Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Predicting drug-target interactions using restricted Boltzmann machines

Predicting drug-target interactions using restricted Boltzmann machines Vol. 29 ISMB/ECCB 2013, pages i126–i134 BIOINFORMATICS doi:10.1093/bioinformatics/btt234 Predicting drug-target interactions using restricted Boltzmann machines 1 2, Yuhao Wang and Jianyang Zeng 1 2 Department of Automation and Institute for Interdisciplinary Information Sciences, Tsinghua University, Beijing 100084, China ABSTRACT urged drug developers to seek new uses for existing or aban- doned drugs (Booth and Zemmel, 2004). Such a new strategy Motivation: In silico prediction of drug-target interactions plays an is also called drug repositioning or drug repurposing. A strong important role toward identifying and developing new uses of existing support for the possibility of drug repositioning is the increas- or abandoned drugs. Network-based approaches have recently ingly accepted concept of ‘polypharmacology’, i.e. individual become a popular tool for discovering new drug-target interactions drugs can interact with multiple targets rather than a single (DTIs). Unfortunately, most of these network-based approaches can target (MacDonald et al., 2006; Xie et al., 2012). For example, only predict binary interactions between drugs and targets, and infor- serotonin and serotonergic drugs can interact with both 5-HT G mation about different types of interactions has not been well ex- protein-coupled receptors and 59HT ion channel proteins, 3A ploited for DTI prediction in previous studies. On the other hand, even though these two target proteins are not related in sequence incorporating additional information about drug-target relationships or structure (Cheng et al., 2012; Keiser et al., 2007; Kroeze et al., or drug modes of action can improve prediction of DTIs. 2002; Roth et al., 2004). This polypharmacological property of a Furthermore, the predicted types of DTIs can broaden our understand- drug enables us to identify more than one target that it can act on ing about the molecular basis of drug action. and hence develop its new uses. Results: We propose a first machine learning approach to integrate In silico prediction of interactions between drugs and target multiple types of DTIs and predict unknown drug-target relationships proteins provides an important tool for drug repositioning, as it or drug modes of action. We cast the new DTI prediction problem into can significantly reduce wet-laboratory work and lower the cost a two-layer graphical model, called restricted Boltzmann machine,and of the experimental determination of new drug-target inter- apply a practical learning algorithm to train our model and make pre- actions (DTIs). Various methods have been proposed for in dictions. Tests on two public databases show that our restricted silico DTI prediction. When 3D structures are available, molecu- Boltzmann machine model can effectively capture the latent features lar docking is commonly used to virtually screen a large number of a DTI network and achieve excellent performance on predicting of compounds against a target protein (Cheng et al., 2007; different types of DTIs, with the area under precision-recall curve up Donald, 2011; Morris et al., 2009). When 3D structures of mol- to 89.6. In addition, we demonstrate that integrating multiple types of ecules are absent, a number of different approaches have been DTIs can significantly outperform other predictions either by simply developed to address the in silico DTI prediction problem. Most mixing multiple types of interactions without distinction or using only of these structure-free methods can be grouped into two classes, a single interaction type. Further tests show that our approach can namely, ligand-based and network-based approaches. A represen- infer a high fraction of novel DTIs that has been validated by known tative ligand-based approach is the similarity ensemble approach experiments in the literature or other databases. These results indicate (Keiser et al., 2007, 2009), which predicts new DTIs using 2D that our approach can have highly practical relevance to DTI predic- structure similarity of ligands. Although ligand-based tion and drug repositioning, and hence advance the drug discovery approaches are able to discover a number of DTIs that have process. been validated experimentally, they have difficulty in identifying Availability: Software and datasets are available on request. drugs with novel scaffolds that differ from those of reference Contact: [email protected] compounds (Yabuuchi et al., 2011). Supplementary information: Supplementary data are available at Numerous network-based approaches have been proposed to Bioinformatics online. exploit latent features of DTI profiles and have recently become a popular tool for DTI prediction and drug repositioning (Bleakley and Yamanishi, 2009; Chen et al., 2012; Cheng et al., 1INTRODUCTION 2012; Mei et al., 2012; van Laarhoven et al., 2011; Xia et al., Drug development currently remains an expensive and time-con- 2010; Xie et al., 2012; Yamanishi et al., 2008). Although these suming process with extremely low success rate: it typically takes network-based approaches have achieved promising results, 10–15 years and $800 million–1 billion to bring a new drug to most of them can only predict binary DTIs, that is, they can market (Dimasi, 2001). In recent decades, the rate of the number only determine whether a drug interacts with a target protein, of new drugs approved by the US Food and Drug but cannot tell how they interact with each other. However, in- Administration versus the amount of money invested in pharma- dividual DTIs generally have different meanings. For example, ceutical research and development has significantly declined drug-target pairs can be described by different relationships,such (Booth and Zemmel, 2004). This productivity problem has as direct and indirect interactions (Gu¨ nther et al., 2008). A direct interaction is usually caused by protein-ligand binding, whereas *To whom correspondence should be addressed. an indirect interaction can be induced by the changed expression The Author 2013. Published by Oxford University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/ by-nc/3.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact [email protected] Predicting DTIs using RBMs level of a protein or active metabolites produced by a drug. In 1.1 Related work addition, DTIs can be annotated with different drug modes of A number of network-based approaches have been proposed for action, e.g. activation and inhibition (Kuhn et al., 2012). Here, predicting unknown interactions between drugs and targets. In ‘drug modes of action’ may be slightly different from the term Yamanishi et al. (2008), a supervised learning framework was used in the literature, e.g. (Iorio et al., 2010). In this article, we developed based on a bipartite graph, which integrates both mainly use this term to represent the following three specific chemical and genomic spaces by mapping them into a unified types of interactions: binding, activation and inhibition. space. Cheng et al. (2012) proposed a network-based inference Hereinafter, we will use ‘types of DTIs’ to represent both drug- approach to predict new DTIs by exploiting the topology simi- target relationships and drug modes of action. A network in larity of the underlying interaction network. In Zhao and Li which links are associated with different meanings is called the (2010), drug phenotypic, chemical indexes and protein–protein multidimensional network (Lu¨ and Zhou, 2011). In our context, interactions in genomic space were integrated into a computa- we call a DTI network where links are annotated with different tional framework for DTI prediction. Chen et al. (2012) pro- types of interactions the multidimensional DTI network.On the posed a random walk approach for DTI prediction based on a one hand, types of DTIs provide additional useful information, heterogeneous network, which integrates drug similarity, target which can be incorporated into a multidimensional network to similarity and DTI similarity. Bleakley and Yamanishi (2009) improve DTI prediction. On the other hand, the predicted types presented a new approach, called Bipartite Local Model of DTIs can extend our understanding about the molecular basis (BLM), to predict unknown DTIs by combining independent of drug action. Despite these positive aspects, current rich infor- drug-based and target-based prediction results using a supervised mation about types of DTIs (Gu¨ nther et al., 2008; Kuhn et al., learning method. Mei et al. (2012) further extended this BLM 2012) has not been well exploited for DTI prediction, and how to approach to incorporate the capacity of learning from neighbors incorporate such information into a multidimensional network to and predict the interactions for new drug or target candidates. In predict different types of DTIs still remains an open question. Xia et al. (2010), a manifold regularization semi-supervised learn- In this article, we propose an effective machine-learning ap- ing method was proposed to integrate heterogenous biological proach to accurately predict different types of DTIs on a multi- data sources for DTI prediction. A regularized least square al- dimensional network. Unlike previous network-based gorithm was proposed in van Laarhoven et al. (2011) for DTI approaches (Bleakley and Yamanishi, 2009; Chen et al., 2012; prediction using a product of kernels derived from DTI profiles. Cheng et al., 2012; Mei et al., 2012; van Laarhoven et al., 2011; In Gottlieb et al. (2011), He et al. (2010) and Perlman et al. Xia et al., 2010; Yamanishi et al., 2008), which only predict (2011), DTI prediction was formulated into a classification prob- binary DTIs, our approach not only identifies new DTIs but lem after defining multiple groups of drug-related and target- also infers their corresponding types of interactions, such as related features, such as drug–drug and gene–gene similarity drug-target relationships or drug modes of action. Our approach measures. In addition to chemical and genomic data, phenotypic uses a generalized version of a two-layer undirected graphical information, such as side-effect profiles (Campillos et al., 2008; model, called restricted Boltzmann machine (RBM) (Hinton Mizutani et al., 2012), transcriptional response data (Iorio et al., and Salakhutdinov, 2006), to represent a multidimensional 2010) and public gene expression data (Dudley et al., 2011; Sirota DTI network, which encodes different types of DTIs. In add- et al., 2011), has also been used for DTI prediction and drug ition, we apply a practical learning algorithm, called Contrastive repositioning. Although previous network-based approaches Divergence (CD) (Hinton, 2002), to train our RBM model and have achieved promising results for DTI prediction and drug predict unknown types of DTIs. repositioning, few of them are specifically designed for integrat- To our knowledge, our work is the first approach to predict ing and predicting different types of DTIs on a multidimensional different types of DTIs on a multidimensional network, which network. not only describes binary DTIs but also encodes their correspond- RBMs, which are used as important learning modules for con- ing types of interactions.We have testedour algorithm ontwo structing deep belief nets (Arel et al., 2010; Bengio, 2009), have public databases, namely, MATADOR (Gu¨ nther et al., 2008) been successfully applied in many fields, such as dimensionality and STITCH (Kuhn et al., 2012), which contain information reduction (Hinton and Salakhutdinov, 2006), classification (Lar- about drug-target relationships and drug modes of action, re- ochelle and Bengio, 2008), collaborative filtering (Salakhutdinov spectively. Our tests demonstrate that our RBM model can be et al., 2007) and computational biology (Eickholt and Cheng, used as a highly powerful tool for integrating different types of 2012). Recently, the predictive power of RBMs has also been DTIs into a multidimensional network and predicting different demonstrated in the Netflix Prize contest (Bell and Koren, types of interactions with high accuracy. In particular, our results 2007; Salakhutdinov et al., 2007), a public competition for de- show that integrating different types of DTIs into prediction with veloping the best collaborative filtering algorithm to predict user distinction can achieve the area under precision-recall curve ratings for movies. To our knowledge, our work is the first ap- (AUPR) up to 89.6, which can significantly outperform other proach to apply RBMs into large-scale DTI prediction. predictions either by simply mixing multiple types of interactions without distinction or using only a single interaction type. Further tests show that our approach can predict a high percent- 2 METHODS age of novel DTIs that has been validated by known experiments 2.1 RBMs for DTI prediction in the literature or other databases. These results indicate that our approach can have potential applications in drug reposition- We use an RBM to formulate the DTI prediction problem on a multidi- ing and hence advance the drug discovery process. mensional network. An RBM is a two-layer graphical model that can be i127 Y.Wang and J.Zeng AB AB Fig. 1. An RBM with binary hidden units representing latent features and visible units encoding observed types of DTIs. (A) Overview of an RBM, where m is the number of hidden units and n is the number of Fig. 2. A toy example for constructing RBMs from a multidimensional visible units. (B) The information encoded in a visible unit DTI network. (A) A simple multidimensional DTI network, where indi- cators x and x are equivalent to 1 if corresponding interaction direct indirect types are present in visible data and 0 otherwise. (B) Constructed RBMs for corresponding targets. The binary numbers inside rectangles represent used to learn a probability distribution over