MaxCLK: discovery of cancer driver genes via maximal clique and information entropy of modules Authors: Jian Liu, Fubin Ma, Yongdi Zhu, Naiqian Zhang, Lingming Kong, Jia Mi, Haiyan Cong, Rui Gao, Mingyi Wang, Yusen Zhang, Jonathan Wren Published: 2023 ABSTRACT AbstractMotivationCancer is caused by the accumulation of somatic mutations in multiple pathways, in which driver mutations are typically of the properties of high coverage and high exclusivity in patients. Identifying cancer driver genes has a pivotal role in understanding the mechanisms of oncogenesis and treatment.ResultsHere, we introduced MaxCLK, an algorithm for identifying cancer driver genes, which was developed by an integrated analysis of somatic mutation data and protein–protein interaction (PPI) networks and further improved by an information entropy index. Tested on pancancer and single cancers, MaxCLK outperformed other existing methods with higher accuracy. About pancancer, we predicted 154 driver genes and 787 driver modules. The analysis of co-occurrence and exclusivity between modules and pathways reveals the correlation of their combinations. Overall, our study has deepened the understanding of driver mechanism in PPI topology and found novel driver genes.Availability and implementationThe source codes for MaxCLK are freely available at https://github.com/ShandongUniversityMasterMa/MaxCLK-main. 1 Introduction Cancer is recognized as one of the most complex diseases threatening human health (Torre et al. 2016). The development of sequencing technology enables us to better understand the mechanism of cancer during its evolution by analyzing gene mutation data from cancer patients (Cancer Genome Atlas Network 2012). Systematic cancer genomics projects such as The Cancer Genome Atlas (TCGA) (Tomczak et al. 2015) and International Cancer Genome Consortium (ICGC) (Zhang et al. 2019) have collected and analyzed a large amount of genomic data on various cancers and brought new insights into the study of cancer driver genes. Kyoto Encyclopedia of Genes and Genomes (KEGG) (Kanehisa et al. 2012) is a biological pathway database that integrates genomic, chemical and systemic functional information, which can help researchers gradually learn about the mechanism of carcinogenesis. The mutations that give a cancer cell a fundamental growth advantage and promote cancer development are called driver mutations and the corresponding genes subject to alteration are called driver genes. On the other hand, there may exist several other mutations that occur randomly in a tumor sample but those are not directly associated with cancer development. Such mutations are called passenger mutations. Moreover, the imbalance in the number of driver and passenger always creates many false positives for the prediction of driver genes. Thus, it is highly challenging to distinguish the real drivers from lots of passengers. There is much available evidence suggesting that a cancer driver module typically has two significant features: high coverage and high exclusivity. The high coverage represents that each patient possesses at least one mutation in the driver module, and the high exclusivity means that any pair of genes of the driver module occur as little co-mutation as possible in the patients covered. Dendrix (Vandin et al. 2012) was specifically proposed for the identification of gene modules satisfying these two features. Multi-Dendrix (Leiserson et al. 2013) was designed to identify multiple mutually exclusive gene modules simultaneously using an integer linear program. MEMo (Ciriello et al. 2012), a statistical algorithm, was developed to identify specific gene modules using three different standards in correlation analysis and statistical tests. Many experiments (Das and Yu 2012, Ryslik et al. 2013) have shown that the driver modules in cancer should be highly mirrored in protein–protein interaction (PPI) networks, such as HotNet2 (Leiserson et al. 2015), which was proposed to search for subnetworks with significant symbiotic mutations and elucidate the rare combinations of symbiotic mutations with PPIs. CovEx (Gao et al. 2017), an algorithm that used binary linear programming to select driver modules and genes, designed an Ex value to measure the exclusivity of driver modules. The integration of multi-omics data and PPI networks is beneficial to capture the mutation patterns of genes more completely. In this study, we