Mapping the Functional Network of Human Cancer Through Machine Learning and Pan-Cancer Proteogenomics

Abstract

Large-scale omics profiling has uncovered a vast array of somatic mutations and cancer-associated proteins, posing significant challenges for their functional interpretation. Here we present a network-based approach centered on FunMap, a pan-cancer functional network constructed using supervised machine learning on extensive proteomics and RNASeq data from 1,194 patients spanning 11 cancer types. Comprising 10,525 protein-coding genes, FunMap connects functionally associated genes with unprecedented precision, surpassing traditional protein-protein interaction maps. Network analysis identifies functional protein modules, reveals a hierarchical structure linked to cancer hallmarks and clinical phenotypes, provides deeper insights into established cancer drivers, and predicts functions for understudied cancer-associated proteins. Additionally, applying graph neural network-based deep learning to FunMap uncovers drivers with low mutation frequency. This study establishes FunMap as a powerful and unbiased tool for interpreting somatic mutations and understudied proteins, with broad implications for advancing cancer biology and informing therapeutic strategies.

INTRODUCTION

Advancements in next-generation sequencing and mass spectrometry have transformed cancer research. Large-scale initiatives like The Cancer Genome Atlas (TCGA), the International Cancer Genome Consortium (ICGC), and the Clinical Proteomic Tumor Analysis Consortium (CPTAC) have significantly deepened our understanding of cancer, revealing a vast array of somatic mutations and cancer-associated proteins. These advancements present new challenges in the functional interpretation of identified mutations and proteins, especially for the numerous low-frequency mutations¹ and understudied proteins².

Protein-protein interaction networks have been instrumental in prioritizing somatic mutations and predicting the functions of uncharacterized proteins^3–5. However, many of the known interactions were identified in non-cancer contexts, limiting their relevance to cancer research. Recent efforts have started to address this gap by mapping interactions for selected proteins in specific cancer cell lines^6,7. Despite these advances, unbiased, genome-scale identification of protein-protein interactions across diverse cancer types remains a daunting task. Moreover, in vitro cell line models have inherent limitations, such as the absence of the tumor microenvironment. mRNA co-expression has also been used to infer functional associations, but with varied success^8,9. Studies have shown that protein expression data are more closely aligned with gene function, and that protein co-expression is a much stronger predictor of functional association than mRNA co-expression^10–14.

In this paper, we introduce FunMap, a functional network of 10,525 genes constructed using a supervised machine learning method that integrates proteomics and RNAseq data from 11 cancer types, recently harmonized by the CPTAC pan-cancer working group¹⁵. FunMap connects functionally related genes with unprecedented precision, surpassing existing protein-protein interaction networks. Through network analysis, FunMap uncovers protein modules and a hierarchical modular organization linked to cancer hallmarks and clinical phenotypes, predicts the functions of understudied cancer proteins, offers deeper insights into established cancer drivers, and identifies drivers with low mutation frequency. To facilitate broader use in cancer research, we provide an interactive web application (https://funmap.linkedomics.org/) and source code (https://github.com/bzhanglab/funmap).

RESULTS

Protein co-expression strongly predicts co-functionality

We used mass spectrometry-based proteomics and RNASeq data from 11 tumor cohorts (Supplementary Table 1) to quantify gene co-expression at the protein and mRNA levels, respectively. Cancer types included breast invasive carcinoma (BRCA), clear cell renal cell carcinoma (CCRCC), colon adenocarcinoma (COAD), glioblastoma (GBM), hepatocellular carcinoma (HCC), head and neck squamous cell carcinoma (HNSCC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LSCC), ovarian serous cystadenocarcinoma (OV), pancreatic ductal adenocarcinoma (PDAC), and uterine corpus endometrial carcinoma (UCEC). Tumor samples ranged from 83 to 159 per cohort, and five cancer types also had sufficient normal samples with proteomics and RNASeq data, leading to 16 proteomics and 16 RNASeq datasets (Figure 1A). Each proteomics dataset included 7,961 to 11,815 genes (Figure 1B), with a median of 10,441 and a union of 14,070 genes, among which 6,602 were identified across all 16 datasets and 10,024 in 10 or more (Figure 1C). Each RNASeq dataset included 17,733 to 19,113 genes (Figure 1B), with a median of 18,740 and a union of 19,855 genes, among which 15,603 were identified across all 16 datasets (Figure 1C).