input data (Hinton and the states of visible variables. The RBMs for both target 1 and target 2 Salakhutdinov, 2006). As shown in Figure 1A, an RBM consists of a share the same parameters layer of visible units and a layer of hidden units. Each visible unit is connected to all hidden units, and no intra-layer connection exists be- joint configuration ðv, hÞ can be defined by tween any pair of visible units or any pair of hidden units. In our RBM model, visible units encode observed types of DTIs, such as drug-target n t m n m t XX X XXX k k k k Eðv, hÞ¼ a v  b h  W h v ð1Þ relationships and drug modes of action (Fig. 1B), and hidden units rep- j j j i i ij i i¼1 j¼1 i¼1 j¼1 k¼1 k¼1 resent latent features describing DTIs. We use a simple example (Fig. 2) to describe how to construct RBMs where n, m and t are the numbers of visible units, hidden units and types from a multidimensional DTI network. Figure 2A shows a toy example of DTI encoded in a visible unit, respectively, v , h are visible variables of a multidimensional DTI network, where each DTI is associated with and hidden units, respectively, and a , b are their corresponding bias binary variables, which represent the states of direct and indirect inter- weights (offsets), and W is the corresponding weight for the connection ij actions, respectively. As shown in Figure 2B, we build a specific RBM for k between visible variable v and hidden variable h . Then, the probability every single target with the same number of hidden units and the same of a joint configuration ðv, hÞ can be defined by definitions of visible units. The binary states of visible variables for those missing DTIs are treated as zero in the constructed RBMs. The con- Prðv, hÞ¼ expðÞ Eðv, hÞ ð2Þ structed RBMs for a multidimensional DTI network are associated with the same parameters. In other words, the constructed RBMs for where Z ¼ ðÞ Eðv, hÞ is called the normalizing constant or partition v, h function. Summing over all possible configurations of h,we obtainthe all targets use the same parameters between hidden and visible layers. following marginal distribution over visible data v: In Figure 2B, both RBMs for both target 1 and target 2 share the same parameters. PrðvÞ¼ expðÞ Eðv, hÞ : ð3Þ Compared with other prediction approaches, our RBM model can capture not only the correlations of drug-target pairs in a DTI network but also the correlations of different types of DTIs. Here, we use the As there is no intra-layer connection between any pair of visible or simple example shown in Figure 2 to illustrate how our RBM model hidden units, we can define the following conditional probabilities: exploits the corrections in a multidimensional network to make predic- k k k tion. Suppose that we have known different types of DTIs for target 1 in Prðv ¼ 1 j hÞ¼ ða þ W hÞð4Þ i i ij this example. To know which states of hidden units our RBM model j¼1 should have, we let the three drugs send messages to hidden units and n t XX update their corresponding states. Conversely, once we know the states of k k Prðh ¼ 1 j vÞ¼ ðb þ W vÞð5Þ j j ij i hidden units for target 1, hidden units then send messages to visible units i¼1 k¼1 for the connected drugs and update their corresponding states. As in other collaborative filtering applications (Salakhutdinov et al., 2007), where n, m and t are the numbers of visible units, hidden units and types after a number of iterations of the aforementioned updates, our RBM of DTI encoded in a visible unit, respectively, and ðxÞ¼ 1=ð1 þ e Þ is the logistic function. These conditional probabilities are important for the model can effectively capture the underlying features encoded in a multi- dimensional network. Based on the trained parameters between layers of iterative updates between hidden and visible layers when training an hidden and visible units, our RBM model can then predict unknown RBM model. types of DTIs based on input visible data. Unlike the conditional probabilities of visible variables in In an RBM, suppose that in total there are n visible units, m hidden Salakhutdinov et al. (2007), which used a conditional multinomial distri- units and t types of DTI encoded in a visible unit. Let binary indicator bution, here, our conditional probability of visible variables in Equation 1 k t ¼ðv , ... , v , ... , v Þ, i ¼ 1, ... , n, denote state of the i-th vis- (4) uses a conditional Bernoulli distribution. This is because in our RBM vector v i i i ible unit, where visible variables v ¼ 1if the k-th type of DTI is observed model, more than one visible variable can be assigned one for a DTI. For in input data, and v ¼ 0 otherwise. Let binary indicator variable h , example, a DTI can be annotated with ‘direct’, ‘binding’ and ‘inhibition’ j ¼ 1, ... , m, denote the state of the j-th hidden unit. Let W be the simultaneously. Our RBM model is data-driven and respect data as much ij weight associated with a connection between visible variable v and as possible. As also demonstrated in Salakhutdinov et al.(2007), RBMs hidden variable h.Let v ¼ðv , ... , v Þ and h ¼ðh , ... , h Þ denote the can capture the nature distribution of data. As we treat each interaction j 1 n 1 m configurations of visible and hidden layers, respectively. Then ðv, hÞ is type as a separate binary variable, it is possible that two inconsistent called a joint configuration of an RBM. As there is no intra-layer connec- interaction types (e.g. ‘direct’ and ‘indirect’) can be both yielded by our tion within both hidden and visible layers in an RBM, the energy of a RBM model. This means that the data implies that these two interaction i128 Predicting DTIs using RBMs types are both possible. In addition, by defining each interaction type as a parameters that have been learned in Section 2.2, and p^ is the conditional separate variable, our RBM model is general and can be easily extended probability of hidden variable h given visible data. After that, we com- to include more interaction types from other data sources without much pute the expectation as our final prediction for the query target. manual inspection on the consistence of different types of DTIs. 2.4 Conditional RBMs 2.2 Training For a drug-target pair, which does not have a connection on a multidi- mensional DTI network, it could be considered either a missing inter- We first use a real example to demonstrate the complexity of parameter action or a negative interaction. In general, a set of known DTIs that training in our RBM model. In the MATADOR-based data that we have have been verified experimentally provide more reliable information than tested (Section 3.1), the number of drugs is n ¼ 784. When only predict- those drug-target pairs, which do not have connections on a multidimen- ing direct and indirect interactions, the number of interaction types that sional DTI network. To incorporate this additional information, we use a we need to consider is t ¼ 2. In our RBM model, we typically set the conditional RBM (Salakhutdinov et al., 2007) to further formulate our number of hidden units m ¼ 100. In such a case, the total number of prediction problem. Let r ¼ðr , ... , r Þ be a binary indicator vector, in parameters in our RBM model is n  t þ m þ n  m  t ¼ 158468. In gen- 1 m which r ¼ 1 if there exists a known DTI between the input target and the eral, training an RBM with such a large number of parameters posts a j j-th drug, and r ¼ 0 otherwise. The states of r can be directly obtained difficult task. Later in the text, we will describe how to perform parameter j j based on the visible DTI data from a constructed DTI network. Here, r is training in an RBM model. a binary indicator vector representing the reliability of observed data. In To train an RBM and learn its parameter, we need to maximize the k k general, DTIs that have been verified experimentally provide more reli- likelihood of visible data with respect to the parameters W , a and b . ij i able information than unknown DTIs. A conditional RBM defines a To achieve this goal, we could perform gradient ascent in the log-likeli- joint distribution over ðv, hÞ conditioned on r.InaconditionalRBM, hood of the training data derived from Equation (3) and incrementally the conditional probabilities of visible and hidden units can be written as: adjust the weights and biases: n t m XX X @ log PrðvÞ k k k k k W ¼  ¼  5v h4 5v h4 ð6Þ j data j model Prðh ¼ 1 j v, rÞ¼  b þ W v þ D r , ð12Þ ij i i j j ij i k ij i @W ij i¼1 k¼1 i¼1 where " is the learning rate, 54 denotes an expectation of the data data where D is a parameter describing the effect of r on h. The parameter D ij ij distribution and 54 denotes an expectation of the distribution model can be also learned using the CD algorithm: defined by the model. D ¼  5h 4 5h4 r : ð13Þ ij j data j T i In Equation (6), 54 can be computed easily based on the fre- data quency information obtained from visible data. Unfortunately, it is gen- erally difficult to compute54 , as it would require exponential time model 2.5 Implementation to do so. To avoid this problem, Hinton (2002) proposed a practical learning algorithm, called Contrastive Divergence (CD), which minimizes We chose a conditional RBM to perform DTI prediction on a multidi- the Kullback–Leibler divergence. In this work, we use a mean-field ver- mensional network that encodes different types of DTIs. The algorithm sion of the CD algorithm (Le Roux and Bengio, 2008; Welling and was implemented in Java, using the jaRBM package (http://sourceforge. Hinton, 2002) to train our RBM model. In particular, we use the follow- net/projects/jarbm/). For the value of m, we chose an empirical range ing procedure in each training pass to incrementally adjust the weights according to the literature. Here, we chose m ¼ 100. We set learning and biases: rate  ¼ 0:01. For other parameters, we chose the default values that k k were defined in the jaRBM package. The initial values of W , b , a k k k j ij i W ¼  5v h4 5v h4 ð7Þ j data j T ij i i and D were sampled from Gaussian distribution with standard deriv- ij ation 0.1. In the CD learning procedure, we ran the mean-field updates k k k a ¼  5v 4 5v 4 ð8Þ data T i i i 100 passes over training data. The training of our RBM model runs in several hours on a typical training dataset. For example, the training of b ¼  5h4 5h4 ð9Þ j j data j T our RBM model on the MATADOR-based data (Supplementary Table S1 in Section S1) takes 5.5 h on a PC machine with a 3.3 GHz Intel core i5 where " is the learning rate, 54 denotes an average value over all data processor and 4 GB memory. input data for each update and 54 denotes the average value over T mean-field iterations, which is considered a good approximation of 54 in the log-likelihood function in Equation (6) (Hinton, 2002; model 3 RESULTS AND DISCUSSION Le Roux and Bengio, 2008; Welling and Hinton, 2002). 3.1 Datasets and evaluation metrics 2.3 Making predictions We tested our RBM model on two datasets, which were derived To predict the unknown types of DTIs for a given target with visible data from MATADOR (Gu¨ nther et al., 2008) and STITCH (Kuhn v, we compute the following probability distribution after one iteration of et al., 2012), respectively. MATADOR is a manually curated the mean-field updates (Salakhutdinov et al., 2007): online database, which mainly describes drug-target relation- n t ships, including direct and indirect DTIs. The list of direct and XX k k p^ ¼ Pr h j v ¼ ðb þ W vÞð10Þ j j j ij i indirect DTIs in MATADOR was first extracted by automated i¼1 k¼1 text-mining and then followed by manual annotation (Gu¨ nther et al., 2008). STITCH provides modes of action for the inter- k k k Prðv ¼ 1 j p^ ,   , p^ Þ¼ ða þ W p^ Þ, ð11Þ 1 m j actions between proteins and chemicals, which were annotated i i ij j¼1 based on evidence derived from known experiments in the litera- ture (Kuhn et al., 2012). For the MATADOR-based dataset, where n, m and t are the numbers of visible units, hidden units and types k k of DTI encoded in a visible unit, respectively, W , a and b are we only extracted those DTI records in which protein and ij i i129 Y.Wang and J.Zeng Table 1. Results on predicting direct and indirect DTIs drug names are present and annotation terms correspond to ‘DIRECT’ or ‘INDIRECT’. For the STITCH-based dataset, we only kept a list of DTI records in which drugs and target Drug-target Test method AUC AUPR proteins overlap those in MATADOR. For the mode of action relationship term in the STITCH-based data, we only considered ‘binding’, ‘activation’ and ‘inhibition’. In summary, the MATADOR- Direct interaction Integrating data with distinction 98.7 89.6 based dataset contains 784 drugs, 2431 protein targets and Mixing data without distinction 98.8 72.1 Using direct interactions only 98.0 78.9 13 064 DTIs, in which 7862 interactions are direct and 5202 inter- Indirect interaction Integrating data with distinction 97.1 80.1 actions are indirect. The STITCH-based dataset contains 598 Mixing data without distinction 97.0 37.8 drugs, 671 protein targets and 3296 DTIs, in which 2589, 945 Using indirect interactions only 94.8 62.4 and 1493 interactions are annotated with ‘binding’, ‘activation’ and ‘inhibition’, respectively. Descriptive statistics about these Note: ‘Integrating data with distinction’ corresponds to the test in which our algo- two datasets can be found in Table S1 in Supplementary rithm integrated both direct and indirect interactions with distinction. ‘Mixing data Material Section S1. without distinction’ corresponds to the test in which our algorithm mixed both direct and indirect interactions without distinction. ‘Using direct (indirect) inter- Receiver Operator Characteristic (ROC) and Precision-Recall actions only’ corresponds to the test in which our algorithm used only direct (PR) curves are commonly used to assess the performance of a (indirect) interactions. The highest AUPR score is shown in bold. prediction model. For highly skewed data, ROC curves can give an overoptimistic picture of an algorithm’s performance (Davis and Goadrich, 2006). In this scenario, PR curves provide a better view of the prediction results than ROC curves. As stated in Bleakley and Yamanishi (2009) and van Laarhoven et al. (2011), when evaluating the DTI prediction results, in which there are usually few positive DTIs, PR curves provide greater biological significance and are considered a better quality meas- ure than ROC curves. Therefore, we mainly used PR curves and the AUPR curve to evaluate the performance of our algorithm, though we also reported the area under the ROC curve (AUC) in our test results. 