosity [40], the density of OCTIN crystals [41], and the bacterial cytoplasm [41], an estimation of sedimentation velocity based on Stokes’ law indicates that bacterial OCTIN oligomers would be too small to function as gravity sensors. The low ratio of particle movement by gravitational force relative to Brownian motion (Péclet number, [42]) for oligomers in this size range further demonstrates that their movements would be dominated by thermal fluctuations (S10 Fig and S1 Text) [22]. While the function of OCTIN in bacteria remains unknown, its ability to form high-order oligomers is likely to have predisposed neo-functionalization towards a role in gravity sensing in the Mucorales. This is likely to have required the accumulation of mutations relating to crystal lattice assembly, vacuole targeting, and proteolytic processing. If primitive assemblies were too small to function as gravity sensors (S10 Fig), what factors could account for the retention of octin? Phycomyces OCTIN crystals are found in clusters (Fig 1A), which increases their effective size and sedimentation velocity [24]. Similarly, early OCTIN oligomers could have acted as sensors by clustering. Other scenarios involving neutral selection or another function could also have played a role in the evolutionary transition. In the latter scenario, we note that presently available information does not preclude an enzymatic activity for OCTIN. The periplasm of gram-negative bacteria and the eukaryotic secretory pathway are both oxidizing environments that share a related machinery for translocation of proteins from the cytoplasm [43]. Indeed, OCTINT-mCherry is targeted to the ER when expressed in mammalian tissue culture cells (Fig 4F). To determine whether Phycomyces OCTIN (OCTINP) can self-assemble upon heterologous expression, we expressed an ER-targeted version in mammalian cells. This version of OCTIN co-localizes with an ER lumenal marker, but does not display a punctate signal, suggesting an absence of self-assembly. Western blotting further shows an absence of proteolytic processing (Fig 4G). This indicates that OCTIN crystal assembly is likely to require taxa-specific processing activities. Many vacuolar hydrolases are synthesized as auto-inhibited precursors, which are activated upon delivery to the vacuole through processing by resident proteases [44]. We speculate that the region between p14 and p55 functions to inhibit crystal lattice formation through an analogous mechanism (see S11 Fig for a model of OCTIN assembly). Phycomyces has yet to be transformed [33], and this limits its use as a model system. Thus, understanding the control of crystal assembly will require the identification of OCTIN processing factors and reconstitution in a genetically amenable model system. Phylogenetic analyses strongly support the acquisition of bacterial OCTIN by the Mucorales ancestor through HGT (Figs 2 and S3, S4 and S5). Through its signal sequence, this protein would have been targeted to the endomembrane system (Fig 4F). In this context, the size constraint on OCTIN oligomers was relieved, allowing eventual increase in assembly scale and emergence of the gravity-sensing novelty. The case of OCTIN exemplifies how HGT of a protein undergoing high-order assembly can lead to a novel function that emerges depending on a combination of cellular potentialities and physiological imperatives. Methods OCTIN identification and phylogenetic analyses P. blakesleeanus wild-type strain NRRL155 [25] and crystal-less mutant strain C2 [24] were grown as previously described [41]. Octahedral crystals were purified as previously described [27]. Bands corresponding to p14 and p55 were analyzed by mass spectrometry and Edman degradation (Alphalyse A/S, Odense, Denmark). Peptides p14, p46, and p55 identified the same P. blakesleeanus protein (National Center for Biotechnology Information [NCBI] accession: XP_018295118.1). The search for OCTIN homologs was performed with BLASTP [45] against the NCBI nonredundant database [46] using the OCTIN-specific N-terminal domain (amino acids 1–500) as the query. HMMER3 [47] was used to confirm the presence of the FGE domain (PF03781) [48] in BLAST hits. The accessions of these hits are reported in S3 Table. The extended bacterial species trees (S2 and S9 Figs) were constructed based on a previously reported microbial phylogeny [49]. The original tree, which contains multiple strains from the same species, was pruned to retain 1 strain per species whose annotated genome is available in the NCBI Reference Sequence Database (RefSeq, ftp.ncbi.nlm.nih.gov/refseq/). PhyloPhlAn [49] was used to insert additional octin-possessing species that are not present in the original tree (S4 Table). All other species trees (S4, S6 and S7 Figs) were constructed from 400 conserved protein sequences by PhyloPhlAn using RefSeq bacterial proteomes. The presence of signal sequence was predicted using Phobius [50]. Phylogenetic trees were visualized with ETE3 [51]. To construct OCTIN protein trees (Figs 2 and S3 and S5), OCTIN sequences from the NCBI reference protein database were used. MAFFT [52] with the option E-INS-i was used to obtain sequence alignments, which were trimmed using Trimal [53] at a gap threshold of 70%. ML bootstrap analysis was performed with RAxML [54] using the automatic bootstrapping option [55] (300 replicates) and the PROTGAMMAILGX model as suggested by ProtTest [56]. The human FGE sequence, which serves as the outgroup (Fig 2), was placed on the ML tree a posteriori using the RAxML option -f v [57]. Bayesian trees were constructed using MrBayes [58], run with 12 chains, temperature 0.05, sampling every 500th generation for 300,000 generations. Convergence was assessed using RWTY [59]. The ML and Bayesian phylogenies, as well as the matrix used to derive them are accessible under the identifier S22330 at TreeBASE (https://treebase.org/). To compare the ML trees with and without the monophyly constraint, the best-scoring tree with monophyly constraint was constructed using RAxML with the same parameters specified above for the construction of unconstrained trees. Phylogenetic hypothesis testing using the resampling estimated log-likelihood (RELL) test, Shimodaira–Hasegawa (SH) test, Kishino–Hasegawa (KH) test, and AU test was then performed with the PAML package ‘codeml’ [60] and CONSEL [61]. Identification of FGE domain-containing proteins The search for FGE domain-containing proteins was performed with HMMER3 [47], using the FGE alignment (PF03781) downloaded from http://pfam.xfam.org. The search was performed on RefSeq proteomes of species present in the bacterial phylogeny shown in S2 Fig. Sequences containing at least 100 amino acids upstream of the FGE domains were selected. Annotated domains within these sequences were identified using the hmmscan function of HMMER3 [47]. Homologs of the gliding motility protein GldK, whose N-terminal domain is not annotated, were manually added based on similiarity to the known GldK sequence from Flavobacterium johnsoniae (NCBI accession: AAW78679.1). Recombinant OCTIN expression T. saanensis octin was codon-optimized for expression in E. coli and the synthetic sequence was obtained from Genscript. Full-length octinP was amplified by reverse transcription polymerase chain reaction (RT-PCR) from Phycomyces sporangiophore total RNA. Octin sequences and mCherry fusions were integrated into the pETDuet-1 vector (Novagen, cat #71146) for transformation in E. coli strain HMS174 (Novagen, cat #69453). Primers used in generating the expression plasmids are listed in S5 Table. E. coli periplasmic extract was obtained following a previously described protocol [62] with modifications. The induced culture was centrifuged at 2,500 x g and 4 oC for 10 minutes. The pellet was then gently resuspended in ice-cold PE buffer (20% sucrose, 1 mM EDTA, 50 mM Tris pH 7.4) and placed on a nutating mixer at 4 oC for 15 minutes. This was followed by centrifugation at 2,500 x g and 4°C for 10 minutes. The supernatant was transferred to a clean tube and supplemented with Halt protease and phosphatase inhibitor cocktail (ThermoFisher 78440). This extract was aliquoted and flash-frozen for disassembly assays and western blot. OCTINT and mCherry variants were detected by western blotting using horseradish-peroxidase–conjugated rat anti-HA antibodies (ROCHE, cat# 12013819001) or mouse anti-mCherry (SAB2702286 SIGMA) and secondary goat anti-mouse IgG (SAB4600004 SIGMA). Blot images were acquired using the ChemiDoc Touch Imaging System (Bio-Rad). OCTIN crystal and bacterial oligomer disassembly Crystals suspended in Tris-buffered saline buffer (TBS; 10 mM Tris pH 7.2, 150 mM NaCl) were mounted on a microscope slide. DTT or SDS was added to one side of the coverslip to a final concentration of 50 mM or 0.1%, respectively. Crystal disassembly was recorded using an epifluorescence microscope (BX51; Olympus) and a digital camera (Coolsnap HQ; Photometrics) controlled by Metamorph. Synergistic disassembly of Phycomyces crystals (Fig 3D) and bacterial OCTIN oligomers (Fig 4E) by SDS and DTT was performed by incubating the crystals or periplasmic extract with the indicated combinations of SDS and DTT for 30 minutes at 25 oC. This was followed by centrifugation at 100,000 x g for 30 minutes at 25 oC. The total sample and the resulting supernatant and pellet fractions were analyzed by SDS-PAGE. E. coli imaging Overnight cultures of transformed HMS174 cells were diluted into fresh media and allowed to grow to OD600 of 0.7 before induction with 1 mM IPTG. After 4 hours, 5 μl of the suspension was diluted into 1 ml of fresh LB media and 300 μl was placed on a 35-mm microscopy dish (Matek P35G-1.5-10-C) that had been pre-treated with 50 μg/ml poly-D-lysine (Sigma P7886). After 1 hour the media was removed and replaced with 2 ml of fresh media. Imaging was carried out with a Leica SP8 inverted laser-scanning confocal microscope fitted with a white-light laser and 100x lens of numerical aperture (NA) 1.4. Each image is composed of 4 averaged frames taken at 1% laser power at 587-nm excitation with a scan speed of 400 MHz. Mammalian cell culture HeLa cells cultured in 6-well dishes or 8-well chamber slides were transiently transfected with the indicated plasmids using lipofectamine 3000 (ThermoFisher) and cultured for 48 hours before fixing for microscopy or harvesting for western blot analysis. Cells were fixed with 4% Paraformaldehyde (EMS #15700) in phosphate buffered saline (PBS) and then kept in 90% glycerol PBS for imaging. Imaging was carried out using a Leica SP8 fitted with a 63x objective NA of 1.4. The white-light laser was set to 488 nm and 587 nm for GFP and mCherry, respectively. Images are a single z plane taken with 8 line averages at 5% laser power, with a scan speed of 200 MHz, 50% gain and a pixel size of 70 nm. To extract protein for western blotting, HeLa cells were lysed in RIPA buffer (50 mM Tris-HCl pH7.4, 150 mM NaCl, 1% Triton-X100, 0.1% Sodium Deoxycholate, 1% SDS) supplemented with Halt protease and phosphatase inhibitor cocktail. Insoluble material was pelleted at 10,000 x g and the supernatant fraction boiled in SDS-PAGE loading dye. 10 μg of total cell extract was run per lane. Western blotting was carried out as stated above. OCTIN identification and phylogenetic analyses P. blakesleeanus wild-type strain NRRL155 [25] and crystal-less mutant strain C2 [24] were grown as previously described [41]. Octahedral crystals were purified as previously described [27]. Bands corresponding to p14 and p55 were analyzed by mass spectrometry and Edman degradation (Alphalyse A/S, Odense, Denmark). Peptides p14, p46, and p55 identified the same P. blakesleeanus protein (National Center for Biotechnology Information [NCBI] accession: XP_018295118.1). The search for OCTIN homologs was performed with BLASTP [45] against the NCBI nonredundant database [46] using the OCTIN-specific N-terminal domain (amino acids 1–500) as the query. HMMER3 [47] was used to confirm the presence of the FGE domain (PF03781) [48] in BLAST hits. The accessions of these hits are reported in S3 Table. The extended bacterial species trees (S2 and S9 Figs) were constructed based on a previously reported microbial phylogeny [49]. The original tree, which contains multiple strains from the same species, was pruned to retain 1 strain per species whose annotated genome is available in the NCBI Reference Sequence Database (RefSeq, ftp.ncbi.nlm.nih.gov/refseq/). PhyloPhlAn [49] was used to insert additional octin-possessing species that are not present in the original tree (S4 Table). All other species trees (S4, S6 and S7 Figs) were constructed from 400 conserved protein sequences by PhyloPhlAn using RefSeq bacterial proteomes. The presence of signal sequence was predicted using Phobius [50]. Phylogenetic trees were visualized with ETE3 [51]. To construct OCTIN protein trees (Figs 2 and S3 and S5), OCTIN sequences from the NCBI reference protein database were used. MAFFT [52] with the option E-INS-i was used to obtain sequence alignments, which were trimmed using Trimal [53] at a gap threshold of 70%. ML bootstrap analysis was performed with RAxML [54] using the automatic bootstrapping option [55] (300 replicates) and the PROTGAMMAILGX model as suggested by ProtTest [56]. The human FGE sequence, which serves as the outgroup (Fig 2), was placed on the ML tree a posteriori using the RAxML option -f v [57]. Bayesian trees were constructed using MrBayes [58], run with 12 chains, temperature 0.05, sampling every 500th generation for 300,000 generations. Convergence was assessed using RWTY [59]. The ML and Bayesian phylogenies, as well as the matrix used to derive them are accessible under the identifier S22330 at TreeBASE (https://treebase.org/). To compare the ML trees with and without the monophyly constraint, the best-scoring tree with monophyly constraint was constructed using RAxML with the same parameters specified above for the construction of unconstrained trees. Phylogenetic hypothesis testing using the resampling estimated log-likelihood (RELL) test, Shimodaira–Hasegawa (SH) test, Kishino–Hasegawa (KH) test, and AU test was then performed with the PAML package ‘codeml’ [60] and CONSEL [61]. Identification of FGE domain-containing proteins The search for FGE domain-containing proteins was performed with HMMER3 [47], using the FGE alignment (PF03781) downloaded from http://pfam.xfam.org. The search was performed on RefSeq proteomes of species present in the bacterial phylogeny shown in S2 Fig. Sequences containing at least 100 amino acids upstream of the FGE domains were selected. Annotated domains within these sequences were identified using the hmmscan function of HMMER3 [47]. Homologs of the gliding motility protein GldK, whose N-terminal domain is not annotated, were manually added based on similiarity to the known GldK sequence from Flavobacterium johnsoniae (NCBI accession: AAW78679.1). Recombinant OCTIN expression T. saanensis octin was codon-optimized for expression in E. coli and the synthetic sequence was obtained from Genscript. Full-length octinP was amplified by reverse transcription polymerase chain reaction (RT-PCR) from Phycomyces sporangiophore total RNA. Octin sequences and mCherry fusions were integrated into the pETDuet-1 vector (Novagen, cat #71146) for transformation in E. coli strain HMS174 (Novagen, cat #69453). Primers used in generating the expression plasmids are listed in S5 Table. E. coli periplasmic extract was obtained following a previously described protocol [62] with modifications. The induced culture was centrifuged at 2,500 x g and 4 oC for 10 minutes. The pellet was then gently resuspended in ice-cold PE buffer (20% sucrose, 1 mM EDTA, 50 mM Tris pH 7.4) and placed on a nutating mixer at 4 oC for 15 minutes. This was followed by centrifugation at 2,500 x g and 4°C for 10 minutes. The supernatant was transferred to a clean tube and supplemented with Halt protease and phosphatase inhibitor cocktail (ThermoFisher 78440). This extract was aliquoted and flash-frozen for disassembly assays and western blot. OCTINT and mCherry variants were detected by western blotting using horseradish-peroxidase–conjugated rat anti-HA antibodies (ROCHE, cat# 12013819001) or mouse anti-mCherry (SAB2702286 SIGMA) and secondary goat anti-mouse IgG (SAB4600004 SIGMA). Blot images were acquired using the ChemiDoc Touch Imaging System (Bio-Rad). OCTIN crystal and bacterial oligomer disassembly Crystals suspended in Tris-buffered saline buffer (TBS; 10 mM Tris pH 7.2, 150 mM NaCl) were mounted on a microscope slide. DTT or SDS was added to one side of the coverslip to a final concentration of 50 mM or 0.1%, respectively. Crystal disassembly was recorded using an epifluorescence microscope (BX51; Olympus) and a digital camera (Coolsnap HQ; Photometrics) controlled by Metamorph. Synergistic disassembly of Phycomyces crystals (Fig 3D) and bacterial OCTIN oligomers (Fig 4E) by SDS and DTT was performed by incubating the crystals or periplasmic extract with the indicated combinations of SDS and DTT for 30 minutes at 25 oC. This was followed by centrifugation at 100,000 x g for 30 minutes at 25 oC. The total sample and the resulting supernatant and pellet fractions were analyzed by SDS-PAGE. E. coli imaging Overnight cultures of transformed HMS174 cells were diluted into fresh media and allowed to grow to OD600 of 0.7 before induction with 1 mM IPTG. After 4 hours, 5 μl of the suspension was diluted into 1 ml of fresh LB media and 300 μl was placed on a 35-mm microscopy dish (Matek P35G-1.5-10-C) that had been pre-treated with 50 μg/ml poly-D-lysine (Sigma P7886). After 1 hour the media was removed and replaced with 2 ml of fresh media. Imaging was carried out with a Leica SP8 inverted laser-scanning confocal microscope fitted with a white-light laser and 100x lens of numerical aperture (NA) 1.4. Each image is composed of 4 averaged frames taken at 1% laser power at 587-nm excitation with a scan speed of 400 MHz. Mammalian cell culture HeLa cells cultured in 6-well dishes or 8-well chamber slides were transiently transfected with the indicated plasmids using lipofectamine 3000 (ThermoFisher) and cultured for 48 hours before fixing for microscopy or harvesting for western blot analysis. Cells were fixed with 4% Paraformaldehyde (EMS #15700) in phosphate buffered saline (PBS) and then kept in 90% glycerol PBS for imaging. Imaging was carried out using a Leica SP8 fitted with a 63x objective NA of 1.4. The white-light laser was set to 488 nm and 587 nm for GFP and mCherry, respectively. Images are a single z plane taken with 8 line averages at 5% laser power, with a scan speed of 200 MHz, 50% gain and a pixel size of 70 nm. To extract protein for western blotting, HeLa cells were lysed in RIPA buffer (50 mM Tris-HCl pH7.4, 150 mM NaCl, 1% Triton-X100, 0.1% Sodium Deoxycholate, 1% SDS) supplemented with Halt protease and phosphatase inhibitor cocktail. Insoluble material was pelleted at 10,000 x g and the supernatant fraction boiled in SDS-PAGE loading dye. 10 μg of total cell extract was run per lane. Western blotting was carried out as stated above. Supporting information S1 Fig. Cell-type–specific octin expression and sequencing of the crystal-less mutant. (A) The full-length octin transcript is expressed exclusively in sporangiophores. RNA was extracted from the indicated cell types and subjected to RT-PCR to amplify the indicated cDNAs. The octin primers are designed to amplify the entire predicted open reading frame. (B) Premature stop codon in the octin open reading frame of the crystal-less mutant. Chromatograms of WT and crystal-less strains and alignment to the reference sequence (XM_018441888.1). Translated sequence is shown below the corresponding nucleotides. The base substitution resulting in the premature stop codon is highlighted in gray. Numbers on the left represent the starting positions of the nucleotide and protein sequences. This figure was generated with Benchling (benchling.com). RT-PCR, reverse transcription polymerase chain reaction; WT, wild type. https://doi.org/10.1371/journal.pbio.2004920.s001 (TIF) S2 Fig. Extended bacterial species tree showing the sporadic distribution of OCTIN-containing species. Octin-possessing species are indicated by red bars. OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s002 (TIF) S3 Fig. ML phylogenetic analysis supports OCTIN HGT between bacteria. Bootstrap supports greater than 50 are shown as node labels. The tree is rooted with the human FGE sequence. Values of 100 are represented by thick horizontal lines. Taxa are color-coded according to the legend. FGE, formylglycine-generating enzyme; HGT, horizontal gene transfer; ML, maximum likelihood; OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s003 (TIF) S4 Fig. Bacterial species tree constructed from 400 conserved protein sequences. Shimodaira–Hasegawa support value is shown at the corresponding branch. Taxa are color-coded according to the legend. https://doi.org/10.1371/journal.pbio.2004920.s004 (TIF) S5 Fig. The Bayesian inference of the OCTIN phylogeny. (A) Bayesian tree constructed from the same eukaryotic and bacterial OCTIN sequences as shown in S3A Fig. Taxa are color-coded according to the legend. (B) Convergence assessment of the Bayesian OCTIN trees performed using RWTY [59]. OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s005 (TIF) S6 Fig. Distribution of acido- and proteobacterial OCTINs suggests an ancient bacterial origin. High density of OCTIN-containing species in an acidobacterial clade. Species containing OCTIN are shown in purple. The species tree was constructed from 400 conserved protein sequences with annotated genomes. The tree is rooted using proteobacteria whose names are in gray. Shimodaira–Hasegawa branch support values are shown as node labels. OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s006 (TIF) S7 Fig. High density of OCTIN-containing species in the proteobacterial clade Xanthomonadales. Species containing OCTIN are in blue. The Xanthomonadales species tree was constructed from 400 conserved protein sequences with annotated genomes. The tree is rooted using proteobacteria whose names are in gray. Shimodaira–Hasegawa branch support values are shown as node labels. OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s007 (TIF) S8 Fig. Alignment of the FGE domain from the indicated species. Catalytic residues required for sulfatase activation by human FGE are highlighted in red and shown in bold font. Mutations resulting in MSD are shown in parentheses above the alignment. Secondary structural features defined by human FGE crystal structure are identified with black bars and labeled. Cysteine residues colored yellow form an intramolecular disulfide bridge in human FGE and pFGE. Residues associated with calcium binding in FGE and pFGE are shown in blue. Species names are colored according to S3 Fig legend. FGE, formylglycine-generating enzyme; MSD, multiple sulfate deficiency; pFGE, FGE paralog. https://doi.org/10.1371/journal.pbio.2004920.s008 (TIF) S9 Fig. The OCTIN superfamily: Distribution of FGE domain-containing protein families in bacteria. The presence of different