proposed a new method, MaxCLK, which combines the coverage and exclusivity of modules to identify driver genes, and further improves its performance by information entropy (IE). MaxCLK substantially outperformed CovEx and HotNet2, two salient tools of a similar type, with accuracy improved by at least 30% on average. Furthermore, applying MaxCLK to pancancer datasets, 154 core-MaxCLK genes were obtained and we discovered 18 core-MaxCLK genes that were identified only by our method, most of which have been reported to potentially have deep associations with cancer. Besides the above methods, MaxCLK was also compared with other popular tools, and its performance was superior to them on pancancer and single cancer datasets. Moreover, we integrated 270 predictions of MaxCLK and then performed Gene Ontology (GO) and KEGG analysis on their consensus results to explore potential carcinogenic pathways. In single cancer datasets, we found that ATM and MDM2 are the drivers of eight and seven cancers, respectively, and then conducted a differential analysis in normal and tumor samples with them. To confirm the driver genes of specific cancers, we identified external studies that verify the driver function of known cancer genes and other core-MaxCLK genes in corresponding cancers. Finally, the modules with all genes belonging to core-MaxCLK genes were extracted in predictions and analyzed with 123 enriched pathways in correlation and they were considered potential driver modules. 2 Materials and methods Here, we introduce a new algorithm, termed MaxCLK, to identify cancer driver genes. It takes the somatic mutation data including 11 datasets (Supplementary Table S1) and PPI data as input. First, MaxCLK creates local networks targeted at any mutated gene in the PPI influence networks (Fig. 1A). And then, exclusive graphs are constructed by combining mutation data with different local networks, respectively (see Section 2.3). In Fig. 1B, the cliques in given exclusive graphs are treated as preselection modules. And by means of IE introduced firstly by our method to describe the exclusivity and coverage of modules, MaxCLK can obtain sufficient candidate modules. A demonstration of the calculation of IE is provided in Supplementary Fig. S1. Next, a minimum set cover model (see Section 2.5) is responsible for extracting the most influential crucial modules to patients from candidates and these modules are seen as predicted driver modules (Fig. 1C). Finally, MaxCLK incorporates all the predicted driver modules into a consensus network to optimize its performance (Fig. 1D). Flowchart of MaxCLK. (A) The construction of exclusive graphs. For each mutated target gene, three local networks with it as the root node are constructed based on the following PPI networks: HINT+HI2012, iRefIndex, and Multinet. The exclusive graphs are formed by integrating local networks with the pancancer dataset. (B) Searching for candidate modules. Cliques of sizes k in exclusive graphs are designated as preselection modules (blue rectangles), which undergo additional filtration with the P-value of IE to identify candidate driver modules. (C) Predictions of MaxCLK. The candidate modules are filtered further using a minimum set cover model to identify crucial modules. P1, P2, and P3 correspond to three PPI networks. (D) Consensus strategy. All predictions are consolidated into a weighted undirected network, determined by the connection statuses of gene pairs within crucial modules. Finally, type-1 and type-2 subnetworks are extracted as consensus outcomes. 2.1 Mutation and reference datasets We obtained three pancancer datasets, A (11 565 genes), B (19 459 genes), and C (11 533 genes), spanning 12 cancer types and including 3281 patients as well as three PPI influence networks: HINT+HI2012, iRefIndex, and Multinet, from HotNet2 (http://compbio-research.cs.brown.edu/pancancer/hotnet2/), which were generated by random walk with restart in real PPIs. Because colon adenocarcinoma (COAD) and rectum adenocarcinoma (READ) are integrated into a dataset, 11 datasets are obtained instead of 12. Moreover, we collected annotated cancer genes as reference datasets from different publicly available sources: 3347 protein coding cancer genes were obtained from Network of Cancer Genes (NCG) (Dressler et al. 2022), 571 genes were downloaded from the Cancer Gene Census (Tate et al. 2019) (CGC), 138 cancer genes were downloaded from a study (Vogelstein et al. 2013), 1064 genes were obtained from the precise oncology knowledge database (OncoKB) (Chakravarty et al. 2017), and 1217 genes were obtained from the tumor suppressor gene database (Zhao et al. 2013) (TSG). NCG is default reference dataset and to avoid ambiguity caused by gene symbol updates, NCGv5.0 is for comparison while NCGv7.0 is for the exploration of novel genes. 