3.2 Predicting direct and indirect DTIs In general, drug-target relationships can be classified into two categories, including direct and indirect interactions. Direct inter- actions are usually caused by protein-ligand binding, whereas indirect interactions can be driven by different mechanisms. For instance, indirect interactions can be induced by metabolites of drugs or changes in gene expressions (Gu¨ nther et al., 2008). To examine whether our algorithm can accurately predict direct and indirect DTIs on a multidimensional network, we tested it on the MATADOR-based data. In particular, we performed the follow- ing three tests: (i) Integrating both direct and indirect DTIs with Fig. 3. PR curves for the direct and indirect DTIs predicted by our RBM distinction, which means that the input data of the RBM is a model. (A) PR curves for the direct DTIs predicted by our model. (B)PR curves for the indirect DTIs predicted by our model multidimensional vector indicating different types of inter- actions; (ii) Mixing both direct and indirect DTIs without dis- tinction, which means that the input data is a one-dimensional with distinction significantly outperformed the other two tests, binary vector indicating whether DTIs are observed; (iii) Using which only considered a single interaction type or simply mixed only a single interaction type (i.e. using only direct or indirect direct and indirect interactions without distinction. In particular, DTIs). For each test, we performed a 10-fold cross-validation the integration of direct and indirect interactions improved the procedure, as described in Supplementary Material Section S3. AUPR score by 410% for direct interaction prediction and The results of the aforementioned three tests are summarized 417% for indirect interaction prediction (Table 1). In PR in Table 1, and their corresponding PR curves are shown in curves, compared with the other two tests, integrating both inter- Figure 3. When integrating both direct and indirect interactions action types with distinction achieved a better precision value with distinction, our algorithm achieved the best performance almost at every recall value (Fig. 3). An interesting result is that with an AUPR score up to 89.6 for direct interaction prediction mixing both interactions without distinction yielded worse per- and 80.1 for indirect interaction prediction. These results demon- strated that our RBM model can effectively integrate different formance than using a single interaction type (Table 1). For ex- drug-target relationships on a multidimensional DTI network to ample, for direct interaction prediction, the AUPR score was accurately predict direct or indirect interactions. In addition, our decreased from 78.9 to 72.1. This result implies that incorporating results showed that integrating both direct and indirect DTIs a new data type into DTI prediction is non-trivial and requires i130 Predicting DTIs using RBMs careful data integration. In addition, our results show that our has been widely used in previous prediction approaches. More RBM model respect the data and can predict a high percentage details of this simple logic-based algorithm can be found of consistent DTIs. For example, when we chose 0.5 as the prob- in Supplementary Material Section S5. As summarized in ability threshold to infer the interaction type,499% of positive Table 2, the comparison results show that our algorithm outper- DTIs predictedbyouralgorithmwereconsistent(i.e.51% of DTIs formed the simple logic based approach. This is expected, as the were predicted as both direct and indirect interactions). simple logic based approach simply uses the closest interaction In our 10-fold cross-validation test (Table 1), the test method type profiles to predict unknown types of DTIs. Such a strategy ‘using direct (indirect) interaction only’ used less training data only exploits a small proportion of training data and cannot be than other two test methods. This may create bias when compar- sufficient enough to capture the deep correlations of different ing two prediction methods using training data with different types of DTIs in a multidimensional network. On the other sizes. To make a fair comparison, we have made an additional hand, our RBM model uses hidden units to represent the intrin- comparison between methods ‘integrating data with distinction’ sic correlations of different types of DTIs in the network and can and ‘using direct (indirect) interaction only’ using training data effectively capture the latent feature of drug-target relationships of the same size (Supplementary Material Section S4). Our new and thus make accurate predictions. comparison results confirmed that integrating data with distinc- Overall, our approach achieved better performance for direct tion outperformed the method that uses a single interaction only. interaction prediction than for indirect interaction prediction To check whether our algorithm can have a wider range of (Table 1). Probably this is because in the MATADOR-based applications, we also performed a 5-fold cross-validation test data, indirect interactions describe many different mechanisms (Supplementary Material Section S3). Compared with the 10- of DTIs and thus provide less predictive power than direct inter- fold cross-validation results, we only found a small decrease in actions. We expect that our method would achieve better per- our algorithm’s performance in the 5-fold cross-validation test. formance for indirect interaction prediction if different To perform more sanity check on our algorithm’s perform- mechanisms of indirect DTIs in the MATADOR-based data ance, we also conducted another test that is similar to leave-one- could be further identified. out cross-validation except that we removed homologous pro- teins from training data. For training data, we only kept those 3.3 Predicting drug modes of action DTIs in which proteins have sequence identity525% to the target In addition to direct and indirect interactions, DTIs can be also of the validating interaction. This process can significantly reduce annotated with different drug modes of action, such as activation the size of dataset. In our test, the training dataset has only 300 or inhibition. To evaluate our algorithm’s performance on pre- targets. This test is more rigorous than previous leave-one-out dicting different drug modes of action, we tested it mainly on the cross-validation (LOOCV) tests performed in (Bleakley and STITCH-based data. We focused on three drug modes of action, Yamanishi, 2009; van Laarhoven et al., 2011; Yamanishi et al., including binding, activation and inhibition. In addition to test- 2008), in that we have removed similar proteins with high sequence ing the STITCH-based data, we included two additional tests, identity from training data and thus reduced impact of protein which further incorporated the MATADOR-based data, that is, homology on the network-based prediction of DTIs. each DTI in the STITCH-based data was also associated with The results of our LOOCV like test show that, our approach direct or indirect interaction derived from the MATADOR- still achieved decent prediction accuracy, with AUPR 79.0 for based data. Overall, we conducted the following five tests: (i) direct interaction prediction, but the AUPR score for indirect integrating both drug-target relationships from the interaction prediction dropped to 59.1. We further investigated MATADOR-based data and drug modes of action from the this problem and found that the average degree of indirect inter- STITCH-based data with distinction; (ii) mixing DTIs from actions with drugs for targets in training data was reduced to both MATADOR-based and STITCH-based data without dis- 1.67. With insufficient number of DTIs, it would be difficult to tinction (i.e. all DTIs were associated with only binary values); train our RBM model and make accurate prediction. (iii) integrating drug modes of action from the STITCH-based To mimic the real situation in which DTI network data are data with distinction; (iv) mixing drug modes of action from the available, we also performed an additional test on those DTIs in STITCH-based data without distinction; and (v) using a singe which the degree of the target is above the average degree of the mode of action from the STITCH-based data. For every test, we multidimensional network (i.e. the degree of the target in the validating interaction is  6). The results of this test show that, our approach achieved decent performance, with AUPR 80.4 for Table 2. Results on comparing our approach with the simple logic based direct interaction prediction and 74.5 for indirect interaction pre- approach diction. These results indicate that in the real situation, our RBM model can make reasonably accurate prediction even homolo- Drug-target Test method AUC AUPR gous proteins are not present in training data. relationship As little work has been developed for predicting unknown types of DTIs in a multidimensional network, it is difficult for Direct interaction Our approach 98.7 89.6 us to directly compare our work with other prediction Simple logic-based approach 92.1 81.6 approaches. Instead, we have compared our algorithm with a Indirect interaction Our approach 97.1 80.1 simple logic-based approach on the MATADOR-based data. Simple logic-based approach 88.4 74.5 This simple logic-based approach follows the basic premise (i.e. similar drugs and targets should have similar interactions) that Note: The highest AUPR score is shown in bold. i131 Y.Wang and J.Zeng also carried out a 10-fold cross-validation procedure as described performed the following test, which is similar to the experiments in Supplementary Material Section S3. conducted in (Bleakley and Yamanishi, 2009; van Laarhoven Both AUC and AUPR scores of the aforementioned five tests et al., 2011). We predicted unknown direct DTIs using the com- are reported in Table 3, and their corresponding PR curves are plete MATADOR-based dataset as training data. For the pre- shown in Figure 4. When integrating different types of DTIs with dicted results, we mainly focused on DTIs involving those distinction, our algorithm can achieve AUC up to 96.9 and proteins that are not listed as drug-metabolizing enzymes or AUPR up to 79.1. The test results showed that our RBM transporter proteins in DrugBank (Knox et al., 2011). We out- model can effectively exploit different types of interactions putted the set of the top 50 scoring predictions that are not pre- encoded in a multidimensional DTI network to predict drug sent in training data. We then checked whether these new modes of action. Among all five tests, integrating different predicted DTIs appear in other databases, including ChEMBL types of DTIs with distinction produced higher AUPR scores (Gaulton et al., 2011), DrugBank (Knox et al., 2011) and than the other tests, which used only a single interaction type STITCH (Kuhn et al., 2012). or simply mixed types of interactions without distinction. When Supplementary Figure S1 in Section S2 visualizes part of the differentiating different types of DTIs, integrating direct and in- DTI network constructed based on the set of the 50 highest direct interactions from the MATADOR-based data slightly im- scoring interactions predicted by our approach using the proved the prediction performance. When comparing the results MATADOR-based data. Table 4 shows the list of the top 10 on predicting different modes of action, binding interactions scoring direct DTI predictions. Among these predicted DTIs, a were predicted with higher accuracy than activation and inhib- high fraction of the predicted interactions (7 of 10) was found in ition interactions. This is probably because in the STITCH-based ChEMBL, DrugBank or STITCH. In the remaining three DTIs data, binding interactions are more informative in predicting that we did not find experimental evidence from the three data- DTIs than other two modes of action. bases, there might still exist interaction for each drug-target pair. As in Section 3.2, we also performed an additional test using For example, although it is known that mifepristone does not training data of the same size for comparing methods ‘integrating bind to estrogen receptor, it may have some effects on the ex- data with distinction’ and ‘using a single data type only’, when pression of estrogen receptor (Jiang et al., 2007). predicting different modes of action. Our new comparison results In Table 4, some of our new predictions are trivial, as they can confirmed that integrating different data types with distinction also be easily derived based on known interactions of similar outperformed the method that uses only a single data type proteins or drugs in the dataset. However, our new predictions (Supplementary Material Section S4). also provided some interesting results. For example, among the top 50 scoring predictions, our algorithm suggested that there 3.4 New predictions may exist interactions between spironolactone and the mem- brane progestin receptor gamma protein and between mesalazine To examine our algorithm’s ability for predicting novel DTIs (also called 5-aminosalicylic acid) and leukotriene-A4 hydrolase that are not present in our data derived from MATADOR, we Table 3. Results on predicting drug modes of action Mode of action Test method AUC AUPR Binding interaction Integrating MATADOR and STITCH with distinction 96.9 79.1 Mixing MATADOR and STITCH without distinction 97.9 53.3 Integrating