Figure 1: Protein co-expression is a strong predictor of gene co-functionality.

A) Proteomics and RNASeq data from tumor (T) and normal (N) samples across 11 cancer cohorts used in this study. Number of samples, n, are indicated in the plot. B) Numbers of quantified proteins and mRNAs in individual datasets. C) Numbers of proteins and mRNAs quantified across datasets. The numbers inside blue shaded boxes indicate the numbers of datasets with quantitative data. D) Log likelihood ratios (LLRs) quantifying functional relevance of the top ranking gene pairs based on Pearson’s correlation coefficient, from top 10k to 300k in each dataset. E) Distributions of LLRs of the gene pairs with a given mRNA co-expression (x axis) and protein co-expression (y axis) pattern in the 11 tumor datasets. The density plots on the top and right visualizes mRNA and protein co-expression distributions, respectively.

To assess the relationship between gene co-expression and co-functionality, we used a previously published “gold standard” derived from the Reactome pathway database¹². This gold standard defines gene pairs co-annotated in the same ‘detailed’ pathway (≤ 200 genes) as positive pairs, and those without shared pathway annotations as negative pairs. It includes 205,284 positive and 11,327,528 negative gene pairs. This extensive dataset allowed us to quantify the functional relevance of any specific set of gene pairs by calculating the log likelihood ratio (LLR), with higher LLRs indicating stronger evidence of functional relevance (Methods).

For each proteomics and RNASeq dataset, we ranked gene pairs by their Pearson’s correlation coefficients (PCCs) and computed LLRs for the top 10,000 to 300,000 pairs. LLRs showed a decreasing trend across all datasets (Figure 1D). In most tumor datasets, proteomics data consistently yielded higher LLRs than RNASeq, indicating greater functional relevance. However, in normal datasets, proteomics LLRs were similar to or lower than RNASeq LLRs. This may be explained by the low inter-sample heterogeneity in normal protein datasets (Extended Data Fig. 1), hindering the detection of correlations between functionally related genes. The low inter-sample heterogeneity likely also contributed to the lower LLRs in normal protein datasets compared to tumor protein datasets. Interestingly, despite lower heterogeneity in tumor protein datasets compared to tumor RNA datasets (Extended Data Fig. 1), the higher LLRs in the protein data suggest that this level of heterogeneity is sufficient for detecting functionally relevant correlations.

To delve deeper into how mRNA and protein co-expression patterns relate to gene co-functionality within the tumor datasets, we grouped gene pairs into 400 two-dimensional bins based on their correlations in both proteomics and RNASeq data and then computed LLRs for each bin (Figure 1E). Gene pairs with higher protein correlation consistently displayed elevated LLR scores, even when mRNA correlation was moderately positive or even negative. While gene pairs with higher mRNA correlation also tended to have higher LLR scores, these higher scores were more frequently observed in areas where there were strong correlations at both mRNA and protein levels. Together, these results demonstrate that while both protein and mRNA correlations indicate gene co-functionality, protein correlation is a much stronger predictor.

A machine learned functional map

We utilized supervised machine learning to integrate the diverse predictive signals from all 32 proteomics and RNASeq datasets to construct a comprehensive functional network. Normal sample datasets were included because they were derived from tumor-adjacent normal tissues, which provide clinically relevant biological information¹⁶. Despite varying magnitudes, each dataset displayed functional relevance (LLR>1, Figure 1D). To account for differences in sample size and inter-sample heterogeneity across datasets, we computed PCC-based mutual rank (MR) scores for all gene pairs within each dataset (Methods), as MR is a robust metric for assessing gene co-expression across diverse datasets¹⁷.

