MGOGP: a gene module-based heuristic algorithm for cancer-related gene prioritization

MGOGP: a gene module-based heuristic algorithm for cancer-related gene prioritization Background: Prioritizing genes according to their associations with a cancer allows researchers to explore genes in more informed ways. By far, Gene-centric or network-centric gene prioritization methods are predominated. Genes and their protein products carry out cellular processes in the context of functional modules. Dysfunctional gene modules have been previously reported to have associations with cancer. However, gene module information has seldom been considered in cancer-related gene prioritization. Results: In this study, we propose a novel method, MGOGP (Module and Gene Ontology-based Gene Prioritization), for cancer-related gene prioritization. Different from other methods, MGOGP ranks genes considering information of both individual genes and their affiliated modules, and utilize Gene Ontology (GO) based fuzzy measure value as well as known cancer-related genes as heuristics. The performance of the proposed method is comprehensively validated by using both breast cancer and prostate cancer datasets, and by comparison with other methods. Results show that MGOGP outperforms other methods, and successfully prioritizes more genes with literature confirmed evidence. Conclusions: This work will aid researchers in the understanding of the genetic architecture of complex diseases, and improve the accuracy of diagnosis and the effectiveness of therapy. Keywords: Gene prioritization, Gene module, Gene ontology, Cancer-related genes Background scores of each annotation of each candidate genes by Discovering cancer-related genes has profound applica- comparing enriched terms in a given set of training genes. tions in modelling, diagnosis, therapeutic intervention, Endeavour [8] prioritizes candidate genes by similarity and in helping researchers get clues on which genes to values between candidate genes and seed genes, by inte- explore [1–3]. Computational approaches are preferred grating more than six types of genomic datasets from over due to their high efficiency and low cost [4, 5]. Many com- a dozen data sources. Methods of the second kind putational methods have been proposed, including: a) prioritize genes using the guilt-by-association principle, gene-based function similarity measure methods [6–9]; b) which means genes interacting with known disease genes biological interaction network-based methods [10–14], are more likely disease-related genes. For instance, PINTA and c) methods based on multiple datasets fusion [15–17]. [10] prioritizes candidate genes by utilizing an underlying Methods of the first kind based on the hypothesis that global protein interaction network. Other methods rank phenotypically similar diseases are caused by functionally candidate genes by exploiting either local or global net- related genes. Based on this hypothesis, many methods work information [2]. Methods of the last kind incorpor- prioritize genes by computing similarity scores between ate datasets such as gene expression, biomedical literature, the candidate genes and the known disease genes. For ex- gene ontology, and PPIs together for gene prioritization. ample, ToppGene [6] ranks genes based on similarity For example, ProphNet [17] integrates information of different types of biological entities in a number of hetero- * Correspondence: lgx1034@163.com; baitian@jlu.edu.cn; 413224445@qq.com geneous data networks. Taking all these methods into con- College of Computer Science and Technology, Jilin University, Changchun sideration, they are either gene-centric or network-centric. 130012, China Full list of author information is available at the end of the article © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Su et al. BMC Bioinformatics (2018) 19:215 Page 2 of 12 However, gene module as a basic functional unit of genes has interaction data and disease associations. Gene Ontology seldom been considered. Consortium describes the functions of specific genes, using Gene module can be defined as a protein complex, a terms known as GO (Gene Ontology). KEGG map genes to pathway, a sub-network of protein interactions. Module pathways while GSEA provides functional gene groups col- detection has long been studied and many useful algo- lected from BioCarta genes sets, KEGG gene sets and rithms have been proposed, such as [18–21]. Although Reactome gene sets. With these annotation information, different methods have different module detection strat- we can easily group genes into functional modules. egies, most of them rely on PPIs network. PPIs network Complex diseases, especially cancer are caused by the suffers from drawbacks as highlighted in [22]. Firstly, the dysfunction of groups of genes and/or gene interactions ra- PPI network is incomplete, which only covers the interac- ther than the mutations of individual genes. Detecting and tions of well-researched proteins. For instance, of the prioritizing cancer-related genes from the perspective of 20,502 genes in the gene expression matrix downloaded gene module is promising. Although some useful work has from The Cancer Genome Atlas (TCGA), only 9078 been conducted [34, 35], the results are still far from being (44.2%) and 2761 (13.4%) genes are included in Human satisfactory. In this study, we take the importance of not Protein Reference Database (HPRD) [23] and Database of only genes but also their affiliated modules into consider- Interacting Proteins (DIP) [24] PPIs networks respectively. ation, and prioritizing genes in a heuristic way. We measure As a result, detected modules are incomplete and their module importance by the number of differential genes accuracy are limited. Secondly, protein interactions in within the module and the number of differential correla- PPIs network suffer from high false positive and negative tions between the module genes. Besides, the number of rates, modules discovered from such PPI data also suffer known cancer-related genes in the module is also consid- from high false rates. All these inherent limitations affect ered. We measure the gene importance by three aspects in- the coverage and accuracy of the inferred modules. formation: a), gene’s differential expression value, b), the Nowadays, numerous public databases of protein and number of differential correlations between the gene and gene annotation information are available, such as Entrez all other module gene. c), the fuzzy measure based similar- Gene [25], Ensembl [26], PIR iProClass [27], GeneCards ity values between the gene and all known cancer-related [28], KEGG [29], Gene Ontology Consortium [30], DAVID genes (if exist) within the module. The global rank of all [31], GSEA [32]and UniProt[33]. For instance, DAVID genes is obtained by utilizing a rank fusion strategy. [31] contains information on over 1.5 million genes from more than 65,000 species, with annotation types, including Methods sequence features, protein domain information, pathway As showninFig.1, MGOGP takes gene expression datasets, maps, enzyme substrates and reaction, protein-protein gene modules, known disease genes and gene ontology Fig. 1 Main components of MGOGP Su et al. BMC Bioinformatics (2018) 19:215 Page 3 of 12 annotation information [36] as input, and the ranked genes we set Se(g ) = 1, which means the gene g is a candidate i i as output. The main parts including: module importance differential expression gene. Se(g ) is defined as follows: measure, module-specific gene importance measure, mod- ule rank and module-specific gene prioritization, and global 0; if padj g > μ Se g ¼ ð1Þ cancer-related gene prioritization. Figure 2 schematically 1; else illustrates these steps in detail. First, obtain functional gene modules; then get the global To further improve the statistical significance of the se- ranking of all modules and the local ranking of all lected candidate differential expression genes, we applied module-specific genes based on their importance; finally, the a multiple random sampling strategy. As defined in Eq. 2. rank fusion algorithm further gives all genes a global rank. > X 0; if Se g < ω Input datasets DEG g ¼ ð2Þ > s¼1 As shown in Fig. 1, MGOGP takes gene expression data- 1; else sets, gene modules, known disease-related genes and gene ontology annotation information as input. In this Where S is the number of sampling; ω is a threshold study, all gene modules are downloaded from GSEA value; if a gene g is selected as a differential expression website (http://software.broadinstitute.org/gsea/down- gene we set DEG(g ) = 1, Otherwise, we set DEG(g )=0. i i loads.jsp). All GO ontologies of genes are downloaded We define Ncr(m ) as the ratio of differential expres- from GeneCards [37, 38]. Information of relationships sion genes in the module m as shown in Eq. 3: between GO terms are got from Gene Ontology Consor- tium website. P DEG g i¼1 Ncr m ¼ ð3Þ j∈1; 2; 3; …; M Module importance measure We measure the importance of a module by: the number of differentially expressed genes in the module, the num- Where, g is the ith gene in the module m ; N is the i j ber of differential correlations between module genes total number of genes in the module m ; DEG(g ) is de- j i and the basic importance of the module itself. fined in Eq. 2. We use DESeq2 for gene differential expression ana- Next, for each pair of genes in the module m , two correl- lysis [3, 35, 39, 40]. If genes with padj(g ) value bigger ation values are calculated using normal and tumor samples than the threshold value μ, we set Se(g ) = 0. Otherwise, respectively. As defined in Eqs. 4 and 5 respectively. Fig. 2 MGOGP processes are illustrated. a Obtain gene modules, b Module importance measure and prioritization, c Module-specific gene importance measure and prioritization, d Compute global gene ranking Su et al. BMC Bioinformatics (2018) 19:215 Page 4 of 12 We define Ecr(m ) as the ratio of differential correla- ðÞ x −xðÞ y −y l¼1 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi r g ; g ¼ ð4Þ N tions among genes in the module m . Ecr(m ) is defined i h j j L 2 2 ðÞ x −x ðÞ y −y l in Eq. 10: l¼1 l r (g , g ) is the Pearson correlation value between gene N i h DEE g ; g g and gene g across all normal samples. L is the normal k¼1 i h i h Ecr m ¼ sample number. ð10Þ NNðÞ −1 K ¼ and i; h∈1; 2; 3; …; N x −x y −y q¼1 q rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi r g ; g ¼ ð5Þ i h x −x y −x K and N is the edge number and the gene number of q¼1 q the module m respectively. We measure the basic importance of a module by cal- r (g , g ) is the Pearson correlation value between gene T i h culating the ratio of known disease genes in a module, g and gene g across all tumor samples. Q is the tumor i h as shown in Eq. 11: sample number. To test whether the correlation coefficient between gene g and gene g is differentially correlated, we test i