data with distinction 95.0 78.1 Mixing data without distinction 95.6 68.0 Using binding interactions only 94.1 74.4 Activation interaction Integrating MATADOR and STITCH with distinction 94.4 67.4 Mixing MATADOR and STITCH without distinction 96.9 35.6 Integrating data with distinction 91.2 65.2 Mixing data without distinction 94.2 50.5 Using activation interactions only 87.7 56.3 Inhibition interaction Integrating MATADOR and STITCH with distinction 94.1 67.1 Mixing MATADOR and STITCH without distinction 96.9 38.6 Integrating data with distinction 92.5 65.2 Mixing data without distinction 93.9 44.3 Using inhibition interactions only 89.5 60.2 Note: ‘Integrating MATADOR and STITCH with distinction’ corresponds to the test in which our algorithm integrated both drug-target relationships from the MATADOR- based data and drug modes of action from the STITCH-based data with distinction. ‘Mixing MATADOR and STITCH without distinction’ corresponds to the test in which our algorithm mixed DTIs from both MATADOR-based and STITCH-based data without distinction. ‘Integrating data with distinction’ corresponds to the test in which our algorithm integrated drug modes of action from the STITCH-based data with distinction. ‘Mixing data without distinction’ corresponds to the test in which our algorithm mixed drug modes of action from the STITCH-based data without distinction. ‘Using binding (activation or inhibition) interactions only’ corresponds to the test in which our algorithm used only binding (activation or inhibition) interactions from the STITCH-based data. The highest AUPR score is shown in bold. i132 Predicting DTIs using RBMs Table 4. Top 10 scoring direct DTIs predicted by our approach using the MATADOR data a b Rank Pair Description Evidence 1 DB00812 Phenylbutazone C, D, S P23219 PTGS1: Prostaglandin G/H synthase 1 2 DB04599 Aniracetam S P42261 GRIA1: Glutamate receptor 1 precursor 3 DB00834 Mifepristone P03372 ESR1: Estrogen receptor 4 DB01392 Yohimbine D, S P28222 HTR1B: 5-hydroxytryptamine 1B receptor 5 DB01297 Practolol S P07550 ADRB2: Beta-2 adrenergic receptor 6 DB01297 Practolol P13945 ADRB3: Beta-3 adrenergic receptor 7 DB00334 Olanzapine D, S P08908 HTR1A: 5-hydroxytryptamine 1 A receptor 8 DB01224 Quetiapine D P21918 DRD5: D(1B) dopamine receptor 9 DB01224 Quetiapine D P21728 DRD1: D(1 A) dopamine receptor 10 DB01233 Metoclopramide P21918 DRD5: D(1B) dopamine receptor Fig. 4. PR curves for the predicted drug modes of action. (A)PR curves for the predicted binding interactions. (B) PR curves for the predicted Drugs and targets are represented by DrugBank IDs and UniProt ID, respectively. activation interactions. (C) PR curves for the predicted inhibition DTIs that are observed in ChEMBL, DrugBank and STITCH are marked with interactions ‘C’, ‘D’ and ‘S’, respectively. (LTA4H) enzyme. Spironolactone is known as a diuretic or anti- different types of DTIs, such as drug-target relationships or hypertensive drug and can act on the aldosterone receptor as a drug modes of action. Tests on two public databases showed competitive antagonist (Macdonald, 1997). Our predicted inter- that our algorithm can achieve excellent prediction performance action between spironolactone and membrane progestin receptor with high AUPR scores. Further tests indicated that our ap- gamma protein indicates that spironolactone may have some proach can infer a list of novel DTIs, which is practically progestogenic function. Interestingly, this hypothesis can be con- useful for drug repositioning. firmed from the clinical studies performed in (Nakhjavani et al., Although our algorithm has been tested only on direct and 2009). In addition, mesalazine is an anti-inflammatory drug that indirect drug-target relationships, and three drug modes of is primarily used to treat inflammatory bowel disease (Sandborn action, it is general and can be easily extended to integrate et al., 2007). LTA4H is an important enzyme that converts leu- other types of DTIs (e.g. phenotypic effects). Current version kotriene-A4 to leukotriene-B4 (Rudberg et al., 2004). It has been of our prediction algorithm only considers connections between proposed that leukotriene-B4 may play an important role in a drugs and targets. In the future, we will extend our approach to number of different acute and chronic inflammatory diseases, exploit the connections within target proteins or drugs. For ex- including inflammatory bowel disease (Haeggstro¨ m, 2000). ample, the sequence similarity scores between target proteins, the These studies imply that our predicted interaction between mesa- substructure similarity scores between drugs or drug–drug inter- lazine and LTA4H is probably true. actions (Gottlieb et al., 2012; Tatonetti et al., 2012) can be also These results on new predictions indicated that our RBM incorporated into our prediction model. As the conventional model is practically useful in predicting novel DTIs and can version of an RBM does not allow the connections within the have potential applications in drug repositioning. same layer, such an extension will require careful thought. Currently, our algorithm has been tested only on two databases (i.e. MATADOR and STITCH). We will test it on more data in 4 CONCLUSION the future. For example, it will have more significance if we can In this article, we proposed a first machine-learning approach to predict DTIs based on human proteins and all molecules in predict different types of DTIs on a multidimensional network. PubChem (Kaiser, 2005) or a similar database. Finally, we are Our approach uses an RBM model to effectively encode multiple also seeking wet-laboratory collaborators to experimentally sources of information about DTIs and accurately predict verify the highest scoring DTIs predicted by our algorithm. i133 Y.Wang and J.Zeng Jiang,J. et al. (2007) Effects of mifepristone on expression of estrogen receptor and ACKNOWLEDGEMENTS progesterone receptor in cultured human eutopic and ectopic endometria. The authors thank Dr L. Wang for the helpful discussions in the Zhonghua Fu Chan Ke Za Zhi, 36, 218–221. Kaiser,J. (2005) Science resources. chemists want NIH to curtail database. Science, early stage of this project. They are grateful to Mr M. Zhou for 308,774. helping us prepare the drug-target interaction data. They thank Keiser,M.J. et al. (2007) Relating protein pharmacology by ligand chemistry. Nat. the anonymous reviewers for their helpful comments and Biotechnol., 25, 197–206. suggestions. Keiser,M.J. et al. (2009) Predicting new molecular targets for known drugs. Nature, 462, 175–181. Funding: This work was supported in part by the National Basic Knox,C. et al. (2011) DrugBank 3.0: a comprehensive resource for ‘omics’ research on drugs. Nucleic Acids Res., 39, D1035–D1041. Research Program of China Grant 2011CBA00300, Kroeze,W.K. et al. (2002) Molecular biology of serotonin receptors structure and 2011CBA00301, the National Natural Science Foundation of function at the molecular level. Curr. Top. Med. Chem., 2, 507–528. China Grant 61033001, 61061130540, 61073174. Kuhn,M. et al. (2012) STITCH 3: zooming in on protein-chemical interactions. Nucleic Acids Res., 40, D876–D880. Conflict of Interest: none declared. Larochelle,H. and Bengio,Y. (2008) Classification using discriminative restricted boltzmann machines. In: Proceedings of the 25th international conference on Machine learning, ICML’08. ACM, New York, NY, USA, pp. 536–543. Le Roux,N. and Bengio,Y. (2008) Representational power of restricted boltzmann REFERENCES machines and deep belief networks. Neural Comput., 20, 1631–1649. Lu¨ ,L. and Zhou,T. (2011) Link prediction in complex networks: a survey. Physica A Arel,I. et al. (2010) Deep machine learning-a new frontier in artificial intelligence Stat. Mech. Appl., 390, 1150–1170. research. IEEE Comput. Intell. Mag., 5, 13–18. Macdonald,F. (1997) Dictionary of Pharmacological Agents. CRC Press, Taylor and Bell,R. and Koren,Y. (2007) Lessons from the netflix prize challenge. ACM Francis Group, LLC, Boca Raton, FL, USA. SIGKDD Explor. Newsl., 9,75–79. MacDonald,M.L. et al. (2006) Identifying off-target effects and hidden phenotypes Bengio,Y. (2009) Learning deep architectures for AI. Found. Trends Mach. Learn., of drugs in human cells. Nat. Chem. Biol., 2, 329–337. 2,1–127. Mei,J.P. et al. (2012) Drug-target interaction prediction by learning from local in- Bleakley,K. and Yamanishi,Y. (2009) Supervised prediction of drug-target inter- formation and neighbors. Bioinformatics, 29, 238–245. actions using bipartite local models. Bioinformatics, 25, 2397–2403. Mizutani,S. et al. (2012) Relating drug-protein interaction network with drug side Booth,B. and Zemmel,R. (2004) Prospects for productivity. Nat. Rev. Drug. Discov., effects. Bioinformatics, 28, i522–i528. 3, 451–456. Morris,G.M. et al. (2009) AutoDock4 and AutoDockTools4: automated docking Campillos,M. et al. (2008) Drug target identification using side-effect similarity. with selective receptor flexibility. J. Comput. Chem., 30, 2785–2791. Science, 321, 263–266. Nakhjavani,M. et al. (2009) Short term effects of spironolactone on blood lipid Chen,X. et al. (2012) Drug-target interaction prediction by random walk on the profile: a 3-month study on a cohort of young women with hirsutism. Br. J. heterogeneous network. Mol. Biosyst., 8, 1970–1978. Clin. Pharmacol., 68, 634–637. Cheng,A.C. et al. (2007) Structure-based maximal affinity model predicts small- Perlman,L. et al. (2011) Combining drug and gene similarity measures for drug- molecule druggability. Nat. Biotechnol., 25, 71–75. target elucidation. J. Comput. Biol., 18, 133–145. Cheng,F. et al. (2012) Prediction of drug-target interactions and drug repositioning Roth,B.L. et al. (2004) Magic shotguns versus magic bullets: selectively non-selective via network-based inference. PLoS Comput. Biol., 8, e1002503. drugs for mood disorders and schizophrenia. Nat. Rev. Drug Discov., 3, Davis,J. and Goadrich,M. (2006) The relationship between precision-recall and roc 353–359. curves. In: Proceedings of the 23rd international conference on Machine learning, Rudberg,P.C. et al. (2004) Leukotriene A4 hydrolase: identification of a common ICML’06. ACM, New York, NY, USA, pp. 233–240. carboxylate recognition site for the epoxide hydrolase and aminopeptidase sub- Dimasi,J.A. (2001) New drug development in the united states from 1963 to 1999. strates. J. Biol. Chem., 279, 27376–27382. Clin. Pharmacol. Ther., 69, 286–296. Salakhutdinov,R. et al. (2007) Restricted boltzmann machines for collaborative Donald,B.R. (2011) Algorithms in Structural Molecular Biology. MIT Press, filtering. In: Proceedings of the 24th international conference on Machine learning, Cambridge, MA, USA. ICML’07. ACM, New York, NY, USA, pp. 791–798. Dudley,J.T. et al. (2011) Computational repositioning of the anticonvulsant topir- Sandborn,W.J. et al. (2007) Medical management of mild to moderate crohn’s dis- amate for inflammatory bowel disease. Sci. Transl. Med., 3, 96ra76. ease: evidence-based treatment algorithms for induction and maintenance of Eickholt,J. and Cheng,J. (2012) Predicting protein residue-residue contacts using remission. Aliment Pharmacol. Ther., 26, 987–1003. deep networks and boosting. Bioinformatics, 28, 3066–3072. Sirota,M. et al. (2011) Discovery and preclinical validation of drug indications using Gaulton,A. et al. (2011) ChEMBL: a large-scale bioactivity database for drug dis- compendia of public gene expression data. Sci. Transl. Med., 3, 96ra77. covery. Nucleic Acids Res., 40, D1100–D1107. Tatonetti,N.P. et al. (2012) Data-driven prediction of drug effects and interactions. Gottlieb,A. et al. (2011) PREDICT: a method for inferring novel drug indications Sci. Transl. Med., 4, 125–131. with application to personalized medicine. Mol. Syst. Biol., 7,496. Van Laarhoven,T. et al. (2011) Gaussian interaction profile kernels for predicting Gottlieb,A. et al. (2012) INDI: a computational framework for inferring drug inter- drug-target interaction. Bioinformatics, 27, 3036–3043. actions and their associated recommendations. Mol. Syst. Biol., 8,592. Welling,M. and Hinton,G. (2002) A new learning algorithm for mean field boltz- Gu¨ nther,S. et al. (2008) SuperTarget and Matador: resources for exploring drug- mann machines. Artif. Neural Netw., 2415, 351–357. target relationships. Nucleic Acids Res., 36, D919–D922. Xia,Z. et al. (2010) Semi-supervised drug-protein interaction prediction from het- Haeggstro¨ m,J.Z. (2000) Structure, function, and regulation of leukotriene A4 erogeneous biological spaces. BMC Syst. Biol., 4 (Suppl. 2),S6. hydrolase. Am. J. Respir. Crit. Care. Med., 161,S25–S31. Xie,L. et al. (2012) Novel computational approaches to polypharmacology as a He,Z. et al. (2010) Predicting drug-target interaction networks based on functional means to define responses to individual drugs. Annu. Rev. Pharmacol. groups and biological features. PLoS One, 5,e9603. Toxicol., 52, 361–379. Hinton,G. and Salakhutdinov,R. (2006) Reducing the dimensionality of data with Yabuuchi,H. et al. (2011) Analysis of multiple compound-protein interactions re- neural networks. Science, 28, 504–507. veals novel bioactive molecules. Mol. Syst. Biol., 7,472. Hinton,G.E. (2002) Training products of experts by minimizing contrastive diver- Yamanishi,Y. et al. (2008) Prediction of drug-target interaction networks from the gence. Neural Comput., 14, 1771–1800. integration of chemical and genomic spaces. Bioinformatics, 24, i232–i240. Iorio,F. et al. (2010) Discovery of drug mode of action and drug repositioning from Zhao,S. and Li,S. (2010) Network-based relating pharmacological and genomic transcriptional responses. Proc. Natl Acad. Sci. USA, 107, 14621–14626. spaces for drug target identification. PLoS One, 5,e11764. i134 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Bioinformatics Pubmed Central

Predicting drug-target interactions using restricted Boltzmann machines

Wang, Yuhao; Zeng, Jianyang
Bioinformatics , Volume 29 (13) – Jun 19, 2013
Download PDF
9 pages

Loading next page...