protein subfamilies is indicated by colored bars. Members of the FGE subfamily possess catalytic residues and do not contain other domains. FGE domain-containing proteins whose N-terminal extension shows similarity to a known domain are color-coded according to the legend. Those containing novel domains are indicated by black bars. All of these lack FGE catalytic cysteine residues. FGE, formylglycine-generating enzyme; OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s009 (TIF) S10 Fig. Estimated sedimentation properties of hypothetical OCTIN assemblies of varying diameters. (A) Sedimentation velocity estimated based on Stokes’ law, taking into account cytoplasmic density and viscosity (S1 Text). The cross indicates the reported sedimentation velocity of Phycomyces crystal clusters [26]. The estimated particle size corresponding to this sedimentation velocity is in agreement with actual cluster size [26]. The inset shows estimated sedimentation velocity of sub-micron particles in μm/minute. Grey region indicates the size range of cytoplasmic particles in OCTIN-possessing bacteria, given the cell diameter range of 0.3–0.8 μm (S1 Text). (B) Péclet number of hypothetical OCTIN assemblies. The cross indicates the Péclet number corresponding to the Phycomyces crystal cluster documented in reference [26]. Note that thermal fluctuations dominate the movement of assemblies in the size range of bacterial cytoplasmic bodies. OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s010 (TIF) S11 Fig. Model for formation of the OCTIN crystal lattice. The boxed cartoon depicts different regions of full-length OCTIN. The formation of a 3-dimensional protein lattice requires a minimum of 3 intermolecular contacts. For simplicity, the disulphide crosslinked p55 sub-assembly is depicted as a trimer and the lattice is depicted in two dimensions. We speculate that non-covalent contacts required for assembly are shielded by the region between p14 and p55, which is removed by proteolytic processing. The folding and processing events could occur simultaneously or in the opposite order to that depicted. Note that the role of p14 in lattice assembly remains unclear. OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s011 (TIF) S1 Movie. Disintegration of Phycomyces OCTIN crystals by DTT. This movie complements Fig 3C. DTT, dithiothreitol; OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s012 (AVI) S2 Movie. Disintegration of Phycomyces OCTIN crystals by SDS. This movie complements Fig 3C. OCTIN, octahedral crystal matrix protein; SDS, sodium dodecyl sulfate. https://doi.org/10.1371/journal.pbio.2004920.s013 (AVI) S1 Text. Biophysical constraints preclude OCTIN’s function in bacterial gravitropism. OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s014 (PDF) S1 Table. Comparison between the best-scoring ML trees constructed from representative bacterial and eukaryotic OCTINs with and without the eukaryote monophyly constraint. ML, maximum likelihood; OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s015 (DOCX) S2 Table. Comparison between the best-scoring ML tree with and without the acidobacteria and proteobacteria monophyly constraint. ML, maximum likelihood. https://doi.org/10.1371/journal.pbio.2004920.s016 (DOCX) S3 Table. NCBI accessions of OCTIN homologs used to construct the OCTIN phylogenies. NCBI, National Center for Biotechnology Information; OCTIN, octahedral crystal matrix protein. https://doi.org/10.1371/journal.pbio.2004920.s017 (XLSX) S4 Table. Accession and assembly details of proteomes imputed into the microbial phylogeny. https://doi.org/10.1371/journal.pbio.2004920.s018 (XLSX) S5 Table. Primers used to generate octin constructs. https://doi.org/10.1371/journal.pbio.2004920.s019 (XLSX) Acknowledgments We thank Alexander Idnurm for providing the Phycomyces crystal-less mutant.

doi: 10.1371/journal.pbio.2003611pmid: 29630591

Introduction Throughout the central nervous system (CNS), the volume transmitter nitric oxide (NO) has been implicated in controlling synaptic function by multiple mechanisms, including modulation of transmitter release, plasticity, or neuronal excitability [1–3]. NO-mediated posttranslational modifications (PTMs) in particular have become increasingly recognized as regulators of specific target proteins [4]. S-nitrosylation is a nonenzymatic and reversible PTM resulting in the addition of a NO group to a cysteine (Cys) thiol/sulfhydryl group, leading to the generation of S-nitrosothiols (SNOs). In spite of the large number of SNO-proteins thus far identified, the functional outcomes and mechanisms of the underlying specificity of S-nitrosylation in terms of target proteins and Cys residues within these proteins are not clear. Synaptic transmitter release is controlled by multiple signaling proteins and involves a cascade of signaling steps [5]. This process requires the assembly of the soluble N-ethyl-maleimide-sensitive fusion protein Attachment Protein Receptor (SNARE) complex and associated proteins, the majority of which can be regulated to modulate synapse function. Regulatory mechanisms include phosphorylation of SNARE proteins [6] as well as SNARE-binding proteins such as complexin (cpx), which have been reported at different synapses such as the Drosophila neuromuscular junction (NMJ) [7] or in the rat CNS [8]. Several contrasting effects on transmitter release are induced by NO-mediated PTMs [9]. Other forms of protein modification to modulate cellular signaling include prenylation, an attachment of a farnesyl or geranyl-geranyl moiety to a Cys residue in proteins harboring a C-terminal CAAX prenylation motif. This process renders proteins attached to endomembrane/endoplasmic reticulum (ER) and Golgi structures until further processing, as shown for Rab GTPases [10–12]. Farnesylation also regulates mouse cpx 3/4 [13] and Drosophila cpx function [14–17]. The Cys within CAAX motifs can also undergo S-nitrosylation, which interferes with the farnesylation signaling [18]; however, direct evidence in a physiological environment is lacking. Cpx function has been studied in many different systems and there is controversy regarding its fusion-clamp activity. Cpx supports Ca2+-triggered exocytosis but also exhibits a clamping function [19–24]. Analysis of mouse cpx double-knockout neurons lacking cpx 1 and 2 found only a facilitating function for cpx on release, and different D. melanogaster and Caenorhabditis elegans cpx mutant lines exhibit altered phenotypes in clamping or priming/fusion function [14–17, 24–27], illustrating the controversial actions of cpx. Here, we investigated the effects of NO on synaptic transmission and found that NO reduces Ca2+-triggered release as well as the size of the functional vesicle pool, which was reversed by glutathione (GSH) signaling. At the same time, spontaneous release rates were negatively affected by NO. We confirmed that cpx is S-nitrosylated and that NO changes the synaptic localization of cpx, as also seen following genetic and pharmacological inhibition of farnesylation. Thus, we propose that the function of cpx is regulated by S-nitrosylation of Cys within the CAAX motif to prevent farnesylation. This increases cpx-SNARE-protein interactions, thereby rendering cpx with a dominant clamping function, which suppresses both spontaneous and evoked release. Results NO-induced suppression of evoked and spontaneous synaptic release is independent of cGMP Previously, we found that enhancing endogenous nitric oxide synthase (NOS) activity induced by overexpression of D. melanogaster NOS (DmNOS) caused a reduction in synaptic strength at the Drosophila NMJ synapse [28]. To examine the effects of NO on glutamatergic transmission in more detail, we exposed wild-type (WT) w1118 control (Ctrl) larvae to NO donors, which provide an estimated NO concentration of about 200 nM [29]. When recording evoked excitatory junction currents (eEJCs) up to 70 min during NO incubation, the amplitudes started to decline significantly after 35 min (Fig 1A and S1 Data, p < 0.05; n = 3 each). Mean eEJC amplitudes and quantal content (QC) at 50 min for Ctrl (122 ± 7 nA, QC: 200 ± 15, n = 20–22) and NO treatment (59 ± 7 nA, QC: 93 ± 10, n = 14) are shown in Fig 1B. As the canonical NO-cGMP pathway is active in Drosophila [30] and potentially responsible for this observation, we blocked the soluble guanylyl cyclase (sGC) with 1H-[1,2,4]oxadiazolo[4,3-a]quinoxalin-1-one (ODQ, 50 μM). Interestingly, ODQ did not prevent the effects of NO, suggesting a cGMP-independent mechanism (amplitudes: Ctrl + ODQ: 127 ± 5 nA, NO + ODQ: 70 ± 7 nA, QC: Ctrl + ODQ: 200 ± 22, NO + ODQ: 130 ± 11, Fig 1B, n = 10–16). As Drosophila has endogenous NO signaling and produces neuronal NO in a Ca2+/calmodulin-dependent manner [31, 32], we used NOS knockout-like (NOS “null”) larvae to assess endogenous NO modulation of release. We used two different lines with strongly reduced DmNOS showing NOS “null” activity (NOSC and NOSΔ15 [33, 34]) and we would expect that lack of endogenous NO generation has the opposite effects on release. When recording eEJCs, both genotypes exhibited a tendency towards larger eEJC amplitudes and QC (Fig 1C) and, in addition, we detected an increased presynaptic release probability (pvr) in NOSC NMJs, as indicated by the reduced paired pulse ratio (PPR) at 20 ms ISI (0.80 ± 0.03 [n = 11], p = 0.002, Student t test) compared to WT Ctrls (0.93 ± 0.03 [n = 17]), indicating endogenous nitrergic effects on release probabilities. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. NO reduces evoked release and frequency of spontaneous release in a cGMP-independent manner. (A) NO suppresses evoked release (eEJC) over a time course of 55 min. Insets show representative single eEJCs at 40 min for both conditions. (B) Mean eEJC amplitudes (left axis) and QC (right axis) of w1118 NMJs are reduced following NO exposure (at 40 min). The sGC inhibitor ODQ (50 μM) did not affect the response to NO. (C) Mean eEJC amplitudes (left axis, black) and QC (right axis, grey) of NOSC and NOSΔ15 NMJs. (D) Raw mEJC recordings of w1118 NMJs and mEJC parameters (top to bottom: amplitude, frequency, decay). Top insets show representative mEJC recordings. Bottom insets show single mEJCs (grey) and averaged mEJC (red) with single exponential fit to the decay. (E) Raw mEJC recordings for both NOSC and NOSΔ15 genotypes. Below, mEJC quantal parameters: amplitude and frequency, Student t test each relative to w1118 Ctrl, *p = 0.04, ***p = 0.001. (F) cGMP content of larval brains under the conditions indicated (NO: 40 min NO exposure, NO + ODQ: 40 min NO exposure in the presence of 50 μM ODQ, NO + Zap: 40 min NO exposure + PDE inhibitor Zap, 20 μM). (G) FlincG3 fluorescence images of a Ctrl and stimulated NMJ (20 Hz for 10 s, duty cycle: 1 min for total of 10 min). (H) Summary of FlincG3 fluorescence (in a.u.’s). The raw data can be found in S1 Data. Data denote mean ± SEM in all graphs, ANOVA with post hoc Tukey-Kramer, *p < 0.05, **p < 0.01, ****p < 0.0001. a.u., arbitrary unit; cGMP, cyclic guanosine monophosphate; Ctrl, control; eEJC, evoked EJC; EJC, excitatory junction current; mEJC, miniature EJC; NMJ, neuromuscular junction; NO, nitric oxide; ODQ, 1H-[1,2,4]oxadiazolo[4,3-a]quinoxalin-1-one; PDE, phosphodiesterase; QC, quantal content; sGC, soluble guanylyl cyclase; Zap, zaprinast. https://doi.org/10.1371/journal.pbio.2003611.g001 To further understand the effects of NO on release, we analyzed miniature EJCs (mEJCs) under the same conditions. NO had no effect on mEJC amplitudes or decay kinetics; however, the frequency was reduced following NO and NO+ODQ incubation (Ctrl: 2.0 ± 0.2 nA [n = 25], NO: 1.1 ± 0.1 nA [n = 16], NO+ODQ: 1.0 ± 0.2 nA [n = 8], ODQ: 1.7 ± 0.2 nA [n = 11], Ctrl versus NO: p < 0.01, Ctrl versus NO+ODQ: p < 0.05, Fig 1D). This suggests that NO is unlikely to affect synaptic vesicle filling or composition/activity and density of postsynaptic D. melanogaster glutamate receptors (DmGluR) [35]. We tested miniature events in the NOS “null” mutants and confirmed a further inhibitory role of NO signaling on release, with mEJC frequencies being significantly enhanced in NOSΔ15 (3.5 ± 0.5 s−1 [n = 4], p = 0.001) and NOSC (3.5 ± 0.4 s−1 [n = 16], p = 0.04) larvae compared to Ctrl (Fig 1E), without affecting mEJC amplitudes (NOSΔ15: 0.8 ± 0.1 nA [n = 13], NOSC: 1.1 ± 0.3 nA [n = 3] Fig 1E) or decay kinetics (NOSΔ15: 8.9 ± 0.6 ms [n = 12], NOSC: 9.4 ± 0.3 ms [n = 4], p > 0.05 versus Ctrl). Thus, reduction of endogenous NOS activity shows opposite effects to elevation of NO levels, confirming the inhibitory action of NO on evoked and spontaneous vesicle release. As the data imply cGMP-independent signaling, we wanted to confirm that cGMP levels are not altered following NO stimulation. Thus, we measured cGMP directly in isolated larval brains. NO application did not raise cGMP levels (at 50 min: Ctrl: 2.4 ± 0.5 pmol/mg, NO: 3.0 ± 0.6 pmol/mg, p > 0.05 [n = 30 each], Fig 1F). Cyclase inhibition in the presence of NO did not significantly reduce cGMP levels, confirming lack of NO-induced neuronal cGMP accumulation. We found that any generated cGMP was broken down by phosphodiesterase DmPDE5/6 [36], as cGMP increased following NO stimulation only with PDE inhibition (20 μM zaprinast [Zap]; NO+Zap: 50.2 ± 8.3 pmol/mg, p < 0.0001), while Zap alone had no effect (Zap: 4.6 ± 2.0 pmol/mg, p > 0.05). To assess whether NO is produced endogenously to induce modulation of synaptic function as observed above, we expressed FlincG3 presynaptically and stimulated NMJs at 20 Hz (for 10 s every minute for 20 min). As shown in Fig 1G, 20 Hz stimulation induced a significant increase in fluorescence, confirming endogenous presynaptic generation of NO (Ctrl: 62 ± 4 arbitrary units [a.u.’s], Stim: 96 ± 8 a.u.’s, Fig 1H [n = 13–15 boutons], p < 0.01). Importantly, addition of the NO donor did not further increase the fluorescence, indicating that activity-induced synaptic NO concentrations reach similar levels (NO: 93 ± 7 a.u.’s). A potential target of NO signaling is mitochondria [37], which are required for the energy to maintain vesicle recycling and synaptic transmission [38]. Thus, we measured mitochondrial activity in third instar larvae under the same conditions (50 min NO incubation) and found that mitochondrial activity was unaffected by NO (S1 Fig and S9 Data), suggesting that the effects of NO on synaptic transmission are not due to ATP depletion. Together, these data suggest that NO has a presynaptic effect on transmitter release, which is independent of cGMP signaling. Ca2+ dependency of evoked release is reduced by NO Several mechanisms contribute to the regulation of synaptic strength [39], including altered pvr, alterations in the number of readily releasable vesicles and release sites (N) or quantal size (q). Alterations in q are likely not involved in the NO-induced effects observed based on our mEJC data above (Fig 1). We next assessed additional release parameters, including pvr, N, vesicle pool size, and Ca2+ dependency of release in NOS “null” and WT NMJs following nitrergic signaling. We determined pool size via a method successfully applied at the Drosophila NMJ, by analyzing the cumulative QC of trains of higher frequency stimulation [40]. Stimulation at 50 Hz for 500 ms in 1.5 mM extracellular calcium concentration ([Ca2+]e) retrieves vesicles from the readily releasable pool (RRP) [41]. This stimulation pattern induced mild depression in Ctrls and strong initial facilitation of trains under NO conditions (Fig 2A and S2 Data). Cumulative QC analysis revealed a pool size of 453 ± 37 (n = 17) in Ctrl and 185 ± 18 in NO-exposed NMJs (n = 16, p < 0.01), suggesting a strong reduction in ready-releasable/recycling vesicles (Fig 2A–2D). Supporting the above data, pool size estimation in the presence of ODQ confirmed cGMP independence (NO+ODQ: 310 ± 33 [n = 9], p < 0.05 versus Ctrl; ODQ alone: 501 ± 34 [n = 9], p > 0.05 versus Ctrl, Fig 2A–2D). And importantly, analysis of the vesicle pool sizes in NOS “null” mutants revealed a strong 2-fold increase compared to w1118 Ctrl and an over 5-fold increase compared to NO application (NOSC: 975 ± 161 [n = 11]; NOSΔ15: 958 ± 139 [n = 4], p < 0.001 versus Ctrl, Fig 2A–2D). To exclude any potential developmental effects caused by NOS deficiency that could account for these strong increases in release, we assessed NMJ morphology and ultrastructure. We analyzed the total volume of NMJs (horseradish peroxidase [HRP] signal) and the number of Bruchpilot (Brp) puncta/NMJ volume of z-stack confocal images (S2A and S2B Fig and S9 Data) and measured the number of AZs, T-bars per Ib bouton, and vesicles within a 250-nm semicircle around the AZ (S2C and S2D Fig). These data indicated that reduced NOS activity has no developmental impact on the structure of NMJs and synaptic boutons and can therefore not explain the physiological differences observed above. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. The size of the RRP and RP is reduced by NO and enhanced in NOS “null” NMJs. (A) Raw eEJC recordings of a 50-Hz train, 500 ms. (B) Mean QC for conditions indicated. (C) Cumulative QC with back extrapolated linear regression to the last 200 ms to time 0, yielding an estimated readily releasable vesicle pool size (RRP, intercept with y-axis, pool sizes in italics). (D) Mean RRP sizes for conditions indicated. NO reduces PPR sizes, with ODQ having no effects. NOS “null” NMJs show potentiated responses (ANOVA with post hoc Tukey-Kramer was used for comparisons). (E) Raw eEJC recordings of a 50-Hz train, 8 s. (F) Mean QC for conditions indicated. (G) Cumulative QC with back extrapolated linear regression to the last 2 s to time 0 (w1118 Ctrl and NO), yielding an estimated RP size (intercept with y-axis, pool sizes in italics). (H) Mean RP sizes for conditions indicated. NO reduces RP sizes similarly to RRP sizes. The raw data can be found in S2 Data. Student t test, **p = 0.0062, data denote mean ± SEM in all graphs. Ctrl, control; eEJC, evoked EJC; EJC, excitatory junctional current; NMJ, neuromuscular junction; NO, nitric oxide; NOS, nitric oxide synthase; ODQ, 1H-[1,2,4]oxadiazolo[4,3-a]quinoxalin-1-one; PPR, paired pulse ratio; QC, quantal content; RP, reserve pool; RRP, readily releasable pool. https://doi.org/10.1371/journal.pbio.2003611.g002 In addition to changes in release, NO could also exert its effects indirectly via modulating transmitter uptake and pool recovery. To exclude this possibility that altered recovery from depression affected the above pool estimations, we examined eEJC recovery. Following depletion of vesicle pools during a 50-Hz train (1 s), we measured the time course of recovery over the following 60 s. NO did not show any effects on the time constant of recovery (S3 Fig and S9 Data). In order to test whether NO acts specifically on RRP or also affects the availability of other pools, we stimulated the NMJ for longer periods (8 s) at 50 Hz. This prolonged stimulation leads to recruitment of vesicles from the reserve pool (RP) [42, 43]. Analysis revealed that NO also caused a strong reduction of release from the RP (Fig 2E–2H, Ctrl: 11,160 ± 1,645 [n = 6]; NO: 5,286 ± 798 [n = 7], p = 0.0062). One important protein that regulates vesicle clustering and release of neurotransmitter is the phosphoprotein synapsin (syn), which regulates recycling of RP vesicles in Drosophila NMJs [43]. We tested whether modulation of syn could be responsible by employing larvae deficient in this protein from the Syn97-null mutation [44]. These larvae did not exhibit any reduction in single-stimulus QC compared to Ctrls, but prolonged recruitment (500 ms at 50 Hz) showed reduced vesicle availabilities. Importantly, incubation of Syn97 larvae with NO led to further reduction of both parameters (S4 Fig and S9 Data), suggesting that NO effects are via a different signaling route. Based on these data, we suggest that NO decreases release of vesicles from the RRP and RP but does not affect the rate of vesicle pool recovery from depletion. The NO-mediated effects appear to be independent of syn, suggesting an event downstream of vesicle recruitment per se. We next applied an independent approach to estimate the synaptic parameters: fluctuation analysis [45] to estimate the number of functional release sites N. eEJCs were elicited at varying calcium concentrations ([Ca2+]e: 0.5–3 mM, 0.2 Hz) and amplitudes were plotted over [Ca2+]e (Fig 3A and 3B and S3 Data). NO exposure led to reduced release across different Ca2+ concentrations (0.75–3 mM). N was estimated from parabolic fits to the variance-mean plots for each NMJ (Fig 3C). This analysis revealed a strong reduction in N following NO exposure (Fig 3D, NCtrl: 630 ± 104 [n = 5], NNO: 117 ± 32 [n = 6], p = 0.0006). The estimation of N from the fluctuation analysis (about 600) in Ctrl is in accordance with previously reported electron microscopy (EM) data showing a number of about 500 vesicles per NMJ [46]. These data confirm that NO most likely reduces the number of releasable vesicles by preventing vesicle fusion at individual release sites. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Ca2+ dependency of evoked release is reduced by NO. (A) Representative raw eEJC recordings at different [Ca2+]e (from 0.5 to 3 mM). (B) Time course of single eEJCs for 2 NMJs (w1118 Ctrl and NO exposure) at indicated [Ca2+]e. (C) Parabolic fits to the variance-mean relationships for the conditions indicated. (D) N estimated from fluctuation analysis in both conditions. (E) Ca2+ cooperativity of evoked release is shown on a double logarithmic plot for w1118 Ctrl and NO. (F) Raw traces for PPRs of a Ctrl and NO-treated NMJ at 1.0 and 1.5 mM [Ca2+]e, illustrating a reduced release probability following NO exposure. (G) Summary of PPR at 1 and 1.5 mM [Ca2+]e for various ISI for Ctrl and NO (n = 5 larvae for Ctrl, n = 3 larvae for NO, with 2–3 NMJs per larva). (H) Images of myrGCaMP-expressing NMJs from a Ctrl (top) and a NO-treated larva (bottom) before, during (at 3 s, 60 Hz), and after the 8-s train. Changes in GCaMP5 fluorescence were analyzed for each bouton to calculate ΔF/F0. The raw data can be found in S3 Data. Values per NMJ were averaged and data denoting mean ± SEM for each condition of 6–8 NMJs (3–4 larvae) are shown in I. *p < 0.05, ***p < 0.001, Student t test. [Ca2+]e, extracellular calcium concentration; Ctrl, control; eEJC, evoked EJC; EJC, excitatory junction current; ISI, interspike interval; myrGCaMP, N-myristoylated GCaMP; N, number of release-ready vesicles; NMJ, neuromuscular junction; NO, nitric oxide; PPR, paired pulse ratio; stim, stimulation. https://doi.org/10.1371/journal.pbio.2003611.g003 The reduced QC seen following NO exposure can also be attributable to a change in the Ca2+ dependency of release, so we determined whether the reduced transmitter release is due to altered Ca2+ cooperativity of release [47]. The Hill slope was strongly reduced by NO (Ctrl: 3.2 ± 0.4 [n = 6], NO: 1.8 ± 0.7 [n = 5], p = 0.0024, Fig 3E); however, the half maximal effective Ca2+ concentration (EC50) was unaltered (Ctrl: 1.0 ± 0.03, NO: 1.0 ± 0.09, p > 0.05, Fig 3E), indicating that sensitivity to Ca2+ was not affected by NO. To further assess nitrergic effects on pvr, we used the PPR approach by delivering two pulses with interspike intervals (ISIs) between 10 and 200 ms at two different [Ca2+]e (1 and 1.5 mM, Fig 3F and 3G) in Ctrl and NO-treated NMJs. Analysis showed that Ctrl NMJs only exhibit slight potentiation at low Ca2+ and high ISI, indicative of low pvr. In contrast, pvr in the presence of NO was decreased, as shown by an increased PPR (potentiation at all ISI at 1 mM Ca2+ and 20 and 40 ms ISI at 1.5 mM Ca2+, p < 0.05, Ctrl versus NO at each ISI), which is also in agreement with elevated pvr in NOS “null” larvae. With about 500 release sites per NMJ and a QC of 200 (Ctrl) and 90 (NO), our data present estimated pvr values of 0.33 (Ctrl) and 0.16 (NO), with Ctrl values similar to estimates made previously in WT larvae [40]. Previously, we have shown that NO signaling can suppress mammalian P/Q and N-type Ca2+ channels [48]. In order to test whether altered Ca2+ influx could cause the observed effects on evoked release at the NMJ, we tested whether NO application for 60 min changed presynaptic Ca2+ levels during a train of synaptic stimulation. GCaMP5 was expressed presynaptically and activity-evoked Ca2+ influx in type 1b NMJ boutons was imaged at different extracellular Ca2+ concentrations (0.25–3 mM). Our data showed that NO had no effect on stimulated Ca2+ levels at any concentration tested (ΔF/F0, myrGCamP5: 3 mM Ca2+: Ctrl: 0.70 ± 0.09, NO: 0.78 ± 0.12 [n = 13–18 boutons from 4–6 NMJs each], p > 0.05; Fig 3H and 3I; GCaMP5: 0.25 mM Ca2+: 0.24 ± 0.03, NO: 0.14 ± 0.03, 0.5 mM Ca2+: Ctrl: 0.42 ± 0.08, NO: 0.50 ± 0.07, 1.5 mM Ca2+: Ctrl: 1.13 ± 0.14, NO: 1.18 ± 0.21 [n = 28–46 boutons from 7–11 NMJs each], p > 0.05; S5 Fig and S9 Data). Together, the data suggest that NO reduced evoked release and the frequency of spontaneous release, likely due to reduced release probability and Ca2+ cooperativity, which manifests itself in reduced vesicle fusion. We showed that the Ca2+ dependence of release, but not Ca2+ entry per se, was reduced by NO, which indicates a possible modulation of SNARE (-associated) protein interactions via NO-mediated PTMs. Enhanced denitrosylation signaling reverses and precludes NO effects S-nitrosylation is a reversible non-enzymatic protein modification, the levels of which can be regulated via S-nitrosoglutathione reductase (GSNOR), the sole alcohol dehydrogenase 5 (ADH-5) isozyme in vertebrate brains [49], which has a homologue in Drosophila (encoded by the formaldehyde dehydrogenase [fdh] gene). This de-nitrosylation process requires GSH. GSH is produced from L-glutamate and Cys via the enzyme glutamate-cysteine ligase (GCL), the rate-limiting step in GSH synthesis in fly [50]. The Drosophila GCL holoenzyme is heterodimeric, consisting of a catalytic (DmGCLc) and a modifier (DmGCLm) subunit, each encoded by a unique gene, and overexpression of either subunit increases cellular GSH levels [50]. In order to assess the contributions of SNO formation to the physiology at the NMJ, we investigated the effects of altering neuronal GSH levels. If NO mediates its observed actions via SNO formation, we should be able to prevent/reduce the effects on transmitter release by providing elevated GSH levels by (i) GSH supplementation, (ii) overexpression of GSNOR (fdh), or (iii) overexpression of GCL (DmGCLm/c) and, inversely, enhance NO effects by using RNA interference (RNAi) expression of the above proteins. We tested first the recovery of NO-mediated reduction of eEJC amplitudes following NO exposure for 50 min by washing out NO. eEJC amplitudes recovered slightly (Fig 4A, green and S4 Data); however, when washing in GSH (150 μM), the amplitudes recovered to control levels after 15 min (GSH [blue] versus NO at 50 min [red], p < 0.05), indicating a GSH-mediated reversal. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Genetic and pharmacological induction of denitrosylation reverses and prevents nitrergic effects on transmitter release and vesicle pools. (A) NO-induced suppression of evoked release (w1118 Ctrl eEJCs data in grey from Fig 1A) can be reversed after 15 min of GSH application (150 μM, blue). Wash out of NO alone shows mild nonsignificant recovery (green). (B) Mean QC for conditions indicated (OE, elav > UAS-fdh31 [GSNOR], elav > UAS-GCLm, elav > UAS-GCLc, all ± NO). NO has no effects on QC in genotypes indicated (w1118 + NO data in grey from Fig 1 for comparisons, ***p < 0.0001 w1118 + NO versus w1118), #p < 0.05, ##p < 0.01, ###p < 0.001, ####p < 0.0001 versus w1118 + NO. (C) Cumulative QC graphs for genotypes indicated (±NO) with linear regression to the last 200 ms, Ctrl in grey taken from Fig 2 for comparison. (D) Mean RRP sizes, NO has no effects on vesicle pool size in tested genotypes (w1118 data in grey from Fig 1, w1118 + NO versus w1118: ***p < 0.0001), ##p < 0.01, ###p < 0.01, ####p < 0.01 versus w1118 + NO. (E) PPR at 20 ms ISI for indicated conditions, ****p < 0.0001 versus Ctrl. (F) mEJCs analysis: left, mean mEJC amplitudes (±NO), middle, mean mEJC frequencies (±NO), right, mean mEJC decay kinetics for genotypes indicated (±NO). OE of GSNOR, GCLm, or GCLc prevents NO effects on the frequency of mEJCs. The raw data can be found in S4 Data. Data denote mean ± SEM for all data comparisons. ANOVA with post hoc Tukey-Kramer. Ctrl, control; eEJC, evoked EJC; EJC, excitatory junction current; fdh, formaldehyde dehydrogenase; GCLc, glutamate-cysteine ligase catalytic subunit C; GCLm, glutamate-cysteine ligase catalytic subunit M; GSH, glutathione; GSNOR, S-nitrosoglutathione reductase; ISI, interspike interval; mEJC, miniature EJC; NO, nitric oxide; n.s., nonsignificant; OE, overexpression; PPR, paired pulse ratio; QC, quantal content; RRP, readily releasable pool. https://doi.org/10.1371/journal.pbio.2003611.g004 To characterize effects of endogenous GSH formation, we used elav-Gal4-driven UAS-fdh31, UAS-DmGCLm, and UAS-DmGCLc overexpression. It has been shown that overexpression of either DmGCLc or DmGCLm results in enhanced enzyme activity and elevated GSH levels [50], GSNOR overexpression (elav > UAS-fdh31) reduces global S-nitrosylation in fly, and conversely, GSNOR-RNAi expression (elav > UAS-fdhri34) elevates SNO protein levels [51]. Overexpression of GSNOR and GCLm/c (Fig 4B–4D) prevented NO effects on QC (GSNOR: 238 ± 20 [n = 11], DmGCLm: 197 ± 32 [n = 7], DmGCLc: 215 ± 39 [n = 6], GSNOR+NO: 223 ± 24 [n = 8], DmGCLm+NO: 177 ± 10 [n = 7], DmGCLc+NO: 329 ± 26 [n = 3], p > 0.05) and vesicle pool sizes (GSNOR: 438 ± 51 [n = 10], DmGCLm: 400 ± 99 [n = 7], DmGCLc: 496 ± 93 [n = 6], GSNOR+NO: 472 ± 34 [n = 8), DmGCLm+NO: 360 ± 40 [n = 7], DmGCLc+NO: 685 ± 148 [n = 3], p > 0.05, Fig 4B–4D). These data confirm that by enhancing GSNOR and GCL activities, thereby elevating intracellular GSH levels, the effects of NO on pool size and pvr (PPR at 20 ms ISI; w1118 Ctrl [0.93 ± 0.03] versus NO [1.2 ± 0.07], p < 0.0001, GSNOR overexpression [0.88 ± 0.02], +NO [0.84 ± 0.05]/GCLm overexpression [0.87 ± 0.04], +NO [0.92 ± 0.03]/GCLc overexpression [0.99 ± 0.07], +NO [0.96 ± 0.02], p > 0.05, Fig 4E) were precluded, suggesting that this was due to reduced SNO formation. Furthermore, overexpression of GSNOR, DmGCLm, and DmGCLc prevented the reduction in mEJC frequency following NO exposure (fGSNOR: 2.4 ± 0.3 s−1 [n = 13]; fDmGCLm: 3.0 ± 0.3 s−1 [n = 13]; fDmGCLc: 1.5 ± 0.42 s−1 [n = 5]; fGSNOR+NO: 1.7 ± 0.2 s−1 [n = 7]; fDmGCLm+NO: 2.9 ± 0.4 s−1 [n = 13]; fDmGCLc+NO: 0.4 ± 0.1 s−1 [n = 3], p > 0.05 versus w1118 Ctrl and versus each Ctrl, Fig 4F) without affecting mEJC amplitudes (GSNOR: −0.6 ± 0.07 nA [n = 13]; DmGCLm: −0.7 ± 0.07 nA [n = 13]; DmGCLc: −0.5 ± 0.07 nA [n = 5]; GSNOR+NO: −0.6 ± 0.07 nA [n = 7]; DmGCLm+NO: −0.6 ± 0.08 nA [n = 13]; DmGCLc+NO: −0.6 ± 0.07 nA [n = 3], p > 0.05 versus w1118 Ctrl and versus each Ctrl, Fig 4F) or decays (GSNOR: 7.5 ± 0.2 ms [n = 13]; DmGCLm: 9.7 ± 0.4 ms [n = 13], DmGCLc: 6.7 ± 0.3 ms [n = 5], GSNOR+NO: 6.2 ± 0.2 ms [n = 7], DmGCLm+NO: 7.9 ± 0.5 ms [n = 13], DmGCLc+NO: 6.2 ± 0.4 ms [n = 3], p > 0.05 versus w1118 Ctrl and versus each Ctrl, Fig 4F). Furthermore, the reduction of endogenous GSNOR and DmGCLm activities (elav > UAS-RNAi) caused partial electrophysiological phenotypes, such as a decrease in eEJC amplitudes, QC, or vesicle pool size compared to w1118 Ctrl, with NO having no further major negative effects (S6 Fig and S9 Data). Nitrergic effects require the presence of cpx We next asked which signaling routes and PTMs are involved in NO modulation of release. The SNARE-binding and fusion-clamp protein cpx regulates not only the Ca2+ cooperativity of evoked release but also spontaneous release [14] as well as release probabilities [52], thereby presenting a strong candidate for mediating the observed NO-induced changes. Cpx acts by binding to the SNARE complex, thereby promoting the clamping of release, and only when replaced by synaptotagmin 1 in response to Ca2+ influx will vesicle fusion be initiated [14, 20]. Dmcpx function can be regulated by protein kinase A (PKA) phosphorylation of serine126 (Ser126) [7] or by prenylation at the C-terminus [15, 16]. In order to test whether cpx is required to exert NO effects, we first used cpx null mutants (cpxSH1, cpx-/-) [14]. In these animals, we detected a strong reduction in evoked release and QC (22.6 ± 3.2 [n = 11], p < 0.0001 versus Ctrl), which was unaffected by NO (13.8 ± 2.0 [n = 4], p > 0.05 versus cpx-/-, p < 0.0001 versus Ctrl, Fig 5A–5D and S5 Data). Similarly, when comparing the vesicle pool size, cpx-/- NMJs showed a strong reduction (26 ± 6 [n = 11], p < 0.0001 versus Ctrl), which again was unaffected by NO (22 ± 5 [n = 4], p > 0.05 versus cpx-/-, p < 0.0001 versus Ctrl, Fig 5A–5D). These data confirm that cpx is required for NO to induce suppression of evoked release and available vesicle pool size and suggest that NO might enhance the clamping function of cpx in WT larvae. We next tested the impact of NO on the clamping ability of cpx by characterizing spontaneous release. Interestingly, the frequency of spontaneous events inversely correlates with endogenous cpx levels [14]. We analyzed mEJCs in cpx-/- muscle 6 (m6), which exhibited an extremely high frequency [14] (>40 × w1118, Fig 5E). NO did not reduce the mEJC frequency in those preparations, although a precise analysis is difficult due to strong overlap of single mEJCs [14]. In order to allow more accurate frequency measurements in cpx-/- animals, we used neighboring muscle 5 (m5), posessing a synapse with approximately 4-fold fewer release sites compared to m6. Similar to m6, cpx-/- increased mEJC frequencies >10-fold compared to Ctrl (m5: w1118: 0.8 ± 0.2 s−1 [n = 3], cpx-/-: 11.6 ± 0.8 s−1 [n = 6], p < 0.0001); however, following NO exposure, this preparation did not show any change in mEJC frequency (m5 cpx-/- + NO: 9.7 ± 0.9 s−1 [n = 5], p > 0.05 versus m5 cpx-/-, Fig 5E and 5F), suggesting the requirement of cpx for the observed nitrergic effects. Nevertheless, we recorded from m6 of heterozygous animals, which exhibit higher frequencies than w1118 but are still accurately quantifiable (m6 cpx+/-: 4.6 ± 0.8 s−1 [n = 5]). Here, NO induced a strong reduction in the frequency (m6 cpx+/- + NO: 0.6 ± 0.2 s−1 [n = 5] #p < 0.05 versus m6 cpx+/- Ctrl, Fig 5E and 5F), similar to that seen in w1118. These data confirm that NO only modulates spontaneous release frequencies in the presence of cpx. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Nitrergic effects on evoked and spontaneous release require cpx. (A) Raw eEJC recordings of a 50-Hz train, 500 ms in cpx-/- larvae (±NO). (B) Mean QC for genotypes indicated, showing the lack of NO effects in cpx-/- larvae (w1118 Ctrl in grey from Fig 2). (C) Cumulative QC for the same conditions as in (B). (D) Mean QC and vesicle pool sizes for the genotypes indicated (w1118 Ctrl data from Figs 1 and 2, ****p < 0.0001 versus w1118 Ctrl). (E) mEJCs recordings from m6 and m5 in cpx-/- and cpx+/- ± NO. (F) Mean mEJC frequencies for conditions and genotypes indicated (****p < 0.0001 versus m5 w1118 Ctrl, #p < 0.05 versus m6 cpx+/- Ctrl). (G) Raw traces of mEJC recordings before (left) and after (right) high frequency stimulation in w1118 (±NO) and NOSC larvae. (H) Average fold change of mEJC frequency during 50 s after stimulation compared to baseline frequency before stimulation for conditions and genotypes indicated (NOS “null” comprised of data from NOSC and NOSΔ15). NO treatment (40 min) suppresses increases in frequency, NOS “null” potentiates relative increases and enhanced denitrosylation (GCLm and GSNOR OE) prevents NO-induced suppression (*p < 0.05, **p < 0.01 versus Ctrl, ##p < 0.01, ####p < 0.001 versus NO). The raw data can be found in S5 Data. Data denote mean ± SEM in all graphs, ANOVA with post hoc Tukey-Kramer. Ctrl, control; cpx, complexin; eEJC, evoked EJC; EJC, excitatory junction current; GCLm, glutamate-cysteine ligase catalytic subunit M; GSNOR, S-nitrosoglutathione reductase; mEJC, miniature EJC; m5, muscle 5; m6, muscle 6; NO, nitric oxide; OE, overexpression; QC, quantal content. https://doi.org/10.1371/journal.pbio.2003611.g005 Together, these data show that in the absence of cpx, NO causes no electrophysiological phenotypes. The NO-mediated reduction of eEJC amplitudes, QC, pool size, and mEJC frequency all require the presence of cpx, suggesting that its modulation might be responsible for the observed nitrergic effects, which could be explained by a gain-of-clamping function [53]. This potential effect was further investigated by using the established paradigm of activity-induced enhancement of spontaneous release at the Drosophila NMJ [7]. We assessed whether NO modulation of release also affects this activity-dependent signaling, which would strengthen the role of cpx as a target for nitrergic regulation and a general regulatory mechanism. PKA has been reported to modulate mEJC frequency potentiation in a cpx overexpression model (Dmcpx 7B, [7]). We confirmed that high frequency stimulation (50 Hz for 3 s) led to an enhanced mEJC frequency in w1118 NMJs relative to baseline (Ctrl: 1.9 ± 0.2-fold [n = 13], Fig 5G and 5H). Interestingly, repeating this protocol in larvae exposed to NO showed a lack of frequency potentiation (NO: 0.8 ± 0.1-fold [n = 14], p < 0.05 versus Ctrl), which was also ODQ independent (NO + ODQ: 1.0 ± 0.1-fold [n = 7], p > 0.05 versus NO, Fig 5G and 5H). To test whether the manipulation of PTMs also affects nitrergic suppression of frequency potentiation, we used larvae overexpressing GCLm and GSNOR and NOS “null” larvae. We found that GCLm and GSNOR overexpression occluded nitrergic effects on suppression of mEJC frequency potentiation, whereas the lack of NO signaling led to enhanced potentiation (GCLm + NO: 2.3 ± 0.4-fold [n = 7], GSNOR + NO: 2.5 ± 0.5-fold [n = 7], NOS “null” [comprised of n = 5 NOSC and n = 3 NOSΔ15]: 3.8 ± 0.3-fold, **p < 0.01 versus Ctrl, ##p < 0.01 versus NO, ####p < 0.001 versus NO, Fig 5G and 5H). These data show that NO suppresses the activity-mediated increase in mEJC frequency and suggest that, similar to phospho-incompetent cpx mutants [7], nitrergic modulation of WT cpx produces an inhibitory action on spontaneous release. The lack of PTM signaling leads to an enhanced frequency potentiation, strengthening the notion that NO-mediated effects are responsible for suppression of synaptic release and our data point towards modulation of cpx as a key signaling mechanism. Nitrergic activity affects farnesylation of cpx and enhances its clamping properties Having shown that cpx signaling is involved in NO-mediated effects on spontaneous and evoked release, we next considered if S-nitrosylation of the Cys residue within the C-terminus of cpx possessing the CAAX motif could explain the observed results. Importantly, prenylation has been studied in several genetically modified cpx proteins in which the CAAX motif was eliminated [15, 16]. These studies suggest that deletion of final parts of the C-terminus/final amino acid affects cpx localization, interactions with SNARE-proteins, and, subsequently, its function. To explore the effects of cpx farnesylation more in detail, we made use of Drosophila lines expressing green fluorescent protein (GFP)-tagged WT and mutant cpx (cpx1257, lacking the final amino acid [16]), referred to as CpxΔX. This mutant has been shown to exhibit altered co-localization with syntaxin at the dorsolongitudinal flight muscle (DLM) neuromuscular synapse. We assessed localizations of WT and mutant cpx at the NMJ (elav > UAS-cpx-GFP, elav > UAS-cpx1257-GFP) with respect to their interaction with the AZ protein, Brp. WT cpx exhibits diffuse localization within boutons (as previously reported [15]) with little co-localization with Brp (Fig 6A and 6B and S6 Data). In contrast, the mutant form, lacking farnesylation, is highly co-localized with Brp, as indicated by the increase in Pearson’s coefficient (Fig 6A and 6B; WT cpx: 0.35 ± 0.30 [n = 9], CpxΔX: 0.65 ± 0.03 [n = 9], p < 0.0001). These data confirm that preventing cpx farnesylation results in enhanced co-localization with AZ. To further support these data, we conducted high-resolution stimulated emission depletion (STED) microscopy [54] and analyzed the Pearson’s coefficient for the co-localization of Brp with cpx. This experiment verified the confocal data showing enhanced co-localization of CpxΔX with Brp versus WT cpx (WT cpx: 0.13 ± 0.02 [n = 25], CpxΔX: 0.27 ± 0.02 [n = 23], p < 0.0001, Fig 6C and 6D). As Dmcpx possesses a predominant clamping function [23], we propose that NO could lead to a reduction in farnesylation, a consequent stronger interaction with the SNARE complex at the AZ, and thereby enhance its clamping function upon transmitter release. To specifically confirm co-localizations, we used the high-resolution proximity ligation assay (PLA), with which we imaged interactions of Brp with cpx. We used both lines, WT cpx-GFP and CpxΔX-GFP expressing larvae, and found that PLA signals are strongly enhanced at NMJs expressing the mutant cpx (Fig 6E and 6F; WT cpx: 0.04 ± 0.004 [n = 9], CpxΔX: 0.12 ± 0.02 [n = 9], p = 0.009). As the co-localization data may depend upon expression of GFP-tagged cpx, we confirmed equal GFP expression levels in both lines by immunoblotting (S9A Fig). These co-localization and PLA experiments confirm an enhanced association of a mutated farnesylation-incompetent cpx with Brp and suggest that lack of farnesylation renders cpx in close proximity to release sites of AZs. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Lack of cpx farnesylation promotes its co-localization with AZs. (A) Representative maximal projection confocal images of GFP-tagged WT Cpx and mutant Cpx1257 (CpxΔX) at the NMJ (green: cpx, red: Brp). (B) Enhanced Cpx-Brp co-localization indicated by higher Pearson’s coefficient for CpxΔX (n–number of NMJs). (C) STED images showing cpx and Brp localization in single boutons in WT and CpxΔX mutants. The mutation (Cpx1257) increases the co-localization of cpx with Brp as indicated by the enhanced Pearson’s coefficient (D) (n–number of boutons). (E) Maximal projection confocal images of NMJs from larvae expressing WT Cpx and CpxΔX. PLA fluorescence shown in red and HRP staining in green. (F) Summated PLA signal volumes relative to NMJ volumes for cpx (GFP)-Brp interactions (n–number of NMJs). The raw data can be found in S6 Data. Data denote mean ± SEM, Student t test, **p = 0.009, ****p < 0.0001. AZ, active zone; Brp, Bruchpilot; cpx, complexin; GFP, green fluorescent protein; HRP, horseradish peroxidase; NMJ, neuromuscular junction; PLA, proximity ligation assay; STED, stimulated emission depletion; WT, wild-type. https://doi.org/10.1371/journal.pbio.2003611.g006 In order to assess this possibility further, we used pharmacological and genetic tools to modulate cpx farnesylation and compared protein localization and synaptic release following farnesyl transferase (FTase) inhibition and NO exposure. Reduced expression of the Drosophila ortholog of FTase or inhibition of FTase by L-744,832 and GGTI-298 have strong effects on fly lethality [55], implicating a crucial function of this signaling in fly. First, we tested how FTase inhibition (20 μM L-744,832 + 10 μM GGTI-298) and NO exposure affect cpx co-localization with the SNARE complex proteins syntaxin and synaptotagmin or Brp, using the PLA. We measured total PLA volume of NMJ z-stacks and normalized PLA signals to NMJ volume. We found that both treatments (depicted as “farnesyl inh” and “NO,” Fig 7A and 7B and S7 Data) led to enhanced co-localization of cpx with syntaxin and Brp (syntaxin-cpx: Ctrl: 0.04 ± 0.007, NO: 0.12 ± 0.02, farnesyl inh: 0.11 ± 0.02, Brp-cpx: 0.02 ± 0.007, NO: 0.08 ± 0.03, farnesyl inh: 0.09 ± 0.05, Fig 7A and 7B; p < 0.01, p < 0.001 versus Ctrl), suggesting that NO PTMs and farnesylation inhibition enrich cpx at the AZ. When analyzing the interactions between the Ca2+ sensor synaptotagmin and cpx, we found that this interaction was completely suppressed following treatments (Ctrl: 0.2 ± 0.06, NO: 0.03 ± 0.006, farnesyl inh: 0.04 ± 0.007, Fig 7A and 7B; p < 0.01 versus Ctrl). The PLA data were further supported by STED imaging studies showing identical changes in protein co-localization, as determined by Pearson’s coefficient analysis (S7 Fig and S9 Data). One possibility to allow for greater amounts of cpx to be available for binding to SNAREs is by enhancing free and soluble cytosolic levels as a consequence of reduced farnesylation. Farnesylation of cpx results in its membrane tethering, and thus protein fractions, which are membrane bound, are less mobile than soluble cytosolic proteins. To assess the mobility of potentially farnesylated versus soluble (non-farnesylated) cpx and thus distinguish between these two pools of cpx, we performed fluorescence recovery after photobleaching (FRAP) analysis of GFP-tagged WT and farnesylation-incompetent cpx (CpxΔX). Although a previous study did not detect differences between farnesylated versus non-farnesylated cpx isoform using this method with a photo-bleaching area of half a bouton [15], we found that accurate FRAP analysis of cpx-GFP mobility can only be performed by using substantially smaller bleaching areas, as reported previously [56] (S8 Fig and S9 Data). Using this approach, we found that bleaching an area of 2.5 μm2 (instead of >10 μm2) generally leads to faster recovery rates (S8 Fig and S9 Data). Our data confirmed that lack of farnesylation (CpxΔX) allows for greater movement of cpx and faster recovery (tau: WT cpx: 18.1 ± 1.7 ms, CpxΔX: 11.9 ± 1.2 ms [p < 0.05], WT Cpx + NO: 8.8 ± 0.8 ms [p < 0.0001], n = 18–20, Fig 7C), as expected for a soluble protein. Our data further show that NO treatment caused the same increase in recovery rates (Fig 7C), suggesting that NO also prevented farnesylation. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Reduced farnesylation of cpx enhances AZ localization and alters interactions with SNARE proteins. (A) Maximal projection confocal images of w1118 NMJs in Ctrl, following 60 min of exposure to NO or farnesylation inhibitor (“farnesyl inh”: 10 μM GGTI-298 + 20 μM L-744,832). PLA fluorescence in red and HRP staining in green for: left, syntaxin-cpx; middle: Brp-cpx; right: synaptotagmin-cpx interactions. (B) Analysis of summated PLA signal volumes relative to NMJ volumes. (C) FRAP experiments were performed at NMJs expressing GFP-cpx, shown as representative images of WT GFP-cpx at different time points (bleaching area: 2.5 μm2, scale bar: 2 μm). Right, mean data showing recovery of WT, CpxΔX, and NO-treated WT cpx, with mean tau values summarized. Note, lack of farnesylation due to the mutation or NO treatment results in faster recovery rates. (D) Representative mEJC recordings following 60 min incubation with GGTI-298 + L-744,832 (“farnesyl inh”) or of a larva expressing FTase-RNAi with mean mEJC frequencies. (E) Trains of 50-Hz stimulation of a larva incubated for 60 min with GGTI-298 + L-744,832 (“farnesyl inh”) or expressing FTase-RNAi with mean eEJC amplitudes and QC. (F) Cumulative QC of 50-Hz trains with mean estimated RRP sizes (right). The raw data can be found in S7 Data. Data denote mean ± SEM, Student t test, or ANOVA with post hoc Tukey-Kramer as indicated, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. AZ, active zone; Brp, Bruchpilot; cpx, complexin; Ctrl; control; eEJC, evoked EJC; EJC, excitatory junction current; FRAP, fluorescence recovery after photobleaching; FTase, farnesyl transferase; GFP, green fluorescent protein; HRP, horseradish peroxidase; mEJC, miniature EJC; NMJ, neuromuscular junction; NO, nitric oxide; PLA, proximity ligation assay; QC, quantal content; RNAi, RNA interference; RRP, readily releasable pool; SNARE, soluble N-ethyl-maleimide-sensitive fusion protein Attachment Protein Receptor; WT, wild-type. https://doi.org/10.1371/journal.pbio.2003611.g007 These data suggest that due to enriched local levels, cpx outcompetes synaptotagmin for SNARE binding at the AZ, thereby displacing synaptotagmin, as reported previously in biochemical studies [53]. Our data show that pharmacological and genetic inhibition of farnesylation promotes cpx co-localization with the AZ and supports the notion that this negatively impacts on synaptotagmin-SNARE complex binding, subsequently reducing release. The specificity of the PLA was corroborated by lack of Brp-cpx PLA signals in cpx-/- larvae (S9B–S9D Fig). Next, we explored the possibility of whether specific inhibition of FTase activity by L-744,832 and GGTI-298 and FTase RNAi mimics the effects of NO on synaptic transmission. We found that, in both conditions, the frequency of mEJCs was reduced to similar values seen following NO exposure (fmEJC: L-744,832 + GGTI-298: 0.7 ± 0.1 s−1 [n = 8], p = 0.0051 versus Ctrl, FTase RNAi: 0.9 ± 0.2 s−1 [n = 9], p = 0.0136 versus Ctrl, Student t test, Fig 7D). Importantly, both L-744,832 + GGTI-298 and FTase RNAi expression reduced evoked transmission and available vesicle pool size to levels similar to those following NO incubation (L-744,832 + GGTI-298: eEJC: 56 ± 5 nA, QC: 80 ± 13 [n = 9], pool size: 180 ± 27 [n = 9], p < 0.0001 versus each w1118 Ctrl; FTase RNAi: eEJC: 75 ± 5 nA, QC: 82 ± 6 [n = 9], pool size: 120 ± 17 [n = 9], p < 0.0001 versus each w1118 Ctrl, Student t test, Fig 7E and 7F). These data suggest that the farnesylation status of cpx mediates nitrergic effects, resulting in changed SNARE protein interactions, which determines the physiological outcome of cpx. To further investigate the effects of NO directly on the prenylation process, we employed the well-characterized GFP-CAAX transfection model [57]. Here, human embryonic kidney (HEK) cells were transfected with GFP-CAAX (K-Ras motif) and the membrane association was assessed in response to prenylation inhibition and NO treatment. In control conditions, GFP exhibited a strong fluorescence signal at the membrane, which disappeared and redistributed into the cytosol following pharmacological inhibition of prenylation (L-744,832 + GGTI-298, p < 0.0001), confirming the prenylation-mediated localization of GFP-CAAX to the membrane (Fig 8A and S8 Data). Importantly, we showed that NO treatment (propylamine propylamine NONOate [PAPA-NONOate], p < 0.0001) induced a similar phenotype, with GFP being localized predominantly in a cytosolic manner—suggesting that NO prevents farnesylation through the same pathway (Fig 8A). To confirm that the Cys within the CAAX motif can undergo S-nitrosylation, we performed the Biotin Switch Assay on cpx-3 from isolated mouse retinas. NO donor incubation induced a >2-fold increase in SNO-cpx (Fig 8B), confirming this PTM on cpx and suggesting that this PTM is responsible for NO-induced changes in localization and function of cpx. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 8. Cpx nitrosylation and block of farnesylation leads to redistribution of GFP-CAAX and enhances Dmcpx localization at AZs. (A) GFP-CAAX expression in HEK cells showing membrane fluorescence signals in Ctrls. Pharmacological inhibition of farnesylation by GGTI-298 + L-744,832 (“farnesyl inh,” p < 0.0001 versus Ctrl, ANOVA) and NO incubation (p < 0.0001 versus Ctrl, ANOVA) result in a redistribution of GFP fluorescence into the cytosol. Fluorescence signals were analyzed by line scan and plotted as intensities (a.u.) over distance across the cell somata, n = 80–104 cells, scale bar: 20 μm. (B) Mouse cpx-3 is S-nitrosylated in response to NO donor application. Immunoblot intensities increased 2.4 ± 0.5-fold following NO application. (C) Representative recordings of a 50-Hz train and spontaneous activity of a larva expressing Dmcpx 7AC140W with mean eEJC amplitudes, QC, vesicle pool size, and mEJC frequency shown in (D). (E) Top, maximal projection confocal images of NMJs expressing a WT cpx or Dmcpx 7AC140W showing the PLA signal in red; bottom, STED images showing cpx and Brp staining in WT and Dmcpx 7AC140W mutants. (F) Analysis of PLA data. (G) Co-localization data with Pearson’s coefficient for interactions of cpx with Brp. The raw data can be found in S8 Data. Data denote mean ± SEM, Student t test (D, F), *p < 0.05, ***p < 0.001, ****p < 0.0001. a.u., arbitrary unit; AZ, active zone; Brp, Bruchpilot; cpx, complexin; Ctrl, control; eEJC, evoked EJC; EJC, excitatory junction current; GFP, green fluorescent protein; HEK, human embryonic kidney; HRP, horseradish peroxidase; mEJC, miniature EJC; NMJ, neuromuscular junction; NO, nitric oxide; PLA, proximity ligation assay; QC, quantal content; STED, stimulated emission depletion; WT, wild-type. https://doi.org/10.1371/journal.pbio.2003611.g008 To specifically confirm the effects of S-nitrosylation and SNO interaction with farnesylation of cpx in Drosophila, we generated and expressed a nitroso-mimetic cpx mutant (Dmcpx 7AC140W) in a cpx null background (cpxSH1) and assessed synaptic responses. The Cys140 of Dmcpx is located within a hydrophobic region, as predicted in the Kyle Doolittle plot, which favors S-nitrosylation [58]. This mutant exhibits reduced evoked responses, QC, and vesicle pool sizes (eEJC: 70 ± 7 nA, QC: 106 ± 8, pool size: 204 ± 23 [n = 15 each], p < 0.0001 versus each w1118 Ctrl, Fig 8C and 8D), indicating that the mimicking of S-nitrosylation and simultaneous lack of farnesylation of cpx caused the observed changes. Importantly, this mutation also induced a reduction in spontaneous activity (fmEJC: 1.3 ± 0.2 s−1 [n = 15], p < 0.05 versus w1118 Ctrl, Fig 8C and 8D), reinforcing the argument of enhanced clamping function due to SNO formation and lack of farnesylation. The expression of WT cpx in the null background did not affect QC, pool size, or mEJC frequency (QC: 167 ± 17 [n = 5]; pool size: 381 ± 76 [n = 5]; fmEJC: 2.4 ± 0.4 s−1 [n = 10 each], p > 0.05 versus each w1118 Ctrl). To confirm changes in localization of Dmcpx 7AC140W, we analyzed PLA signals and found that Dmcpx 7AC140W highly co-localizes with Brp, in strong contrast to WT cpx (WT: 0.025 ± 0.013, Dmcpx 7AC140W: 0.17 ± 0.03 [n = 6–7], p < 0.0001, both expressed in cpx-/- background, Fig 8E and 8F). The data from the PLA experiments were confirmed by STED confocal microscopy, showing significantly higher Pearson’s coefficients for the co-localization of the cpx mutant C140W with Brp relative to the interaction of WT cpx with Brp (WT cpx: 0.13 ± 0.03, Dmcpx 7AC140W: 0.34 ± 0.02 [n = 20–24], p < 0.0001; Fig 8E and 8G). These data demonstrate that independent approaches to block farnesylation (and mimic of cpx-SNO) recapitulate nitrergic modulation of release and protein localization and therefore link for the first time NO-induced PTM and farnesylation signaling of cpx. We propose that S-nitrosylation acts as a novel endogenous pathway to alter cpx farnesylation signaling and protein–protein interactions and thereby allows a fine-tuning of synaptic function. NO-induced suppression of evoked and spontaneous synaptic release is independent of cGMP Previously, we found that enhancing endogenous nitric oxide synthase (NOS) activity induced by overexpression of D. melanogaster NOS (DmNOS) caused a reduction in synaptic strength at the Drosophila NMJ synapse [28]. To examine the effects of NO on glutamatergic transmission in more detail, we exposed wild-type (WT) w1118 control (Ctrl) larvae to NO donors, which provide an estimated NO concentration of about 200 nM [29]. When recording evoked excitatory junction currents (eEJCs) up to 70 min during NO incubation, the amplitudes started to decline significantly after 35 min (Fig 1A and S1 Data, p < 0.05; n = 3 each). Mean eEJC amplitudes and quantal content (QC) at 50 min for Ctrl (122 ± 7 nA, QC: 200 ± 15, n = 20–22) and NO treatment (59 ± 7 nA, QC: 93 ± 10, n = 14) are shown in Fig 1B. As the canonical NO-cGMP pathway is active in Drosophila [30] and potentially responsible for this observation, we blocked the soluble guanylyl cyclase (sGC) with 1H-[1,2,4]oxadiazolo[4,3-a]quinoxalin-1-one (ODQ, 50 μM). Interestingly, ODQ did not prevent the effects of NO, suggesting a cGMP-independent mechanism (amplitudes: Ctrl + ODQ: 127 ± 5 nA, NO + ODQ: 70 ± 7 nA, QC: Ctrl + ODQ: 200 ± 22, NO + ODQ: 130 ± 11, Fig 1B, n = 10–16). As Drosophila has endogenous NO signaling and produces neuronal NO in a Ca2+/calmodulin-dependent manner [31, 32], we used NOS knockout-like (NOS “null”) larvae to assess endogenous NO modulation of release. We used two different lines with strongly reduced DmNOS showing NOS “null” activity (NOSC and NOSΔ15 [33, 34]) and we would expect that lack of endogenous NO generation has the opposite effects on release. When recording eEJCs, both genotypes exhibited a tendency towards larger eEJC amplitudes and QC (Fig 1C) and, in addition, we detected an increased presynaptic release probability (pvr) in NOSC NMJs, as indicated by the reduced paired pulse ratio (PPR) at 20 ms ISI (0.80 ± 0.03 [n = 11], p = 0.002, Student t test) compared to WT Ctrls (0.93 ± 0.03 [n = 17]), indicating endogenous nitrergic effects on release probabilities. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. NO reduces evoked release and frequency of spontaneous release in a cGMP-independent manner. (A) NO suppresses evoked release (eEJC) over a time course of 55 min. Insets show representative single eEJCs at 40 min for both conditions. (B) Mean eEJC amplitudes (left axis) and QC (right axis) of w1118 NMJs are reduced following NO exposure (at 40 min). The sGC inhibitor ODQ (50 μM) did not affect the response to NO. (C) Mean eEJC amplitudes (left axis, black) and QC (right axis, grey) of NOSC and NOSΔ15 NMJs. (D) Raw mEJC recordings of w1118 NMJs and mEJC parameters (top to bottom: amplitude, frequency, decay). Top insets show representative mEJC recordings. Bottom insets show single mEJCs (grey) and averaged mEJC (red) with single exponential fit to the decay. (E) Raw mEJC recordings for both NOSC and NOSΔ15 genotypes. Below, mEJC quantal parameters: amplitude and frequency, Student t test each relative to w1118 Ctrl, *p = 0.04, ***p = 0.001. (F) cGMP content of larval brains under the conditions indicated (NO: 40 min NO exposure, NO + ODQ: 40 min NO exposure in the presence of 50 μM ODQ, NO + Zap: 40 min NO exposure + PDE inhibitor Zap, 20 μM). (G) FlincG3 fluorescence images of a Ctrl and stimulated NMJ (20 Hz for 10 s, duty cycle: 1 min for total of 10 min). (H) Summary of FlincG3 fluorescence (in a.u.’s). The raw data can be found in S1 Data. Data denote mean ± SEM in all graphs, ANOVA with post hoc Tukey-Kramer, *p < 0.05, **p < 0.01, ****p < 0.0001. a.u., arbitrary unit; cGMP, cyclic guanosine monophosphate; Ctrl, control; eEJC, evoked EJC; EJC, excitatory junction current; mEJC, miniature EJC; NMJ, neuromuscular junction; NO, nitric oxide; ODQ, 1H-[1,2,4]oxadiazolo[4,3-a]quinoxalin-1-one; PDE, phosphodiesterase; QC, quantal content; sGC, soluble guanylyl cyclase; Zap, zaprinast. https://doi.org/10.1371/journal.pbio.2003611.g001 To further understand the effects of NO on release, we analyzed miniature EJCs (mEJCs) under the same conditions. NO had no effect on mEJC amplitudes or decay kinetics; however, the frequency was reduced following NO and NO+ODQ incubation (Ctrl: 2.0 ± 0.2 nA [n = 25], NO: 1.1 ± 0.1 nA [n = 16], NO+ODQ: 1.0 ± 0.2 nA [n = 8], ODQ: 1.7 ± 0.2 nA [n = 11], Ctrl versus NO: p < 0.01, Ctrl versus NO+ODQ: p < 0.05, Fig 1D). This suggests that NO is unlikely to affect synaptic vesicle filling or composition/activity and density of postsynaptic D. melanogaster glutamate receptors (DmGluR) [35]. We tested miniature events in the NOS “null” mutants and confirmed a further inhibitory role of NO signaling on release, with mEJC frequencies being significantly enhanced in NOSΔ15 (3.5 ± 0.5 s−1 [n = 4], p = 0.001) and NOSC (3.5 ± 0.4 s−1 [n = 16], p = 0.04) larvae compared to Ctrl (Fig 1E), without affecting mEJC amplitudes (NOSΔ15: 0.8 ± 0.1 nA [n = 13], NOSC: 1.1 ± 0.3 nA [n = 3] Fig 1E) or decay kinetics (NOSΔ15: 8.9 ± 0.6 ms [n = 12], NOSC: 9.4 ± 0.3 ms [n = 4], p > 0.05 versus Ctrl). Thus, reduction of endogenous NOS activity shows opposite effects to elevation of NO levels, confirming the inhibitory action of NO on evoked and spontaneous vesicle release. As the data imply cGMP-independent signaling, we wanted to confirm that cGMP levels are not altered following NO stimulation. Thus, we measured cGMP directly in isolated larval brains. NO application did not raise cGMP levels (at 50 min: Ctrl: 2.4 ± 0.5 pmol/mg, NO: 3.0 ± 0.6 pmol/mg, p > 0.05 [n = 30 each], Fig 1F). Cyclase inhibition in the presence of NO did not significantly reduce cGMP levels, confirming lack of NO-induced neuronal cGMP accumulation. We found that any generated cGMP was broken down by phosphodiesterase DmPDE5/6 [36], as cGMP increased following NO stimulation only with PDE inhibition (20 μM zaprinast [Zap]; NO+Zap: 50.2 ± 8.3 pmol/mg, p < 0.0001), while Zap alone had no effect (Zap: 4.6 ± 2.0 pmol/mg, p > 0.05). To assess whether NO is produced endogenously to induce modulation of synaptic function as observed above, we expressed FlincG3 presynaptically and stimulated NMJs at 20 Hz (for 10 s every minute for 20 min). As shown in Fig 1G, 20 Hz stimulation induced a significant increase in fluorescence, confirming endogenous presynaptic generation of NO (Ctrl: 62 ± 4 arbitrary units [a.u.’s], Stim: 96 ± 8 a.u.’s, Fig 1H [n = 13–15 boutons], p < 0.01). Importantly, addition of the NO donor did not further increase the fluorescence, indicating that activity-induced synaptic NO concentrations reach similar levels (NO: 93 ± 7 a.u.’s). A potential target of NO signaling is mitochondria [37], which are required for the energy to maintain vesicle recycling and synaptic transmission [38]. Thus, we measured mitochondrial activity in third instar larvae under the same conditions (50 min NO incubation) and found that mitochondrial activity was unaffected by NO (S1 Fig and S9 Data), suggesting that the effects of NO on synaptic transmission are not due to ATP depletion. Together, these data suggest that NO has a presynaptic effect on transmitter release, which is independent of cGMP signaling. Ca2+ dependency of evoked release is reduced by NO Several mechanisms contribute to the regulation of synaptic strength [39], including altered pvr, alterations in the number of readily releasable vesicles and release sites (N) or quantal size (q). Alterations in q are likely not involved in the NO-induced effects observed based on our mEJC data above (Fig 1). We next assessed additional release parameters, including pvr, N, vesicle pool size, and Ca2+ dependency of release in NOS “null” and WT NMJs following nitrergic signaling. We determined pool size via a method successfully applied at the Drosophila NMJ, by analyzing the cumulative QC of trains of higher frequency stimulation [40]. Stimulation at 50 Hz for 500 ms in 1.5 mM extracellular calcium concentration ([Ca2+]e) retrieves vesicles from the readily releasable pool (RRP) [41]. This stimulation pattern induced mild depression in Ctrls and strong initial facilitation of trains under NO conditions (Fig 2A and S2 Data). Cumulative QC analysis revealed a pool size of 453 ± 37 (n = 17) in Ctrl and 185 ± 18 in NO-exposed NMJs (n = 16, p < 0.01), suggesting a strong reduction in ready-releasable/recycling vesicles (Fig 2A–2D). Supporting the above data, pool size estimation in the presence of ODQ confirmed cGMP independence (NO+ODQ: 310 ± 33 [n = 9], p < 0.05 versus Ctrl; ODQ alone: 501 ± 34 [n = 9], p > 0.05 versus Ctrl, Fig 2A–2D). And importantly, analysis of the vesicle pool sizes in NOS “null” mutants revealed a strong 2-fold increase compared to w1118 Ctrl and an over 5-fold increase compared to NO application (NOSC: 975 ± 161 [n = 11]; NOSΔ15: 958 ± 139 [n = 4], p < 0.001 versus Ctrl, Fig 2A–2D). To exclude any potential developmental effects caused by NOS deficiency that could account for these strong increases in release, we assessed NMJ morphology and ultrastructure. We analyzed the total volume of NMJs (horseradish peroxidase [HRP] signal) and the number of Bruchpilot (Brp) puncta/NMJ volume of z-stack confocal images (S2A and S2B Fig and S9 Data) and measured the number of AZs, T-bars per Ib bouton, and vesicles within a 250-nm semicircle around the AZ (S2C and S2D Fig). These data indicated that reduced NOS activity has no developmental impact on the structure of NMJs and synaptic boutons and can therefore not explain the physiological differences observed above. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. The size of the RRP and RP is reduced by NO and enhanced in NOS “null” NMJs. (A) Raw eEJC recordings of a 50-Hz train, 500 ms. (B) Mean QC for conditions indicated. (C) Cumulative QC with back extrapolated linear regression to the last 200 ms to time 0, yielding an estimated readily releasable vesicle pool size (RRP, intercept with y-axis, pool sizes in italics). (D) Mean RRP sizes for conditions indicated. NO reduces PPR sizes, with ODQ having no effects. NOS “null” NMJs show potentiated responses (ANOVA with post hoc Tukey-Kramer was used for comparisons). (E) Raw eEJC recordings of a 50-Hz train, 8 s. (F) Mean QC for conditions indicated. (G) Cumulative QC with back extrapolated linear regression to the last 2 s to time 0 (w1118 Ctrl and NO), yielding an estimated RP size (intercept with y-axis, pool sizes in italics). (H) Mean RP sizes for conditions indicated. NO reduces RP sizes similarly to RRP sizes. The raw data can be found in S2 Data. Student t test, **p = 0.0062, data denote mean ± SEM in all graphs. Ctrl, control; eEJC, evoked EJC; EJC, excitatory junctional current; NMJ, neuromuscular junction; NO, nitric oxide; NOS, nitric oxide synthase; ODQ, 1H-[1,2,4]oxadiazolo[4,3-a]quinoxalin-1-one; PPR, paired pulse ratio; QC, quantal content; RP, reserve pool; RRP, readily releasable pool. https://doi.org/10.1371/journal.pbio.2003611.g002 In addition to changes in release, NO could also exert its effects indirectly via modulating transmitter uptake and pool recovery. To exclude this possibility that altered recovery from depression affected the above pool estimations, we examined eEJC recovery. Following depletion of vesicle pools during a 50-Hz train (1 s), we measured the time course of recovery over the following 60 s. NO did not show any effects on the time constant of recovery (S3 Fig and S9 Data). In order to test whether NO acts specifically on RRP or also affects the availability of other pools, we stimulated the NMJ for longer periods (8 s) at 50 Hz. This prolonged stimulation leads to recruitment of vesicles from the reserve pool (RP) [42, 43]. Analysis revealed that NO also caused a strong reduction of release from the RP (Fig 2E–2H, Ctrl: 11,160 ± 1,645 [n = 6]; NO: 5,286 ± 798 [n = 7], p = 0.0062). One important protein that regulates vesicle clustering and release of neurotransmitter is the phosphoprotein synapsin (syn), which regulates recycling of RP vesicles in Drosophila NMJs [43]. We tested whether modulation of syn could be responsible by employing larvae deficient in this protein from the Syn97-null mutation [44]. These larvae did not exhibit any reduction in single-stimulus QC compared to Ctrls, but prolonged recruitment (500 ms at 50 Hz) showed reduced vesicle availabilities. Importantly, incubation of Syn97 larvae with NO led to further reduction of both parameters (S4 Fig and S9 Data), suggesting that NO effects are via a different signaling route. Based on these data, we suggest that NO decreases release of vesicles from the RRP and RP but does not affect the rate of vesicle pool recovery from depletion. The NO-mediated effects appear to be independent of syn, suggesting an event downstream of vesicle recruitment per se. We next applied an independent approach to estimate the synaptic parameters: fluctuation analysis [45] to estimate the number of functional release sites N. eEJCs were elicited at varying calcium concentrations ([Ca2+]e: 0.5–3 mM, 0.2 Hz) and amplitudes were plotted over [Ca2+]e (Fig 3A and 3B and S3 Data). NO exposure led to reduced release across different Ca2+ concentrations (0.75–3 mM). N was estimated from parabolic fits to the variance-mean plots for each NMJ (Fig 3C). This analysis revealed a strong reduction in N following NO exposure (Fig 3D, NCtrl: 630 ± 104 [n = 5], NNO: 117 ± 32 [n = 6], p = 0.0006). The estimation of N from the fluctuation analysis (about 600) in Ctrl is in accordance with previously reported electron microscopy (EM) data showing a number of about 500 vesicles per NMJ [46]. These data confirm that NO most likely reduces the number of releasable vesicles by preventing vesicle fusion at individual release sites. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Ca2+ dependency of evoked release is reduced by NO. (A) Representative raw eEJC recordings at different [Ca2+]e (from 0.5 to 3 mM). (B) Time course of single eEJCs for 2 NMJs (w1118 Ctrl and NO exposure) at indicated [Ca2+]e. (C) Parabolic fits to the variance-mean relationships for the conditions indicated. (D) N estimated from fluctuation analysis in both conditions. (E) Ca2+ cooperativity of evoked release is shown on a double logarithmic plot for w1118 Ctrl and NO. (F) Raw traces for PPRs of a Ctrl and NO-treated NMJ at 1.0 and 1.5 mM [Ca2+]e, illustrating a reduced release probability following NO exposure. (G) Summary of PPR at 1 and 1.5 mM [Ca2+]e for various ISI for Ctrl and NO (n = 5 larvae for Ctrl, n = 3 larvae for NO, with 2–3 NMJs per larva). (H) Images of myrGCaMP-expressing NMJs from a Ctrl (top) and a NO-treated larva (bottom) before, during (at 3 s, 60 Hz), and after the 8-s train. Changes in GCaMP5 fluorescence were analyzed for each bouton to calculate ΔF/F0. The raw data can be found in S3 Data. Values per NMJ were averaged and data denoting mean ± SEM for each condition of 6–8 NMJs (3–4 larvae) are shown in I. *p < 0.05, ***p < 0.001, Student t test. [Ca2+]e, extracellular calcium concentration; Ctrl, control; eEJC, evoked EJC; EJC, excitatory junction current; ISI, interspike interval; myrGCaMP, N-myristoylated GCaMP; N, number of release-ready vesicles; NMJ, neuromuscular junction; NO, nitric oxide; PPR, paired pulse ratio; stim, stimulation. https://doi.org/10.1371/journal.pbio.2003611.g003 The reduced QC seen following NO exposure can also be attributable to a change in the Ca2+ dependency of release, so we determined whether the reduced transmitter release is due to altered Ca2+ cooperativity of release [47]. The Hill slope was strongly reduced by NO (Ctrl: 3.2 ± 0.4 [n = 6], NO: 1.8 ± 0.7 [n = 5], p = 0.0024, Fig 3E); however, the half maximal effective Ca2+ concentration (EC50) was unaltered (Ctrl: 1.0 ± 0.03, NO: 1.0 ± 0.09, p > 0.05, Fig 3E), indicating that sensitivity to Ca2+ was not affected by NO. To further assess nitrergic effects on pvr, we used the PPR approach by delivering two pulses with interspike intervals (ISIs) between 10 and 200 ms at two different [Ca2+]e (1 and 1.5 mM, Fig 3F and 3G) in Ctrl and NO-treated NMJs. Analysis showed that Ctrl NMJs only exhibit slight potentiation at low Ca2+ and high ISI, indicative of low pvr. In contrast, pvr in the presence of NO was decreased, as shown by an increased PPR (potentiation at all ISI at 1 mM Ca2+ and 20 and 40 ms ISI at 1.5 mM Ca2+, p < 0.05, Ctrl versus NO at each ISI), which is also in agreement with elevated pvr in NOS “null” larvae. With about 500 release sites per NMJ and a QC of 200 (Ctrl) and 90 (NO), our data present estimated pvr values of 0.33 (Ctrl) and 0.16 (NO), with Ctrl values similar to estimates made previously in WT larvae [40]. Previously, we have shown that NO signaling can suppress mammalian P/Q and N-type Ca2+ channels [48]. In order to test whether altered Ca2+ influx could cause the observed effects on evoked release at the NMJ, we tested whether NO application for 60 min changed presynaptic Ca2+ levels during a train of synaptic stimulation. GCaMP5 was expressed presynaptically and activity-evoked Ca2+ influx in type 1b NMJ boutons was imaged at different extracellular Ca2+ concentrations (0.25–3 mM). Our data showed that NO had no effect on stimulated Ca2+ levels at any concentration tested (ΔF/F0, myrGCamP5: 3 mM Ca2+: Ctrl: 0.70 ± 0.09, NO: 0.78 ± 0.12 [n = 13–18 boutons from 4–6 NMJs each], p > 0.05; Fig 3H and 3I; GCaMP5: 0.25 mM Ca2+: 0.24 ± 0.03, NO: 0.14 ± 0.03, 0.5 mM Ca2+: Ctrl: 0.42 ± 0.08, NO: 0.50 ± 0.07, 1.5 mM Ca2+: Ctrl: 1.13 ± 0.14, NO: 1.18 ± 0.21 [n = 28–46 boutons from 7–11 NMJs each], p > 0.05; S5 Fig and S9 Data). Together, the data suggest that NO reduced evoked release and the frequency of spontaneous release, likely due to reduced release probability and Ca2+ cooperativity, which manifests itself in reduced vesicle fusion. We showed that the Ca2+ dependence of release, but not Ca2+ entry per se, was reduced by NO, which indicates a possible modulation of SNARE (-associated) protein interactions via NO-mediated PTMs. Enhanced denitrosylation signaling reverses and precludes NO effects S-nitrosylation is a reversible non-enzymatic protein modification, the levels of which can be regulated via S-nitrosoglutathione reductase (GSNOR), the sole alcohol dehydrogenase 5 (ADH-5) isozyme in vertebrate brains [49], which has a homologue in Drosophila (encoded by the formaldehyde dehydrogenase [fdh] gene). This de-nitrosylation process requires GSH. GSH is produced from L-glutamate and Cys via the enzyme glutamate-cysteine ligase (GCL), the rate-limiting step in GSH synthesis in fly [50]. The Drosophila GCL holoenzyme is heterodimeric, consisting of a catalytic (DmGCLc) and a modifier (DmGCLm) subunit, each encoded by a unique gene, and overexpression of either subunit increases cellular GSH levels [50]. In order to assess the contributions of SNO formation to the physiology at the NMJ, we investigated the effects of altering neuronal GSH levels. If NO mediates its observed actions via SNO formation, we should be able to prevent/reduce the effects on transmitter release by providing elevated GSH levels by (i) GSH supplementation, (ii) overexpression of GSNOR (fdh), or (iii) overexpression of GCL (DmGCLm/c) and, inversely, enhance NO effects by using RNA interference (RNAi) expression of the above proteins. We tested first the recovery of NO-mediated reduction of eEJC amplitudes following NO exposure for 50 min by washing out NO. eEJC amplitudes recovered slightly (Fig 4A, green and S4 Data); however, when washing in GSH (150 μM), the amplitudes recovered to control levels after 15 min (GSH [blue] versus NO at 50 min [red], p < 0.05), indicating a GSH-mediated reversal. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Genetic and pharmacological induction of denitrosylation reverses and prevents nitrergic effects on transmitter release and vesicle pools. (A) NO-induced suppression of evoked release (w1118 Ctrl eEJCs data in grey from Fig 1A) can be reversed after 15 min of GSH application (150 μM, blue). Wash out of NO alone shows mild nonsignificant recovery (green). (B) Mean QC for conditions indicated (OE, elav > UAS-fdh31 [GSNOR], elav > UAS-GCLm, elav > UAS-GCLc, all ± NO). NO has no effects on QC in genotypes indicated (w1118 + NO data in grey from Fig 1 for comparisons, ***p < 0.0001 w1118 + NO versus w1118), #p < 0.05, ##p < 0.01, ###p < 0.001, ####p < 0.0001 versus w1118 + NO. (C) Cumulative QC graphs for genotypes indicated (±NO) with linear regression to the last 200 ms, Ctrl in grey taken from Fig 2 for comparison. (D) Mean RRP sizes, NO has no effects on vesicle pool size in tested genotypes (w1118 data in grey from Fig 1, w1118 + NO versus w1118: ***p < 0.0001), ##p < 0.01, ###p < 0.01, ####p < 0.01 versus w1118 + NO. (E) PPR at 20 ms ISI for indicated conditions, ****p < 0.0001 versus Ctrl. (F) mEJCs analysis: left, mean mEJC amplitudes (±NO), middle, mean mEJC frequencies (±NO), right, mean mEJC decay kinetics for genotypes indicated (±NO). OE of GSNOR, GCLm, or GCLc prevents NO effects on the frequency of mEJCs. The raw data can be found in S4 Data. Data denote mean ± SEM for all data comparisons. ANOVA with post hoc Tukey-Kramer. Ctrl, control; eEJC, evoked EJC; EJC, excitatory junction current; fdh, formaldehyde dehydrogenase; GCLc, glutamate-cysteine ligase catalytic subunit C; GCLm, glutamate-cysteine ligase catalytic subunit M; GSH, glutathione; GSNOR, S-nitrosoglutathione reductase; ISI, interspike interval; mEJC, miniature EJC; NO, nitric oxide; n.s., nonsignificant; OE, overexpression; PPR, paired pulse ratio; QC, quantal content; RRP, readily releasable pool. https://doi.org/10.1371/journal.pbio.2003611.g004 To characterize effects of endogenous GSH formation, we used elav-Gal4-driven UAS-fdh31, UAS-DmGCLm, and UAS-DmGCLc overexpression. It has been shown that overexpression of either DmGCLc or DmGCLm results in enhanced enzyme activity and elevated GSH levels [50], GSNOR overexpression (elav > UAS-fdh31) reduces global S-nitrosylation in fly, and conversely, GSNOR-RNAi expression (elav > UAS-fdhri34) elevates SNO protein levels [51]. Overexpression of GSNOR and GCLm/c (Fig 4B–4D) prevented NO effects on QC (GSNOR: 238 ± 20 [n = 11], DmGCLm: 197 ± 32 [n = 7], DmGCLc: 215 ± 39 [n = 6], GSNOR+NO: 223 ± 24 [n = 8], DmGCLm+NO: 177 ± 10 [n = 7], DmGCLc+NO: 329 ± 26 [n = 3], p > 0.05) and vesicle pool sizes (GSNOR: 438 ± 51 [n = 10], DmGCLm: 400 ± 99 [n = 7], DmGCLc: 496 ± 93 [n = 6], GSNOR+NO: 472 ± 34 [n = 8), DmGCLm+NO: 360 ± 40 [n = 7], DmGCLc+NO: 685 ± 148 [n = 3], p > 0.05, Fig 4B–4D). These data confirm that by enhancing GSNOR and GCL activities, thereby elevating intracellular GSH levels, the effects of NO on pool size and pvr (PPR at 20 ms ISI; w1118 Ctrl [0.93 ± 0.03] versus NO [1.2 ± 0.07], p < 0.0001, GSNOR overexpression [0.88 ± 0.02], +NO [0.84 ± 0.05]/GCLm overexpression [0.87 ± 0.04], +NO [0.92 ± 0.03]/GCLc overexpression [0.99 ± 0.07], +NO [0.96 ± 0.02], p > 0.05, Fig 4E) were precluded, suggesting that this was due to reduced SNO formation. Furthermore, overexpression of GSNOR, DmGCLm, and DmGCLc prevented the reduction in mEJC frequency following NO exposure (fGSNOR: 2.4 ± 0.3 s−1 [n = 13]; fDmGCLm: 3.0 ± 0.3 s−1 [n = 13]; fDmGCLc: 1.5 ± 0.42 s−1 [n = 5]; fGSNOR+NO: 1.7 ± 0.2 s−1 [n = 7]; fDmGCLm+NO: 2.9 ± 0.4 s−1 [n = 13]; fDmGCLc+NO: 0.4 ± 0.1 s−1 [n = 3], p > 0.05 versus w1118 Ctrl and versus each Ctrl, Fig 4F) without affecting mEJC amplitudes (GSNOR: −0.6 ± 0.07 nA [n = 13]; DmGCLm: −0.7 ± 0.07 nA [n = 13]; DmGCLc: −0.5 ± 0.07 nA [n = 5]; GSNOR+NO: −0.6 ± 0.07 nA [n = 7]; DmGCLm+NO: −0.6 ± 0.08 nA [n = 13]; DmGCLc+NO: −0.6 ± 0.07 nA [n = 3], p > 0.05 versus w1118 Ctrl and versus each Ctrl, Fig 4F) or decays (GSNOR: 7.5 ± 0.2 ms [n = 13]; DmGCLm: 9.7 ± 0.4 ms [n = 13], DmGCLc: 6.7 ± 0.3 ms [n = 5], GSNOR+NO: 6.2 ± 0.2 ms [n = 7], DmGCLm+NO: 7.9 ± 0.5 ms [n = 13], DmGCLc+NO: 6.2 ± 0.4 ms [n = 3], p > 0.05 versus w1118 Ctrl and versus each Ctrl, Fig 4F). Furthermore, the reduction of endogenous GSNOR and DmGCLm activities (elav > UAS-RNAi) caused partial electrophysiological phenotypes, such as a decrease in eEJC amplitudes, QC, or vesicle pool size compared to w1118 Ctrl, with NO having no further major negative effects (S6 Fig and S9 Data). Nitrergic effects require the presence of cpx We next asked which signaling routes and PTMs are involved in NO modulation of release. The SNARE-binding and fusion-clamp protein cpx regulates not only the Ca2+ cooperativity of evoked release but also spontaneous release [14] as well as release probabilities [52], thereby presenting a strong candidate for mediating the observed NO-induced changes. Cpx acts by binding to the SNARE complex, thereby promoting the clamping of release, and only when replaced by synaptotagmin 1 in response to Ca2+ influx will vesicle fusion be initiated [14, 20]. Dmcpx function can be regulated by protein kinase A (PKA) phosphorylation of serine126 (Ser126) [7] or by prenylation at the C-terminus [15, 16]. In order to test whether cpx is required to exert NO effects, we first used cpx null mutants (cpxSH1, cpx-/-) [14]. In these animals, we detected a strong reduction in evoked release and QC (22.6 ± 3.2 [n = 11], p < 0.0001 versus Ctrl), which was unaffected by NO (13.8 ± 2.0 [n = 4], p > 0.05 versus cpx-/-, p < 0.0001 versus Ctrl, Fig 5A–5D and S5 Data). Similarly, when comparing the vesicle pool size, cpx-/- NMJs showed a strong reduction (26 ± 6 [n = 11], p < 0.0001 versus Ctrl), which again was unaffected by NO (22 ± 5 [n = 4], p > 0.05 versus cpx-/-, p < 0.0001 versus Ctrl, Fig 5A–5D). These data confirm that cpx is required for NO to induce suppression of evoked release and available vesicle pool size and suggest that NO might enhance the clamping function of cpx in WT larvae. We next tested the impact of NO on the clamping ability of cpx by characterizing spontaneous release. Interestingly, the frequency of spontaneous events inversely correlates with endogenous cpx levels [14]. We analyzed mEJCs in cpx-/- muscle 6 (m6), which exhibited an extremely high frequency [14] (>40 × w1118, Fig 5E). NO did not reduce the mEJC frequency in those preparations, although a precise analysis is difficult due to strong overlap of single mEJCs [14]. In order to allow more accurate frequency measurements in cpx-/- animals, we used neighboring muscle 5 (m5), posessing a synapse with approximately 4-fold fewer release sites compared to m6. Similar to m6, cpx-/- increased mEJC frequencies >10-fold compared to Ctrl (m5: w1118: 0.8 ± 0.2 s−1 [n = 3], cpx-/-: 11.6 ± 0.8 s−1 [n = 6], p < 0.0001); however, following NO exposure, this preparation did not show any change in mEJC frequency (m5 cpx-/- + NO: 9.7 ± 0.9 s−1 [n = 5], p > 0.05 versus m5 cpx-/-, Fig 5E and 5F), suggesting the requirement of cpx for the observed nitrergic effects. Nevertheless, we recorded from m6 of heterozygous animals, which exhibit higher frequencies than w1118 but are still accurately quantifiable (m6 cpx+/-: 4.6 ± 0.8 s−1 [n = 5]). Here, NO induced a strong reduction in the frequency (m6 cpx+/- + NO: 0.6 ± 0.2 s−1 [n = 5] #p < 0.05 versus m6 cpx+/- Ctrl, Fig 5E and 5F), similar to that seen in w1118. These data confirm that NO only modulates spontaneous release frequencies in the presence of cpx. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Nitrergic effects on evoked and spontaneous release require cpx. (A) Raw eEJC recordings of a 50-Hz train, 500 ms in cpx-/- larvae (±NO). (B) Mean QC for genotypes indicated, showing the lack of NO effects in cpx-/- larvae (w1118 Ctrl in grey from Fig 2). (C) Cumulative QC for the same conditions as in (B). (D) Mean QC and vesicle pool sizes for the genotypes indicated (w1118 Ctrl data from Figs 1 and 2, ****p < 0.0001 versus w1118 Ctrl). (E) mEJCs recordings from m6 and m5 in cpx-/- and cpx+/- ± NO. (F) Mean mEJC frequencies for conditions and genotypes indicated (****p < 0.0001 versus m5 w1118 Ctrl, #p < 0.05 versus m6 cpx+/- Ctrl). (G) Raw traces of mEJC recordings before (left) and after (right) high frequency stimulation in w1118 (±NO) and NOSC larvae. (H) Average fold change of mEJC frequency during 50 s after stimulation compared to baseline frequency before stimulation for conditions and genotypes indicated (NOS “null” comprised of data from NOSC and NOSΔ15). NO treatment (40 min) suppresses increases in frequency, NOS “null” potentiates relative increases and enhanced denitrosylation (GCLm and GSNOR OE) prevents NO-induced suppression (*p < 0.05, **p < 0.01 versus Ctrl, ##p < 0.01, ####p < 0.001 versus NO). The raw data can be found in S5 Data. Data denote mean ± SEM in all graphs, ANOVA with post hoc Tukey-Kramer. Ctrl, control; cpx, complexin; eEJC, evoked EJC; EJC, excitatory junction current; GCLm, glutamate-cysteine ligase catalytic subunit M; GSNOR, S-nitrosoglutathione reductase; mEJC, miniature EJC; m5, muscle 5; m6, muscle 6; NO, nitric oxide; OE, overexpression; QC, quantal content. https://doi.org/10.1371/journal.pbio.2003611.g005 Together, these data show that in the absence of cpx, NO causes no electrophysiological phenotypes. The NO-mediated reduction of eEJC amplitudes, QC, pool size, and mEJC frequency all require the presence of cpx, suggesting that its modulation might be responsible for the observed nitrergic effects, which could be explained by a gain-of-clamping function [53]. This potential effect was further investigated by using the established paradigm of activity-induced enhancement of spontaneous release at the Drosophila NMJ [7]. We assessed whether NO modulation of release also affects this activity-dependent signaling, which would strengthen the role of cpx as a target for nitrergic regulation and a general regulatory mechanism. PKA has been reported to modulate mEJC frequency potentiation in a cpx overexpression model (Dmcpx 7B, [7]). We confirmed that high frequency stimulation (50 Hz for 3 s) led to an enhanced mEJC frequency in w1118 NMJs relative to baseline (Ctrl: 1.9 ± 0.2-fold [n = 13], Fig 5G and 5H). Interestingly, repeating this protocol in larvae exposed to NO showed a lack of frequency potentiation (NO: 0.8 ± 0.1-fold [n = 14], p < 0.05 versus Ctrl), which was also ODQ independent (NO + ODQ: 1.0 ± 0.1-fold [n = 7], p > 0.05 versus NO, Fig 5G and 5H). To test whether the manipulation of PTMs also affects nitrergic suppression of frequency potentiation, we used larvae overexpressing GCLm and GSNOR and NOS “null” larvae. We found that GCLm and GSNOR overexpression occluded nitrergic effects on suppression of mEJC frequency potentiation, whereas the lack of NO signaling led to enhanced potentiation (GCLm + NO: 2.3 ± 0.4-fold [n = 7], GSNOR + NO: 2.5 ± 0.5-fold [n = 7], NOS “null” [comprised of n = 5 NOSC and n = 3 NOSΔ15]: 3.8 ± 0.3-fold, **p < 0.01 versus Ctrl, ##p < 0.01 versus NO, ####p < 0.001 versus NO, Fig 5G and 5H). These data show that NO suppresses the activity-mediated increase in mEJC frequency and suggest that, similar to phospho-incompetent cpx mutants [7], nitrergic modulation of WT cpx produces an inhibitory action on spontaneous release. The lack of PTM signaling leads to an enhanced frequency potentiation, strengthening the notion that NO-mediated effects are responsible for suppression of synaptic release and our data point towards modulation of cpx as a key signaling mechanism. Nitrergic activity affects farnesylation of cpx and enhances its clamping properties Having shown that cpx signaling is involved in NO-mediated effects on spontaneous and evoked release, we next considered if S-nitrosylation of the Cys residue within the C-terminus of cpx possessing the CAAX motif could explain the observed results. Importantly, prenylation has been studied in several genetically modified cpx proteins in which the CAAX motif was eliminated [15, 16]. These studies suggest that deletion of final parts of the C-terminus/final amino acid affects cpx localization, interactions with SNARE-proteins, and, subsequently, its function. To explore the effects of cpx farnesylation more in detail, we made use of Drosophila lines expressing green fluorescent protein (GFP)-tagged WT and mutant cpx (cpx1257, lacking the final amino acid [16]), referred to as CpxΔX. This mutant has been shown to exhibit altered co-localization with syntaxin at the dorsolongitudinal flight muscle (DLM) neuromuscular synapse. We assessed localizations of WT and mutant cpx at the NMJ (elav > UAS-cpx-GFP, elav > UAS-cpx1257-GFP) with respect to their interaction with the AZ protein, Brp. WT cpx exhibits diffuse localization within boutons (as previously reported [15]) with little co-localization with Brp (Fig 6A and 6B and S6 Data). In contrast, the mutant form, lacking farnesylation, is highly co-localized with Brp, as indicated by the increase in Pearson’s coefficient (Fig 6A and 6B; WT cpx: 0.35 ± 0.30 [n = 9], CpxΔX: 0.65 ± 0.03 [n = 9], p < 0.0001). These data confirm that preventing cpx farnesylation results in enhanced co-localization with AZ. To further support these data, we conducted high-resolution stimulated emission depletion (STED) microscopy [54] and analyzed the Pearson’s coefficient for the co-localization of Brp with cpx. This experiment verified the confocal data showing enhanced co-localization of CpxΔX with Brp versus WT cpx (WT cpx: 0.13 ± 0.02 [n = 25], CpxΔX: 0.27 ± 0.02 [n = 23], p < 0.0001, Fig 6C and 6D). As Dmcpx possesses a predominant clamping function [23], we propose that NO could lead to a reduction in farnesylation, a consequent stronger interaction with the SNARE complex at the AZ, and thereby enhance its clamping function upon transmitter release. To specifically confirm co-localizations, we used the high-resolution proximity ligation assay (PLA), with which we imaged interactions of Brp with cpx. We used both lines, WT cpx-GFP and CpxΔX-GFP expressing larvae, and found that PLA signals are strongly enhanced at NMJs expressing the mutant cpx (Fig 6E and 6F; WT cpx: 0.04 ± 0.004 [n = 9], CpxΔX: 0.12 ± 0.02 [n = 9], p = 0.009). As the co-localization data may depend upon expression of GFP-tagged cpx, we confirmed equal GFP expression levels in both lines by immunoblotting (S9A Fig). These co-localization and PLA experiments confirm an enhanced association of a mutated farnesylation-incompetent cpx with Brp and suggest that lack of farnesylation renders cpx in close proximity to release sites of AZs. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Lack of cpx farnesylation promotes its co-localization with AZs. (A) Representative maximal projection confocal images of GFP-tagged WT Cpx and mutant Cpx1257 (CpxΔX) at the NMJ (green: cpx, red: Brp). (B) Enhanced Cpx-Brp co-localization indicated by higher Pearson’s coefficient for CpxΔX (n–number of NMJs). (C) STED images showing cpx and Brp localization in single boutons in WT and CpxΔX mutants. The mutation (Cpx1257) increases the co-localization of cpx with Brp as indicated by the enhanced Pearson’s coefficient (D) (n–number of boutons). (E) Maximal projection confocal images of NMJs from larvae expressing WT Cpx and CpxΔX. PLA fluorescence shown in red and HRP staining in green. (F) Summated PLA signal volumes relative to NMJ volumes for cpx (GFP)-Brp interactions (n–number of NMJs). The raw data can be found in S6 Data. Data denote mean ± SEM, Student t test, **p = 0.009, ****p < 0.0001. AZ, active zone; Brp, Bruchpilot; cpx, complexin; GFP, green fluorescent protein; HRP, horseradish peroxidase; NMJ, neuromuscular junction; PLA, proximity ligation assay; STED, stimulated emission depletion; WT, wild-type. https://doi.org/10.1371/journal.pbio.2003611.g006 In order to assess this possibility further, we used pharmacological and genetic tools to modulate cpx farnesylation and compared protein localization and synaptic release following farnesyl transferase (FTase) inhibition and NO exposure. Reduced expression of the Drosophila ortholog of FTase or inhibition of FTase by L-744,832 and GGTI-298 have strong effects on fly lethality [55], implicating a crucial function of this signaling in fly. First, we tested how FTase inhibition (20 μM L-744,832 + 10 μM GGTI-298) and NO exposure affect cpx co-localization with the SNARE complex proteins syntaxin and synaptotagmin or Brp, using the PLA. We measured total PLA volume of NMJ z-stacks and normalized PLA signals to NMJ volume. We found that both treatments (depicted as “farnesyl inh” and “NO,” Fig 7A and 7B and S7 Data) led to enhanced co-localization of cpx with syntaxin and Brp (syntaxin-cpx: Ctrl: 0.04 ± 0.007, NO: 0.12 ± 0.02, farnesyl inh: 0.11 ± 0.02, Brp-cpx: 0.02 ± 0.007, NO: 0.08 ± 0.03, farnesyl inh: 0.09 ± 0.05, Fig 7A and 7B; p < 0.01, p < 0.001 versus Ctrl), suggesting that NO PTMs and farnesylation inhibition enrich cpx at the AZ. When analyzing the interactions between the Ca2+ sensor synaptotagmin and cpx, we found that this interaction was completely suppressed following treatments (Ctrl: 0.2 ± 0.06, NO: 0.03 ± 0.006, farnesyl inh: 0.04 ± 0.007, Fig 7A and 7B; p < 0.01 versus Ctrl). The PLA data were further supported by STED imaging studies showing identical changes in protein co-localization, as determined by Pearson’s coefficient analysis (S7 Fig and S9 Data). One possibility to allow for greater amounts of cpx to be available for binding to SNAREs is by enhancing free and soluble cytosolic levels as a consequence of reduced farnesylation. Farnesylation of cpx results in its membrane tethering, and thus protein fractions, which are membrane bound, are less mobile than soluble cytosolic proteins. To assess the mobility of potentially farnesylated versus soluble (non-farnesylated) cpx and thus distinguish between these two pools of cpx, we performed fluorescence recovery after photobleaching (FRAP) analysis of GFP-tagged WT and farnesylation-incompetent cpx (CpxΔX). Although a previous study did not detect differences between farnesylated versus non-farnesylated cpx isoform using this method with a photo-bleaching area of half a bouton [15], we found that accurate FRAP analysis of cpx-GFP mobility can only be performed by using substantially smaller bleaching areas, as reported previously [56] (S8 Fig and S9 Data). Using this approach, we found that bleaching an area of 2.5 μm2 (instead of >10 μm2) generally leads to faster recovery rates (S8 Fig and S9 Data). Our data confirmed that lack of farnesylation (CpxΔX) allows for greater movement of cpx and faster recovery (tau: WT cpx: 18.1 ± 1.7 ms, CpxΔX: 11.9 ± 1.2 ms [p < 0.05], WT Cpx + NO: 8.8 ± 0.8 ms [p < 0.0001], n = 18–20, Fig 7C), as expected for a soluble protein. Our data further show that NO treatment caused the same increase in recovery rates (Fig 7C), suggesting that NO also prevented farnesylation. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Reduced farnesylation of cpx enhances AZ localization and alters interactions with SNARE proteins. (A) Maximal projection confocal images of w1118 NMJs in Ctrl, following 60 min of exposure to NO or farnesylation inhibitor (“farnesyl inh”: 10 μM GGTI-298 + 20 μM L-744,832). PLA fluorescence in red and HRP staining in green for: left, syntaxin-cpx; middle: Brp-cpx; right: synaptotagmin-cpx interactions. (B) Analysis of summated PLA signal volumes relative to NMJ volumes. (C) FRAP experiments were performed at NMJs expressing GFP-cpx, shown as representative images of WT GFP-cpx at different time points (bleaching area: 2.5 μm2, scale bar: 2 μm). Right, mean data showing recovery of WT, CpxΔX, and NO-treated WT cpx, with mean tau values summarized. Note, lack of farnesylation due to the mutation or NO treatment results in faster recovery rates. (D) Representative mEJC recordings following 60 min incubation with GGTI-298 + L-744,832 (“farnesyl inh”) or of a larva expressing FTase-RNAi with mean mEJC frequencies. (E) Trains of 50-Hz stimulation of a larva incubated for 60 min with GGTI-298 + L-744,832 (“farnesyl inh”) or expressing FTase-RNAi with mean eEJC amplitudes and QC. (F) Cumulative QC of 50-Hz trains with mean estimated RRP sizes (right). The raw data can be found in S7 Data. Data denote mean ± SEM, Student t test, or ANOVA with post hoc Tukey-Kramer as indicated, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. AZ, active zone; Brp, Bruchpilot; cpx, complexin; Ctrl; control; eEJC, evoked EJC; EJC, excitatory junction current; FRAP, fluorescence recovery after photobleaching; FTase, farnesyl transferase; GFP, green fluorescent protein; HRP, horseradish peroxidase; mEJC, miniature EJC; NMJ, neuromuscular junction; NO, nitric oxide; PLA, proximity ligation assay; QC, quantal content; RNAi, RNA interference; RRP, readily releasable pool; SNARE, soluble N-ethyl-maleimide-sensitive fusion protein Attachment Protein Receptor; WT, wild-type. https://doi.org/10.1371/journal.pbio.2003611.g007 These data suggest that due to enriched local levels, cpx outcompetes synaptotagmin for SNARE binding at the AZ, thereby displacing synaptotagmin, as reported previously in biochemical studies [53]. Our data show that pharmacological and genetic inhibition of farnesylation promotes cpx co-localization with the AZ and supports the notion that this negatively impacts on synaptotagmin-SNARE complex binding, subsequently reducing release. The specificity of the PLA was corroborated by lack of Brp-cpx PLA signals in cpx-/- larvae (S9B–S9D Fig). Next, we explored the possibility of whether specific inhibition of FTase activity by L-744,832 and GGTI-298 and FTase RNAi mimics the effects of NO on synaptic transmission. We found that, in both conditions, the frequency of mEJCs was reduced to similar values seen following NO exposure (fmEJC: L-744,832 + GGTI-298: 0.7 ± 0.1 s−1 [n = 8], p = 0.0051 versus Ctrl, FTase RNAi: 0.9 ± 0.2 s−1 [n = 9], p = 0.0136 versus Ctrl, Student t test, Fig 7D). Importantly, both L-744,832 + GGTI-298 and FTase RNAi expression reduced evoked transmission and available vesicle pool size to levels similar to those following NO incubation (L-744,832 + GGTI-298: eEJC: 56 ± 5 nA, QC: 80 ± 13 [n = 9], pool size: 180 ± 27 [n = 9], p < 0.0001 versus each w1118 Ctrl; FTase RNAi: eEJC: 75 ± 5 nA, QC: 82 ± 6 [n = 9], pool size: 120 ± 17 [n = 9], p < 0.0001 versus each w1118 Ctrl, Student t test, Fig 7E and 7F). These data suggest that the farnesylation status of cpx mediates nitrergic effects, resulting in changed SNARE protein interactions, which determines the physiological outcome of cpx. To further investigate the effects of NO directly on the prenylation process, we employed the well-characterized GFP-CAAX transfection model [57]. Here, human embryonic kidney (HEK) cells were transfected with GFP-CAAX (K-Ras motif) and the membrane association was assessed in response to prenylation inhibition and NO treatment. In control conditions, GFP exhibited a strong fluorescence signal at the membrane, which disappeared and redistributed into the cytosol following pharmacological inhibition of prenylation (L-744,832 + GGTI-298, p < 0.0001), confirming the prenylation-mediated localization of GFP-CAAX to the membrane (Fig 8A and S8 Data). Importantly, we showed that NO treatment (propylamine propylamine NONOate [PAPA-NONOate], p < 0.0001) induced a similar phenotype, with GFP being localized predominantly in a cytosolic manner—suggesting that NO prevents farnesylation through the same pathway (Fig 8A). To confirm that the Cys within the CAAX motif can undergo S-nitrosylation, we performed the Biotin Switch Assay on cpx-3 from isolated mouse retinas. NO donor incubation induced a >2-fold increase in SNO-cpx (Fig 8B), confirming this PTM on cpx and suggesting that this PTM is responsible for NO-induced changes in localization and function of cpx. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 8. Cpx nitrosylation and block of farnesylation leads to redistribution of GFP-CAAX and enhances Dmcpx localization at AZs. (A) GFP-CAAX expression in HEK cells showing membrane fluorescence signals in Ctrls. Pharmacological inhibition of farnesylation by GGTI-298 + L-744,832 (“farnesyl inh,” p < 0.0001 versus Ctrl, ANOVA) and NO incubation (p < 0.0001 versus Ctrl, ANOVA) result in a redistribution of GFP fluorescence into the cytosol. Fluorescence signals were analyzed by line scan and plotted as intensities (a.u.) over distance across the cell somata, n = 80–104 cells, scale bar: 20 μm. (B) Mouse cpx-3 is S-nitrosylated in response to NO donor application. Immunoblot intensities increased 2.4 ± 0.5-fold following NO application. (C) Representative recordings of a 50-Hz train and spontaneous activity of a larva expressing Dmcpx 7AC140W with mean eEJC amplitudes, QC, vesicle pool size, and mEJC frequency shown in (D). (E) Top, maximal projection confocal images of NMJs expressing a WT cpx or Dmcpx 7AC140W showing the PLA signal in red; bottom, STED images showing cpx and Brp staining in WT and Dmcpx 7AC140W mutants. (F) Analysis of PLA data. (G) Co-localization data with Pearson’s coefficient for interactions of cpx with Brp. The raw data can be found in S8 Data. Data denote mean ± SEM, Student t test (D, F), *p < 0.05, ***p < 0.001, ****p < 0.0001. a.u., arbitrary unit; AZ, active zone; Brp, Bruchpilot; cpx, complexin; Ctrl, control; eEJC, evoked EJC; EJC, excitatory junction current; GFP, green fluorescent protein; HEK, human embryonic kidney; HRP, horseradish peroxidase; mEJC, miniature EJC; NMJ, neuromuscular junction; NO, nitric oxide; PLA, proximity ligation assay; QC, quantal content; STED, stimulated emission depletion; WT, wild-type. https://doi.org/10.1371/journal.pbio.2003611.g008 To specifically confirm the effects of S-nitrosylation and SNO interaction with farnesylation of cpx in Drosophila, we generated and expressed a nitroso-mimetic cpx mutant (Dmcpx 7AC140W) in a cpx null background (cpxSH1) and assessed synaptic responses. The Cys140 of Dmcpx is located within a hydrophobic region, as predicted in the Kyle Doolittle plot, which favors S-nitrosylation [58]. This mutant exhibits reduced evoked responses, QC, and vesicle pool sizes (eEJC: 70 ± 7 nA, QC: 106 ± 8, pool size: 204 ± 23 [n = 15 each], p < 0.0001 versus each w1118 Ctrl, Fig 8C and 8D), indicating that the mimicking of S-nitrosylation and simultaneous lack of farnesylation of cpx caused the observed changes. Importantly, this mutation also induced a reduction in spontaneous activity (fmEJC: 1.3 ± 0.2 s−1 [n = 15], p < 0.05 versus w1118 Ctrl, Fig 8C and 8D), reinforcing the argument of enhanced clamping function due to SNO formation and lack of farnesylation. The expression of WT cpx in the null background did not affect QC, pool size, or mEJC frequency (QC: 167 ± 17 [n = 5]; pool size: 381 ± 76 [n = 5]; fmEJC: 2.4 ± 0.4 s−1 [n = 10 each], p > 0.05 versus each w1118 Ctrl). To confirm changes in localization of Dmcpx 7AC140W, we analyzed PLA signals and found that Dmcpx 7AC140W highly co-localizes with Brp, in strong contrast to WT cpx (WT: 0.025 ± 0.013, Dmcpx 7AC140W: 0.17 ± 0.03 [n = 6–7], p < 0.0001, both expressed in cpx-/- background, Fig 8E and 8F). The data from the PLA experiments were confirmed by STED confocal microscopy, showing significantly higher Pearson’s coefficients for the co-localization of the cpx mutant C140W with Brp relative to the interaction of WT cpx with Brp (WT cpx: 0.13 ± 0.03, Dmcpx 7AC140W: 0.34 ± 0.02 [n = 20–24], p < 0.0001; Fig 8E and 8G). These data demonstrate that independent approaches to block farnesylation (and mimic of cpx-SNO) recapitulate nitrergic modulation of release and protein localization and therefore link for the first time NO-induced PTM and farnesylation signaling of cpx. We propose that S-nitrosylation acts as a novel endogenous pathway to alter cpx farnesylation signaling and protein–protein interactions and thereby allows a fine-tuning of synaptic function. Discussion NO regulates a multitude of physiological and pathological pathways in neuronal function via generation of cGMP, thiol-nitrosylation, and 3-nitrotyrosination in health and disease [59]. Here, we show by employing biochemical and genetic tools in Drosophila, mouse, and HEK cells that NO can S-nitrosylate cpx and modulate—in a cGMP-independent manner—neurotransmitter release at the NMJ by interfering with its prenylation status, thereby affecting the localization and function of this fusion-clamp protein. We found that these nitrergic effects are reversed by GSH application or overexpression of GSH-liberating and de-nitrosylating enzymes (GCLm/c, GSNOR). GSH is the major endogenous scavenger for the NO moiety by the formation of S-nitrosoglutathione (GSNO) and consequently reduces protein-SNO levels via trans- and de-nitrosylation. The suppression of NOS activity facilitates synaptic function and the data support the notion that endogenous or exogenous NO enhances S-nitrosylation, reduces cpx farnesylation, and diminishes release. Of the numerous synaptic molecules involved in release, cpx in particular has been implicated in the regulation of both evoked and spontaneous release due to its fusion-clamp activity. Despite the seemingly simple structure of cpx, its physiological function is highly controversial, as this small SNARE-complex binding protein can both facilitate but also diminish fast Ca2+-dependent and spontaneous release, depending on the system studied [22, 25, 53, 60]. In addition, there are different mammalian isoforms of cpx (1–4), which differ in their C-terminal region, with only cpx 3/4 containing the CAAX prenylation motif. Farnesylation in general determines protein membrane association and protein–protein interactions [61], and some cpx isoforms, such as muscpx 3/4 and Dmcpx 7A, are regulated in this manner [13, 23, 62]. However, muscpx 1/2 does not possess a CAAX motif, suggesting differential regulatory pathways to modulate cpx function. In Drosophila, there are alternative splice variants resulting from a single cpx gene, but the predominant isoform contains the CAAX motif (Dmcpx 7A), implicating the importance of this signaling molecule [15, 16]. The other splice isoform (Dmcpx 7B) lacks the CAAX motif and is expressed at about 1,000-fold lower levels at the larval stage [15], thus making Dmcpx 7A the dominant isoform to be regulated by farnesylation. However, the lack of Dmcpx 7B phosphorylation by PKA induces similar phenotypes as seen in our experiments when assessed following an induction of activity-dependent potentiation of mEJC frequency [7], which also may involve cpx–synaptotagmin 1 interactions. Interestingly, both depletion and excessive levels of cpx suppress Ca2+-dependent and -independent exocytosis [63]. Cpx may promote SNARE complex assembly and simultaneously block completion of fusion by retaining it in a highly fusogenic state. Ca2+-dependent fusion is promoted below a concentration of 100 nM of cpx, whereas above 200 nM, it exhibits a clamping function resulting in a bell-shaped response curve [64]. Previous work suggests that synaptotagmin 1, once bound to Ca2+, relieves the cpx block and allows fusion. Another study reported that selective competition between cpx and synaptotagmin 1 for SNARE binding allows regulation of release [53]. Our data are in agreement with the latter findings, as we observed reduced synaptotagmin 1–cpx interactions following the block of farnesylation (Fig 7), indicating fewer synaptotagmin molecules binding to the SNARE complex to displace cpx. This limited replacement of cpx by synaptotagmin has been implicated in biochemical studies showing that local excess of cpx inhibits release, presumably by outcompeting synaptotagmin binding [53, 60]. Thus, synaptotagmin-SNARE binding is strongly dependent upon the local concentration of cpx [53]. Alternatively, and we cannot exclude this possibility, the modulation of cpx may simply alter its binding to the SNARE complex without directly displacing synaptotagmin, but interpretation of the data from our assays (PLA, co-localization) would not allow us to distinguish between these possibilities. Our data are compatible with the idea that cpx binds to the SNARE complex, facilitates assembly, and then exerts its clamping function by preventing full fusion due to SNARE complex stabilization and subsequent increased energy barrier to allow fusion. Our model could provide an explanation of how cpx can be regulated to signal downstream to modulate transmitter release. So far, there are no data available, apart from mutation studies, as to how cpx function can be altered. We provide data indicating a physiologically relevant mechanism to adjust cpx function, possibly to the requirements of the neuron to adjust synaptic transmission. This likely occurs due to Cys S-nitrosylation and suppression of farnesylation, allowing greater amounts of hydrophilic cpx, not bound to endomembranes, to be available for binding with the SNARE complex in an altered configuration. This cross signaling between nitrosylation/farnesylation has been proposed to act as a molecular switch to modulate Ras activity [65]. Our data show that enhanced nitrergic activity and blocking farnesylation, either genetically (CpxΔX) or pharmacologically, alters the localization of cpx at the Drosophila NMJ and that of GFP-CAAX in HEK cells (Figs 6–8). Furthermore, by using a nitroso-mimetic cpx mutant, we found enhanced co-localization of cpx with the AZ protein Brp, implying a localization-function relationship (Fig 8). This consequently increases the net-clamping function because of elevated local concentrations of cpx. Dmcpx specifically exhibits a strong clamping function, as shown following overexpression in hippocampal neurons, which causes suppression of evoked and spontaneous release accompanied by a reduction of the release probability [23] or reduced vesicle fusion efficiency in in vitro assays [64]. Two independent studies eliminating the CAAX motif in Dmcpx (cpx572 and cpx1257) investigated localization-function interactions and showed disagreeing effects on both release and cpx localization [15, 16]. In particular, it has also been shown that the truncated cpx (cpx572, lacking the last 25 amino acids) does not co-localize with Brp [15]. Interestingly, this mutant causes a strong decrease in C-terminal hydrophobicity and a modest physiological response (increased mini frequency, decreased evoked amplitudes equivalent to a loss of clamping and loss of fusion function) relative to the total knock-out (KO). In contrast, the cpx mutant with single amino acid deletion (cpx1275) causes no effect on evoked but identical effects on the frequency of spontaneous release, suggesting a lack of clamping but no lack of fusion function. In addition, this mutant now co-localizes with the AZ at the NMJ [16]. These two studies indicate that the different mutations cause contrasting electrophysiological and morphological phenotypes, indicating that it is due to the nature of the mutation (lack of the last 25 amino acids versus 1 amino acid), which highlights the importance of a functional C-terminus. More recent studies have shown that deletions of the final amino acids (6 or 12 residues) completely abolished the membrane binding of cpx-1, impairing its inhibitory function and confirming the requirement of an intact C-terminus for inhibition of release [66, 67]. Here, we use an endogenous cpx with intact hydrophobic C-terminus, allowing physiological membrane binding. This is essential for inhibitory function, as the C-terminus is required for selective binding to highly curved membranes, such as those of vesicles [68]. Thus, as we used different approaches to alter farnesylation and generated a single amino acid mutant cpx (Dmcpx 7AC140W), leaving the C-terminus intact, our studies were performed under conditions of endogenous regulation of cpx function and thus provide new functional data on cpx signaling. Importantly, our data show that this regulation alters cpx function, and this is the first study to provide an explanation for the differential effects observed using cpx mutants or even cpx protein fragments in mammals, worm, and fly in various cross-species rescue experiments [20, 23]. Our data are in agreement with a model that non-farnesylated hydrophilic and soluble cytosolic cpx binds to the vesicular membrane via its C-terminal interactions, thereby exerting its inhibitory effect. When proteins are farnesylated, they are likely tethered to endomembranes, other than vesicle membranes [12]. It has to be distinguished between cpx interaction with the vesicle membrane as a result of the hydrophobic C-terminus, allowing cpx to become in close proximity to the AZ, and cpx endomembrane binding following farnesylation, which prevents cpx interactions with the AZ. However, in our case, SNO modification may enhance the binding to other proteins (e.g., SNAREs), thereby augmenting the effects. These additional interactions with unknown binding partners may affect proper cpx function and explain some of the discrepancies seen in studies using other genetically altered cpxs. In summary, our study provides new data to illustrate a potential mechanism to regulate cpx function in a physiological environment, and we showed that NO acts as an endogenous signaling molecule that regulates synaptic membrane targeting of cpx, a pathway that may reconcile some of the controversial findings regarding cpx function. We suggest that increased S-nitrosylation and consequent lack of farnesylation leads to enhanced cytosolic levels of a soluble hydrophilic cpx and less endomembrane-bound fractions (Fig 9), because farnesylation-incompetent proteins remain in the cytosol [12]. These novel observations advance our understanding of similar nitrergic regulation of farnesylation that may be relevant for mammalian cpx-dependent synaptic transmission at the retina ribbon synapse and other brain regions [13]. Finally, this work has broader implications for physiological or pathological regulation of the prenylation pathway not only during neurodegeneration and aging, when enhanced S-nitrosylation might contribute to abnormal farnesylation signaling [69, 70], but also in other biological systems in which nitrergic activity and prenylation have important regulatory functions such as in cardio-vasculature or cancer signaling [71]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 9. Effects of SNO formation on farnesylation and cpx function. (A) The farnesyl transferase facilitates the addition of a farnesyl group to cpx (green) containing the CAAX motif. Upon S-nitrosylation, this motif is not recognized and the protein will not be farnesylated, resulting in lack of endomembrane targeting. (B) The schematic shows the effects of S-nitrosylation of cpx Cys within CAAX, resulting in fewer proteins being farnesylated and tethered to endomembranes. This allows a greater proportion of cytosolic cpx to be able to bind to the SNARE complex (including syntaxin [yellow]/Munc-18 [red], SNAP-25 [dark purple], synaptobrevin [purple]) to compete with synaptotagmin [blue] binding and prevent fusion because of its clamping function, which results in reduced transmitter release. This process is reversible and depends on the availability of GSH-dependent de-nitrosylation of cpx. GSH is generated by GCL and GSNO is broken down into G-SS-G and NH2OH by GSNOR activity. Cys, cysteine; cpx, complexin; G-SS-G, glutathione disulphide; GCL, glutamate-cysteine ligase; GluR, glutamate receptor; GSH, glutathione; GSNO, S-nitrosoglutathione; GSNOR, S-nitrosoglutathione reductase; NH2OH, hydroxylamine; SNARE, soluble N-ethyl-maleimide-sensitive fusion protein Attachment Protein Receptor; SNO, S-nitrosothiol. https://doi.org/10.1371/journal.pbio.2003611.g009 Methods Fly stocks Flies were raised on standard maize media at 25 °C at a 12-h LD cycle. The elav-Gal4 [C155] driver was obtained from the Bloomington Stock Center (Indiana, US). The UAS-RNAi lines (GCLm [CG4919], GCLc [CG2259], and Fnta [CG2976]) were purchased form the Vienna Drosophila Resource Centre (VDRC). The use of the UAS-Gal4 bipartite expression system to drive pan-neuronal expression excludes potential postsynaptic effects. The elav-Gal4 driver (female flies) and the UAS responder lines (male flies) were crossed to obtain offspring expressing the genes of interest and w1118 were used as Ctrls. The fluorescent Ca2+ sensor GCaMP5 was tethered to the plasma membrane with an N-terminal myristoylation (myr) sequence as described previously [72]. The UAS-myrGCaMP5 and cpxSH1 null mutant lines were provided by Troy Littleton (MIT, Cambridge, MA) [73]. GCaMP5 was expressed in glutamatergic neurons (OK371-Gal4; UAS-GCaMP5). cpx expression levels are shown for w1118 and cpxSH1 larvae in S9B and S9C Fig. UAS-fdh31 (expression of fdh homologue of mammalian GSNOR/ADH-5) and UAS-fdhri34/25 (expression of fdh RNAi) mutant transgenic lines were kindly provided by Li Liu Institute of Biophysics, Chinese Academy of Sciences, Beijing, China) [51]. NOSΔ15/NOSC lines