h info m ¼ num d þ 1 =N ð11Þ j j whether r (g , g ) and r (g , g ) are significantly different. T i h N i h The two correlation coefficients are changed to Z (g , g ) N i h and Z (g , g ) respectively. num(d ) is the number of known disease genes in the T i h j module m ; N is the number of genes in the module m . j j 1 þ r g ; g i h The module importance is defined in Eq. 12. Z g ; g ¼ log ð6Þ i h 1−r g ; g i h pm ¼ Ncr m þ Ecr m =2  info m Similarly, r (g , g ) is changed to Z (g , g ) as Eq. (6). j j j j T i h T i h j∈1; 2; 3…; M The differential correlation is tested based on Fisher’s z-test [41]. As defined in Eq. (7): ð12Þ Z g ; g −Z g ; g N T i h i h Z ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð7Þ where m means the jth module; M is the total number 1 1 of modules. L−3 Q−3 The Z value has an approximately Gaussian distribu- Module-specific gene importance measure tion under the null hypothesis [41]. If the fdr value of a We measure the importance of a gene (p(g )) in the gene is bigger than the threshold value υ, we set Sc(g , module by measuring: the gene’s differential expression g ) = 0, otherwise we set Sc(g , g ) = 1, which means the h i h value, the number of differential correlations between correlation coefficient is a potential differential correl- the gene and all other module genes and the basic im- ation. Sc(g , g ) is defined as follows: i h portance of the gene itself. The number of differential correlations (CorC(g )) be- 0; if fdr g ; g > υ i h Sc g ; g ¼ ð8Þ i h tween the gene g and all other genes in the same mod- 1; else ule is calculated as in Eq. 13. Where fdr(g , g ) is the local false-discovery rate (fdr) i h derived from fdrtool package [42]; υ is a threshold value. P N−1 Sc g ; g i h h¼1;h≠i As the way we find differential expression genes, we CorC g ¼ N−1 retain only those significantly changed correlations. As ð13Þ i; h∈1; 2; 3; …; N; g ∈m defined in Eq. 9: j∈1; 2; 3; …; M > X 0; if Sc g ; g < δ i h DEE g ; g ¼ ð9Þ N is the number of genes in the module m ; M is the s j i h > s¼1 total module number. 1; else Finally, the basic importance of a gene is determined Where S is the number of sampling; δ is a threshold by the gene ontology-based fuzzy measure similarity value; we set DEE(g , g ) = 1 if the gene g and g are dif- values between the gene and all known disease gene (if i h i h ferentially correlated. Otherwise, we set DEE(g , g )= 0. exist) in the same module. As shown in Eq. 14. i h Su et al. BMC Bioinformatics (2018) 19:215 Page 5 of 12 pTðÞ¼ info m g ¼ countðÞ T þ children of T in corpus k k 0; if num m d ¼ 0 > j h ð18Þ countðÞ all GOtermsincorpus 1; if g isaknown disease gene itself X 1≤k ≤ j T j >   g num m d ðÞ j h i S g ; m d =num m d ; else FMS j h j h h¼1 The importance of gene (p(g )) in a module is defined ð14Þ in Eq. 19. num(m _d ) is the number of known disease genes j h pg ¼ padj g þ CorC g þ info g in the module m.If num(m _d ) = 0, which means no i i i i j j h ð19Þ i∈1; 2; 3; …; N; g ∈m known diseasegeneinthe module m,weset j info(m _g )= 0. If g itself is a known disease gene, we j i i set info(m _g ) = 1. Otherwise, we calculate the gene N is the number of genes in the module m . j i j importance value based on the fuzzy similarity meas- ure between the gene and all the known disease gene Global gene ranking in the module m . S (m _g ,m _d ) is defined in Eq. Most genes deploy their functions in the context of j FMS j i j h 15,asin[43]: sophisticated functional modules [45, 46]. Therefore, the global rank of a gene need be decided by its own importance and the importance of its affiliated Sm T ∩T þ Sm T ∩T i m g m d h m g m d j j h j j h i i S m g ; m d ¼ FMS j j h module. As in [34], a rank fusion strategy is used to fuse the local rank of genes in each module into a ð15Þ global rank. The rank fusion strategy is a recursive process. It Where Sm is the Sugeno measure [43] defined on GO decides the rank of the nth gene based on all the terms of gene m _g and Sm is the Sugeno measure j i h top-ranked n − 1 genes. We define i as the number of defined on GO terms of module disease gene m _d . j h genes having already obtained their global ranking in Let T is the set of GO annotation terms of gene m g the recursive process of rank fusion, m(i, j)asthe m _g , Sm , is a real value function, satisfying [44]: j i i number of top i genes located in the module j after having determined the top i genes. t(i, j)asthe ex- 1) Sm ðT Þ¼ 0; if T ¼ ∅; elseSm ðT Þ¼ 1: i m g m g i m g j i j i j i pectation of the number of top i genes located in the 2) Sm ðT Þ≤SmðT Þif T ⊆T i m g m d m g m d j i j h j i j h module j. e(i, j) as the expectation of probability that 3) For all T ; T ⊆T with T ∩ T = Φ A B m g A B j i the i + 1 globally ranked genes come from the module Sm ðT ∪T Þ¼ Sm ðT Þþ Sm ðT Þ i A B i A i B j.Weuse themoduleimportance value p(m)asthe þλSm ðT ÞSm ðT Þ; λ > −1 i A i B probability of a disease-related gene comes from it. The relationship between i, m(i, j), t(i, j)and p(m)is For a given gene annotation set T , the parameter λ m g showninEq. 20: of its Sugeno fuzzy measure can be uniquely solved as in Eq. 16: tiðÞ ; j ¼ ip m ð20Þ eiðÞ ; j ¼ tiðÞ þ 1; j −miðÞ ; j ðÞ 1 þ λ¼ ðÞ 1 þ λSm ð16Þ Initially, the first ranked gene in the module with high- i¼1 est importance value is chosen as the top 1 gene in the gene’s global rank, because all genes in each module This equation has a unique solution for λ > −1. Let have been ranked from big to small according to their Sm = Sm({T }). The mapping T → Sm is called a fuzzy k k k k importance value. Let i as the number of genes having density function. The fuzzy density value, Sm , is inter- obtained their global ranking, to decide the i + 1 ranked preted as the importance of the single information gene, we need to find the module with the biggest e(i, j) source T in determining the similarity of two genes. As value, because e(i, j) indicates the expectation of prob- defined in Eq. 17: ability that the i + 1 globally ranked genes from module j. So the genes ranked m(i, j) + 1 in the module j will be chosen as the top i + 1 ranked gene, because in the mod- Sm ¼ − ln pTðÞ= max − ln pT ð17Þ k k j T ∈T j g ule j, top m(i, j) genes has obtained the global ranking. Repeat the process until all genes get ranked. As shown Where p(T ) is defined in Eq. 18: in Fig. 3 (in Additional file 1). k Su et al. BMC Bioinformatics (2018) 19:215 Page 6 of 12 Fig. 3 Rank fusion process. N is the number of genes in the module j, M is the total module number As shown in Table 2, all the six genes are ranked on Results Both raw count and normalized gene expression data- average within top10% of all the candidate genes, which sets are downloaded from TCGA (http://cancergen- indicates the superiority of MGOGP to other three algo- ome.nih.gov/)[47], which include expression values of rithms. For further comparison, we put these 21 genes 20,503 genes across 102 normal samples and 779 together, each time we randomly select 20 different tumor samples. Besides, gene expression datasets of genes as known disease genes and the remaining 1 gene Prostate adenocarcinoma containing 483 tumor sam- ples and 51 normal samples are also downloaded Table 1 Known prostate cancer genes retrieved from the OMIM from TCGA. Four thousand seven hundred twenty-six Gene Gene Gene name gene modules are downloaded from the website of ID Symbol GSEA (in Additional file 2). 367 AR Androgen receptor Firstly, the performance of MGOGP is validated 675 BRCA2 Breast cancer type 2 susceptibility protein by comparing it with three module based 3732 CD82 CD82 antigen cancer-related gene prioritization methods (MEN- DEAVOUR, MDK and MRWR) proposed in [34]. 11200 CHEK2 Serine/threonine-protein kinase Chk2 For comparison, the same prostate cancer network 60528 ELAC2 Zinc phosphodiesterase ELAC protein 2 used in [34] are used, which consists of 233 genes 2048 EPHB2 Ephrin type-B receptor 2 precursor and 1218 interactions. Modules are obtained by 3092 HIP1 Huntingtin-interacting protein 1 picking out all the GSEA modules that contain 1316 KLF6 Krueppel-like factor 6 more than three genes in the prostate network after 8379 MAD1L1 Mitotic spindle assembly checkpoint removing irrelevant module genes. Irrelevant genes proteinMAD are genes that are included in GSEA modules but 4481 MSR1 Macrophage scavenger receptor types I and II are not included in these 233 genes. Fifteen known prostate cancer genes are obtained from OMIM 4601 MXI1 MAX-interacting protein 1 (Table 1). Six genes (BRCA1, TP53, EP300, STAT3, 7834 PCAP Predisposing for prostate cancer ZFHX3, HNF1B), which are confirmed have associa- 5728 PTEN Phosphatase and tensin homolog tions with prostate cancer by Genetics Home Refer- 6041 RNASEL 2-5A-dependent ribonuclease ence (https://ghr.nlm.nih.gov/)are used as test 5513 HPC1 Hereditary prostate cancer 1 genes. Results are shown in Table 2. Su et al. BMC Bioinformatics (2018) 19:215 Page 7 of 12 Table 2 Ranks of six test genes in prostate cancer gene Next, we use MGOGP for genome-wide breast cancer network. They are prioritized by MDK, MRWR, Endeavour and gene prioritization. We use 328 breast disease-related MGOGP genes downloaded from SNP4Disease (http://snp4disea- Gene MDK MRWR Endeavour MGOGP se.mpibn.mpg.de/result.php) as seed genes (see BRCA1 29 6 58 63 Additional file 3). Ten well-known breast cancer-related genes (shown in Table 4, which are not contained in the TP53 104 132 85 24 328 genes) are used to validate the effectiveness of our EP300 83 70 90 11 method. All GSEA gene modules are pre-processed by STAT3 39 41 88 17 removing all the genes which do not have gene expres- ZFHX3 174 174 34 19 sion information (the final module list is supplied in HNF1B 44 190 109 26 Additional file 4). The result is shown in Fig. 4. Average Rank 78 102 77 26 As showninFig. 