2.2 PPI local networks Previous studies have shown the complex interactions of driver genes in PPI networks (Davoli et al. 2013). To quickly obtain information on close genetic relationships, MaxCLK draws many local networks from PPI networks based on a breadth-first search strategy. Each local network consists of a target mutated gene and all its (k-1)-order neighbors. Thus, the identification of possible driver genes and modules is limited to closely interacting genes rather than other unconnected genes. In practice, the maximum size of a local network is 300, which allow the local network to cover the (k-1)-order neighborhoods of the target mutated gene (k = 2, 3, or 4). 2.3 Exclusive graphs To explore the high coverage and high exclusivity of driver modules in patients, we defined the exclusivity matrices and the exclusivity graphs to explain the frequency, exclusivity, and co-mutation of mutated genes. Mapping a local network to the mutation matrix, a submatrix having the same mutated genes with the local network, denoted as A∈Rm×n, can be obtained. Then, genes with mutation frequency less than minimum mutation number (MMN) are removed from submatrix A. To reduce noise and mark gene exclusivity, an exclusive matrix is defined by E=ATA, which has the following properties: (i) Eii is the number of patients with mutated gene gi (i=1, 2,…,n); (ii) Eij represents the number of patients whose genes gi and gj are mutated (i≠j). The exclusive graph, denoted as GE, has nodes representing genes in E and its edge connects nodes gi and gj if and only if Eij 0.1. In addition, to make the results more accurate and reliable, MaxCLK takes a Monte Carlo test for the IE values of the preselection modules, randomly sampling the possible modules 106 times in original PPI networks. Then, the modules with P value < 0.05 and Ex≥0.8 are selected as the candidate modules. To screen the most noteworthy modules from so many candidates, a minimum set cover model (Supplementary Table S3) is applied to heuristically search for the crucial modules in the population level (Supplementary Text S1). Here, a module is termed an s-module if its addition to the minimal covering set results in less than s new patients being covered, and the elimination of s-modules will improve the accuracy of MaxCLK. 2.6 Consensus method To obtain more stable driver genes, MaxCLK was used in three PPI networks and introduced a consensus method to aggregate all predictions. We defined three matrices, denoted M1, M2, and M3, reflecting the frequency for each pair of genes in a single module based on each PPI network. Concretely, M1(i, j)=1 if genes gi and gj fall into a certain single module obtained by MaxCLK based on the PPI network HINT+HI2012; otherwise, M1(i, j)=0. M2 and M3 can be defined similarly by applying MaxCLK to iRefIndex and Multinet, respectively. Let MC = M1 + M2+ M3. We denote the weighted graph of MC as GC, where wGC(i, j)=MC(i, j) represents the weight of an edge (i, j) and the weight of a node is defined as the maximum weight of all edges incident to the node, ie, wGC(i)=maxj⁡{wGC(i, j)}. A maximal subnetwork with all its edges and nodes being of weight 3 is defined as a core subnetwork of GC. The genes corresponding to the nodes with weight 3 is called core-MaxCLK genes. Furthermore, to explore potential driver genes, a core subnetwork is extended by absorbing all weight 2 nodes connected to it, which is designated the type-1 subnetwork. The potential driver genes in the core/type-1 subnetwork, in our minding, will have latent capacity to promote the development of cancer treatment and be worthy of studying more deeply. In addition, after removing nodes of weight 3 in GC, the subnetwork induced by nodes and edges with weight 2 in the remaining graph is designated the type-2 subnetwork (Fig. 1). 