were provided by Patrick O’Farrell (UCSF, San Francisco, CA). NOSΔ15 deletion removes sequences encoding residues 1–757, encompassing the entire oxygenase domain and including regions that bind the catalytic heme and the substrate rendering the lines NOS “null” [33, 34]. The syn-null mutant transgenic line (Syn97) was generously provided by Erich Buchner (Universitätsklinikum Würzburg, Germany) [44]. UAS-EGFP-cpx and UAS-EGFP-cpx1257 transgenic lines were kindly provided by Fumiko Kawasaki (Penn State University, PA) [16]. UAS-GCLm and UAS-GCLc transgenic lines were provided by William C. Orr (Southern Methodist University Dallas, TX). Cloning cDNAs encoding for cpx 7A was a gift from Troy Littleton and used as a template for downstream PCRs. Cys 140 of cpx7A isoform was mutated to tryptophan to generate S-nitrosylation mimic mutant. PCR products, which include XhoI and XbaI restriction sites, were cloned into the pJFRC2 vector [74]—a gift from Gerald Rubin (Addgene plasmid no. 26214)—by standard methods. The resulting constructs were injected into attP40 Drosophila strains. The resulting transgenic lines (Dmcpx7AC140W and WT Dmcpx) were crossed into a cpxSH1 background [7] using standard balancing techniques. The FlincG3 ORF was amplified from pTriEx4-H6-FGAm (FlincG3) (Addgene plasmid no. 49202) and the resultant PCR product cloned into pUASTattB by the Protein Expression Laboratory (PROTEX), University of Leicester. Microinjection of the pUASTattB plasmid was performed by the University of Cambridge, Department of Genetics Fly Facility. Electrophysiology TEVC recordings were performed as described previously [75]. Sharp-electrode recordings were made from ventral longitudinal m6 in abdominal segments 2 and 3 of third instar larvae using pClamp 10, an Axoclamp 900A amplifier and Digidata 1440A (Molecular Devices, US) in hemolymph-like solution 3 (HL-3) [76]. Recording electrodes (20–50 MΩ) were filled with 3 M KCl. mEJCs were recorded in the presence of 0.5 μM tetrodotoxin (Tocris, UK). All synaptic responses were recorded from muscles with input resistances ≥4 MΩ, holding currents <4 nA at −60 mV and resting potentials more negative than −60 mV at 25 °C, as differences in recording temperature cause changes in glutamate receptor kinetics and amplitudes [77]. Holding potentials were −60 mV. The extracellular HL-3 contained (in mM): 70 NaCl, 5 KCl, 20 MgCl2, 10 NaHCO3, 115 sucrose, 5 trehalose, 5 HEPES, and 1.5 CaCl2 (0.5–3.0 mM in Fig 3 and S3 Data, as specified). Average single eEJC amplitudes (stimulus: 0.1 ms, 1–5 V) are based on the mean peak eEJC amplitude in response to 10 presynaptic stimuli (recorded at 0.2 Hz). Nerve stimulation was performed with an isolated stimulator (DS2A, Digitimer). Paired-pulse experiments were performed by applying 5 repetitive stimuli (0.2 Hz) at different intervals (20, 40, 100, 200 ms) for each cell at each ISI. All data were digitized at 10 kHz and for miniature recordings, 200-s recordings, we analyzed to obtain mean mEJC amplitudes, decay, and frequency (f) values. QC was estimated for each recording by calculating the ratio of eEJC amplitude/average mEJC amplitude, followed by averaging recordings across all NMJs for a given genotype. mEJC and eEJC recordings were off-line low-pass filtered at 500 Hz and 1 kHz, respectively. Materials were purchased from Sigma-Aldrich (UK) unless otherwise stated. Variance-mean analysis of eEJCs Approximately 40 eEJCs were elicited at different [Ca2+]e, ranging from 0.5 to 3 mM to give mean eEJC amplitudes (I). The mean eEJC is given by I = Npvrq [45], with N being the number of independent release-ready vesicles, pvr the vesicular release probability, and q the quantal size at each given [Ca2+]e. The eEJC variance was calculated as previously described [45]. The plots of the variance-mean were obtained for each cell and fitted with the parabolic function Var(I) = I2/N + qI that was a constraint to pass through the origin. Upon fitting the parabola, pvr and q were calculated using the equations: q = A/(1+CV2) and pvr = I(B/A)(1+CV2) where CV2 is the coefficient of variation of the eEJC amplitudes at a given [Ca2+]e concentration calculated as CV2 = (eEJCs standard deviation/mean amplitude)2; A and B were obtained from the fitting parameters. Estimated values were not corrected for variability in mEJC amplitude distributions or latency fluctuations. Ca2+ cooperativity was assessed by plotting eEJC amplitudes over [Ca2+]e and fitted with the Hill equation (mean eEJC amplitude plotted versus different [Ca2+]e: eEJC([Ca2+]) = eEJCmax[1+(EC50/[Ca2+])slope]−1), yielding the Hill slope as a measure of Ca2+ cooperativity. Cumulative postsynaptic current analysis The apparent size of the RRP was probed by the method of cumulative eEJC amplitudes [78]. Muscles were clamped to −60 mV and eEJC amplitudes during a stimulus train (50 Hz, 500 ms [of a 1-s train]) were calculated as the difference between peak and baseline before stimulus onset of a given eEJC. Receptor desensitization was not blocked as it did not affect eEJC amplitudes, because a comparison of the decay of the first and the last eEJC within a train did not reveal any significant difference in decay kinetics. The number of release-ready vesicles (N) was obtained by back extrapolating a line fit to the linear phase of the 500-ms cumulative eEJC plot (the last 200 ms of the train) to time zero. N was then obtained by dividing the cumulative eEJC amplitude at time zero by the mean mEJC amplitude recorded in the same cell. To calculate the QC in the train, we used mean mEJC amplitudes measured before the train. Immunohistochemistry Third instar larvae were dissected in ice-cold PBS then fixed in 4% paraformaldehyde. After permeabilization with PBS-0.1% Triton (PBS-T) and blocking with PBS-T containing 0.2% bovine serum albumin (BSA) and 2% normal goat serum, larval fillets were incubated at 4 °C overnight in solutions of primary antibody. The following antibody dilutions were used: NC82 (supernatant) anti-Brp (Bruchpilot) 1:200, cpx (1:500), syntaxin (1:200), synaptotagmin (1:200), and GFP (1:200). After 3 × 10 min washes in PBS-T, larvae were incubated with AlexaFluor 488 goat anti-HRP (Jackson Immuno Research) and AlexaFluor 546 goat anti-mouse 1:500 dilution for 90 min at room temperature. Larvae were mounted using Vectashield mounting medium (Vector Labs) and NMJ 6/7 (segments A2 and A3) images were acquired with a Zeiss laser-scanning confocal microscope (LSM 510, Zeiss). Image analysis was performed with ZEN (Zeiss) and Volocity 6.3 software. STED microscopy Images were acquired on a Leica TCS SP8 system attached to a Leica DMi8 inverted microscope (Leica Microsystems). Excitation light (488 nm for AlexaFluor488 or 561 nm for AlexaFluor568) was provided by a white light laser with a repetition rate of 80 MHz. Images were acquired using a 100× 1.4 NA oil immersion objective and fluorescence was detected through a bandpass of 495–550 nm (AlexaFluor488 detection) or 570–650 nm (AlexaFluor 561 detection). Gated STED imaging of samples was achieved through use of 592-nm and 660-nm depletion lasers with a time gate set to 1.8–8 ns using the Leica STED 3X system. All images were acquired with 32-line averages and 22 × 22 nm pixel size. FRAP imaging Images were taken using an LSM 510 confocal microscope (Zeiss). The size of the bleaching area was optimized as shown previously [56]. Bleaching areas were selected within each bouton (about 2.5 μm2) and images acquired every 10 s. Data were fitted with a single exponential to reveal tau values of fluorescence recoveries. PLA The assay was performed as described [79]. Briefly, dissected third instar larvae were fixed in Bouin’s solution for 15 mins on ice, washed in PBT (PBS with 0.1% Triton) 3 times for 10 min each and blocked in PBT/1% BSA for 1 h. Larvae were incubated overnight at 4 °C in mouse and rabbit antibodies against the 2 proteins of interest, diluted in PBT/1% BSA. Primary antibodies used were anti-rabbit cpx (Littleton), anti-rabbit GFP (Abcam), anti-mouse Brp (Developmental Studies Hybridoma Bank [DSHB]), anti-mouse syntaxin (DSHB), and anti-mouse Synaptotagmin (DSHB). All antibodies were used at 1:200 dilution. The next day, PLA probe binding, ligation, and amplification steps were performed as described [79]. Before mounting, larvae were counterstained with AlexaFluor 488 goat anti-HRP (Jackson Immuno Research) at 1:500 dilution for 40 mins. PLA signals were only measured within the HRP signals. PLA signal and NMJ volumes of z-stack images were analyzed in Volocity 6.3. PLA signals were only measured within the HRP signals. All PLA signals were expressed relative to total NMJ volume (S10 Fig and S9 Data). HEK cell transfection A plasma membrane targeted eYFP CAAX protein was constructed by fusing the last 15 amino acids of Human K-Ras isoform b with the C-terminus of eYFP. A short linker sequence GTMASNNTASG was inserted between the last amino acid of eYFP and the membrane targeting CAAX sequence. The resulting construct was subcloned into expression vector pcDNA5 frt and verified by DNA sequencing. HEK293 FT cells were plated on poly-d-lysine coated glass coverslips in 6 well plates and transfected with 0.5 g eYFP CAAX per well using polyethylenimine (PEI) at a ratio of 1 g DNA to 6 g PEI. Prior to imaging, cells were treated for 12 h with the NO donor DETA-NONOate or a combination of the farnesyl transferase inhibitor L-744,832 (20 μM) and the geranylgeranyltransferase I inhibitor GGTI-298 (20 μM). Cells were then washed 3 times with PBS and fixed for 15 min with 4% paraformaldehyde. Coverslips were mounted on glass microscope slides with VectaShield H1500 and observed using a Zeiss laser scanning confocal microscope. Biotin switch assay Animals were kept in the dark 3 h before removing the retinas in order to decrease basal levels of protein nitrosylation. Retinas were kept in DMEM (Gibco 31053–028) with protease inhibitors (Complete) and treated with NO donors (GSNO and PAPA-NONOate, 20 μM) for 40 min at room temperature and protected from light. The biotin-switch assay was performed with the S-nitrosylated Protein Detection kit (Cayman Chemical, 10006518) in the dark. Bradford assay was performed and equal amount of proteins were incubated with Streptavidin beads (Sigma) overnight. Western blot was performed with cpx 3 antibody (Synaptic Systems), 1:1,000. Electron microscopy Third instar larvae were “filleted” in phosphate-buffered saline at room temperature and then fixed in 2% (wt/vol) glutaraldehyde in 0.1 M sodium cacodylate buffer (pH 7.4) at 4 °C overnight. They were postfixed with 1% (wt/vol) osmium tetroxide/1% (wt/vol) potassium ferrocyanide for 1 h at room temperature and then stained en bloc, overnight, with 5% (wt/vol) aqueous uranyl acetate at 4 °C, dehydrated and embedded in Taab epoxy resin (Taab Laboratories Equipment Ltd, Aldermaston, UK). Semi-thin sections, stained with toluidine blue, were used to identify areas containing synaptic regions (m6/7 in regions A2/A3). Ultra-thin sections were cut from these areas, counterstained with lead citrate, and examined in an FEI Talos transmission electron microscope (FEI Company [Thermo Fisher Scientific Inc.], Hillsboro, OR). Images were recorded using an FEI Ceta-16M CCD camera with 4k × 4k pixels. SV measurements were made using ImageJ software. A total of about 500–900 SVs were measured in 5–10 boutons from 3 animals per genotype. GCaMP imaging Wandering third instar larvae expressing presynaptic UAS-myrGCaMP5 or UAS-GCaMP5 using the pan-neuronal C155 or glutamatergic neuronal OK371 driver, respectively, were dissected in low Ca2+ HL-3 saline (0.2 mM CaCl2) at room temperature. The motor nerves were carefully snipped below the ventral nerve cord, and the CNS was removed. The preparation was washed several times with HL-3 containing 1.5 mM Ca2+. Nerve stimulation was performed with an isolated stimulator (DS2A, Digitimer) and images were recorded before, during (2–6 s in a train at 60 Hz) and after the stimulation period (8 s) in HL-3 containing 3 mM Ca2+ or during 15 s in a 20-s train at 20 Hz at indicated Ca2+ levels in the presence of 5 mM L-glutamic acid. We acquired images at a rate of 1 image per 4 s using a Zeiss laser-scanning confocal microscope (LSM 510 Meta; Zeiss) with a 63× 1.0 NA water immersion objective (Zeiss). Excitation was set at 488 nm (Argon laser) using a dichroic mirror 490 nm and a bandpass filter 500–550 nm. Low sampling rates were sufficient to investigate Ca2+ plateau levels during the 8-s stimulation periods [80]. A single confocal plane of muscle pair 6/7 NMJ in segments A2 or A3 was imaged to establish a baseline. Small z-drifts were manually corrected during the imaging session. Imaging sessions in which significant movement of the muscle occurred were discarded. Images were analyzed using Volocity 6.3 Image Analysis software (PerkinElmer). Single bouton fluorescence intensities were measured (average within a bouton) and bouton ΔF/F0 values were averaged for each NMJ. FlincG3 imaging NMJs of larvae expressing UAS-FlincG3 presynaptically were imaged as described above to measure GCaMP fluorescence. To prevent cGMP breakdown by PDE activity, preparations were incubated with 10μM Zap prior to imaging. Mitochondrial respiratory activity assay High resolution respirometry was performed with an Oroboros O2K oxygraph (Oroboros Instruments Ltd.). For each measurement, 3 third instar larvae were homogenized in 100 μL of respiration buffer MiR05 [81]. Leak state respiration was measured after adding 5 mM of pyruvate, 2 mM of malate, and 10 mM of glutamate. Oxphos capacity supported by Complex I was measured after addition of 1.25 mM ADP. After addition of 10 mM succinate, Oxphos capacity supported by both Complex I and Complex II were measured. Free Oxphos capacity was calculated as the difference Oxphos–Leak. Respiratory Ctrl ratios (RCRs) were calculated as the ratio Oxphos/Leak. cGMP radio immunoassay Larval brains (30 per condition) were isolated and assessed for cGMP production. Briefly, brain extracts were diluted 5-fold in 100 mM sodium acetate, pH 6.2, and acetylated by consecutive addition of triethylamine (10 μL) and acetic anhydride (5 μL) and used in the radioimmunoassay [82] within 60 min. Cyclic GMP standards (100 μL; 0–4 nM) were treated identically. Acetylated samples (100 μL) were mixed with 2′-O-succinyl 3-[125I]-iodotyrosine methyl ester cyclic GMP (GE Healthcare, IM107) (50 μL, about 3,000 d.p.m. made up in 50 mM sodium acetate, 0.2% BSA, pH 6.2), and 100 μL of anti-cyclic GMP antibody (GE Healthcare, TRK500; diluted in 50 mM sodium acetate, 0.2% BSA, pH 6.2). Samples were intermittently vortex mixed during a 4-h incubation at 4 °C. Free and bound cyclic GMP was separated by charcoal precipitation with 500 μL of a charcoal suspension (1% [w/v] activated charcoal in 100 mM potassium phosphate, 0.2% BSA, pH 6.2). After vortex mixing for 5 min, samples were centrifuged (13,000 × g, 4 min, 4 °C) and radioactivity determined in an aliquot of supernatant (600 μL). Unknown values were determined from the cyclic GMP standard curve using GraphPad Prism 7 (GraphPad Software Inc., San Diego, CA). Data points represent 2 measurements of 30 brains for each condition. Drug applications NO donor solutions were made freshly from stock solutions on the day and working solutions (200 μM sodium nitroprusside [SNP] and 5 μM PAPA-NONOate, each releasing about 200 nM NO) [29] were kept on ice for up to 6 h. All experiments to assess NO signaling were made between 40 and 60 min of NO exposure (NO: 200 μM SNP, 20 μM PAPA-NONOate; presented data comprise responses following incubation with either donor as they are not different from each other [Student t test, p > 0.05]; 500 μM SNP was used in Figs 7A, 7B, 8E and 8F [S7 and S8 Data]). Incubations with drugs: Zap (PDE inhibitor), ODQ (sGC inhibitor), L-744,832, and GGTI-298 (FTase inhibitors) incubation for 1 h; both block FTase with an IC50: 1.8 nM and IC50: 203 nM, respectively [83]. Drugs were purchased from Tocris or Sigma. Statistics Statistical analysis was performed with Prism 6.3 and 7 and InStat 3 (Graphpad Software Inc., San Diego, CA). Statistical tests were carried out using an ANOVA test when applicable with a posteriori test (1-way ANOVA with Tukey’s multiple comparisons test) or unpaired Student t test, as indicated. Data are expressed as mean ± SEM where n is the number of boutons, NMJs, or larvae as indicated and significance is shown as *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001. Fly stocks Flies were raised on standard maize media at 25 °C at a 12-h LD cycle. The elav-Gal4 [C155] driver was obtained from the Bloomington Stock Center (Indiana, US). The UAS-RNAi lines (GCLm [CG4919], GCLc [CG2259], and Fnta [CG2976]) were purchased form the Vienna Drosophila Resource Centre (VDRC). The use of the UAS-Gal4 bipartite expression system to drive pan-neuronal expression excludes potential postsynaptic effects. The elav-Gal4 driver (female flies) and the UAS responder lines (male flies) were crossed to obtain offspring expressing the genes of interest and w1118 were used as Ctrls. The fluorescent Ca2+ sensor GCaMP5 was tethered to the plasma membrane with an N-terminal myristoylation (myr) sequence as described previously [72]. The UAS-myrGCaMP5 and cpxSH1 null mutant lines were provided by Troy Littleton (MIT, Cambridge, MA) [73]. GCaMP5 was expressed in glutamatergic neurons (OK371-Gal4; UAS-GCaMP5). cpx expression levels are shown for w1118 and cpxSH1 larvae in S9B and S9C Fig. UAS-fdh31 (expression of fdh homologue of mammalian GSNOR/ADH-5) and UAS-fdhri34/25 (expression of fdh RNAi) mutant transgenic lines were kindly provided by Li Liu Institute of Biophysics, Chinese Academy of Sciences, Beijing, China) [51]. NOSΔ15/NOSC lines were provided by Patrick O’Farrell (UCSF, San Francisco, CA). NOSΔ15 deletion removes sequences encoding residues 1–757, encompassing the entire oxygenase domain and including regions that bind the catalytic heme and the substrate rendering the lines NOS “null” [33, 34]. The syn-null mutant transgenic line (Syn97) was generously provided by Erich Buchner (Universitätsklinikum Würzburg, Germany) [44]. UAS-EGFP-cpx and UAS-EGFP-cpx1257 transgenic lines were kindly provided by Fumiko Kawasaki (Penn State University, PA) [16]. UAS-GCLm and UAS-GCLc transgenic lines were provided by William C. Orr (Southern Methodist University Dallas, TX). Cloning cDNAs encoding for cpx 7A was a gift from Troy Littleton and used as a template for downstream PCRs. Cys 140 of cpx7A isoform was mutated to tryptophan to generate S-nitrosylation mimic mutant. PCR products, which include XhoI and XbaI restriction sites, were cloned into the pJFRC2 vector [74]—a gift from Gerald Rubin (Addgene plasmid no. 26214)—by standard methods. The resulting constructs were injected into attP40 Drosophila strains. The resulting transgenic lines (Dmcpx7AC140W and WT Dmcpx) were crossed into a cpxSH1 background [7] using standard balancing techniques. The FlincG3 ORF was amplified from pTriEx4-H6-FGAm (FlincG3) (Addgene plasmid no. 49202) and the resultant PCR product cloned into pUASTattB by the Protein Expression Laboratory (PROTEX), University of Leicester. Microinjection of the pUASTattB plasmid was performed by the University of Cambridge, Department of Genetics Fly Facility. Electrophysiology TEVC recordings were performed as described previously [75]. Sharp-electrode recordings were made from ventral longitudinal m6 in abdominal segments 2 and 3 of third instar larvae using pClamp 10, an Axoclamp 900A amplifier and Digidata 1440A (Molecular Devices, US) in hemolymph-like solution 3 (HL-3) [76]. Recording electrodes (20–50 MΩ) were filled with 3 M KCl. mEJCs were recorded in the presence of 0.5 μM tetrodotoxin (Tocris, UK). All synaptic responses were recorded from muscles with input resistances ≥4 MΩ, holding currents <4 nA at −60 mV and resting potentials more negative than −60 mV at 25 °C, as differences in recording temperature cause changes in glutamate receptor kinetics and amplitudes [77]. Holding potentials were −60 mV. The extracellular HL-3 contained (in mM): 70 NaCl, 5 KCl, 20 MgCl2, 10 NaHCO3, 115 sucrose, 5 trehalose, 5 HEPES, and 1.5 CaCl2 (0.5–3.0 mM in Fig 3 and S3 Data, as specified). Average single eEJC amplitudes (stimulus: 0.1 ms, 1–5 V) are based on the mean peak eEJC amplitude in response to 10 presynaptic stimuli (recorded at 0.2 Hz). Nerve stimulation was performed with an isolated stimulator (DS2A, Digitimer). Paired-pulse experiments were performed by applying 5 repetitive stimuli (0.2 Hz) at different intervals (20, 40, 100, 200 ms) for each cell at each ISI. All data were digitized at 10 kHz and for miniature recordings, 200-s recordings, we analyzed to obtain mean mEJC amplitudes, decay, and frequency (f) values. QC was estimated for each recording by calculating the ratio of eEJC amplitude/average mEJC amplitude, followed by averaging recordings across all NMJs for a given genotype. mEJC and eEJC recordings were off-line low-pass filtered at 500 Hz and 1 kHz, respectively. Materials were purchased from Sigma-Aldrich (UK) unless otherwise stated. Variance-mean analysis of eEJCs Approximately 40 eEJCs were elicited at different [Ca2+]e, ranging from 0.5 to 3 mM to give mean eEJC amplitudes (I). The mean eEJC is given by I = Npvrq [45], with N being the number of independent release-ready vesicles, pvr the vesicular release probability, and q the quantal size at each given [Ca2+]e. The eEJC variance was calculated as previously described [45]. The plots of the variance-mean were obtained for each cell and fitted with the parabolic function Var(I) = I2/N + qI that was a constraint to pass through the origin. Upon fitting the parabola, pvr and q were calculated using the equations: q = A/(1+CV2) and pvr = I(B/A)(1+CV2) where CV2 is the coefficient of variation of the eEJC amplitudes at a given [Ca2+]e concentration calculated as CV2 = (eEJCs standard deviation/mean amplitude)2; A and B were obtained from the fitting parameters. Estimated values were not corrected for variability in mEJC amplitude distributions or latency fluctuations. Ca2+ cooperativity was assessed by plotting eEJC amplitudes over [Ca2+]e and fitted with the Hill equation (mean eEJC amplitude plotted versus different [Ca2+]e: eEJC([Ca2+]) = eEJCmax[1+(EC50/[Ca2+])slope]−1), yielding the Hill slope as a measure of Ca2+ cooperativity. Cumulative postsynaptic current analysis The apparent size of the RRP was probed by the method of cumulative eEJC amplitudes [78]. Muscles were clamped to −60 mV and eEJC amplitudes during a stimulus train (50 Hz, 500 ms [of a 1-s train]) were calculated as the difference between peak and baseline before stimulus onset of a given eEJC. Receptor desensitization was not blocked as it did not affect eEJC amplitudes, because a comparison of the decay of the first and the last eEJC within a train did not reveal any significant difference in decay kinetics. The number of release-ready vesicles (N) was obtained by back extrapolating a line fit to the linear phase of the 500-ms cumulative eEJC plot (the last 200 ms of the train) to time