4, all the 10 breast cancer-related genes are ranked within the top5% of the gene prioritization results. During the process, we set S = 1000, ω=0.9 and δ=0.9 (which means of the 1000 sampling results, over 90% fulfill for test. Each run we compared the ranked positions of the filter criteria). We set υ =0.05 and μ =0.01 as most the 1 test gene between our method and Endeavour. others do [39, 41]. The performance of MGOGP under Results are shown in Table 3. In Table 3 some genes do different parameter settings are supplied in Additional file 5. not exist, because they don’t exist in our GSEA gene Thetop 10 ranked modulesinthiscasestudy areshown in modules or not exist in Endeavour database. According Table 5. to Table 3, 11 of the 13 known prostate cancer-related As can be seen from Table 5,many top-ranked genes and 4 of the 6 test genes have much higher ranks modules are included in well-known breast cancer path- than these of the Endeavour. Moreover, the average ways, such as PI3K/AKT [48]pathway andVEGF ranking of these genes is 51 by MGOGP, which is better ligand-receptor pathway. The VEGF family of ligands than 82 by Endeavour. and receptors are intimately involved in tumor angio- genesis, lymphangiogenesis, and metastasis [49]. More importantly, of the 100 genes in the top 10 ranked modules, 20 of them are contained in the KEGG breast Table 3 Ranks of each validation gene cancer pathway (hsa05224), which is an indication of Gene MGOGP Endeavour the good performance of MGOGP for cancer gene AR 32 30 prioritization. BRCA2 29 40 Next, we validate the performance of MGOGP by com- CD82 169 211 paring the gene prioritization results with results obtained CHEK2 19 35 by methods: Endeavour [8], GeneFriends [50], PINTA ELAC2 64 176 [10], TOPPGene [6] and TOPNet [13]. All the methods use the same datasets and under their default parameter EPHB2 45 165 HIP1 91 111 Table 4 Ten well-known breast cancer genes KLF6 88 72 Gene Gene symbol Gene name MAD1L1 78 194 ID MSR1 60 190 672 BRCA1 Breast Cancer 1, Early Onset MXI1 92 89 675 BRCA2 Breast Cancer 2, Early Onset PCAP Not Exist Not Exist 7157 TP53 Tumor Protein P53 PTEN 24 94 5728 PTEN Phosphatase And Tensin Homolog RNASEL 67 83 841 CASP8 Caspase 8, Apoptosis-Related Cysteine Peptidase HPC1 Not Exist Not Exist 2263 FGFR2 Fibroblast Growth Factor Receptor 2 BRCA1 46 16 4214 MAP3K1 Mitogen-Activated Protein Kinase Kinase TP53 5 5 Kinase 1, E3 Ubiquitin Protein Ligas EP300 11 12 11200 CHEK2 Checkpoint Kinase 2 STAT3 17 23 472 ATM ATM Serine/Threonine Kinase ZFHX3 59 68 83990 BRIP1 BRCA1 Interacting Protein C-Terminal HNF1B 26 12 Helicase 1 Su et al. BMC Bioinformatics (2018) 19:215 Page 8 of 12 Fig. 4 Known cancer-related gene prioritization result settings. The results are shown in Fig. 5. Brief descriptions others are left for test (each kind of selection repeat of these methods are provided in Additional file 6. Core 100 times). We count the average number of test sourcecode of MGOGP is provided in Additional file 7. genes appear in Top 200 gene prioritization results. Other source codes are available from the corresponding Results are shown in Fig. 6. author on reasonable request. Finally, to further validate our method, we get the top In Fig. 5, we count the number of breast 10 ranked genes of each method in Fig. 5. The results cancer-related genes in the gene prioritization results. are shown in Table 6. As is shown in Fig. 5, MGOGP outperforms other In Fig. 7, the number of Known Disease Gene is the methods in detecting cancer-related genes. We use all number of genes supplied for training each method that the 328 breast disease related genes as known disease fall within the top 10. For example, in Table 6, PTEN, gene (Endeavour and GeneFriends used the same gene VEGFB, and MCM2 are three genes fall within the top sets) and count the number of known disease genes ap- 10 of the gene ranking result, so the number of Known pear in top 100–1000 prioritization results. Disease Gene of MGOGP in Fig. 7 is 3. For each gene To do comparison more rigorously, we further com- within the top 10 gene ranking results of each method, pare MGOGP to Endeavour, TOPNet and TOPPGene. we search the number of articles in PubMed mention Each time we randomly select 100, 150 and 200 dif- the association between the gene and breast cancer. We ferent known disease genes from the 328 breast count the number of genes has more than 10 PubMed disease-related genes for known disease genes and article reference. As shown in Fig. 7, genes detected by MGOGP have more article supports than other methods. Table 5 Top 10 ranked modules Rank Module name Gene number Importance value Discussion and conclusion 1 zerbini_response_to_sulindac_dn 6 0.542 Results of omics experiments commonly consist of 2 reichert_g1s_regulators_as_pi3k_ 8 0.523 a large set of genes, while researchers usually only targets care about the behaviour of several genes. In this 3 sa_g2_and_m_phases 8 0.492 paper, a heuristic algorithm is proposed for priori- tizing disease-associated genes by utilizing gene 4 reactome_vegf_ligand_receptor_ 10 0.478 interactions ontology annotation information and known 5 biocarta_srcrptp_pathway 11 0.461 disease-related genes as heuristic information. Dif- ferent from existing methods, we propose to rank 6 honrado_breast_cancer_brca1_ 18 0.447 vs_brca2 genes considering the importance of both individual genes and their affiliated modules, and utilize Gene 7 tcga_glioblastoma_mutated 8 0.445 Ontology (GO) based fuzzy measure value as well 8 pid_vegf_vegfr_pathway 10 0.444 as known disease genes as heuristics, and use rank 9 liang_silenced_by_methylation_dn 11 0.411 fusion strategy to obtain the global gene 10 agarwal_akt_pathway_targets 10 0.410 prioritization. Results show that MGOGP Su et al. BMC Bioinformatics (2018) 19:215 Page 9 of 12 Fig. 5 Comparison results between 6 methods. Endeavour, GeneFriends, PINTA, TOPPGene, TOppNet, and MGOGP outperforms many other methods in cancer-related number of interactions among well-researched genes gene prioritization. maybemuchmorethanthatofnewly discovered Different from other module-based gene prioritization genes). methods, where modules are detected by partitioning Different from module-based methods in [34], the network using the network clustering methods, we MGOGP ranks modules considering three aspects of obtain modules through gene function annotation, that information: module-specific gene importance, differ- is, functionally related genes are grouped into the same ential correlations, and importance of the module it- modules. Because gene interaction networks often suffer self. In [34], the author considers the importance of a from the problems of high rates of false positive/negative module by considering only the number of disease interactions, and modules detected by network cluster- genes and the size of the module, which may bias to- ing algorithms often have limited accuracy, so our ward big modules. Furthermore, gene as the major method is more advanced. One important difference component of the module whose importance is not between modules used in this study and modules de- considered when measuring the importance of a mod- tected through network partition is that no edges in ule in [34]. While in our method, when measuring our module. Instead, we use statistical methods de- the importance of a module, we consider: the import- tecting differential correlationsbetween geneswithin ance of the module itself, the importance of module a module, which could help avoid the preference of contained genes as well as differential correlations genes or modules that are well-researched (because within the module, which are the main improvements currently obtained network is far from complete, the of our method. Fig. 6 Comparison results between MGOGP, Endeavour, TOPPGene, and TOPNet with different number of known disease genes as input Su et al. BMC Bioinformatics (2018) 19:215 Page 10 of 12 Table 6 Top 10 ranked genes of each method MGOGP Endeavor GeneFriends PINTA ToppGene ToppNet Top 10 gene CCNB1IP1 SNRPF LURAP1L MGP RAD51 APP CCNE2 BUB3 PVRL2 EEF1A1 APEX1 ELAVL1 NEK1 MSH2 CYFIP1 TPT1 SIRT2 NTRK1 NRP1 SSBP1 FAM120A RPS6 NOC2L RPA1 CDC25C RFC4 IL13RA1 RPL3 NEDD1 XPO1 VIM EZH2 MYO1B RPS27 TERT EED PTEN CENPF BCL9L ACTB EPN3 CUL3 VEGFB BLMH NQO1 SCGB2A2 PPARGC1A BARD1 MCM2 KIF20B RIN2 RPL11 NBN HSP90AA1 PTGS2 BAZ1A SDC4 PIP ATR NXF1 Known disease genes PTEN MSH2 NQO1 SCGB2A2 RAD5 BARD1 fall in the top 10 gene VEGFB EZH2 PIP TERT MCM2 NBN ATR In Table 6, each method is run with default parameter settings and use same training genes. Top 10 gene means the top 10 genes prioritized by each method and Known disease genes fall in the top 10 gene means genes supplied for training each method falls in the top 10 genes. Detail statistic results are shownin Fig. 7 Compared with other non-module-based prioritization top-ranked modules are included in well-known can- methods, our algorithm also has obvious advantages. cer pathways, and top-ranked genes have more sup- First, it is easier to find the potential pathogenic genes porting PubMed articles. All of the results show that that cause the disease from the point of view of gene our methods perform better than the state of the art modules. Second, it takes cross-validation strategy which methods. could guarantee the stability of the recognition re- Prioritization methods are useful for assisting sci- sults. And our method works with heuristic informa- entists at early research stages, and to formulate tion which could effectively avoid the blindness of novel hypotheses of interest. In the future, one of the search. our main goals is to see how our method behaves in By applying MGOGP on different datasets, we dem- other prioritization problems when using different onstrate that MGOGP performs better than previous entities and sources of data sets not covered in this gene or network-centric methods in terms of poten- study. Furthermore, we plan to study in more detail tial disease-related genes prediction. Firstly, the per- the quality of the datasets and their influence on al- formance of MGOGP is validated by comparing it gorithm performance, and design new methods to with three module based cancer-related gene try to improve the results. As we all know that the prioritization methods. Results show that all test methods become more mature the results will be- genes are ranked on average within top10% of all the come increasingly accurate and more biologically candidate genes. According to our results, many meaningful. Fig. 7 Detail statistic results of results in Table 6 Su et al. BMC Bioinformatics (2018) 19:215 Page 11 of 12 Additional files Received: 13 March 2017 Accepted: 23 May 2018 Additional file 1: A step by step example of Rank Fusion process. This file provides an example of how to get the final gene rank. (DOCX 275 kb) References 1. Gill N, Singh S, Aseri TC. Computational disease gene prioritization: an Additional file 2: GSEA gene module. This file is all the gene modules appraisal. J Comput Biol. 2014;21(6):456–65. downloaded from GSEA website. (TXT 2837 kb) 2. Moreau Y, Tranchevent LC. Computational tools for prioritizing candidate Additional file 3: Breast-Cancer-Gene. This is the known breast cancer- genes: boosting disease gene discovery. Nat Rev Genet. 2012;13(8):523–36. related genes downloaded from SNP4Disease. (TXT 2 kb) 3. Cruz-Monteagudo M, Borges F, Paz YMC, Cordeiro MN, Rebelo I, Perez- Additional file 4: Final module list. This is the refined module list after Castillo Y, Helguera AM, Sanchez-Rodriguez A, Tejera E. Efficient and removing irrelevant genes. (TXT 2736 kb) biologically relevant consensus strategy for Parkinson’s disease gene prioritization. BMC Med Genet. 