 
/lp/pubmed-central/predicting-drug-target-interactions-using-restricted-boltzmann-KxwHg5lnz9

References (59)

Publisher
Pubmed Central
Copyright
© The Author 2013. Published by Oxford University Press.
ISSN
1367-4803
eISSN
1367-4811
DOI
10.1093/bioinformatics/btt234
Publisher site
See Article on Publisher Site

Abstract

Vol. 29 ISMB/ECCB 2013, pages i126–i134 BIOINFORMATICS doi:10.1093/bioinformatics/btt234 Predicting drug-target interactions using restricted Boltzmann machines 1 2, Yuhao Wang and Jianyang Zeng 1 2 Department of Automation and Institute for Interdisciplinary Information Sciences, Tsinghua University, Beijing 100084, China ABSTRACT urged drug developers to seek new uses for existing or aban- doned drugs (Booth and Zemmel, 2004). Such a new strategy Motivation: In silico prediction of drug-target interactions plays an is also called drug repositioning or drug repurposing. A strong important role toward identifying and developing new uses of existing support for the possibility of drug repositioning is the increas- or abandoned drugs. Network-based approaches have recently ingly accepted concept of ‘polypharmacology’, i.e. individual become a popular tool for discovering new drug-target interactions drugs can interact with multiple targets rather than a single (DTIs). Unfortunately, most of these network-based approaches can target (MacDonald et al., 2006; Xie et al., 2012). For example, only predict binary interactions between drugs and targets, and infor- serotonin and serotonergic drugs can interact with both 5-HT G mation about different types of interactions has not been well ex- protein-coupled receptors and 59HT ion channel proteins, 3A ploited for DTI prediction in previous studies. On the other hand, even though these two target proteins are not related in sequence incorporating additional information about drug-target relationships or structure (Cheng et al., 2012; Keiser et al., 2007; Kroeze et al., or drug modes of action can improve prediction of DTIs. 2002; Roth et al., 2004). This polypharmacological property of a Furthermore, the predicted types of DTIs can broaden our understand- drug enables us to identify more than one target that it can act on ing about the molecular basis of drug action. and hence develop its new uses. Results: We propose a first machine learning approach to integrate In silico prediction of interactions between drugs and target multiple types of DTIs and predict unknown drug-target relationships proteins provides an important tool for drug repositioning, as it or drug modes of action. We cast the new DTI prediction problem into can significantly reduce wet-laboratory work and lower the cost a two-layer graphical model, called restricted Boltzmann machine,and of the experimental determination of new drug-target inter- apply a practical learning algorithm to train our model and make pre- actions (DTIs). Various methods have been proposed for in dictions. Tests on two public databases show that our restricted silico DTI prediction. When 3D structures are available, molecu- Boltzmann machine model can effectively capture the latent features lar docking is commonly used to virtually screen a large number of a DTI network and achieve excellent performance on predicting of compounds against a target protein (Cheng et al., 2007; different types of DTIs, with the area under precision-recall curve up Donald, 2011; Morris et al., 2009). When 3D structures of mol- to 89.6. In addition, we demonstrate that integrating multiple types of ecules are absent, a number of different approaches have been DTIs can significantly outperform other predictions either by simply developed to address the in silico DTI prediction problem. Most mixing multiple types of interactions without distinction or using only of these structure-free methods can be grouped into two classes, a single interaction type. Further tests show that our approach can namely, ligand-based and network-based approaches. A represen- infer a high fraction of novel DTIs that has been validated by known tative ligand-based approach is the similarity ensemble approach experiments in the literature or other databases. These results indicate (Keiser et al., 2007, 2009), which predicts new DTIs using 2D that our approach can have highly practical relevance to DTI predic- structure similarity of ligands. Although ligand-based tion and drug repositioning, and hence advance the drug discovery approaches are able to discover a number of DTIs that have process. been validated experimentally, they have difficulty in identifying Availability: Software and datasets are available on request. drugs with novel scaffolds that differ from those of reference Contact: [email protected] compounds (Yabuuchi et al., 2011). Supplementary information: Supplementary data are available at Numerous network-based approaches have been proposed to Bioinformatics online. exploit latent features of DTI profiles and have recently become a popular tool for DTI prediction and drug repositioning (Bleakley and Yamanishi, 2009; Chen et al., 2012; Cheng et al., 1INTRODUCTION 2012; Mei et al., 2012; van Laarhoven et al., 2011; Xia et al., Drug development currently remains an expensive and time-con- 2010; Xie et al., 2012; Yamanishi et al., 2008). Although these suming process with extremely low success rate: it typically takes network-based approaches have achieved promising results, 10–15 years and $800 million–1 billion to bring a new drug to most of them can only predict binary DTIs, that is, they can market (Dimasi, 2001). In recent decades, the rate of the number only determine whether a drug interacts with a target protein, of new drugs approved by the US Food and Drug but cannot tell how they interact with each other. However, in- Administration versus the amount of money invested in pharma- dividual DTIs generally have different meanings. For example, ceutical research and development has significantly declined drug-target pairs can be described by different relationships,such (Booth and Zemmel, 2004). This productivity problem has as direct and indirect interactions (Gu¨ nther et al., 2008). A direct interaction is usually caused by protein-ligand binding, whereas *To whom correspondence should be addressed. an indirect interaction can be induced by the changed expression The Author 2013. Published by Oxford University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/ by-nc/3.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact [email protected] Predicting DTIs using RBMs level of a protein or active metabolites produced by a drug. In 1.1 Related work addition, DTIs can be annotated with different drug modes of A number of network-based approaches have been proposed for action, e.g. activation and inhibition (Kuhn et al., 2012). Here, predicting unknown interactions between drugs and targets. In ‘drug modes of action’ may be slightly different from the term Yamanishi et al. (2008), a supervised learning framework was used in the literature, e.g. (Iorio et al., 2010). In this article, we developed based on a bipartite graph, which integrates both mainly use this term to represent the following three specific chemical and genomic spaces by mapping them into a unified types of interactions: binding, activation and inhibition. space. Cheng et al. (2012) proposed a network-based inference Hereinafter, we will use ‘types of DTIs’ to represent both drug- approach to predict new DTIs by exploiting the topology simi- target relationships and drug modes of action. A network in larity of the underlying interaction network. In Zhao and Li which links are associated with different meanings is called the (2010), drug phenotypic, chemical indexes and protein–protein multidimensional network (Lu¨ and Zhou, 2011). In our context, interactions in genomic space were integrated into a computa- we call a DTI network where links are annotated with different tional framework for DTI prediction. Chen et al. (2012) pro- types of interactions the multidimensional DTI network.On the posed a random walk approach for DTI prediction based on a one hand, types of DTIs provide additional useful information, heterogeneous network, which integrates drug similarity, target which can be incorporated into a multidimensional network to similarity and DTI similarity. Bleakley and Yamanishi (2009) improve DTI prediction. On the other hand, the predicted types presented a new approach, called Bipartite Local Model of DTIs can extend our understanding about the molecular basis (BLM), to predict unknown DTIs by combining independent of drug action. Despite these positive aspects, current rich infor- drug-based and target-based prediction results using a supervised mation about types of DTIs (Gu¨ nther et al., 2008; Kuhn et al., learning method. Mei et al. (2012) further extended this BLM 2012) has not been well exploited for DTI prediction, and how to approach to incorporate the capacity of learning from neighbors incorporate such information into a multidimensional network to and predict the interactions for new drug or target candidates. In predict different types of DTIs still remains an open question. Xia et al. (2010), a manifold regularization semi-supervised learn- In this article, we propose an effective machine-learning ap- ing method was proposed to integrate heterogenous biological proach to accurately predict different types of DTIs on a multi- data sources for DTI prediction. A regularized least square al- dimensional network. Unlike previous network-based gorithm was proposed in van Laarhoven et al. (2011) for DTI approaches (Bleakley and Yamanishi, 2009; Chen et al., 2012; prediction using a product of kernels derived from DTI profiles. Cheng et al., 2012; Mei et al., 2012; van Laarhoven et al., 2011; In Gottlieb et al. (2011), He et al. (2010) and Perlman et al. Xia et al., 2010; Yamanishi et al., 2008), which only predict (2011), DTI prediction was formulated into a classification prob- binary DTIs, our approach not only identifies new DTIs but lem after defining multiple groups of drug-related and target- also infers their corresponding types of interactions, such as related features, such as drug–drug and gene–gene similarity drug-target relationships or drug modes of action. Our approach measures. In addition to chemical and genomic data, phenotypic uses a generalized version of a two-layer undirected graphical information, such as side-effect profiles (Campillos et al., 2008; model, called restricted Boltzmann machine (RBM) (Hinton Mizutani et al., 2012), transcriptional response data (Iorio et al., and Salakhutdinov, 2006), to represent a multidimensional 2010) and public gene expression data (Dudley et al., 2011; Sirota DTI network, which encodes different types of DTIs. In add- et al., 2011), has also been used for DTI prediction and drug ition, we apply a practical learning algorithm, called Contrastive repositioning. Although previous network-based approaches Divergence (CD) (Hinton, 2002), to train our RBM model and have achieved promising results for DTI prediction and drug predict unknown types of DTIs. repositioning, few of them are specifically designed for integrat- To our knowledge, our work is the first approach to predict ing and predicting different types of DTIs on a multidimensional different types of DTIs on a multidimensional network, which network. not only describes binary DTIs but also encodes their correspond- RBMs, which are used as important learning modules for con- ing types of interactions.We have testedour algorithm ontwo structing deep belief nets (Arel et al., 2010; Bengio, 2009), have public databases, namely, MATADOR (Gu¨ nther et al., 2008) been successfully applied in many fields, such as dimensionality and STITCH (Kuhn et al., 2012), which contain information reduction (Hinton and Salakhutdinov, 2006), classification (Lar- about drug-target relationships and drug modes of action, re- ochelle and Bengio, 2008), collaborative filtering (Salakhutdinov spectively. Our tests demonstrate that our RBM model can be et al., 2007) and computational biology (Eickholt and Cheng, used as a highly powerful tool for integrating different types of 2012). Recently, the predictive power of RBMs has also been DTIs into a multidimensional network and predicting different demonstrated in the Netflix Prize contest (Bell and Koren, types of interactions with high accuracy. In particular, our results 2007; Salakhutdinov et al., 2007), a public competition for de- show that integrating different types of DTIs into prediction with veloping the best collaborative filtering algorithm to predict user distinction can achieve the area under precision-recall curve ratings for movies. To our knowledge, our work is the first ap- (AUPR) up to 89.6, which can significantly outperform other proach to apply RBMs into large-scale DTI prediction. predictions either by simply mixing multiple types of interactions without distinction or using only a single interaction type. Further tests show that our approach can predict a high percent- 2 METHODS age of novel DTIs that has been validated by known experiments 2.1 RBMs for DTI prediction in the literature or other databases. These results indicate that our approach can have potential applications in drug reposition- We use an RBM to formulate the DTI prediction problem on a multidi- ing and hence advance the drug discovery process. mensional network. An RBM is a two-layer graphical model that can be i127 Y.Wang and J.Zeng AB AB Fig. 1. An RBM with binary hidden units representing latent features and visible units encoding observed types of DTIs. (A) Overview of an RBM, where m is the number of hidden units and n is the number of Fig. 2. A toy example for constructing RBMs from a multidimensional visible units. (B) The information encoded in a visible unit DTI network. (A) A simple multidimensional DTI network, where indi- cators x and x are equivalent to 1 if corresponding interaction direct indirect types are present in visible data and 0 