3 Results 3.1 Preliminary predictions of MaxCLK Three pancancer datasets A, B, and C, generated in different filtering criteria, were obtained from HotNet2, including 12 cancer types and 3281 samples. Default dataset is A filtered by removing hypermutated samples and genes with low expression. Sensitivity is defined as the number of identified known cancer genes and accuracy is the ratio of the number of cancer genes identified to the number of the genes predicted, respectively. MaxCLK involves three variable parameters, MMN (= 40, 50, 60,70, 80), errEx (= 1, 2, 3, 4, 5, 6), and k (=2, 3, 4). MMN represents the minimum mutation number of genes to eliminate low mutation genes. Parameter errEx controls the generation of exclusive graphs. Parameter k denotes the size of cliques (Supplementary Text S2). MaxCLK was applied to three PPI influence networks, HINT+HI2012, iRefIndex, and Multinet respectively and 270 in total predictions were obtained (Supplementary Tables S4-1–S4-3). For k = 2 (Supplementary Table S5-1), the number of modules ranges from 1 to 8, while considering k = 3 (Supplementary Table S5-2) and k = 4 (Supplementary Table S5-3), their range is 120–339 and 75–479, respectively. About sensitivity and accuracy, considering k = 3, 30 predictions were involved for each PPI, and MaxCLK reached its minimum sensitivities of 43, 8 and 47 with corresponding accuracies of 58.11%, 50.00%, and 51.65% and maximum sensitivities of 143, 124, and 132 with corresponding accuracies of 74.48%, 73.37%, and 67.01% on HINT+HI2012, iRefIndex, and Multinet, respectively. Furthermore, the minimum and maximum accuracies in all 90 predictions with k = 4 were 44.87% and 70.83%, respectively. The performance of MaxCLK under other reference datasets was also evaluated (Supplementary Data1). Different parameter combinations have undulant outcomes (Supplementary Fig. S3) and they will be integrated by the consensus method below. 3.2 Consensus network To obtain robust predictions, a consensus method (Fig. 1D) was used to enhance the identification of driver genes. The genes with maximum weight in consensus network are called core-MaxCLK genes and ultimately regarded as predicted driver genes. Default s = 1, a type-1 consensus network (Supplementary Table S6) was obtained containing 154 core-MaxCLK genes in which 126 genes were from NCGv7.0 with an accuracy of 81.82% and covering 94.66% patients. 15 type-2 subnetworks (Supplementary Table S7) were obtained containing 46 genes in total in which 24 genes were related to cancer. Fifty-five of 154 core-MaxCLK genes were also found by CovEx and 81 of remaining 99 genes were identified in all reference datasets (Supplementary Table S8). The main functions of the additional 18 genes were summarized in Supplementary Table S9 and some studies proved that they have participated in the generation and development of tumor or its environment (Table 1). Table 1.Novel potential driver genes and their literature.aGenePMIDCoverageA2M96976961.29%ASAP1201547191.42%CBX833417156 and 30718464 and 251973523.61%CELSR229489009 and 322933431.61%CELSR316273260 and 33811453 and 322265011.80%CFH15342420 and 157745751.87%ITPR1257147782.25%LTBP1322168151.90%NAV2309975691.61%NEDD925577245 and 324249792.48%PAF132781084 and 324715083.35%POLB336736902.29%RIN128612496 and 18089779 and 330696950.32%aPubMed IDs of literature and the coverage of 18 core-MaxCLK genes in pancancer. Novel potential driver genes and their literature.a PubMed IDs of literature and the coverage of 18 core-MaxCLK genes in pancancer. Moreover, the accuracy of the consensus results could be improved if s-modules (Materials and Methods, Section 2.5) were eliminated. For example, 93 core-MaxCLK genes were predicted with an accuracy of 92.47% with s = 6. We reached similar conclusions under other reference datasets (Supplementary Tables S10-1–S10-5). Meanwhile, it shows that a relatively high s value can be conducive to improving the accuracy but some crucial modules would be ignored. Experimentally, with s = 1, 2, 3, MaxCLK provided effective consensus predictions. 