2016;9:12. Additional file 5: Parameters discussion. This file discusses the 4. Bromberg Y. Chapter 15: disease gene prioritization. PLoS Comput Biol. performance of MGOGP under different parameter settings. (DOCX 65 kb) 2013;9(4):e1002902. https://doi.org/10.1371/journal.pcbi.1002902. Additional file 6: Brief description of gene prioritization methods. This 5. Doncheva NT, Kacprowski T, Albrecht M. Recent approaches to the file provides the short description of comparison methods, including prioritization of candidate disease genes. Wiley Interdiscip Rev Syst Biol their input datasets, limitations, and type. (DOCX 17 kb) Med. 2012;4(5):429–42. Additional file 7: Sourcecode. Some core code of our method. (TXT 5 kb) 6. Chen J, Bardes EE, Aronow BJ, Jegga AG. ToppGene Suite for gene list enrichment analysis and candidate gene prioritization. Nucleic Acids Res. 2009;37:W305–11. Acknowledgments 7. Schlicker A, Lengauer T, Albrecht M. Improving disease gene prioritization The results here are in whole or part based upon datasets generated by the using the semantic similarity of Gene Ontology terms. Bioinformatics. 2010; TCGA Research Network: http://cancergenome.nih.gov/. 26(18):i561–7. 8. Tranchevent LC, Barriot R, Yu S, Van Vooren S, Van Loo P, Coessens B, De Funding Moor B, Aerts S, Moreau Y. ENDEAVOUR update: a web resource for gene This work is supported by Graduate Innovation Fund of Jilin University (No. prioritization in multiple species. Nucleic Acids Res. 2008;36:W377–84. 2016031); The National Nature Science Foundation of China (No. 61373051, 9. Yu W, Wulf A, Liu T, Khoury MJ, Gwinn M. Gene Prospector: an evidence No. 61502343, No. 61772226 and No. 61702214); Science and Technology gateway for evaluating potential susceptibility genes and interacting risk Development Program of Jilin Province (No. 20140204004GX); The Science factors for human diseases. BMC Bioinformatics. 2008;9:528. Research Funds for the Guangxi Universities (No. KY2015ZD122); The Science 10. Nitsch D, Tranchevent LC, Goncalves JP, Vogt JK, Madeira SC, Moreau Y. Research Funds for the Wuzhou University (2014A002); Project of Science PINTA: a web server for network-based gene prioritization from expression and Technology Innovation Platform of Computing and Software Science data. Nucleic Acids Res. 2011;39(Web Server issue):W334–8. (985 Engineering); The Key Laboratory for Symbol Computation and 11. Xie B, Agam G, Balasubramanian S, Xu J, Gilliam TC, Maltsev N, Bornigen D. Knowledge Engineering of the National Education Ministry of China; The Disease gene prioritization using network and feature. J Comput Biol. 2015; Fundamental Research Funds for the Central. China Postdoctoral Science 22(4):313–23. Foundation (No. 2014M561293); Development Project of Jilin Province of 12. Navlakha S, Kingsford C. The power of protein interaction networks for China (No. 20150520064JH). associating genes with diseases. Bioinformatics. 2010;26(8):1057–63. 13. Chen J, Aronow BJ, Jegga AG. Disease candidate gene identification and prioritization using protein interaction networks. BMC Availability of data and materials Bioinformatics. 2009;10:73. The datasets used and/or analyzed during the current study are available 14. Erten S, Bebek G, Ewing RM, Koyuturk M. DADA: degree-aware algorithms from the corresponding author on reasonable request. for network-based disease gene prioritization. BioData mining. 2011;4(19). https://doi.org/10.1186/1756-0381-4-19. Authors’ contributions 15. Wu C, Zhu J, Zhang X. Integrating gene expression and protein-protein LS made contributions to method design and data analysis, and a major interaction network to prioritize cancer-associated genes. BMC contributor in writing the manuscript. GL involved in drafting the manuscript Bioinformatics. 2012;13:182. and revision. TB analyzed the results and made contributions to method 16. Simoes SN, Martins DC Jr, Pereira CA, Hashimoto RF, Brentani H. NERI: network- implementation. XM performed comparative analysis. QM made medicine based integrative approach for disease gene prioritization by relative contributions to results interpretation and also involved in data acquisition importance. BMC Bioinformatics. 2015;16(Suppl 19):S9. and manuscript writing. All authors read and approved the final manuscript. 17. Martínez V, Cano C, Blanco A. ProphNet: a generic prioritization method through propagation of information. BMC Bioinformatics. 2014;15(Suppl 1): Ethics approval and consent to participate S5. doi:https://doi.org/10.1186/1471-2105-15-S1-S5. In this study, all gene expression datasets were downloaded from TCGA 18. Zhang Y, Lin H, Yang Z, Wang J. Integrating experimental and literature database (https://tcga-data.nci.nih.gov/tcga/). There are no restrictions on the protein-protein interaction data for protein complex prediction. BMC use of TCGA data for research and data analysis purposes. All datasets can Genomics. 2015;16(Suppl 2):S4. be downloaded and used freely, and not require an ethics statement. 19. Srihari S, Yong CH, Patil A, Wong L. Methods for protein complex prediction and their contributions towards understanding the organisation, function Competing interests and dynamics of complexes. FEBS Lett. 2015;589(19 Pt A):2590–602. The authors declare that they have no competing interests. 20. Su L, Liu G, Wang H, Tian Y, Zhou Z, Han L, Yan L. GECluster: a novel protein complex prediction method. Biotechnol Biotechnol Equip. 2014;28(4):753–61. 21. Bader GD, Hogue CW. An automated method for finding molecular complexes Publisher’sNote in large protein interaction networks. BMC Bioinformatics. 2003;4:2. Springer Nature remains neutral with regard to jurisdictional claims in 22. Ramaprasad A, Pain A, Ravasi T. Defining the protein interaction network of published maps and institutional affiliations. human malaria parasite plasmodium falciparum. Genomics. 2012;99(2):69–75. 23. Keshava Prasad TS, Goel R, Kandasamy K, Keerthikumar S, Kumar S, Author details Mathivanan S, Telikicherla D, Raju R, Shafreen B, Venugopal A, et al. Human College of Computer Science and Technology, Jilin University, Changchun protein reference database–2009 update. Nucleic Acids Res. 2009; 130012, China. Key Laboratory of Symbolic Computation and Knowledge 37(Database):D767–72. Engineering of Ministry of Education, Jilin University, Changchun 130012, 24. Xenarios I, Salwinski L, Duan XJ, Higney P, Kim SM, Eisenberg D. DIP, the China. The First Clinical Hospital of Jilin University, Changchun 130021, Database of Interacting Proteins: a research tool for studying cellular China. networks of protein interactions. Nucleic Acids Res. 2002;30(1):303–5. Su et al. BMC Bioinformatics (2018) 19:215 Page 12 of 12 25. Maglott D, Ostell J, Pruitt KD, Tatusova T. Entrez gene: gene-centered information at NCBI. Nucleic Acids Res. 2011;39:D52–7. 26. Flicek P, Amode MR, Barrell D, Beal K, Brent S, Chen Y, Clapham P, Coates G, Fairley S, Fitzgerald S, et al. Ensembl 2011. Nucleic Acids Res. 2011; 39(Database):D800–6. 27. Wu CH, Huang H, Nikolskaya A, Hu Z, Barker WC. The iProClass integrated database for protein functional analysis. Comput Biol Chem. 2004;28(1):87–96. 28. Rebhan M, Chalifa-Caspi V, Prilusky J, Lancet D. GeneCards: integrating information about genes, proteins and diseases. Trends Genet. 1997;13(4):163. 29. Kanehisa M, Sato Y, Kawashima M, Furumichi M, Tanabe M. KEGG as a reference resource for gene and protein annotation. Nucleic Acids Res. 2016;44(D1):D457–62. 30. Gene Ontology C. Gene ontology consortium: going forward. Nucleic Acids Res. 2015;43(Database issue):D1049–56. 31. Huang DW, Sherman BT, Tan Q, Kir J, Liu D, Bryant D, Guo Y, Stephens R, Baseler MW, Lane HC, et al. DAVID Bioinformatics Resources: expanded annotation database and novel algorithms to better extract biology from large gene lists. Nucleic Acids Res. 2007;35(Web Server issue):W169–75. 32. Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci U S A. 2005;102(43):15545–50. 33. UniProt C. UniProt: a hub for protein information. Nucleic Acids Res. 2015; 43(Database issue):D204–12. 34. Chen X, Yan GY, Liao XP. A novel candidate disease genes prioritization method based on module partition and rank fusion. OMICS. 2010;14(4):337–56. 35. Liu X, Liu ZP, Zhao XM, Chen L. Identifying disease genes and module biomarkers by differential interactions. J Am Med Inform Assoc. 2012; 19(2):241–8. 36. Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, et al. Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet. 2000;25(1):25–9. 37. Belinky F, Nativ N, Stelzer G, et al. PathCards: multi-source consolidation of human biological pathways. Database: J Biol Databases and Curation. 2015; 2015:bav006. doi:https://doi.org/10.1093/database/bav006. 38. Rappaport N, Twik M, Nativ N, Stelzer G, Bahir I, Stein TI, Safran M, Lancet D. MalaCards: a comprehensive automatically-mined database of human diseases. Curr Protoc Bioinformatics/editoral board, Andreas D Baxevanis [et al]. 2014;47:1.24.21–19. 39. Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 2014;15(12). https:// doi.org/10.1101/002832. 40. Wen Z, Liu ZP, Liu Z, Zhang Y, Chen L. An integrated approach to identify causal network modules of complex diseases with application to colorectal cancer. J Am Med Inform Assoc. 2013;20(4):659–67. 41. Fukushima A. DiffCorr: an R package to analyze and visualize differential correlations in biological networks. Gene. 2013;518(1):209–14. 42. Strimmer K. fdrtool: a versatile R package for estimating local and tail area- based false discovery rates. Bioinformatics. 2008;24(12):1461–2. 43. Popescu M, Keller JM, Mitchell JA. Fuzzy measures on the Gene Ontology for gene product similarity. IEEE/ACM Trans Comput Biol Bioinform. 2006;3(3):263–74. 44. Chen J, Xu H, Aronow BJ, Jegga AG. Improved human disease candidate gene prioritization using mouse phenotype. BMC Bioinformatics. 2007;8:392. 45. Vanunu O, Magger O, Ruppin E, Shlomi T, Sharan R. Associating genes and protein complexes with disease via network propagation. PLoS Comput Biol. 2010;6(1):e1000641. 46. Wang L, Sun FZ, Chen T. Prioritizing functional modules mediating genetic perturbations and their phenotypic effects: a global strategy. Genome Biol. 2008;9(12):R174. doi:https://doi.org/10.1186/gb-2008-9-12-r174. 47. Zhu Y, Qiu P, Ji Y. TCGA-assembler: open-source software for retrieving and processing TCGA data. Nat Methods. 2014;11(6):599–600. 48. Mukohara T. PI3K mutations in breast cancer: prognostic and therapeutic implications. Breast Cancer (Dove Med Press). 2015;7:111–23. 49. Eppenberger M, Zlobec I, Baumhoer D, Terracciano L, Lugli A. Role of the VEGF ligand to receptor ratio in the progression of mismatch repair- proficient colorectal cancer. BMC Cancer. 2010;10:93. 50. van Dam S, Craig T, de Magalhaes JP. GeneFriends: a human RNA-seq-based gene and transcript co-expression database. Nucleic Acids Res. 2015; 43(Database issue):D1124–32. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png BMC Bioinformatics Springer Journals

MGOGP: a gene module-based heuristic algorithm for cancer-related gene prioritization

Free
12 pages
Loading next page...
 
/lp/springer_journal/mgogp-a-gene-module-based-heuristic-algorithm-for-cancer-related-gene-gpAVxN7BXy
Publisher
BioMed Central
Copyright
Copyright © 2018 by The Author(s).