zero. N was then obtained by dividing the cumulative eEJC amplitude at time zero by the mean mEJC amplitude recorded in the same cell. To calculate the QC in the train, we used mean mEJC amplitudes measured before the train. Immunohistochemistry Third instar larvae were dissected in ice-cold PBS then fixed in 4% paraformaldehyde. After permeabilization with PBS-0.1% Triton (PBS-T) and blocking with PBS-T containing 0.2% bovine serum albumin (BSA) and 2% normal goat serum, larval fillets were incubated at 4 °C overnight in solutions of primary antibody. The following antibody dilutions were used: NC82 (supernatant) anti-Brp (Bruchpilot) 1:200, cpx (1:500), syntaxin (1:200), synaptotagmin (1:200), and GFP (1:200). After 3 × 10 min washes in PBS-T, larvae were incubated with AlexaFluor 488 goat anti-HRP (Jackson Immuno Research) and AlexaFluor 546 goat anti-mouse 1:500 dilution for 90 min at room temperature. Larvae were mounted using Vectashield mounting medium (Vector Labs) and NMJ 6/7 (segments A2 and A3) images were acquired with a Zeiss laser-scanning confocal microscope (LSM 510, Zeiss). Image analysis was performed with ZEN (Zeiss) and Volocity 6.3 software. STED microscopy Images were acquired on a Leica TCS SP8 system attached to a Leica DMi8 inverted microscope (Leica Microsystems). Excitation light (488 nm for AlexaFluor488 or 561 nm for AlexaFluor568) was provided by a white light laser with a repetition rate of 80 MHz. Images were acquired using a 100× 1.4 NA oil immersion objective and fluorescence was detected through a bandpass of 495–550 nm (AlexaFluor488 detection) or 570–650 nm (AlexaFluor 561 detection). Gated STED imaging of samples was achieved through use of 592-nm and 660-nm depletion lasers with a time gate set to 1.8–8 ns using the Leica STED 3X system. All images were acquired with 32-line averages and 22 × 22 nm pixel size. FRAP imaging Images were taken using an LSM 510 confocal microscope (Zeiss). The size of the bleaching area was optimized as shown previously [56]. Bleaching areas were selected within each bouton (about 2.5 μm2) and images acquired every 10 s. Data were fitted with a single exponential to reveal tau values of fluorescence recoveries. PLA The assay was performed as described [79]. Briefly, dissected third instar larvae were fixed in Bouin’s solution for 15 mins on ice, washed in PBT (PBS with 0.1% Triton) 3 times for 10 min each and blocked in PBT/1% BSA for 1 h. Larvae were incubated overnight at 4 °C in mouse and rabbit antibodies against the 2 proteins of interest, diluted in PBT/1% BSA. Primary antibodies used were anti-rabbit cpx (Littleton), anti-rabbit GFP (Abcam), anti-mouse Brp (Developmental Studies Hybridoma Bank [DSHB]), anti-mouse syntaxin (DSHB), and anti-mouse Synaptotagmin (DSHB). All antibodies were used at 1:200 dilution. The next day, PLA probe binding, ligation, and amplification steps were performed as described [79]. Before mounting, larvae were counterstained with AlexaFluor 488 goat anti-HRP (Jackson Immuno Research) at 1:500 dilution for 40 mins. PLA signals were only measured within the HRP signals. PLA signal and NMJ volumes of z-stack images were analyzed in Volocity 6.3. PLA signals were only measured within the HRP signals. All PLA signals were expressed relative to total NMJ volume (S10 Fig and S9 Data). HEK cell transfection A plasma membrane targeted eYFP CAAX protein was constructed by fusing the last 15 amino acids of Human K-Ras isoform b with the C-terminus of eYFP. A short linker sequence GTMASNNTASG was inserted between the last amino acid of eYFP and the membrane targeting CAAX sequence. The resulting construct was subcloned into expression vector pcDNA5 frt and verified by DNA sequencing. HEK293 FT cells were plated on poly-d-lysine coated glass coverslips in 6 well plates and transfected with 0.5 g eYFP CAAX per well using polyethylenimine (PEI) at a ratio of 1 g DNA to 6 g PEI. Prior to imaging, cells were treated for 12 h with the NO donor DETA-NONOate or a combination of the farnesyl transferase inhibitor L-744,832 (20 μM) and the geranylgeranyltransferase I inhibitor GGTI-298 (20 μM). Cells were then washed 3 times with PBS and fixed for 15 min with 4% paraformaldehyde. Coverslips were mounted on glass microscope slides with VectaShield H1500 and observed using a Zeiss laser scanning confocal microscope. Biotin switch assay Animals were kept in the dark 3 h before removing the retinas in order to decrease basal levels of protein nitrosylation. Retinas were kept in DMEM (Gibco 31053–028) with protease inhibitors (Complete) and treated with NO donors (GSNO and PAPA-NONOate, 20 μM) for 40 min at room temperature and protected from light. The biotin-switch assay was performed with the S-nitrosylated Protein Detection kit (Cayman Chemical, 10006518) in the dark. Bradford assay was performed and equal amount of proteins were incubated with Streptavidin beads (Sigma) overnight. Western blot was performed with cpx 3 antibody (Synaptic Systems), 1:1,000. Electron microscopy Third instar larvae were “filleted” in phosphate-buffered saline at room temperature and then fixed in 2% (wt/vol) glutaraldehyde in 0.1 M sodium cacodylate buffer (pH 7.4) at 4 °C overnight. They were postfixed with 1% (wt/vol) osmium tetroxide/1% (wt/vol) potassium ferrocyanide for 1 h at room temperature and then stained en bloc, overnight, with 5% (wt/vol) aqueous uranyl acetate at 4 °C, dehydrated and embedded in Taab epoxy resin (Taab Laboratories Equipment Ltd, Aldermaston, UK). Semi-thin sections, stained with toluidine blue, were used to identify areas containing synaptic regions (m6/7 in regions A2/A3). Ultra-thin sections were cut from these areas, counterstained with lead citrate, and examined in an FEI Talos transmission electron microscope (FEI Company [Thermo Fisher Scientific Inc.], Hillsboro, OR). Images were recorded using an FEI Ceta-16M CCD camera with 4k × 4k pixels. SV measurements were made using ImageJ software. A total of about 500–900 SVs were measured in 5–10 boutons from 3 animals per genotype. GCaMP imaging Wandering third instar larvae expressing presynaptic UAS-myrGCaMP5 or UAS-GCaMP5 using the pan-neuronal C155 or glutamatergic neuronal OK371 driver, respectively, were dissected in low Ca2+ HL-3 saline (0.2 mM CaCl2) at room temperature. The motor nerves were carefully snipped below the ventral nerve cord, and the CNS was removed. The preparation was washed several times with HL-3 containing 1.5 mM Ca2+. Nerve stimulation was performed with an isolated stimulator (DS2A, Digitimer) and images were recorded before, during (2–6 s in a train at 60 Hz) and after the stimulation period (8 s) in HL-3 containing 3 mM Ca2+ or during 15 s in a 20-s train at 20 Hz at indicated Ca2+ levels in the presence of 5 mM L-glutamic acid. We acquired images at a rate of 1 image per 4 s using a Zeiss laser-scanning confocal microscope (LSM 510 Meta; Zeiss) with a 63× 1.0 NA water immersion objective (Zeiss). Excitation was set at 488 nm (Argon laser) using a dichroic mirror 490 nm and a bandpass filter 500–550 nm. Low sampling rates were sufficient to investigate Ca2+ plateau levels during the 8-s stimulation periods [80]. A single confocal plane of muscle pair 6/7 NMJ in segments A2 or A3 was imaged to establish a baseline. Small z-drifts were manually corrected during the imaging session. Imaging sessions in which significant movement of the muscle occurred were discarded. Images were analyzed using Volocity 6.3 Image Analysis software (PerkinElmer). Single bouton fluorescence intensities were measured (average within a bouton) and bouton ΔF/F0 values were averaged for each NMJ. FlincG3 imaging NMJs of larvae expressing UAS-FlincG3 presynaptically were imaged as described above to measure GCaMP fluorescence. To prevent cGMP breakdown by PDE activity, preparations were incubated with 10μM Zap prior to imaging. Mitochondrial respiratory activity assay High resolution respirometry was performed with an Oroboros O2K oxygraph (Oroboros Instruments Ltd.). For each measurement, 3 third instar larvae were homogenized in 100 μL of respiration buffer MiR05 [81]. Leak state respiration was measured after adding 5 mM of pyruvate, 2 mM of malate, and 10 mM of glutamate. Oxphos capacity supported by Complex I was measured after addition of 1.25 mM ADP. After addition of 10 mM succinate, Oxphos capacity supported by both Complex I and Complex II were measured. Free Oxphos capacity was calculated as the difference Oxphos–Leak. Respiratory Ctrl ratios (RCRs) were calculated as the ratio Oxphos/Leak. cGMP radio immunoassay Larval brains (30 per condition) were isolated and assessed for cGMP production. Briefly, brain extracts were diluted 5-fold in 100 mM sodium acetate, pH 6.2, and acetylated by consecutive addition of triethylamine (10 μL) and acetic anhydride (5 μL) and used in the radioimmunoassay [82] within 60 min. Cyclic GMP standards (100 μL; 0–4 nM) were treated identically. Acetylated samples (100 μL) were mixed with 2′-O-succinyl 3-[125I]-iodotyrosine methyl ester cyclic GMP (GE Healthcare, IM107) (50 μL, about 3,000 d.p.m. made up in 50 mM sodium acetate, 0.2% BSA, pH 6.2), and 100 μL of anti-cyclic GMP antibody (GE Healthcare, TRK500; diluted in 50 mM sodium acetate, 0.2% BSA, pH 6.2). Samples were intermittently vortex mixed during a 4-h incubation at 4 °C. Free and bound cyclic GMP was separated by charcoal precipitation with 500 μL of a charcoal suspension (1% [w/v] activated charcoal in 100 mM potassium phosphate, 0.2% BSA, pH 6.2). After vortex mixing for 5 min, samples were centrifuged (13,000 × g, 4 min, 4 °C) and radioactivity determined in an aliquot of supernatant (600 μL). Unknown values were determined from the cyclic GMP standard curve using GraphPad Prism 7 (GraphPad Software Inc., San Diego, CA). Data points represent 2 measurements of 30 brains for each condition. Drug applications NO donor solutions were made freshly from stock solutions on the day and working solutions (200 μM sodium nitroprusside [SNP] and 5 μM PAPA-NONOate, each releasing about 200 nM NO) [29] were kept on ice for up to 6 h. All experiments to assess NO signaling were made between 40 and 60 min of NO exposure (NO: 200 μM SNP, 20 μM PAPA-NONOate; presented data comprise responses following incubation with either donor as they are not different from each other [Student t test, p > 0.05]; 500 μM SNP was used in Figs 7A, 7B, 8E and 8F [S7 and S8 Data]). Incubations with drugs: Zap (PDE inhibitor), ODQ (sGC inhibitor), L-744,832, and GGTI-298 (FTase inhibitors) incubation for 1 h; both block FTase with an IC50: 1.8 nM and IC50: 203 nM, respectively [83]. Drugs were purchased from Tocris or Sigma. Statistics Statistical analysis was performed with Prism 6.3 and 7 and InStat 3 (Graphpad Software Inc., San Diego, CA). Statistical tests were carried out using an ANOVA test when applicable with a posteriori test (1-way ANOVA with Tukey’s multiple comparisons test) or unpaired Student t test, as indicated. Data are expressed as mean ± SEM where n is the number of boutons, NMJs, or larvae as indicated and significance is shown as *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001. Supporting information S1 Fig. Activities of mitochondrial complex I and II are not affected by NO signaling. Larval preparations were incubated either in Ctrl HL-3 or HL-3 + NO donors for 50 min and mitochondrial function was assessed by high resolution respirometry using an Oroboros Oxygraph-2K. (A) Oxygen fluxes (pmol/(s*ml) of Ctrl [red] and NO-treated [green] larvae; arrows below indicate additions of the following substrates: pyruvate, malate, glutamate, and succinate. (B) Summary of O2 flux measurements, left (Complex I: Ctrl: 23 ± 5, NO: 23 ± 3, Complex I+II: Ctrl: 32 ± 7, NO: 32 ± 3). Right, respiratory control ratios: oxidative phosphorylation/leak (Complex I: Ctrl: 10 ± 1, NO: 9 ± 2, Complex I+II: Ctrl: 14 ± 1 NO: 13 ± 2). The raw data for this figure can be found in S9 Data. Data denote mean ± SEM, p > 0.05, ANOVA with post hoc Tukey-Kramer was used for comparisons: Ctrl versus NO, n = 6 larvae each. ADP, adenosine diphosphate; Ctrl, control; D, ADP; G, glutamate; M, malate; NO, nitric oxide; PYR, pyruvate; S, succinate. https://doi.org/10.1371/journal.pbio.2003611.s001 (TIF) S2 Fig. NOS null background does not cause developmental effects on NMJ size and ultrastructure (related to Fig 2). (A) NMJ volume was calculated from z-stack confocal images (HRP) in both genotypes (NMJ volume: WT: 3,183 ± 287 μm3 [n = 46 NMJs], NOSΔ15: 3,178 ± 284 μm3 [n = 11 NMJs], NOSC: 3,282 ± 494 μm3 [n = 5 NMJs], p > 0.05, ANOVA; Brp puncta/NMJ volume: WT: 0.07 ± 0.01 [n = 6 NMJs], NOSΔ15: 0.11 ± 0.01 [n = 7 NMJs], NOSC: 0.11 ± 0.01 [n = 5 NMJs], p > 0.05, ANOVA with post hoc Tukey-Kramer was used for comparisons). (C) Representative electron microscopy images of 1b boutons from each genotype, with red semicircles indicating the area for vesicle counts. (D) Mean values for the number of synapses (AZ), number of T-bars and number of vesicles within a 250-nm semicircle radius from the center of the AZ for each genotype (number of synapses: WT: 3.8 ± 0.3, NOSC: 4.2 ± 0.5; number of T-bars: WT: 1.2 ± 0.1, NOSC: 1.7 ± 0.3, number of vesicles: WT: 23 ± 1, NOSC: 25 ± 1 [n = 24 and n = 16 boutons per genotype], p > 0.05 for all comparisons, Student t test). The raw data for this figure can be found in S9 Data. Data denote mean ± SEM. AZ, active zone; Brp, Bruchpilot; HRP, horseradish peroxidase; NMJ, neuromuscular junction; WT, wild-type. https://doi.org/10.1371/journal.pbio.2003611.s002 (TIF) S3 Fig. Recovery from depletion is not affected by NO signaling. (A) Recordings showing a 1-s 50-Hz train of eEJCs with subsequent single eEJCs at various time points. Note the broken trace to accommodate for the long intervals. (B) Mean eEJC amplitudes at increasing intervals after high-frequency stimulation (50Hz, 1s) for Ctrl (black, n = 16 NMJs), NO (red, n = 15 NMJs) and NO+ODQ (orange, n = 7 NMJs) with single exponential fits to the data points. (C) Mean time constant for recovery tau values for the conditions indicated (9.6 ± 1.2 s [Ctrl], 10.4 ± 1.1 s [NO], and 9.4 ± 1.9 s [NO + ODQ]). The raw data for this figure can be found in S9 Data. Data denote mean ± SEM, p > 0.05; ANOVA with post hoc Tukey-Kramer was used for comparisons. Ctrl, control; eEJC, evoked EJC; EJC, excitatory junction current; NMJ, neuromuscular junction; NO, nitric oxide; ODQ, 1H-[1,2,4]oxadiazolo[4,3-a]quinoxalin-1-one. https://doi.org/10.1371/journal.pbio.2003611.s003 (TIF) S4 Fig. NO supresses evoked release and vesicle pools in syn knock-out NMJs. (A) Representative 50Hz trains of synaptic stimuli in Ctrl syn null mutant (Syn97 Ctrl) and NO-treated syn null mutant (Syn97 + NO) NMJs. (B) Cumulative QC for both conditions, showing the reduced available pool size following NO incubation. (C) Mean QC and vesicle pool sizes for conditions indicated (QC: 157 ± 20, NO: 84 ± 16; pool size: Ctrl: 288 ± 39, NO: 134 ± 23). The raw data for this figure can be found in S9 Data. Data denote mean ± SEM, *p < 0.05 versus w1118 Ctrl, #p < 0.05 versus its Ctrl, &&p < 0.01 versus its Ctrl, ANOVA for each comparing QC and vesicle size, n = 5 NMJs each. Ctrl, control; NMJ, neuromuscular junction; NO, nitric oxide; QC, quantal content; syn, synapsin. https://doi.org/10.1371/journal.pbio.2003611.s004 (TIF) S5 Fig. Activity-induced intracellular Ca2+ levels are not affected by NO (related to Fig 3). (A) Confocal images of GCaMP5 expressing NMJs. GCaMP5 was expressed in motor neurons and fluorescence was imaged during a train of synaptic stimulation at 20 Hz. Experiments were performed at various extracellular Ca2+ concentrations (0.25–1.5 mM, as indicated) in Ctrl (representative images) and NO-treated NMJs. (B) Summary of ΔF/F0 for conditions indicated. Mean ΔF/F0: 0.25 Ca2+: Ctrl: 0.24 ± 0.04, NO: 0.14 ± 0.03, 0.5 Ca2+: Ctrl: 0.42 ± 0.09, NO: 0.5 ± 0.07, 1.5 Ca2+: Ctrl: 1.1 ± 0.1, NO: 1.2 ± 0.2. NO does not affect the presynaptic Ca2+ levels at any concentration tested. p > 0.05, ANOVA with post hoc Tukey-Kramer was used for comparisons. n = 7–11 NMJs per condition. The raw data for this figure can be found in S9 Data. Ctrl, control; NMJ, neuromuscular junction; NO, nitric oxide. https://doi.org/10.1371/journal.pbio.2003611.s005 (TIF) S6 Fig. Reduction of endogenous de-nitrosylation capacity partially supresses synaptic release (related to Fig 4). (A) Mean eEJC amplitudes (GSNOR RNAi: Ctrl: 95 ± 10 nA, NO: 85 ± 7 nA, GCLm RNAi: Ctrl: 118 ± 11 nA, NO: 78 ± 4 nA). (B) QC (GSNOR RNAi: Ctrl: 170 ± 8, NO: 139 ± 10, GCLm RNAi: Ctrl: 129 ± 17, NO: 97 ± 12) and (C) vesicle pool sizes estimated by back extrapolation from cumulative QCs (GSNOR RNAi: Ctrl: 223 ± 38, NO: 214 ± 48, GCLm RNAi: Ctrl: 228 ± 38, NO: 98 ± 24), all for genotypes indicated. (E) Western blot analysis of GSNOR (fdh31, RNAi-fdh25) expression in genotypes indicated. The raw data for this figure can be found in S9 Data. Data denote mean ± SEM, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 versus w1118 Ctrl, #p < 0.05 versus its Ctrl, ANOVA with post hoc Tukey-Kramer was used for comparisons, n = 10–8 NMJs. Ctrl, control; eEJC, evoked EJC; EJC, excitatory junction current; fdh, formaldehyde dehydrogenase; GCLm, glutamate-cysteine ligase modifier subunit M; GSNOR, S-nitrosoglutathione reductase; NMJ, neuromuscular junction; NO, nitric oxide; QC, quantal content; RNAi, RNA interference. https://doi.org/10.1371/journal.pbio.2003611.s006 (TIF) S7 Fig. Nitrergic activity alters cpx localization relative to SNARE and Brp proteins (related to Fig 7). (A) Representative STED confocal images of boutons from Ctrl larvae and those exposed to NO donor. (B) Co-localization analysis reveals Pearson’s coefficients for indicated conditions (Cpx-Brp: Ctrl: 0.24 ± 0.04, NO: 0.37 ± 0.02, farnesyl inh: 0.47 ± 0.02, Cpx-Syx: Ctrl: 0.12 ± 0.04, NO: 0.26 ± 0.03, farnesyl inh: 0.30 ± 0.04, Cpx-Synap: Ctrl: 0.42 ± 0.04, NO: 0.18 ± 0.03, farnesyl inh: 0.18 ± 0.04), n = 8–30 boutons; data denote mean ± SEM, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. ANOVA with post hoc Tukey-Kramer was used for comparisons. The raw data for this figure can be found in S9 Data. Brp, Bruchpilot; cpx, complexin; Ctrl, conrol; NO, nitric oxide; SNARE, soluble N-ethyl-maleimide-sensitive fusion protein Attachment Protein Receptor; STED, stimulated emission depletion; Synap, synaptotagmin; Syx, syntaxin. https://doi.org/10.1371/journal.pbio.2003611.s007 (TIF) S8 Fig. FRAP analysis reveals differences in recovery depending on the bleaching area (related to Fig 7). (A) Representative images of boutons expressing WT cpx-GFP before bleaching and at different time points after photo bleaching. Top row shows recordings with a bleaching area roughly the size of half a bouton (>10 μm2); bottom row shows images using a bleaching area of 2.5 μm2. (B) Analysis of recovery from bleach shows faster time constants using the smaller bleaching area compared to half-bouton bleach. Using the smaller bleaching areas, there is a pronounced difference between WT and mutant cpx. Synaptobrevin was used as a control for photo bleaching associated with vesicular movement. Note, images and analysis for the 2.5 μm2 bleaching areas are the same as in Fig 7C; scale bar: 2 μm. The raw data for this figure can be found in S9 Data. cpx, complexin; FRAP, fluorescence recovery after photobleaching; GFP, green fluorescent protein; WT, wild-type. https://doi.org/10.1371/journal.pbio.2003611.s008 (TIF) S9 Fig. Confirmation of cpx expression. (A) Western blot analysis shows expression levels of cpx, GFP, and β-actin in w1118, WT cpx-GFP and cpxΔX-GFP lines. (B) Representative images of Brp-Cpx PLA in 2 example cpxSH1 NMJs show no PLA signal. (C) Confocal single plane images of a w1118 NMJ stained for cpx and Brp and (D) confocal single plane image of a cpxSH1 larva that does not express cpx. Brp, Bruchpilot; cpx, complexin; GFP, green fluorescent protein; NMJ, neuromuscular junction; PLA, proximity ligation assay; WT, wild-type. https://doi.org/10.1371/journal.pbio.2003611.s009 (TIF) S10 Fig. NMJ volumes are not different across genotypes. NMJs for different genotypes and conditions have similar volumes compared to Ctrls (WT Cpx-GFP [Cpx2A]: 2,857 ± 261 μm3, CpxΔX-GFP [Cpx1257]: 3,395 ± 787 μm3, NO treatment: 3,400 ± 228 μm3, farnesyl inh: 3,973 ± 306 μm3, Cpx WT: 3,617 ± 875 μm3, CpxC140W: 3,508 ± 707 μm3). Data denote mean ± SEM; ANOVA with post hoc Tukey-Kramer was used for comparisons versus w1118 WT, p > 0.05. The raw data for this figure can be found in S9 Data. Cpx, complexin; Ctrl, control; GFP, green fluorescent protein; NMJ, neuromuscular junction; NO, nitric oxide; WT, wild-type. https://doi.org/10.1371/journal.pbio.2003611.s010 (TIF) S1 Data. Raw values used to generate graphs in Fig 1. The raw data presented in worksheets 1–5 serve as underlying data for Fig 1A–1F and 1G. https://doi.org/10.1371/journal.pbio.2003611.s011 (XLSX) S2 Data. Raw values used to generate graphs in Fig 2. The raw data presented in worksheets 1–4 serve as underlying data for Fig 2B–2D and 2F–2H. https://doi.org/10.1371/journal.pbio.2003611.s012 (XLSX) S3 Data. Raw values used to generate graphs in Fig 3. The raw data presented in worksheets 1–5 serve as underlying data for Fig 3C–3E, 3G and 3I. https://doi.org/10.1371/journal.pbio.2003611.s013 (XLSX) S4 Data. Raw values used to generate graphs in Fig 4. The raw data presented in worksheets 1–4 serve as underlying data for Fig 4A–4F. https://doi.org/10.1371/journal.pbio.2003611.s014 (XLSX) S5 Data. Raw values used to generate graphs in Fig 5. The raw data presented in worksheets 1–2 serve as underlying data for Fig 5B–5D and 5F. https://doi.org/10.1371/journal.pbio.2003611.s015 (XLSX) S6 Data. Raw values used to generate graphs in Fig 6. The raw data presented in worksheets 1–3 serve as underlying data for Fig 6B, 6D and 6F. https://doi.org/10.1371/journal.pbio.2003611.s016 (XLSX) S7 Data. Raw values used to generate graphs in Fig 7. The raw data presented in worksheets 1–5 serve as underlying data for Fig 7B–7F. https://doi.org/10.1371/journal.pbio.2003611.s017 (XLSX) S8 Data. Raw values used to generate graphs in Fig 8. The raw data presented in worksheets 1–5 serve as underlying data for Fig 8A, 8B, 8D, 8F and 8G. https://doi.org/10.1371/journal.pbio.2003611.s018 (XLSX) S9 Data. Raw values used to generate graphs in S1–S8 and S10 Figs. The raw data presented in worksheets 1–9 serve as underlying data for S1–S8 and S10 Figs. https://doi.org/10.1371/journal.pbio.2003611.s019 (XLSX) Acknowledgments Some fly stocks and antibody were kindly provided by T. Littleton and other fly lines by W. Orr, L. Liu, P. O’Farrell, R. Baines, and E. Buchner. We also thank M. Jepson and D. Alibhai (Wolfson Bioimaging Facility, University of Bristol, UK) for conducting the STED microscopy experiments. Some fly lines and cpx–GFP constructs were kindly provided by F. Kawasaki (Penn State University, PA). The University of Iowa Developmental Studies Hybridoma Bank (DSHB) provided essential reagents. Many thanks to Dr. D. Read (MRC) for support in confocal microscopy and to the Department of Genetics, Fly Facility University of Cambridge, UK.