3.3 Pathogenic pathways of core-MaxCLK genes Using an R package clusterProfiler (v4.2.2) (Yu et al. 2012), 123 KEGG pathways were significantly (P < 0.05, q < 0.05) enriched with core-MaxCLK genes (Supplementary Data2, Sheet1). The top 40 most significant pathways are known or potential cancer-related pathways (Fig. 2A). For example, hepatocellular carcinoma (P = 1.63×10-24), prostate cancer (P = 1.35×10-20), chronic myeloid leukemia (P = 5.17×10-19), breast cancer (P = 5.17×10-19), and endocrine resistance (P = 9.90×10-17) are well-known cancer pathways. For others, the PI3K-Akt signaling pathway (P = 7.14×10-14) was recently shown to be related to tumor development and treatment (Shi et al. 2019, Mardanshahi et al. 2021). Aberrant Ras signaling (P = 3.23×10-9) conduction will induce cancer (Degirmenci et al. 2020, Moon 2021) and coordinates (Soleimani et al. 2019) with the MAPK signaling pathway (P = 3.47×10-9). The enrichment of Gene Ontology (GO) terms of core-MaxCLK genes is also shown in Fig. 2B and 1470 GO terms were significantly enriched (Supplementary Data2, Sheet2). Similarly, GO enrichment (Fig. 2C and D) was conducted with the genes of type-2 subnetworks. KEGG and GO enrichment of core-MaxCLK genes. (A) Significantly (P < 1.11×10-7) enriched top 40 KEGG pathways. The horizontal axis represents the ratio of genes in each enriched KEGG pathway and the vertical axis represents significantly enriched pathways. (B) Top 10 significantly enriched GO terms in biological processes (BP, P < 1.05×10-10), cellular components (CC, P < 2.64×10-4), and molecular functions (MF, P < 1.22×10-6). (C, D) The top 10 terms in BP and MF enrichment of 46 genes from type-2 subnetworks. 3.4 Comparison with other methods We compared MaxCLK with CovEx and HotNet2. In dataset A, MaxCLK predicted 154 driver genes with an accuracy of 65.58%. CovEx predicted 236 genes with an accuracy of 52.54%. HotNet2, which is a popular software of the same type, identified 125 genes with an accuracy of 40.80% (Table 2). In contrast, MaxCLK identified 45 unique genes and additional 33 genes common to other methods (Fig. 3A). Table 2.Comparison of MaxCLK, CovEx and HotNet2.aCancerNMSMAMNCSCACNHSHAHDataset A15410165.58%23612452.54%1255140.80%Dataset B15910062.89%23411348.29%1475436.73%Dataset C1559863.23%26112246.74%994545.45%BLCA312477.42%301653.33%1474228.57%BRCA623861.29%1165345.69%501836.00%COADREAD523567.31%191368.42%761925.00%GBM624166.13%432455.81%251144.00%HNSC261973.08%663045.45%932324.73%KIRC683754.41%763242.11%23834.78%LAML432660.47%733852.05%422866.67%LUAD382155.26%522853.85%2405020.83%LUSC211466.67%271866.67%1032726.21%OV634165.08%281553.57%25728.00%UCEC573866.67%201575.00%732736.99%aThe columns NM, SM, AM, NC, SC, AC and NH, SH, AH correspond to the number of predicted genes, sensitivity and accuracy of MaxCLK, CovEx, and HotNet2, respectively. The cancer gene benchmark is NCGv5.0. Comparison of MaxCLK, CovEx and HotNet2.a The columns NM, SM, AM, NC, SC, AC and NH, SH, AH correspond to the number of predicted genes, sensitivity and accuracy of MaxCLK, CovEx, and HotNet2, respectively. The cancer gene benchmark is NCGv5.0. Comparison of MaxCLK and other methods. (A–D) depict Venn diagrams comparing MaxCLK with other methods or driver gene datasets. (E) Setting s = 1, the horizontal axis represents different cancers, and the vertical axis represents accuracy (%). The polylines with different colors denote the accuracy of different methods. (F) Similar to (E), it demonstrates the sensitivity of different methods. (G, H) The accuracy and sensitivity of MaxCLK for various s values. MaxCLK was also compared with other current methods, including MutPanning (Dietlein et al. 2020), dNdScv (Martincorena et al. 2018), OncodriveCLUSTL (Arnedo-Pac et al. 2019), OncodriveFML (Mularoni et al. 2016), MutSigCV (Lawrence et al. 2013), and MEXCOwalk (Ahmed et al. 2020). We used the number of core-MaxCLK genes to intercept gene rankings of these methods and compared their results with that of MaxCLK (Fig. 3B and C). Figure 3D denoted the comparison of MaxCLK with some known driver genes. In addition, Fig. 3E and F respectively presented their accuracy and sensitivity with a default parameter s = 1. Other outcomes were detailed in Supplementary Fig. S4. Furthermore, the accuracy and sensitivity of MaxCLK were illustrated in Fig. 3G and H under various s values across separate datasets. 