Subject
Life Sciences; Bioinformatics; Microarrays; Computational Biology/Bioinformatics; Computer Appl. in Life Sciences; Algorithms
eISSN
1471-2105
D.O.I.
10.1186/s12859-018-2216-0
Publisher site
See Article on Publisher Site

Abstract

Background: Prioritizing genes according to their associations with a cancer allows researchers to explore genes in more informed ways. By far, Gene-centric or network-centric gene prioritization methods are predominated. Genes and their protein products carry out cellular processes in the context of functional modules. Dysfunctional gene modules have been previously reported to have associations with cancer. However, gene module information has seldom been considered in cancer-related gene prioritization. Results: In this study, we propose a novel method, MGOGP (Module and Gene Ontology-based Gene Prioritization), for cancer-related gene prioritization. Different from other methods, MGOGP ranks genes considering information of both individual genes and their affiliated modules, and utilize Gene Ontology (GO) based fuzzy measure value as well as known cancer-related genes as heuristics. The performance of the proposed method is comprehensively validated by using both breast cancer and prostate cancer datasets, and by comparison with other methods. Results show that MGOGP outperforms other methods, and successfully prioritizes more genes with literature confirmed evidence. Conclusions: This work will aid researchers in the understanding of the genetic architecture of complex diseases, and improve the accuracy of diagnosis and the effectiveness of therapy. Keywords: Gene prioritization, Gene module, Gene ontology, Cancer-related genes Background scores of each annotation of each candidate genes by Discovering cancer-related genes has profound applica- comparing enriched terms in a given set of training genes. tions in modelling, diagnosis, therapeutic intervention, Endeavour [8] prioritizes candidate genes by similarity and in helping researchers get clues on which genes to values between candidate genes and seed genes, by inte- explore [1–3]. Computational approaches are preferred grating more than six types of genomic datasets from over due to their high efficiency and low cost [4, 5]. Many com- a dozen data sources. Methods of the second kind putational methods have been proposed, including: a) prioritize genes using the guilt-by-association principle, gene-based function similarity measure methods [6–9]; b) which means genes interacting with known disease genes biological interaction network-based methods [10–14], are more likely disease-related genes. For instance, PINTA and c) methods based on multiple datasets fusion [15–17]. [10] prioritizes candidate genes by utilizing an underlying Methods of the first kind based on the hypothesis that global protein interaction network. Other methods rank phenotypically similar diseases are caused by functionally candidate genes by exploiting either local or global net- related genes. Based on this hypothesis, many methods work information [2]. Methods of the last kind incorpor- prioritize genes by computing similarity scores between ate datasets such as gene expression, biomedical literature, the candidate genes and the known disease genes. For ex- gene ontology, and PPIs together for gene prioritization. ample, ToppGene [6] ranks genes based on similarity For example, ProphNet [17] integrates information of different types of biological entities in a number of hetero- * Correspondence: lgx1034@163.com; baitian@jlu.edu.cn; 413224445@qq.com geneous data networks. Taking all these methods into con- College of Computer Science and Technology, Jilin University, Changchun sideration, they are either gene-centric or network-centric. 130012, China Full list of author information is available at the end of the article © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Su et al. BMC Bioinformatics (2018) 19:215 Page 2 of 12 However, gene module as a basic functional unit of genes has interaction data and disease associations. Gene Ontology seldom been considered. Consortium describes the functions of specific genes, using Gene module can be defined as a protein complex, a terms known as GO (Gene Ontology). KEGG map genes to pathway, a sub-network of protein interactions. Module pathways while GSEA provides functional gene groups col- detection has long been studied and many useful algo- lected from BioCarta genes sets, KEGG gene sets and rithms have been proposed, such as [18–21]. Although Reactome gene sets. With these annotation information, different methods have different module detection strat- we can easily group genes into functional modules. egies, most of them rely on PPIs network. PPIs network Complex diseases, especially cancer are caused by the suffers from drawbacks as highlighted in [22]. Firstly, the dysfunction of groups of genes and/or gene interactions ra- PPI network is incomplete, which only covers the interac- ther than the mutations of individual genes. Detecting and tions of well-researched proteins. For instance, of the prioritizing cancer-related genes from the perspective of 20,502 genes in the gene expression matrix downloaded gene module is promising. Although some useful work has from The Cancer Genome Atlas (TCGA), only 9078 been conducted [34, 35], the results are still far from being (44.2%) and 2761 (13.4%) genes are included in Human satisfactory. In this study, we take the importance of not Protein Reference Database (HPRD) [23] and Database of only genes but also their affiliated modules into consider- Interacting Proteins (DIP) [24] PPIs networks respectively. ation, and prioritizing genes in a heuristic way. We measure As a result, detected modules are incomplete and their module importance by the number of differential genes accuracy are limited. Secondly, protein interactions in within the module and the number of differential correla- PPIs network suffer from high false positive and negative tions between the module genes. Besides, the number of rates, modules discovered from such PPI data also suffer known cancer-related genes in the module is also consid- from high false rates. All these inherent limitations affect ered. We measure the gene importance by three aspects in- the coverage and accuracy of the inferred modules. formation: a), gene’s differential expression value, b), the Nowadays, numerous public databases of protein and number of differential correlations between the gene and gene annotation information are available, such as Entrez all other module gene. c), the fuzzy measure based similar- Gene [25], Ensembl [26], PIR iProClass [27], GeneCards ity values between the gene and all known cancer-related [28], KEGG [29], Gene Ontology Consortium [30], DAVID genes (if exist) within the module. The global rank of all [31], GSEA [32]and UniProt[33]. For instance, DAVID genes is obtained by utilizing a rank fusion strategy. [31] contains information on over 1.5 million genes from more than 65,000 species, with annotation types, including Methods sequence features, protein domain information, pathway As showninFig.1, MGOGP takes gene expression datasets, maps, enzyme substrates and reaction, protein-protein gene modules, known disease genes and gene ontology Fig. 1 Main components of MGOGP Su et al. BMC Bioinformatics (2018) 19:215 Page 3 of 12 annotation information [36] as input, and the ranked genes we set Se(g ) = 1, which means the gene g is a candidate i i as output. The main parts including: module importance differential expression gene. Se(g ) is defined as follows: measure, module-specific gene importance measure, mod- ule rank and module-specific gene prioritization, and global 0; if padj g > μ Se g ¼ ð1Þ cancer-related gene prioritization. Figure 2 schematically 1; else illustrates these steps in detail. First, obtain functional gene modules; then get the global To further improve the statistical significance of the se- ranking of all modules and the local ranking of all lected candidate differential expression genes, we applied module-specific genes based on their importance; finally, the a multiple random sampling strategy. As defined in Eq. 2. rank fusion algorithm further gives all genes a global rank. > X 0; if Se g < ω Input datasets DEG g ¼ ð2Þ > s¼1 As shown in Fig. 1, MGOGP takes gene expression data- 1; else sets, gene modules, known disease-related genes and gene ontology annotation information as input. In this Where S is the number of sampling; ω is a threshold study, all gene modules are downloaded from GSEA value; if a gene g is selected as a differential expression website (http://software.broadinstitute.org/gsea/down- gene we set DEG(g ) = 1, Otherwise, we set DEG(g )=0. i i loads.jsp). All GO ontologies of genes are downloaded We define Ncr(m ) as the ratio of differential expres- from GeneCards [37, 38]. Information of relationships sion genes in the module m as shown in Eq. 3: between GO terms are got from Gene Ontology Consor- tium website. P DEG g i¼1 Ncr m ¼ ð3Þ j∈1; 2; 3; …; M Module importance measure We measure the importance of a module by: the number of differentially expressed genes in the module, the num- Where, g is the ith gene in the module m ; N is the i j ber of differential correlations between module genes total number of genes in the module m ; DEG(g ) is de- j i and the basic importance of the module itself. fined in Eq. 2. We use DESeq2 for gene differential expression ana- Next, for each pair of genes in the module m , two correl- lysis [3, 35, 39, 40]. If genes with padj(g ) value bigger ation values are calculated using normal and tumor samples than the threshold value μ, we set Se(g ) = 0. Otherwise, respectively. As defined in Eqs. 4 and 5 respectively. Fig. 2 MGOGP processes are illustrated. a Obtain gene modules, b Module importance measure and prioritization, c Module-specific gene importance measure and prioritization, d Compute global gene ranking Su et al. BMC Bioinformatics (2018) 19:215 Page 4 of 12 We define Ecr(m ) as the ratio of differential correla- ðÞ x −xðÞ y −y l¼1 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi r g ; g ¼ ð4Þ N tions among genes in the module m . Ecr(m ) is defined i h j j L 2 2 ðÞ x −x ðÞ y −y l in Eq. 10: l¼1 l r (g , g ) is the Pearson correlation value between gene N i h DEE g ; g g and gene g across all normal samples. L is the normal k¼1 i h i h Ecr m ¼ sample number. ð10Þ NNðÞ −1 K ¼ and i; h∈1; 2; 3; …; N x −x y −y q¼1 q rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi r g ; g ¼ ð5Þ i h x −x y −x K and N is the edge number and the gene number of q¼1 q the module m respectively. We measure the basic importance of a module by cal- r (g , g ) is the Pearson correlation value between gene T i h culating the ratio of known disease genes in a module, g and gene g across all tumor samples. Q is the tumor i h as shown in Eq. 11: sample number. To