otherwise. (B) Constructed RBMs for corresponding targets. The binary numbers inside rectangles represent used to learn a probability distribution over input data (Hinton and the states of visible variables. The RBMs for both target 1 and target 2 Salakhutdinov, 2006). As shown in Figure 1A, an RBM consists of a share the same parameters layer of visible units and a layer of hidden units. Each visible unit is connected to all hidden units, and no intra-layer connection exists be- joint configuration ðv, hÞ can be defined by tween any pair of visible units or any pair of hidden units. In our RBM model, visible units encode observed types of DTIs, such as drug-target n t m n m t XX X XXX k k k k Eðv, hÞ¼ a v  b h  W h v ð1Þ relationships and drug modes of action (Fig. 1B), and hidden units rep- j j j i i ij i i¼1 j¼1 i¼1 j¼1 k¼1 k¼1 resent latent features describing DTIs. We use a simple example (Fig. 2) to describe how to construct RBMs where n, m and t are the numbers of visible units, hidden units and types from a multidimensional DTI network. Figure 2A shows a toy example of DTI encoded in a visible unit, respectively, v , h are visible variables of a multidimensional DTI network, where each DTI is associated with and hidden units, respectively, and a , b are their corresponding bias binary variables, which represent the states of direct and indirect inter- weights (offsets), and W is the corresponding weight for the connection ij actions, respectively. As shown in Figure 2B, we build a specific RBM for k between visible variable v and hidden variable h . Then, the probability every single target with the same number of hidden units and the same of a joint configuration ðv, hÞ can be defined by definitions of visible units. The binary states of visible variables for those missing DTIs are treated as zero in the constructed RBMs. The con- Prðv, hÞ¼ expðÞ Eðv, hÞ ð2Þ structed RBMs for a multidimensional DTI network are associated with the same parameters. In other words, the constructed RBMs for where Z ¼ ðÞ Eðv, hÞ is called the normalizing constant or partition v, h function. Summing over all possible configurations of h,we obtainthe all targets use the same parameters between hidden and visible layers. following marginal distribution over visible data v: In Figure 2B, both RBMs for both target 1 and target 2 share the same parameters. PrðvÞ¼ expðÞ Eðv, hÞ : ð3Þ Compared with other prediction approaches, our RBM model can capture not only the correlations of drug-target pairs in a DTI network but also the correlations of different types of DTIs. Here, we use the As there is no intra-layer connection between any pair of visible or simple example shown in Figure 2 to illustrate how our RBM model hidden units, we can define the following conditional probabilities: exploits the corrections in a multidimensional network to make predic- k k k tion. Suppose that we have known different types of DTIs for target 1 in Prðv ¼ 1 j hÞ¼ ða þ W hÞð4Þ i i ij this example. To know which states of hidden units our RBM model j¼1 should have, we let the three drugs send messages to hidden units and n t XX update their corresponding states. Conversely, once we know the states of k k Prðh ¼ 1 j vÞ¼ ðb þ W vÞð5Þ j j ij i hidden units for target 1, hidden units then send messages to visible units i¼1 k¼1 for the connected drugs and update their corresponding states. As in other collaborative filtering applications (Salakhutdinov et al., 2007), where n, m and t are the numbers of visible units, hidden units and types after a number of iterations of the aforementioned updates, our RBM of DTI encoded in a visible unit, respectively, and ðxÞ¼ 1=ð1 þ e Þ is the logistic function. These conditional probabilities are important for the model can effectively capture the underlying features encoded in a multi- dimensional network. Based on the trained parameters between layers of iterative updates between hidden and visible layers when training an hidden and visible units, our RBM model can then predict unknown RBM model. types of DTIs based on input visible data. Unlike the conditional probabilities of visible variables in In an RBM, suppose that in total there are n visible units, m hidden Salakhutdinov et al. (2007), which used a conditional multinomial distri- units and t types of DTI encoded in a visible unit. Let binary indicator bution, here, our conditional probability of visible variables in Equation 1 k t ¼ðv , ... , v , ... , v Þ, i ¼ 1, ... , n, denote state of the i-th vis- (4) uses a conditional Bernoulli distribution. This is because in our RBM vector v i i i ible unit, where visible variables v ¼ 1if the k-th type of DTI is observed model, more than one visible variable can be assigned one for a DTI. For in input data, and v ¼ 0 otherwise. Let binary indicator variable h , example, a DTI can be annotated with ‘direct’, ‘binding’ and ‘inhibition’ j ¼ 1, ... , m, denote the state of the j-th hidden unit. Let W be the simultaneously. Our RBM model is data-driven and respect data as much ij weight associated with a connection between visible variable v and as possible. As also demonstrated in Salakhutdinov et al.(2007), RBMs hidden variable h.Let v ¼ðv , ... , v Þ and h ¼ðh , ... , h Þ denote the can capture the nature distribution of data. As we treat each interaction j 1 n 1 m configurations of visible and hidden layers, respectively. Then ðv, hÞ is type as a separate binary variable, it is possible that two inconsistent called a joint configuration of an RBM. As there is no intra-layer connec- interaction types (e.g. ‘direct’ and ‘indirect’) can be both yielded by our tion within both hidden and visible layers in an RBM, the energy of a RBM model. This means that the data implies that these two interaction i128 Predicting DTIs using RBMs types are both possible. In addition, by defining each interaction type as a parameters that have been learned in Section 2.2, and p^ is the conditional separate variable, our RBM model is general and can be easily extended probability of hidden variable h given visible data. After that, we com- to include more interaction types from other data sources without much pute the expectation as our final prediction for the query target. manual inspection on the consistence of different types of DTIs. 2.4 Conditional RBMs 2.2 Training For a drug-target pair, which does not have a connection on a multidi- mensional DTI network, it could be considered either a missing inter- We first use a real example to demonstrate the complexity of parameter action or a negative interaction. In general, a set of known DTIs that training in our RBM model. In the MATADOR-based data that we have have been verified experimentally provide more reliable information than tested (Section 3.1), the number of drugs is n ¼ 784. When only predict- those drug-target pairs, which do not have connections on a multidimen- ing direct and indirect interactions, the number of interaction types that sional DTI network. To incorporate this additional information, we use a we need to consider is t ¼ 2. In our RBM model, we typically set the conditional RBM (Salakhutdinov et al., 2007) to further formulate our number of hidden units m ¼ 100. In such a case, the total number of prediction problem. Let r ¼ðr , ... , r Þ be a binary indicator vector, in parameters in our RBM model is n  t þ m þ n  m  t ¼ 158468. In gen- 1 m which r ¼ 1 if there exists a known DTI between the input target and the eral, training an RBM with such a large number of parameters posts a j j-th drug, and r ¼ 0 otherwise. The states of r can be directly obtained difficult task. Later in the text, we will describe how to perform parameter j j based on the visible DTI data from a constructed DTI network. Here, r is training in an RBM model. a binary indicator vector representing the reliability of observed data. In To train an RBM and learn its parameter, we need to maximize the k k general, DTIs that have been verified experimentally provide more reli- likelihood of visible data with respect to the parameters W , a and b . ij i able information than unknown DTIs. A conditional RBM defines a To achieve this goal, we could perform gradient ascent in the log-likeli- joint distribution over ðv, hÞ conditioned on r.InaconditionalRBM, hood of the training data derived from Equation (3) and incrementally the conditional probabilities of visible and hidden units can be written as: adjust the weights and biases: n t m XX X @ log PrðvÞ k k k k k W ¼  ¼  5v h4 5v h4 ð6Þ j data j model Prðh ¼ 1 j v, rÞ¼  b þ W v þ D r , ð12Þ ij i i j j ij i k ij i @W ij i¼1 k¼1 i¼1 where " is the learning rate, 54 denotes an expectation of the data data where D is a parameter describing the effect of r on h. The parameter D ij ij distribution and 54 denotes an expectation of the distribution model can be also learned using the CD algorithm: defined by the model. D ¼  5h 4 5h4 r : ð13Þ ij j data j T i In Equation (6), 54 can be computed easily based on the fre- data quency information obtained from visible data. Unfortunately, it is gen- erally difficult to compute54 , as it would require exponential time model 2.5 Implementation to do so. To avoid this problem, Hinton (2002) proposed a practical learning algorithm, called Contrastive Divergence (CD), which minimizes We chose a conditional RBM to perform DTI prediction on a multidi- the Kullback–Leibler divergence. In this work, we use a mean-field ver- mensional network that encodes different types of DTIs. The algorithm sion of the CD algorithm (Le Roux and Bengio, 2008; Welling and was implemented in Java, using the jaRBM package (http://sourceforge. Hinton, 2002) to train our RBM model. In particular, we use the follow- net/projects/jarbm/). For the value of m, we chose an empirical range ing procedure in each training pass to incrementally adjust the weights according to the literature. Here, we chose m ¼ 100. We set learning and biases: rate  ¼ 0:01. For other parameters, we chose the default values that k k were defined in the jaRBM package. The initial values of W , b , a k k k j ij i W ¼  5v h4 5v h4 ð7Þ j data j T ij i i and D were sampled from Gaussian distribution with standard deriv- ij ation 0.1. In the CD learning procedure, we ran the mean-field updates k k k a ¼  5v 4 5v 4 ð8Þ data T i i i 100 passes over training data. The training of our RBM model runs in several hours on a typical training dataset. For example, the training of b ¼  5h4 5h4 ð9Þ j j data j T our RBM model on the MATADOR-based data (Supplementary Table S1 in Section S1) takes 5.5 h on a PC machine with a 3.3 GHz Intel core i5 where " is the learning rate, 54 denotes an average value over all data processor and 4 GB memory. input data for each update and 54 denotes the average value over T mean-field iterations, which is considered a good approximation of 54 in the log-likelihood function in Equation (6) (Hinton, 2002; model 3 RESULTS AND DISCUSSION Le Roux and Bengio, 2008; Welling and Hinton, 2002). 3.1 Datasets and evaluation metrics 2.3 Making predictions We tested our RBM model on two datasets, which were derived To predict the unknown types of DTIs for a given target with visible data from MATADOR (Gu¨ nther et al., 2008) and STITCH (Kuhn v, we compute the following probability distribution after one iteration of et al., 2012), respectively. MATADOR is a manually curated the mean-field updates (Salakhutdinov et al., 2007): online database, which mainly describes drug-target relation- n t ships, including direct and indirect DTIs. The list of direct and XX k k p^ ¼ Pr h j v ¼ ðb þ W vÞð10Þ j j j ij i indirect DTIs in MATADOR was first extracted by automated i¼1 k¼1 text-mining and then followed by manual annotation (Gu¨ nther et al., 2008). STITCH provides modes of action for the inter- k k k Prðv ¼ 1 j p^ ,   , p^ Þ¼ ða þ W p^ Þ, ð11Þ 1 m j actions between proteins and chemicals, which were annotated i i ij j¼1 based on evidence derived from known experiments in the litera- ture (Kuhn et al., 2012). For the MATADOR-based dataset, where n, m and t are the numbers of visible units, hidden units and types k k of DTI encoded in a visible unit, respectively, W , a and b are we only extracted those DTI records in which protein and ij i i129 Y.Wang and J.Zeng Table 1. Results on predicting direct and indirect DTIs drug names are present and annotation terms correspond to ‘DIRECT’ or ‘INDIRECT’. For the STITCH-based dataset, we only kept a list of DTI records in which drugs and target Drug-target Test method AUC AUPR proteins overlap those in MATADOR. For the mode of action relationship term in the STITCH-based data, we only considered ‘binding’, ‘activation’ and ‘inhibition’. In summary, the MATADOR- Direct interaction Integrating data with distinction 98.7 89.6 based dataset contains 784 drugs, 2431 protein targets and Mixing data without distinction 98.8 72.1 Using direct interactions only 98.0 78.9 13 064 DTIs, in which 7862 interactions are direct and 5202 inter- Indirect interaction Integrating data with distinction 97.1 80.1 actions are indirect. The STITCH-based dataset contains 598 Mixing data without distinction 97.0 37.8 drugs, 671 protein targets and 3296 DTIs, in which 2589, 945 Using indirect interactions only 94.8 62.4 and 1493 interactions are annotated with ‘binding’, ‘activation’ and ‘inhibition’, respectively. Descriptive statistics about these Note: ‘Integrating data with distinction’ corresponds to the test in which our algo- two datasets can be found in Table S1 in Supplementary rithm integrated both direct and indirect interactions with distinction. ‘Mixing data Material Section S1. without distinction’ corresponds to the test in which our algorithm mixed both direct and indirect interactions without distinction. ‘Using direct (indirect) inter- Receiver Operator Characteristic (ROC) and Precision-Recall actions only’ corresponds to the test in which our algorithm used only direct (PR) curves are commonly used to assess the performance of a (indirect) interactions. The highest AUPR score is shown in bold. prediction model. For highly skewed data, ROC curves can give an overoptimistic picture of an algorithm’s performance (Davis and Goadrich, 2006). In this scenario, PR curves provide a better view of the prediction results than ROC curves. As stated in Bleakley and Yamanishi (2009) and van Laarhoven et al. (2011), when evaluating the DTI prediction results, in which there are usually few positive DTIs, PR curves provide greater biological significance and are considered a better quality meas- ure than ROC curves. Therefore, we mainly used PR curves and the AUPR curve to evaluate the performance of our algorithm, though we also reported the area under the ROC curve (AUC) in our test results. 