3.5 Results of MaxCLK on single cancers MaxCLK was applied to the mutation data of each cancer in pancancer dataset A. 317 genes were obtained in total, of which 94 were from at least two cancers and 51 were from at least three cancers. In particular, ATM and MDM2 was identified in 8 and 7 cancers, respectively. Four genes, CHD4, EGFR, ERBB2, and SMARCA2, were identified in 6 cancers (Supplementary Data2, Sheet 3). Using GEPIA2 (Tang et al. 2019), the differential expression of ATM and MDM2 in tumor and normal tissues was visualized (Fig. 4A and C). In addition, the differential expression of genes that appeared more frequently in the single cancer results is also shown in Fig. 4B. The expression levels of CHD4, EGFR, ERBB2, and SMARCA2 were also discrepant in corresponding cancers (Supplementary Fig. S5). ATM showed significant differences for both disease-free survival (Fig. 4D) and overall survival (Fig. 4E) in different expression groups, and MDM2 produced similar results (Fig. 4F and G). The differential expression and survival curve of genes, and the exclusivity or co-occurrence of pathway-modules. (A) Differential expression of ATM in tumor (T, red box) and normal (N, blue box) samples with P < 0.01. The horizontal axis represents cancers, and the vertical axis represents the ATM expression. (B) Differential expression heatmap of predicted driver genes of MaxCLK with frequency ≥5 in 11 datasets. (C) Expression of MDM2 in tumor and normal samples. (D, E) The survival curves of ATM. Similarly, (F, G) The survival curves of MDM2. (H) Corresponding to Supplementary Table S13, the significant (P<0.01) correlations between partial modules and pathways are shown. The horizontal axis denotes pathways, the vertical axis represents modules, and the color value is -log10⁡P, in which positive and negative values indicate co-occurrence (marked +) and mutual exclusivity (marked –), respectively. To discover driver genes for specific cancer, we extracted NCGv7.0 genes from core-MaxCLK genes of each cancer and used recent studies to verify gene-cancer relationship (Supplementary Data2, Sheet 4, 5). MaxCLK outperformed other methods (Table 2 and Fig. 3) and performed well in other reference datasets (Supplementary Table S11). In addition, by comparing the core-MaxCLK genes of each cancer with all the reference datasets, some unrecorded genes (Supplementary Table S12) were found, and some studies have demonstrated their role (Table 3). Table 3.Novel potential driver genes and their literature in single cancer.aCancerGenePMIDCoverageUCECFOSL2115705753.80%23339184OVNEDD9297739029.81%LUADF13A1241426435.41%ITGA2323176292.16%LAMLMAGI3232796160.52%RIN1211024290.52%BRCAA2M332738501.05%NEDD9314628982.88%GBMITGB2343138211.72%KIRCANKHD1296955081.92%aPubMed IDs of literature and the coverage of core-MaxCLK genes in single cancer. Novel potential driver genes and their literature in single cancer.a PubMed IDs of literature and the coverage of core-MaxCLK genes in single cancer. 3.6 Co-occurrence and mutual exclusivity of module-pathway pairs To explore the potential relationship among core-MaxCLK genes of pancancer, for modules identified by MaxCLK with size 4 (3), we selected 424 (363) preliminary prediction modules, which all genes were contained in core-MaxCLK genes (Supplementary Data2, Sheet 6, 7). The top 10 and bottom 10 modules are shown in Supplementary Table S13. Based on different datasets, the Cochran-Mantel-Haenszel (CMH) test on 11 layers is performed in each module-pathway pair. In Fig. 4H, CMH is used to measure the co-occurrence and mutual exclusivity between these modules and 123 enrichment pathways. 53 (15) modules of size 4 (3) have co-occurrence or mutual exclusivity relationship (P < 0.01) with no <90 pathways. For example, modules MYC, NRAS, PIK3CG, PLCE1 (k = 4, index 407), and CBX8, PIK3R1, POLB (k = 3, index 19) were significantly correlated with 112 and 111 pathways, respectively. The top 1 module ERBB4, FLT3, MAP3K1, NEDD9 has the highest IE value and furthermore NEDD9, which is not included NCG or other references, was proven to set permissive conditions for HER2-induced tumorigenesis by a recent study (Purazo et al. 2023). 4 Discussion To find functional driver genes or modules in the large-scale genome of different cancer types, we proposed a new method called MaxCLK for comprehensive consideration of cancer genome data and PPI network to identify cancer driver genes. MaxCLK is based on three basic premises. First, the genes in the identified driver modules are topology related in the PPI network, and driver modules can be found more accurately based on this correlation. Second, the driver modules should include most patients (high coverage). Meanwhile, most patients should have only one mutation in a driver module (high exclusivity). Using three annotated PPI influence networks, HINT + HI2012, iRefIndex and Multinet as inputs, MaxCLK achieves 270 predictions in total with different parameters. Furthermore, the consensus method is