We used 50% of the gold standard positive and negative gene pairs as training data to build an Extreme Gradient Boosting (XGBoost) model, using MR scores from the 32 datasets as features to distinguish the positive and negative gene pairs (Methods). Feature importance analysis revealed that tumor protein features contributed the most (61.5%), followed by tumor RNA (20.7%), normal RNA (9.0%), and normal protein (8.8%) (Extended Data Fig. 2). Among individual datasets, the tumor protein data from LSCC contributed the highest.

The trained model was applied to all 98,975,415 gene pairs, which were then sorted by predicted probabilities. LLRs were computed using the remaining 50% set-aside gold standard gene pairs for the top-ranked gene pairs, from the top 50,000 to 250,000 (Figure 2A). Similarly, we trained two additional XGBoost models using only the 16 proteomics datasets or the 16 RNASeq datasets and plotted the LLR curves. For comparison, we included LLR curves from a baseline method based on average PCCs across the 32 datasets and the LSCC tumor protein data alone. Interestingly, the LSCC tumor proteomics dataset performed as well as or better than the combined RNASeq datasets, underscoring the pivotal role of protein-level regulation in coordinating gene function. The XGBoost model combining all datasets clearly outperformed the baseline method based on average PCCs, highlighting the advantage offered by machine learning. It also outperformed the model combining only the proteomics datasets, which in turn outperformed the model combining only the RNASeq datasets or the LSCC tumor proteomics data alone, demonstrating the value of data integration in gene co-functionality prediction.

Figure 2: FunMap has high functional relevance, deep proteome coverage, and a scale-free, modular, and small-world network topology.

A) A supervised machine learning model combining all 32 datasets (ALL_RNA_PRO (xgboost)) achieved higher log likelihood ratios (LLRs) across the whole range of top scoring gene pair numbers from 50K to 250K compared with the models combining only proteomics datasets (ALL_PRO (xgboost)), only RNASeq datasets (ALL-RNA (xgboost)), using average Pearson’s correlation coefficient (PCC) across the 32 datasets (ALL_RNA_PRO (average PCC)), or using the PCCs from the LSCC tumor proteomics data alone (LSCC-T_Pro (PCC)). Applying an LLR cutoff of 3.912 (LR=50) to results from the all inclusive model produced a network with 10,525 genes and 196,800 edges, which was named FunMap. B) Scatter plot comparing functional relevance (y axis) and proteome coverage (x axis) of FunMap and other networks. C) Gene overlap between FunMap and other networks. D) Edge overlap between FunMap and other networks. E) Boxplots depicting proportion of edges connecting proteins with consistent significant over or under-expression in tumors vs normal samples (n = 5 cohorts) for FunMap and other networks. For boxplots, centerline indicates the median, box limits indicate upper and lower quartiles, whiskers indicate the 1.5 interquartile range. P-values derived from paired t-test followed by adjustment based on Holm’s method. F) Degree distribution of FunMap. G-I) Plots comparing average clustering coefficient (F), density (G), and average shortest path length (H) of FunMap and other networks.

Applying an LLR cutoff of 3.912, i.e., likelihood ratio (LR) of 50, to the results from the XGBoost model combining all 32 datasets yielded a functional association network with 10,525 genes and 196,800 edges, which was named FunMap (Supplementary Table 2). With an LR of 50, edges are 50 times more likely to connect functionally associated gene pairs than unrelated pairs. We compared FunMap’s functional relevance and proteome coverage to other networks used in systems biology studies (Figure 2B). FunMap and the ProHD¹², both based primarily on protein co-expression, showed similar LR scores (50 and 56, respectively), though ProHD covered only 2,680 genes. These scores were much higher than those of BioPlex¹⁸ (LR=28), HuRI¹⁹ (LR=10), HI-Union¹⁹ (LR=10), and BioGRID²⁰ (LR=14), networks based on experimentally obtained protein-protein interaction data or curated protein and genetic interaction data. While FunMap showed higher proteome coverage than HuRI and HI-Union, BioPlex and BioGRID covered more genes (13,854 and 17,259, respectively). The STRING network²¹ had the highest LR score (LR=187) and deep coverage of 16,351 genes, but unlike the other purely data-driven networks, it incorporated existing knowledge during network construction, including that used for our evaluation.