3.2 Predicting direct and indirect DTIs In general, drug-target relationships can be classified into two categories, including direct and indirect interactions. Direct inter- actions are usually caused by protein-ligand binding, whereas indirect interactions can be driven by different mechanisms. For instance, indirect interactions can be induced by metabolites of drugs or changes in gene expressions (Gu¨ nther et al., 2008). To examine whether our algorithm can accurately predict direct and indirect DTIs on a multidimensional network, we tested it on the MATADOR-based data. In particular, we performed the follow- ing three tests: (i) Integrating both direct and indirect DTIs with Fig. 3. PR curves for the direct and indirect DTIs predicted by our RBM distinction, which means that the input data of the RBM is a model. (A) PR curves for the direct DTIs predicted by our model. (B)PR curves for the indirect DTIs predicted by our model multidimensional vector indicating different types of inter- actions; (ii) Mixing both direct and indirect DTIs without dis- tinction, which means that the input data is a one-dimensional with distinction significantly outperformed the other two tests, binary vector indicating whether DTIs are observed; (iii) Using which only considered a single interaction type or simply mixed only a single interaction type (i.e. using only direct or indirect direct and indirect interactions without distinction. In particular, DTIs). For each test, we performed a 10-fold cross-validation the integration of direct and indirect interactions improved the procedure, as described in Supplementary Material Section S3. AUPR score by 410% for direct interaction prediction and The results of the aforementioned three tests are summarized 417% for indirect interaction prediction (Table 1). In PR in Table 1, and their corresponding PR curves are shown in curves, compared with the other two tests, integrating both inter- Figure 3. When integrating both direct and indirect interactions action types with distinction achieved a better precision value with distinction, our algorithm achieved the best performance almost at every recall value (Fig. 3). An interesting result is that with an AUPR score up to 89.6 for direct interaction prediction mixing both interactions without distinction yielded worse per- and 80.1 for indirect interaction prediction. These results demon- strated that our RBM model can effectively integrate different formance than using a single interaction type (Table 1). For ex- drug-target relationships on a multidimensional DTI network to ample, for direct interaction prediction, the AUPR score was accurately predict direct or indirect interactions. In addition, our decreased from 78.9 to 72.1. This result implies that incorporating results showed that integrating both direct and indirect DTIs a new data type into DTI prediction is non-trivial and requires i130 Predicting DTIs using RBMs careful data integration. In addition, our results show that our has been widely used in previous prediction approaches. More RBM model respect the data and can predict a high percentage details of this simple logic-based algorithm can be found of consistent DTIs. For example, when we chose 0.5 as the prob- in Supplementary Material Section S5. As summarized in ability threshold to infer the interaction type,499% of positive Table 2, the comparison results show that our algorithm outper- DTIs predictedbyouralgorithmwereconsistent(i.e.51% of DTIs formed the simple logic based approach. This is expected, as the were predicted as both direct and indirect interactions). simple logic based approach simply uses the closest interaction In our 10-fold cross-validation test (Table 1), the test method type profiles to predict unknown types of DTIs. Such a strategy ‘using direct (indirect) interaction only’ used less training data only exploits a small proportion of training data and cannot be than other two test methods. This may create bias when compar- sufficient enough to capture the deep correlations of different ing two prediction methods using training data with different types of DTIs in a multidimensional network. On the other sizes. To make a fair comparison, we have made an additional hand, our RBM model uses hidden units to represent the intrin- comparison between methods ‘integrating data with distinction’ sic correlations of different types of DTIs in the network and can and ‘using direct (indirect) interaction only’ using training data effectively capture the latent feature of drug-target relationships of the same size (Supplementary Material Section S4). Our new and thus make accurate predictions. comparison results confirmed that integrating data with distinc- Overall, our approach achieved better performance for direct tion outperformed the method that uses a single interaction only. interaction prediction than for indirect interaction prediction To check whether our algorithm can have a wider range of (Table 1). Probably this is because in the MATADOR-based applications, we also performed a 5-fold cross-validation test data, indirect interactions describe many different mechanisms (Supplementary Material Section S3). Compared with the 10- of DTIs and thus provide less predictive power than direct inter- fold cross-validation results, we only found a small decrease in actions. We expect that our method would achieve better per- our algorithm’s performance in the 5-fold cross-validation test. formance for indirect interaction prediction if different To perform more sanity check on our algorithm’s perform- mechanisms of indirect DTIs in the MATADOR-based data ance, we also conducted another test that is similar to leave-one- could be further identified. out cross-validation except that we removed homologous pro- teins from training data. For training data, we only kept those 3.3 Predicting drug modes of action DTIs in which proteins have sequence identity525% to the target In addition to direct and indirect interactions, DTIs can be also of the validating interaction. This process can significantly reduce annotated with different drug modes of action, such as activation the size of dataset. In our test, the training dataset has only 300 or inhibition. To evaluate our algorithm’s performance on pre- targets. This test is more rigorous than previous leave-one-out dicting different drug modes of action, we tested it mainly on the cross-validation (LOOCV) tests performed in (Bleakley and STITCH-based data. We focused on three drug modes of action, Yamanishi, 2009; van Laarhoven et al., 2011; Yamanishi et al., including binding, activation and inhibition. In addition to test- 2008), in that we have removed similar proteins with high sequence ing the STITCH-based data, we included two additional tests, identity from training data and thus reduced impact of protein which further incorporated the MATADOR-based data, that is, homology on the network-based prediction of DTIs. each DTI in the STITCH-based data was also associated with The results of our LOOCV like test show that, our approach direct or indirect interaction derived from the MATADOR- still achieved decent prediction accuracy, with AUPR 79.0 for based data. Overall, we conducted the following five tests: (i) direct interaction prediction, but the AUPR score for indirect integrating both drug-target relationships from the interaction prediction dropped to 59.1. We further investigated MATADOR-based data and drug modes of action from the this problem and found that the average degree of indirect inter- STITCH-based data with distinction; (ii) mixing DTIs from actions with drugs for targets in training data was reduced to both MATADOR-based and STITCH-based data without dis- 1.67. With insufficient number of DTIs, it would be difficult to tinction (i.e. all DTIs were associated with only binary values); train our RBM model and make accurate prediction. (iii) integrating drug modes of action from the STITCH-based To mimic the real situation in which DTI network data are data with distinction; (iv) mixing drug modes of action from the available, we also performed an additional test on those DTIs in STITCH-based data without distinction; and (v) using a singe which the degree of the target is above the average degree of the mode of action from the STITCH-based data. For every test, we multidimensional network (i.e. the degree of the target in the validating interaction is  6). The results of this test show that, our approach achieved decent performance, with AUPR 80.4 for Table 2. Results on comparing our approach with the simple logic based direct interaction prediction and 74.5 for indirect interaction pre- approach diction. These results indicate that in the real situation, our RBM model can make reasonably accurate prediction even homolo- Drug-target Test method AUC AUPR gous proteins are not present in training data. relationship As little work has been developed for predicting unknown types of DTIs in a multidimensional network, it is difficult for Direct interaction Our approach 98.7 89.6 us to directly compare our work with other prediction Simple logic-based approach 92.1 81.6 approaches. Instead, we have compared our algorithm with a Indirect interaction Our approach 97.1 80.1 simple logic-based approach on the MATADOR-based data. Simple logic-based approach 88.4 74.5 This simple logic-based approach follows the basic premise (i.e. similar drugs and targets should have similar interactions) that Note: The highest AUPR score is shown in bold. i131 Y.Wang and J.Zeng also carried out a 10-fold cross-validation procedure as described performed the following test, which is similar to the experiments in Supplementary Material Section S3. conducted in (Bleakley and Yamanishi, 2009; van Laarhoven Both AUC and AUPR scores of the aforementioned five tests et al., 2011). We predicted unknown direct DTIs using the com- are reported in Table 3, and their corresponding PR curves are plete MATADOR-based dataset as training data. For the pre- shown in Figure 4. When integrating different types of DTIs with dicted results, we mainly focused on DTIs involving those distinction, our algorithm can achieve AUC up to 96.9 and proteins that are not listed as drug-metabolizing enzymes or AUPR up to 79.1. The test results showed that our RBM transporter proteins in DrugBank (Knox et al., 2011). We out- model can effectively exploit different types of interactions putted the set of the top 50 scoring predictions that are not pre- encoded in a multidimensional DTI network to predict drug sent in training data. We then checked whether these new modes of action. Among all five tests, integrating different predicted DTIs appear in other databases, including ChEMBL types of DTIs with distinction produced higher AUPR scores (Gaulton et al., 2011), DrugBank (Knox et al., 2011) and than the other tests, which used only a single interaction type STITCH (Kuhn et al., 2012). or simply mixed types of interactions without distinction. When Supplementary Figure S1 in Section S2 visualizes part of the differentiating different types of DTIs, integrating direct and in- DTI network constructed based on the set of the 50 highest direct interactions from the MATADOR-based data slightly im- scoring interactions predicted by our approach using the proved the prediction performance. When comparing the results MATADOR-based data. Table 4 shows the list of the top 10 on predicting different modes of action, binding interactions scoring direct DTI predictions. Among these predicted DTIs, a were predicted with higher accuracy than activation and inhib- high fraction of the predicted interactions (7 of 10) was found in ition interactions. This is probably because in the STITCH-based ChEMBL, DrugBank or STITCH. In the remaining three DTIs data, binding interactions are more informative in predicting that we did not find experimental evidence from the three data- DTIs than other two modes of action. bases, there might still exist interaction for each drug-target pair. As in Section 3.2, we also performed an additional test using For example, although it is known that mifepristone does not training data of the same size for comparing methods ‘integrating bind to estrogen receptor, it may have some effects on the ex- data with distinction’ and ‘using a single data type only’, when pression of estrogen receptor (Jiang et al., 2007). predicting different modes of action. Our new comparison results In Table 4, some of our new predictions are trivial, as they can confirmed that integrating different data types with distinction also be easily derived based on known interactions of similar outperformed the method that uses only a single data type proteins or drugs in the dataset. However, our new predictions (Supplementary Material Section S4). also provided some interesting results. For example, among the top 50 scoring predictions, our algorithm suggested that there 3.4 New predictions may exist interactions between spironolactone and the mem- brane progestin receptor gamma protein and between mesalazine To examine our algorithm’s ability for predicting novel DTIs (also called 5-aminosalicylic acid) and leukotriene-A4 hydrolase that are not present in our data derived from MATADOR, we Table 3. Results on predicting drug modes of action Mode of action Test