used to construct consensus networks based on all the crucial modules identified by MaxCLK in 270 independent predictions and extracts the core-MaxCLK driver genes, which indicate the existence of information communication of different functional modules. We found that the accuracy of core-MaxCLK genes was superior to that of all type-2 subnetworks, and the reason may be that the experimental results of the three PPI networks were not completely consistent. Thus, the accuracy of genes simultaneously identified by the three networks was higher. The core-MaxCLK gene symbols under different s were provided in Supplementary Data3. In addition, the comprehensive indicator IE overcomes the deficiencies of other methods in interpretability. As a higher IE value indicates a more complicated mutation situation, a module with high IE value tends to have superior exclusivity and coverage. With core-MaxCLK genes, we performed KEGG pathway and GO analyses and found that many genes were repeatedly scattered in different pathways (Supplementary Fig. S6), which indicated a complex association among them. In addition, applied to a single cancer, the identification performance of MaxCLK was better than other similar or computational methods. The modules in core-MaxCLK genes were performed CMH test with enriched pathways and many interesting pathway-module pairs were found. While MaxCLK has found some proven driver genes, there are still some potential driver genes undetected, due to the insufficiency of characteristics or other regulatory mechanisms of driver genes. Fortunately, ever-increasing genetic genomics and PPIs data will provide a lasting opportunity to improve and rectify our method. In future, we will consider the fusion of multiomics data of tumor and more detailed tumor subtypes and extend the application of MaxCLK into lineage-specific or individual patient rather than current pancancer field. Our research also represents considerable innovation in the method of finding the driver module compared with previous research. For example, CovEx applies binary linear programming (BLP) to find a globally optimal solution with mutation data. However, MaxCLK uses a modified maximal clique algorithm to find all feasible solutions, which is much more efficient than BLP. In our research, for the same data, MaxCLK obtained more novel driver modules, which were ignored by previous methods. Besides of accuracy and sensitivity, F1 score (Supplementary Text S3) was used to measure the identification ability of driver genes and MaxCLK outperformed CovEx and HotNet2 again (Supplementary Table S14). Supplementary Material Click here for additional data file. Acknowledgements Our results are part based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga. We extend special thanks to Professor Guojun Li from Shandong University for his guidance on paper writing. Contributor Information Jian Liu, School of Mathematics and Statistics, Shandong University at Weihai, Weihai, Shandong 264209, China. Fubin Ma, School of Mathematics and Statistics, Shandong University at Weihai, Weihai, Shandong 264209, China. Yongdi Zhu, School of Mathematics and Statistics, Shandong University at Weihai, Weihai, Shandong 264209, China. Naiqian Zhang, School of Mathematics and Statistics, Shandong University at Weihai, Weihai, Shandong 264209, China. Lingming Kong, Marine College, Shandong University at Weihai, Weihai, Shandong 264209, China. Jia Mi, Precision Medicine Research Center, School of Pharmacy, Binzhou Medical University, Yantai, Shandong 264003, China. Haiyan Cong, Department of Central Lab, Weihai Municipal Hospital, Weihai, Shandong 264209, China. Rui Gao, School of Control Science and Engineering, Shandong University, Jinan, Shandong 250100, China. Mingyi Wang, Department of Central Lab, Weihai Municipal Hospital, Weihai, Shandong 264209, China; Department of Central Lab, Weihai Municipal Hospital, Cheeloo College of Medicine, Shandong University, Weihai, Shandong 264200, China. Yusen Zhang, School of Mathematics and Statistics, Shandong University at Weihai, Weihai, Shandong 264209, China. Supplementary data Supplementary data are available at Bioinformatics online. Conflict of interest None declared. Funding This work was supported by the National Natural Science Foundation of China [61877064, U1806202, 61533011]. References Associated Data Supplementary Materials Click here for additional data file.