test whether the correlation coefficient between gene g and gene g is differentially correlated, we test i h info m ¼ num d þ 1 =N ð11Þ j j whether r (g , g ) and r (g , g ) are significantly different. T i h N i h The two correlation coefficients are changed to Z (g , g ) N i h and Z (g , g ) respectively. num(d ) is the number of known disease genes in the T i h j module m ; N is the number of genes in the module m . j j 1 þ r g ; g i h The module importance is defined in Eq. 12. Z g ; g ¼ log ð6Þ i h 1−r g ; g i h pm ¼ Ncr m þ Ecr m =2  info m Similarly, r (g , g ) is changed to Z (g , g ) as Eq. (6). j j j j T i h T i h j∈1; 2; 3…; M The differential correlation is tested based on Fisher’s z-test [41]. As defined in Eq. (7): ð12Þ Z g ; g −Z g ; g N T i h i h Z ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð7Þ where m means the jth module; M is the total number 1 1 of modules. L−3 Q−3 The Z value has an approximately Gaussian distribu- Module-specific gene importance measure tion under the null hypothesis [41]. If the fdr value of a We measure the importance of a gene (p(g )) in the gene is bigger than the threshold value υ, we set Sc(g , module by measuring: the gene’s differential expression g ) = 0, otherwise we set Sc(g , g ) = 1, which means the h i h value, the number of differential correlations between correlation coefficient is a potential differential correl- the gene and all other module genes and the basic im- ation. Sc(g , g ) is defined as follows: i h portance of the gene itself. The number of differential correlations (CorC(g )) be- 0; if fdr g ; g > υ i h Sc g ; g ¼ ð8Þ i h tween the gene g and all other genes in the same mod- 1; else ule is calculated as in Eq. 13. Where fdr(g , g ) is the local false-discovery rate (fdr) i h derived from fdrtool package [42]; υ is a threshold value. P N−1 Sc g ; g i h h¼1;h≠i As the way we find differential expression genes, we CorC g ¼ N−1 retain only those significantly changed correlations. As ð13Þ i; h∈1; 2; 3; …; N; g ∈m defined in Eq. 9: j∈1; 2; 3; …; M > X 0; if Sc g ; g < δ i h DEE g ; g ¼ ð9Þ N is the number of genes in the module m ; M is the s j i h > s¼1 total module number. 1; else Finally, the basic importance of a gene is determined Where S is the number of sampling; δ is a threshold by the gene ontology-based fuzzy measure similarity value; we set DEE(g , g ) = 1 if the gene g and g are dif- values between the gene and all known disease gene (if i h i h ferentially correlated. Otherwise, we set DEE(g , g )= 0. exist) in the same module. As shown in Eq. 14. i h Su et al. BMC Bioinformatics (2018) 19:215 Page 5 of 12 pTðÞ¼ info m g ¼ countðÞ T þ children of T in corpus k k 0; if num m d ¼ 0 > j h ð18Þ countðÞ all GOtermsincorpus 1; if g isaknown disease gene itself X 1≤k ≤ j T j >   g num m d ðÞ j h i S g ; m d =num m d ; else FMS j h j h h¼1 The importance of gene (p(g )) in a module is defined ð14Þ in Eq. 19. num(m _d ) is the number of known disease genes j h pg ¼ padj g þ CorC g þ info g in the module m.If num(m _d ) = 0, which means no i i i i j j h ð19Þ i∈1; 2; 3; …; N; g ∈m known diseasegeneinthe module m,weset j info(m _g )= 0. If g itself is a known disease gene, we j i i set info(m _g ) = 1. Otherwise, we calculate the gene N is the number of genes in the module m . j i j importance value based on the fuzzy similarity meas- ure between the gene and all the known disease gene Global gene ranking in the module m . S (m _g ,m _d ) is defined in Eq. Most genes deploy their functions in the context of j FMS j i j h 15,asin[43]: sophisticated functional modules [45, 46]. Therefore, the global rank of a gene need be decided by its own importance and the importance of its affiliated Sm T ∩T þ Sm T ∩T i m g m d h m g m d j j h j j h i i S m g ; m d ¼ FMS j j h module. As in [34], a rank fusion strategy is used to fuse the local rank of genes in each module into a ð15Þ global rank. The rank fusion strategy is a recursive process. It Where Sm is the Sugeno measure [43] defined on GO decides the rank of the nth gene based on all the terms of gene m _g and Sm is the Sugeno measure j i h top-ranked n − 1 genes. We define i as the number of defined on GO terms of module disease gene m _d . j h genes having already obtained their global ranking in Let T is the set of GO annotation terms of gene m g the recursive process of rank fusion, m(i, j)asthe m _g , Sm , is a real value function, satisfying [44]: j i i number of top i genes located in the module j after having determined the top i genes. t(i, j)asthe ex- 1) Sm ðT Þ¼ 0; if T ¼ ∅; elseSm ðT Þ¼ 1: i m g m g i m g j i j i j i pectation of the number of top i genes located in the 2) Sm ðT Þ≤SmðT Þif T ⊆T i m g m d m g m d j i j h j i j h module j. e(i, j) as the expectation of probability that 3) For all T ; T ⊆T with T ∩ T = Φ A B m g A B j i the i + 1 globally ranked genes come from the module Sm ðT ∪T Þ¼ Sm ðT Þþ Sm ðT Þ i A B i A i B j.Weuse themoduleimportance value p(m)asthe þλSm ðT ÞSm ðT Þ; λ > −1 i A i B probability of a disease-related gene comes from it. The relationship between i, m(i, j), t(i, j)and p(m)is For a given gene annotation set T , the parameter λ m g showninEq. 20: of its Sugeno fuzzy measure can be uniquely solved as in Eq. 16: tiðÞ ; j ¼ ip m ð20Þ eiðÞ ; j ¼ tiðÞ þ 1; j −miðÞ ; j ðÞ 1 þ λ¼ ðÞ 1 þ λSm ð16Þ Initially, the first ranked gene in the module with high- i¼1 est importance value is chosen as the top 1 gene in the gene’s global rank, because all genes in each module This equation has a unique solution for λ > −1. Let have been ranked from big to small according to their Sm = Sm({T }). The mapping T → Sm is called a fuzzy k k k k importance value. Let i as the number of genes having density function. The fuzzy density value, Sm , is inter- obtained their global ranking, to decide the i + 1 ranked preted as the importance of the single information gene, we need to find the module with the biggest e(i, j) source T in determining the similarity of two genes. As value, because e(i, j) indicates the expectation of prob- defined in Eq. 17: ability that the i + 1 globally ranked genes from module j. So the genes ranked m(i, j) + 1 in the module j will be chosen as the top i + 1 ranked gene, because in the mod- Sm ¼ − ln pTðÞ= max − ln pT ð17Þ k k j T ∈T j g ule j, top m(i, j) genes has obtained the global ranking. Repeat the process until all genes get ranked. As shown Where p(T ) is defined in Eq. 18: in Fig. 3 (in Additional file 1). k Su et al. BMC Bioinformatics (2018) 19:215 Page 6 of 12 Fig. 3 Rank fusion process. N is the number of genes in the module j, M is the total module number As shown in Table 2, all the six genes are ranked on Results Both raw count and normalized gene expression data- average within top10% of all the candidate genes, which sets are downloaded from TCGA (http://cancergen- indicates the superiority of MGOGP to other three algo- ome.nih.gov/)[47], which include expression values of rithms. For further comparison, we put these 21 genes 20,503 genes across 102 normal samples and 779 together, each time we randomly select 20 different tumor samples. Besides, gene expression datasets of genes as known disease genes and the remaining 1 gene Prostate adenocarcinoma containing 483 tumor sam- ples and 51 normal samples are also downloaded Table 1 Known prostate cancer genes retrieved from the OMIM from TCGA. Four thousand seven hundred twenty-six Gene Gene Gene name gene modules are downloaded from the website of ID Symbol GSEA (in Additional file 2). 367 AR Androgen receptor Firstly, the performance of MGOGP is validated 675 BRCA2 Breast cancer type 2 susceptibility protein by comparing it with three module based 3732 CD82 CD82 antigen cancer-related gene prioritization methods (MEN- DEAVOUR, MDK and MRWR) proposed in [34]. 11200 CHEK2 Serine/threonine-protein kinase Chk2 For comparison, the same prostate cancer network 60528 ELAC2 Zinc phosphodiesterase ELAC protein 2 used in [34] are used, which consists of 233 genes 2048 EPHB2 Ephrin type-B receptor 2 precursor and 1218 interactions. Modules are obtained by 3092 HIP1 Huntingtin-interacting protein 1 picking out all the GSEA modules that contain 1316 KLF6 Krueppel-like factor 6 more than three genes in the prostate network after 8379 MAD1L1 Mitotic spindle assembly checkpoint removing irrelevant module genes. Irrelevant genes proteinMAD are genes that are included in GSEA modules but 4481 MSR1 Macrophage scavenger receptor types I and II are not included in these 233 genes. Fifteen known prostate cancer genes are obtained from OMIM 4601 MXI1 MAX-interacting protein 1 (Table 1). Six genes (BRCA1, TP53, EP300, STAT3, 7834 PCAP Predisposing for prostate cancer ZFHX3, HNF1B), which are confirmed have associa- 5728 PTEN Phosphatase and tensin homolog tions with prostate cancer by Genetics Home Refer- 6041 RNASEL 2-5A-dependent ribonuclease ence (https://ghr.nlm.nih.gov/)are used as test 5513 HPC1 Hereditary prostate cancer 1 genes. Results are shown in Table 2. Su et al. BMC Bioinformatics (2018) 19:215 Page 7 of 12 Table 2 Ranks of six test genes in prostate cancer gene Next, we use MGOGP for genome-wide breast cancer network. They are prioritized by MDK, MRWR, Endeavour and gene prioritization. We use 328 breast disease-related MGOGP genes downloaded from SNP4Disease (http://snp4disea- Gene MDK MRWR Endeavour MGOGP se.mpibn.mpg.de/result.php) as seed genes (see BRCA1 29 6 58 63 Additional file 3). Ten well-known breast cancer-related genes (shown in Table 4, which are not contained in the TP53 104 132 85 24 328 genes) are used to validate the effectiveness of our EP300 83 70 90 11 method. All GSEA gene modules are pre-processed by STAT3 39 41 88 17 removing all the genes which do not have gene expres- ZFHX3 174 174 34 19 sion information (the final module list is supplied in HNF1B 44 190 109 26 Additional file 4). The result is shown in Fig. 4. Average Rank 78 102 77 26 As showninFig. 