method AUC AUPR Binding interaction Integrating MATADOR and STITCH with distinction 96.9 79.1 Mixing MATADOR and STITCH without distinction 97.9 53.3 Integrating data with distinction 95.0 78.1 Mixing data without distinction 95.6 68.0 Using binding interactions only 94.1 74.4 Activation interaction Integrating MATADOR and STITCH with distinction 94.4 67.4 Mixing MATADOR and STITCH without distinction 96.9 35.6 Integrating data with distinction 91.2 65.2 Mixing data without distinction 94.2 50.5 Using activation interactions only 87.7 56.3 Inhibition interaction Integrating MATADOR and STITCH with distinction 94.1 67.1 Mixing MATADOR and STITCH without distinction 96.9 38.6 Integrating data with distinction 92.5 65.2 Mixing data without distinction 93.9 44.3 Using inhibition interactions only 89.5 60.2 Note: ‘Integrating MATADOR and STITCH with distinction’ corresponds to the test in which our algorithm integrated both drug-target relationships from the MATADOR- based data and drug modes of action from the STITCH-based data with distinction. ‘Mixing MATADOR and STITCH without distinction’ corresponds to the test in which our algorithm mixed DTIs from both MATADOR-based and STITCH-based data without distinction. ‘Integrating data with distinction’ corresponds to the test in which our algorithm integrated drug modes of action from the STITCH-based data with distinction. ‘Mixing data without distinction’ corresponds to the test in which our algorithm mixed drug modes of action from the STITCH-based data without distinction. ‘Using binding (activation or inhibition) interactions only’ corresponds to the test in which our algorithm used only binding (activation or inhibition) interactions from the STITCH-based data. The highest AUPR score is shown in bold. i132 Predicting DTIs using RBMs Table 4. Top 10 scoring direct DTIs predicted by our approach using the MATADOR data a b Rank Pair Description Evidence 1 DB00812 Phenylbutazone C, D, S P23219 PTGS1: Prostaglandin G/H synthase 1 2 DB04599 Aniracetam S P42261 GRIA1: Glutamate receptor 1 precursor 3 DB00834 Mifepristone P03372 ESR1: Estrogen receptor 4 DB01392 Yohimbine D, S P28222 HTR1B: 5-hydroxytryptamine 1B receptor 5 DB01297 Practolol S P07550 ADRB2: Beta-2 adrenergic receptor 6 DB01297 Practolol P13945 ADRB3: Beta-3 adrenergic receptor 7 DB00334 Olanzapine D, S P08908 HTR1A: 5-hydroxytryptamine 1 A receptor 8 DB01224 Quetiapine D P21918 DRD5: D(1B) dopamine receptor 9 DB01224 Quetiapine D P21728 DRD1: D(1 A) dopamine receptor 10 DB01233 Metoclopramide P21918 DRD5: D(1B) dopamine receptor Fig. 4. PR curves for the predicted drug modes of action. (A)PR curves for the predicted binding interactions. (B) PR curves for the predicted Drugs and targets are represented by DrugBank IDs and UniProt ID, respectively. activation interactions. (C) PR curves for the predicted inhibition DTIs that are observed in ChEMBL, DrugBank and STITCH are marked with interactions ‘C’, ‘D’ and ‘S’, respectively. (LTA4H) enzyme. Spironolactone is known as a diuretic or anti- different types of DTIs, such as drug-target relationships or hypertensive drug and can act on the aldosterone receptor as a drug modes of action. Tests on two public databases showed competitive antagonist (Macdonald, 1997). Our predicted inter- that our algorithm can achieve excellent prediction performance action between spironolactone and membrane progestin receptor with high AUPR scores. Further tests indicated that our ap- gamma protein indicates that spironolactone may have some proach can infer a list of novel DTIs, which is practically progestogenic function. Interestingly, this hypothesis can be con- useful for drug repositioning. firmed from the clinical studies performed in (Nakhjavani et al., Although our algorithm has been tested only on direct and 2009). In addition, mesalazine is an anti-inflammatory drug that indirect drug-target relationships, and three drug modes of is primarily used to treat inflammatory bowel disease (Sandborn action, it is general and can be easily extended to integrate et al., 2007). LTA4H is an important enzyme that converts leu- other types of DTIs (e.g. phenotypic effects). Current version kotriene-A4 to leukotriene-B4 (Rudberg et al., 2004). It has been of our prediction algorithm only considers connections between proposed that leukotriene-B4 may play an important role in a drugs and targets. In the future, we will extend our approach to number of different acute and chronic inflammatory diseases, exploit the connections within target proteins or drugs. For ex- including inflammatory bowel disease (Haeggstro¨ m, 2000). ample, the sequence similarity scores between target proteins, the These studies imply that our predicted interaction between mesa- substructure similarity scores between drugs or drug–drug inter- lazine and LTA4H is probably true. actions (Gottlieb et al., 2012; Tatonetti et al., 2012) can be also These results on new predictions indicated that our RBM incorporated into our prediction model. As the conventional model is practically useful in predicting novel DTIs and can version of an RBM does not allow the connections within the have potential applications in drug repositioning. same layer, such an extension will require careful thought. Currently, our algorithm has been tested only on two databases (i.e. MATADOR and STITCH). We will test it on more data in 4 CONCLUSION the future. For example, it will have more significance if we can In this article, we proposed a first machine-learning approach to predict DTIs based on human proteins and all molecules in predict different types of DTIs on a multidimensional network. PubChem (Kaiser, 2005) or a similar database. Finally, we are Our approach uses an RBM model to effectively encode multiple also seeking wet-laboratory collaborators to experimentally sources of information about DTIs and accurately predict verify the highest scoring DTIs predicted by our algorithm. i133 Y.Wang and J.Zeng Jiang,J. et al. (2007) Effects of mifepristone on expression of estrogen receptor and ACKNOWLEDGEMENTS progesterone receptor in cultured human eutopic and ectopic endometria. The authors thank Dr L. Wang for the helpful discussions in the Zhonghua Fu Chan Ke Za Zhi, 36, 218–221. Kaiser,J. (2005) Science resources. chemists want NIH to curtail database. Science, early stage of this project. They are grateful to Mr M. Zhou for 308,774. helping us prepare the drug-target interaction data. They thank Keiser,M.J. et al. (2007) Relating protein pharmacology by ligand chemistry. Nat. the anonymous reviewers for their helpful comments and Biotechnol., 25, 197–206. suggestions. Keiser,M.J. et al. (2009) Predicting new molecular targets for known drugs. Nature, 462, 175–181. Funding: This work was supported in part by the National Basic Knox,C. et al. (2011) DrugBank 3.0: a comprehensive resource for ‘omics’ research on drugs. Nucleic Acids Res., 39, D1035–D1041. Research Program of China Grant 2011CBA00300, Kroeze,W.K. et al. (2002) Molecular biology of serotonin receptors structure and 2011CBA00301, the National Natural Science Foundation of function at the molecular level. Curr. Top. Med. Chem., 2, 507–528. China Grant 61033001, 61061130540, 61073174. Kuhn,M. et al. (2012) STITCH 3: zooming in on protein-chemical interactions. Nucleic Acids Res., 40, D876–D880. Conflict of Interest: none declared. Larochelle,H. and Bengio,Y. (2008) Classification using discriminative restricted boltzmann machines. In: Proceedings of the 25th international conference on Machine learning, ICML’08. ACM, New York, NY, USA, pp. 536–543. Le Roux,N. and Bengio,Y. (2008) Representational power of restricted boltzmann REFERENCES machines and deep belief networks. Neural Comput., 20, 1631–1649. Lu¨ ,L. and Zhou,T. (2011) Link prediction in complex networks: a survey. Physica A Arel,I. et al. (2010) Deep machine learning-a new frontier in artificial intelligence Stat. Mech. Appl., 390, 1150–1170. research. IEEE Comput. Intell. Mag., 5, 13–18. Macdonald,F. (1997) Dictionary of Pharmacological Agents. CRC Press, Taylor and Bell,R. and Koren,Y. (2007) Lessons from the netflix prize challenge. ACM Francis Group, LLC, Boca Raton, FL, USA. SIGKDD Explor. Newsl., 9,75–79. MacDonald,M.L. et al. (2006) Identifying off-target effects and hidden phenotypes Bengio,Y. (2009) Learning deep architectures for AI. Found. Trends Mach. Learn., of drugs in human cells. Nat. Chem. Biol., 2, 329–337. 2,1–127. Mei,J.P. et al. (2012) Drug-target interaction prediction by learning from local in- Bleakley,K. and Yamanishi,Y. (2009) Supervised prediction of drug-target inter- formation and neighbors. Bioinformatics, 29, 238–245. actions using bipartite local models. Bioinformatics, 25, 2397–2403. Mizutani,S. et al. (2012) Relating drug-protein interaction network with drug side Booth,B. and Zemmel,R. (2004) Prospects for productivity. Nat. Rev. Drug. Discov., effects. Bioinformatics, 28, i522–i528. 3, 451–456. Morris,G.M. et al. (2009) AutoDock4 and AutoDockTools4: automated docking Campillos,M. et al. (2008) Drug target identification using side-effect similarity. with selective receptor flexibility. J. Comput. Chem., 30, 2785–2791. Science, 321, 263–266. Nakhjavani,M. et al. (2009) Short term effects of spironolactone on blood lipid Chen,X. et al. (2012) Drug-target interaction prediction by random walk on the profile: a 3-month study on a cohort of young women with hirsutism. Br. J. heterogeneous network. Mol. Biosyst., 8, 1970–1978. Clin. Pharmacol., 68, 634–637. Cheng,A.C. et al. (2007) Structure-based maximal affinity model predicts small- Perlman,L. et al. (2011) Combining drug and gene similarity measures for drug- molecule druggability. Nat. Biotechnol., 25, 71–75. target elucidation. J. Comput. Biol., 18, 133–145. Cheng,F. et al. (2012) Prediction of drug-target interactions and drug repositioning Roth,B.L. et al. (2004) Magic shotguns versus magic bullets: selectively non-selective via network-based inference. PLoS Comput. Biol., 8, e1002503. drugs for mood disorders and schizophrenia. Nat. Rev. Drug Discov., 3, Davis,J. and Goadrich,M. (2006) The relationship between precision-recall and roc 353–359. curves. In: Proceedings of the 23rd international conference on Machine learning, Rudberg,P.C. et al. (2004) Leukotriene A4 hydrolase: identification of a common ICML’06. ACM, New York, NY, USA, pp. 233–240. carboxylate recognition site for the epoxide hydrolase and aminopeptidase sub- Dimasi,J.A. (2001) New drug development in the united states from 1963 to 1999. strates. J. Biol. Chem., 279, 27376–27382. Clin. Pharmacol. Ther., 69, 286–296. Salakhutdinov,R. et al. (2007) Restricted boltzmann machines for collaborative Donald,B.R. (2011) Algorithms in Structural Molecular Biology. MIT Press, filtering. In: Proceedings of the 24th international conference on Machine learning, Cambridge, MA, USA. ICML’07. ACM, New York, NY, USA, pp. 791–798. Dudley,J.T. et al. (2011) Computational repositioning of the anticonvulsant topir- Sandborn,W.J. et al. (2007) Medical management of mild to moderate crohn’s dis- amate for inflammatory bowel disease. Sci. Transl. Med., 3, 96ra76. ease: evidence-based treatment algorithms for induction and maintenance of Eickholt,J. and Cheng,J. (2012) Predicting protein residue-residue contacts using remission. Aliment Pharmacol. Ther., 26, 987–1003. deep networks and boosting. Bioinformatics, 28, 3066–3072. Sirota,M. et al. (2011) Discovery and preclinical validation of drug indications using Gaulton,A. et al. (2011) ChEMBL: a large-scale bioactivity database for drug dis- compendia of public gene expression data. Sci. Transl. Med., 3, 96ra77. covery. Nucleic Acids Res., 40, D1100–D1107. Tatonetti,N.P. et al. (2012) Data-driven prediction of drug effects and interactions. Gottlieb,A. et al. (2011) PREDICT: a method for inferring novel drug indications Sci. Transl. Med., 4, 125–131. with application to personalized medicine. Mol. Syst. Biol., 7,496. Van Laarhoven,T. et al. (2011) Gaussian interaction profile kernels for predicting Gottlieb,A. et al. (2012) INDI: a computational framework for inferring drug inter- drug-target interaction. Bioinformatics, 27, 3036–3043. actions and their associated recommendations. Mol. Syst. Biol., 8,592. Welling,M. and Hinton,G. (2002) A new learning algorithm for mean field boltz- Gu¨ nther,S. et al. (2008) SuperTarget and Matador: resources for exploring drug- mann machines. Artif. Neural Netw., 2415, 351–357. target relationships. Nucleic Acids Res., 36, D919–D922. Xia,Z. et al. (2010) Semi-supervised drug-protein interaction prediction from het- Haeggstro¨ m,J.Z. (2000) Structure, function, and regulation of leukotriene A4 erogeneous biological spaces. BMC Syst. Biol., 4 (Suppl. 2),S6. hydrolase. Am. J. Respir. Crit. Care. Med., 161,S25–S31. Xie,L. et al. (2012) Novel computational approaches to polypharmacology as a He,Z. et al. (2010) Predicting drug-target interaction networks based on functional means to define responses to individual drugs. Annu. Rev. Pharmacol. groups and biological features. PLoS One, 5,e9603. Toxicol., 52, 361–379. Hinton,G. and Salakhutdinov,R. (2006) Reducing the dimensionality of data with Yabuuchi,H. et al. (2011) Analysis of multiple compound-protein interactions re- neural networks. Science, 28, 504–507. veals novel bioactive molecules. Mol. Syst. Biol., 7,472. Hinton,G.E. (2002) Training products of experts by minimizing contrastive diver- Yamanishi,Y. et al. (2008) Prediction of drug-target interaction networks from the gence. Neural Comput., 14, 1771–1800. integration of chemical and genomic spaces. Bioinformatics, 24, i232–i240. Iorio,F. et al. (2010) Discovery of drug mode of action and drug repositioning from Zhao,S. and Li,S. (2010) Network-based relating pharmacological and genomic transcriptional responses. Proc. Natl Acad. Sci. USA, 107, 14621–14626. spaces for drug target identification. PLoS One, 5,e11764. i134

Journal

Bioinformatics – Pubmed Central

Published: Jun 19, 2013

There are no references for this article.