4, all the 10 breast cancer-related genes are ranked within the top5% of the gene prioritization results. During the process, we set S = 1000, ω=0.9 and δ=0.9 (which means of the 1000 sampling results, over 90% fulfill for test. Each run we compared the ranked positions of the filter criteria). We set υ =0.05 and μ =0.01 as most the 1 test gene between our method and Endeavour. others do [39, 41]. The performance of MGOGP under Results are shown in Table 3. In Table 3 some genes do different parameter settings are supplied in Additional file 5. not exist, because they don’t exist in our GSEA gene Thetop 10 ranked modulesinthiscasestudy areshown in modules or not exist in Endeavour database. According Table 5. to Table 3, 11 of the 13 known prostate cancer-related As can be seen from Table 5,many top-ranked genes and 4 of the 6 test genes have much higher ranks modules are included in well-known breast cancer path- than these of the Endeavour. Moreover, the average ways, such as PI3K/AKT [48]pathway andVEGF ranking of these genes is 51 by MGOGP, which is better ligand-receptor pathway. The VEGF family of ligands than 82 by Endeavour. and receptors are intimately involved in tumor angio- genesis, lymphangiogenesis, and metastasis [49]. More importantly, of the 100 genes in the top 10 ranked modules, 20 of them are contained in the KEGG breast Table 3 Ranks of each validation gene cancer pathway (hsa05224), which is an indication of Gene MGOGP Endeavour the good performance of MGOGP for cancer gene AR 32 30 prioritization. BRCA2 29 40 Next, we validate the performance of MGOGP by com- CD82 169 211 paring the gene prioritization results with results obtained CHEK2 19 35 by methods: Endeavour [8], GeneFriends [50], PINTA ELAC2 64 176 [10], TOPPGene [6] and TOPNet [13]. All the methods use the same datasets and under their default parameter EPHB2 45 165 HIP1 91 111 Table 4 Ten well-known breast cancer genes KLF6 88 72 Gene Gene symbol Gene name MAD1L1 78 194 ID MSR1 60 190 672 BRCA1 Breast Cancer 1, Early Onset MXI1 92 89 675 BRCA2 Breast Cancer 2, Early Onset PCAP Not Exist Not Exist 7157 TP53 Tumor Protein P53 PTEN 24 94 5728 PTEN Phosphatase And Tensin Homolog RNASEL 67 83 841 CASP8 Caspase 8, Apoptosis-Related Cysteine Peptidase HPC1 Not Exist Not Exist 2263 FGFR2 Fibroblast Growth Factor Receptor 2 BRCA1 46 16 4214 MAP3K1 Mitogen-Activated Protein Kinase Kinase TP53 5 5 Kinase 1, E3 Ubiquitin Protein Ligas EP300 11 12 11200 CHEK2 Checkpoint Kinase 2 STAT3 17 23 472 ATM ATM Serine/Threonine Kinase ZFHX3 59 68 83990 BRIP1 BRCA1 Interacting Protein C-Terminal HNF1B 26 12 Helicase 1 Su et al. BMC Bioinformatics (2018) 19:215 Page 8 of 12 Fig. 4 Known cancer-related gene prioritization result settings. The results are shown in Fig. 5. Brief descriptions others are left for test (each kind of selection repeat of these methods are provided in Additional file 6. Core 100 times). We count the average number of test sourcecode of MGOGP is provided in Additional file 7. genes appear in Top 200 gene prioritization results. Other source codes are available from the corresponding Results are shown in Fig. 6. author on reasonable request. Finally, to further validate our method, we get the top In Fig. 5, we count the number of breast 10 ranked genes of each method in Fig. 5. The results cancer-related genes in the gene prioritization results. are shown in Table 6. As is shown in Fig. 5, MGOGP outperforms other In Fig. 7, the number of Known Disease Gene is the methods in detecting cancer-related genes. We use all number of genes supplied for training each method that the 328 breast disease related genes as known disease fall within the top 10. For example, in Table 6, PTEN, gene (Endeavour and GeneFriends used the same gene VEGFB, and MCM2 are three genes fall within the top sets) and count the number of known disease genes ap- 10 of the gene ranking result, so the number of Known pear in top 100–1000 prioritization results. Disease Gene of MGOGP in Fig. 7 is 3. For each gene To do comparison more rigorously, we further com- within the top 10 gene ranking results of each method, pare MGOGP to Endeavour, TOPNet and TOPPGene. we search the number of articles in PubMed mention Each time we randomly select 100, 150 and 200 dif- the association between the gene and breast cancer. We ferent known disease genes from the 328 breast count the number of genes has more than 10 PubMed disease-related genes for known disease genes and article reference. As shown in Fig. 7, genes detected by MGOGP have more article supports than other methods. Table 5 Top 10 ranked modules Rank Module name Gene number Importance value Discussion and conclusion 1 zerbini_response_to_sulindac_dn 6 0.542 Results of omics experiments commonly consist of 2 reichert_g1s_regulators_as_pi3k_ 8 0.523 a large set of genes, while researchers usually only targets care about the behaviour of several genes. In this 3 sa_g2_and_m_phases 8 0.492 paper, a heuristic algorithm is proposed for priori- tizing disease-associated genes by utilizing gene 4 reactome_vegf_ligand_receptor_ 10 0.478 interactions ontology annotation information and known 5 biocarta_srcrptp_pathway 11 0.461 disease-related genes as heuristic information. Dif- ferent from existing methods, we propose to rank 6 honrado_breast_cancer_brca1_ 18 0.447 vs_brca2 genes considering the importance of both individual genes and their affiliated modules, and utilize Gene 7 tcga_glioblastoma_mutated 8 0.445 Ontology (GO) based fuzzy measure value as well 8 pid_vegf_vegfr_pathway 10 0.444 as known disease genes as heuristics, and use rank 9 liang_silenced_by_methylation_dn 11 0.411 fusion strategy to obtain the global gene 10 agarwal_akt_pathway_targets 10 0.410 prioritization. Results show that MGOGP Su et al. BMC Bioinformatics (2018) 19:215 Page 9 of 12 Fig. 5 Comparison results between 6 methods. Endeavour, GeneFriends, PINTA, TOPPGene, TOppNet, and MGOGP outperforms many other methods in cancer-related number of interactions among well-researched genes gene prioritization. maybemuchmorethanthatofnewly discovered Different from other module-based gene prioritization genes). methods, where modules are detected by partitioning Different from module-based methods in [34], the network using the network clustering methods, we MGOGP ranks modules considering three aspects of obtain modules through gene function annotation, that information: module-specific gene importance, differ- is, functionally related genes are grouped into the same ential correlations, and importance of the module it- modules. Because gene interaction networks often suffer self. In [34], the author considers the importance of a from the problems of high rates of false positive/negative module by considering only the number of disease interactions, and modules detected by network cluster- genes and the size of the module, which may bias to- ing algorithms often have limited accuracy, so our ward big modules. Furthermore, gene as the major method is more advanced. One important difference component of the module whose importance is not between modules used in this study and modules de- considered when measuring the importance of a mod- tected through network partition is that no edges in ule in [34]. While in our method, when measuring our module. Instead, we use statistical methods de- the importance of a module, we consider: the import- tecting differential correlationsbetween geneswithin ance of the module itself, the importance of module a module, which could help avoid the preference of contained genes as well as differential correlations genes or modules that are well-researched (because within the module, which are the main improvements currently obtained network is far from complete, the of our method. Fig. 6 Comparison results between MGOGP, Endeavour, TOPPGene, and TOPNet with different number of known disease genes as input Su et al. BMC Bioinformatics (2018) 19:215 Page 10 of 12 Table 6 Top 10 ranked genes of each method MGOGP Endeavor GeneFriends PINTA ToppGene ToppNet Top 10 gene CCNB1IP1 SNRPF LURAP1L MGP RAD51 APP CCNE2 BUB3 PVRL2 EEF1A1 APEX1 ELAVL1 NEK1 MSH2 CYFIP1 TPT1 SIRT2 NTRK1 NRP1 SSBP1 FAM120A RPS6 NOC2L RPA1 CDC25C RFC4 IL13RA1 RPL3 NEDD1 XPO1 VIM EZH2 MYO1B RPS27 TERT EED PTEN CENPF BCL9L ACTB EPN3 CUL3 VEGFB BLMH NQO1 SCGB2A2 PPARGC1A BARD1 MCM2 KIF20B RIN2 RPL11 NBN HSP90AA1 PTGS2 BAZ1A SDC4 PIP ATR NXF1 Known disease genes PTEN MSH2 NQO1 SCGB2A2 RAD5 BARD1 fall in the top 10 gene VEGFB EZH2 PIP TERT MCM2 NBN ATR In Table 6, each method is run with default parameter settings and use same training genes. Top 10 gene means the top 10 genes prioritized by each method and Known disease genes fall in the top 10 gene means genes supplied for training each method falls in the top 10 genes. Detail statistic results are shownin Fig. 7 Compared with other non-module-based prioritization top-ranked modules are included in well-known can- methods, our algorithm also has obvious advantages. cer pathways, and top-ranked genes have more sup- First, it is easier to find the potential pathogenic genes porting PubMed articles. All of the results show that that cause the disease from the point of view of gene our methods perform better than the state of the art modules. Second, it takes cross-validation strategy which methods. could guarantee the stability of the recognition re- Prioritization methods are useful for assisting sci- sults. And our method works with heuristic informa- entists at early research stages, and to formulate tion which could effectively avoid the blindness of novel hypotheses of interest. In the future, one of the search. our main goals is to see how our method behaves in By applying MGOGP on different datasets, we dem- other prioritization problems when using different onstrate that MGOGP performs better than previous entities and sources of data sets not covered in this gene or network-centric methods in terms of poten- study. Furthermore, we plan to study in more detail tial disease-related genes prediction. Firstly, the per- the quality of the datasets and their influence on al- formance of MGOGP is validated by comparing it gorithm performance, and design new methods to with three module based cancer-related gene try to improve the results. As we all know that the prioritization methods. Results show that all test methods become more mature the results will be- genes are ranked on average within top10% of all the come increasingly accurate and more biologically candidate genes. According to our results, many meaningful. Fig. 7 Detail statistic results of results in Table 6 Su et al. BMC Bioinformatics (2018) 19:215 Page 11 of 12 Additional files Received: 13 March 2017 Accepted: 23 May 2018 Additional file 1: A step by step example of Rank Fusion process. This file provides an example of how to get the final gene rank. (DOCX 275 kb) References 1. Gill N, Singh S, Aseri TC. Computational disease gene prioritization: an Additional file 2: GSEA gene module. This file is all the gene modules appraisal. J Comput Biol. 2014;21(6):456–65. downloaded from GSEA website. (TXT 2837 kb) 2. Moreau Y, Tranchevent LC. Computational tools for prioritizing candidate Additional file 3: Breast-Cancer-Gene. This is the known breast cancer- genes: boosting disease gene discovery. Nat Rev Genet. 2012;13(8):523–36. related genes downloaded from SNP4Disease. (TXT 2 kb) 3. Cruz-Monteagudo M, Borges F, Paz YMC, Cordeiro MN, Rebelo I, Perez- Additional file 4: Final module list. This is the refined module list after Castillo Y, Helguera AM, Sanchez-Rodriguez A, Tejera E. Efficient and removing irrelevant genes. (TXT 2736 kb) biologically relevant consensus strategy for Parkinson’s disease gene prioritization. BMC Med Genet. 2016;9:12. Additional file 5: Parameters discussion. This file discusses the 4. Bromberg Y. Chapter 15: disease gene prioritization. PLoS Comput Biol. performance of MGOGP under different parameter settings. (DOCX 65 kb) 2013;9(4):e1002902. https://doi.org/10.1371/journal.pcbi.1002902. Additional file 6: Brief description of gene prioritization methods. This 5. Doncheva NT, Kacprowski T, Albrecht M. Recent approaches to the file provides the short description of comparison methods, including prioritization of candidate disease genes. Wiley Interdiscip Rev Syst Biol their input datasets, limitations, and type. (DOCX 17 kb) Med. 2012;4(5):429–42. Additional file 7: Sourcecode. Some core code of our method. (TXT 5 kb) 6. Chen J, Bardes EE, Aronow BJ, Jegga AG. ToppGene Suite for gene list enrichment analysis and candidate gene prioritization. Nucleic Acids Res. 2009;37:W305–11. Acknowledgments 7. Schlicker A, Lengauer T, Albrecht M. Improving disease gene prioritization The results here are in whole or part based upon datasets generated by the using the semantic similarity of Gene Ontology terms. Bioinformatics. 2010; TCGA Research Network: http://cancergenome.nih.gov/. 26(18):i561–7. 8. Tranchevent LC, Barriot R, Yu S, Van Vooren S, Van Loo P, Coessens B, De Funding Moor B, Aerts S, Moreau Y. ENDEAVOUR update: a web resource for gene This work is supported by Graduate Innovation Fund of Jilin University (No. prioritization in multiple species. Nucleic Acids Res. 2008;36:W377–84. 2016031); The National Nature Science Foundation of China (No. 61373051, 9. Yu W, Wulf A, Liu T, Khoury MJ, Gwinn M. Gene Prospector: an evidence No. 61502343, No. 61772226 and No. 61702214); Science and Technology gateway for evaluating potential susceptibility genes and interacting risk Development Program of Jilin Province (No. 20140204004GX); The Science factors for human diseases. BMC Bioinformatics. 2008;9:528. Research Funds for the Guangxi Universities (No. KY2015ZD122); The Science 10. Nitsch D, Tranchevent LC, Goncalves JP, Vogt JK, Madeira SC, Moreau Y. Research Funds for the Wuzhou University (2014A002); Project of Science PINTA: a web server for network-based gene prioritization from expression and Technology Innovation Platform of Computing and Software Science data. Nucleic Acids Res. 2011;39(Web Server issue):W334–8. (985 Engineering); The Key Laboratory for Symbol Computation and 11. Xie B, Agam G, Balasubramanian S, Xu J, Gilliam TC, Maltsev N, Bornigen D. Knowledge Engineering of the National Education Ministry of China; The Disease gene prioritization using network and feature. J Comput Biol. 2015; Fundamental Research Funds for the Central. China Postdoctoral Science 22(4):313–23. Foundation (No. 2014M561293); Development Project of Jilin Province of 12. Navlakha S, Kingsford C. The power of protein interaction networks for China (No. 20150520064JH). associating genes with diseases. Bioinformatics. 2010;26(8):1057–63. 13. Chen J, Aronow BJ, Jegga AG. Disease candidate gene identification and prioritization using protein interaction networks. BMC Availability of data and materials Bioinformatics. 2009;10:73. The datasets used and/or analyzed during the current study are available 14. Erten S, Bebek G, Ewing RM, Koyuturk M. DADA: degree-aware algorithms from the corresponding author on reasonable request. for network-based disease gene prioritization. BioData mining. 2011;4(19). https://doi.org/10.1186/1756-0381-4-19. Authors’ contributions 15. Wu C, Zhu J, Zhang X. Integrating gene expression and protein-protein LS made contributions to method design and data analysis, and a major interaction network to prioritize cancer-associated genes. BMC contributor in writing the manuscript. GL involved in drafting the manuscript Bioinformatics. 2012;13:182. and revision. TB analyzed the results and made contributions to method 16. Simoes SN, Martins DC Jr, Pereira CA, Hashimoto RF, Brentani H. NERI: network- implementation. XM performed comparative analysis. QM made medicine based integrative approach for disease gene prioritization by relative contributions to results interpretation and also involved in data acquisition importance. BMC Bioinformatics. 2015;16(Suppl 19):S9. and manuscript writing. All authors read and approved the final manuscript. 17. Martínez V, Cano C, Blanco A. ProphNet: a generic prioritization method through propagation of information. BMC Bioinformatics. 2014;15(Suppl 1): Ethics approval and consent to participate S5. doi:https://doi.org/10.1186/1471-2105-15-S1-S5. In this study, all gene expression datasets were downloaded from TCGA 18. Zhang Y, Lin H, Yang Z, Wang J. Integrating experimental and literature database (https://tcga-data.nci.nih.gov/tcga/). There are no restrictions on the protein-protein interaction data for protein complex prediction. BMC use of TCGA data for research and data analysis purposes. All datasets can Genomics. 2015;16(Suppl 2):S4. be downloaded and used freely, and not require an ethics statement. 19. Srihari S, Yong CH, Patil A, Wong L. Methods for protein complex prediction and their contributions towards understanding the organisation, function Competing interests and dynamics of complexes. FEBS Lett. 2015;589(19 Pt A):2590–602. The authors declare that they have no competing interests. 20. Su L, Liu G, Wang H, Tian Y, Zhou Z, Han L, Yan L. GECluster: a novel protein complex prediction method. Biotechnol Biotechnol Equip. 2014;28(4):753–61. 21. Bader GD, Hogue CW. An automated method for finding molecular complexes Publisher’sNote in large protein interaction networks. BMC Bioinformatics. 2003;4:2. Springer Nature remains neutral with regard to jurisdictional claims in 22. Ramaprasad A, Pain A, Ravasi T. Defining the protein interaction network of published maps and institutional affiliations. human malaria parasite plasmodium falciparum. Genomics. 2012;99(2):69–75. 23. Keshava Prasad TS, Goel R, Kandasamy K, Keerthikumar S, Kumar S, Author details Mathivanan S, Telikicherla D, Raju R, Shafreen B, Venugopal A, et al. Human College of Computer Science and Technology, Jilin University, Changchun protein reference database–2009 update. Nucleic Acids Res. 2009; 130012, China. Key Laboratory of Symbolic Computation and Knowledge 37(Database):D767–72. Engineering of Ministry of Education, Jilin University, Changchun 130012, 24. Xenarios I, Salwinski L, Duan XJ, Higney P, Kim SM, Eisenberg D. DIP, the China. The First Clinical Hospital of Jilin University, Changchun 130021, Database of Interacting Proteins: a research tool for studying cellular China. networks of protein interactions. Nucleic Acids Res. 2002;30(1):303–5. Su et al. BMC Bioinformatics (2018) 19:215 Page 12 of 12 25. Maglott D, Ostell J, Pruitt KD, Tatusova T. Entrez gene: gene-centered information at NCBI. Nucleic Acids Res. 2011;39:D52–7. 26. Flicek P, Amode MR, Barrell D, Beal K, Brent S, Chen Y, Clapham P, Coates G, Fairley S, Fitzgerald S, et al. Ensembl 2011. Nucleic Acids Res. 2011; 39(Database):D800–6. 27. Wu CH, Huang H, Nikolskaya A, Hu Z, Barker WC. The iProClass integrated database for protein functional analysis. Comput Biol Chem. 2004;28(1):87–96. 28. Rebhan M, Chalifa-Caspi V, Prilusky J, Lancet D. GeneCards: integrating information about genes, proteins and diseases. Trends Genet. 1997;13(4):163. 29. Kanehisa M, Sato Y, Kawashima M, Furumichi M, Tanabe M. KEGG as a reference resource for gene and protein annotation. Nucleic Acids Res. 2016;44(D1):D457–62. 30. Gene Ontology C. Gene ontology consortium: going forward. Nucleic Acids Res. 2015;43(Database issue):D1049–56. 31. Huang DW, Sherman BT, Tan Q, Kir J, Liu D, Bryant D, Guo Y, Stephens R, Baseler MW, Lane HC, et al. DAVID Bioinformatics Resources: expanded annotation database and novel algorithms to better extract biology from large gene lists. Nucleic Acids Res. 2007;35(Web Server issue):W169–75. 32. Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci U S A. 2005;102(43):15545–50. 33. UniProt C. UniProt: a hub for protein information. Nucleic Acids Res. 2015; 43(Database issue):D204–12. 34. Chen X, Yan GY, Liao XP. A novel candidate disease genes prioritization method based on module partition and rank fusion. OMICS. 2010;14(4):337–56. 35. Liu X, Liu ZP, Zhao XM, Chen L. Identifying disease genes and module biomarkers by differential interactions. J Am Med Inform Assoc. 2012; 19(2):241–8. 36. Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, et al. Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet. 2000;25(1):25–9. 37. Belinky F, Nativ N, Stelzer G, et al. PathCards: multi-source consolidation of human biological pathways. Database: J Biol Databases and Curation. 2015; 2015:bav006. doi:https://doi.org/10.1093/database/bav006. 38. Rappaport N, Twik M, Nativ N, Stelzer G, Bahir I, Stein TI, Safran M, Lancet D. MalaCards: a comprehensive automatically-mined database of human diseases. Curr Protoc Bioinformatics/editoral board, Andreas D Baxevanis [et al]. 2014;47:1.24.21–19. 39. Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 2014;15(12). https:// doi.org/10.1101/002832. 40. Wen Z, Liu ZP, Liu Z, Zhang Y, Chen L. An integrated approach to identify causal network modules of complex diseases with application to colorectal cancer. J Am Med Inform Assoc. 2013;20(4):659–67. 41. Fukushima A. DiffCorr: an R package to analyze and visualize differential correlations in biological networks. Gene. 2013;518(1):209–14. 42. Strimmer K. fdrtool: a versatile R package for estimating local and tail area- based false discovery rates. Bioinformatics. 2008;24(12):1461–2. 43. Popescu M, Keller JM, Mitchell JA. Fuzzy measures on the Gene Ontology for gene product similarity. IEEE/ACM Trans Comput Biol Bioinform. 2006;3(3):263–74. 44. Chen J, Xu H, Aronow BJ, Jegga AG. Improved human disease candidate gene prioritization using mouse phenotype. BMC Bioinformatics. 2007;8:392. 45. Vanunu O, Magger O, Ruppin E, Shlomi T, Sharan R. Associating genes and protein complexes with disease via network propagation. PLoS Comput Biol. 2010;6(1):e1000641. 46. Wang L, Sun FZ, Chen T. Prioritizing functional modules mediating genetic perturbations and their phenotypic effects: a global strategy. Genome Biol. 2008;9(12):R174. doi:https://doi.org/10.1186/gb-2008-9-12-r174. 47. Zhu Y, Qiu P, Ji Y. TCGA-assembler: open-source software for retrieving and processing TCGA data. Nat Methods. 2014;11(6):599–600. 48. Mukohara T. PI3K mutations in breast cancer: prognostic and therapeutic implications. Breast Cancer (Dove Med Press). 2015;7:111–23. 49. Eppenberger M, Zlobec I, Baumhoer D, Terracciano L, Lugli A. Role of the VEGF ligand to receptor ratio in the progression of mismatch repair- proficient colorectal cancer. BMC Cancer. 2010;10:93. 50. van Dam S, Craig T, de Magalhaes JP. GeneFriends: a human RNA-seq-based gene and transcript co-expression database. Nucleic Acids Res. 2015; 43(Database issue):D1124–32.

Journal

BMC BioinformaticsSpringer Journals

Published: Jun 5, 2018

References

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off