Genes in FunMap overlapped significantly with those in other networks (Figure 2C), but its edges showed limited overlaps (Figure 2D), indicating a substantial number of additional functional associations. While tumor versus normal differences were not used in FunMap’s construction, analysis of the five cancer types with normal samples revealed that 60% to 74% of FunMap edges connected genes with consistent significant over- or under-expression in tumors (adjusted p < 0.01, Wilcoxon Rank Sum test, Figure 2E). These percentages significantly exceeded those found in the other networks (p < 0.001, Wilcoxon Signed Rank test) (Figure 2E), suggesting a stronger connection of FunMap to cancer.

FunMap showed a power-law degree distribution (Figure 2F), indicating a scale-free topology with highly connected hubs. Compared to other networks, FunMap was characterized by relatively higher average clustering coefficient (similar to STRING), relatively higher density (similar to BioGRID), and the highest average shortest path length (Figure 2G-I). Together, these results suggest high functional relevance, cancer relevance, and proteome coverage of FunMap, as well as its scale-free, modular, and small-world properties.

Cancer-associated dense modules

A high clustering coefficient of FunMap suggests that genes in the network tend to form clusters or modules. To assess FunMap’s ability to connect genes encoding proteins in the same functional module, we used the CORUM database²², which contains 5,204 manually annotated mammalian protein complexes involving 5,299 genes. Among the 196,800 edges in FunMap, 14,401 (7.3%) connected genes encoding proteins in the same CORUM complex (Figure 3A). Strikingly, both the absolute count and percentage of the edges overlapping with CORUM in FunMap were higher than those in the BioPlex network (6,747, 4.4%) (Figure 3A). As BioPlex was designed to experimentally identify protein complexes through affinity purification-mass spectrometry (AP-MS), these results underscore FunMap’s potential in unveiling tightly co-regulated functional modules.

Figure 3: FunMap reveals known and previously unidentified dense modules associated with cancer biology and clinical phenotype.

A) Overlap among gene pairs in FunMap, BioPlex, and gene pairs encoding proteins in the same CORUM complex. B) Examples of CORUM complexes displaying robust connectivity among their complex members in FunMap but not in BioPlex. C) Numbers of the de novo predicted FunMap dense modules with a significant overlap with CORUM complex, BioPlex complex, or Gene Ontology term (p<0.05, Fisher’s exact test, blue shaded sections). D) A tumor overexpressed, ECM-associated dense module (Clique 160). Edge color indicates lack of overlap in BioGRID, BioPlex, HI-union, STRING, and CORUM (pink) or overlap in any of these resources (gray). E) Boxplots comparing average protein abundance of Clique 160 in tumor and normal samples demonstrating tumor overexpression in five cancer cohorts. Number of samples, n, are indicated in parenthesis. P-values determined by two-sided Wilcoxon rank-sum test. F) Kaplan-Meier plots depicting overall survival (OS) difference in CCRCC, HCC, and LUAD patients stratified by median value of the average abundance of proteins in Clique 160. Number of samples, n, are indicated on each plot. Logrank p-values and hazard ratio (HR) shown with 95% confidence intervals derived from Cox-proportional hazard models. G) A tumor underexpressed, cell adhesion-associated dense module (Clique 46). Edge color same as (D). H) Boxplots comparing average protein abundance of Clique 46 in tumor and normal samples demonstrating tumor underexpression in five cancer cohorts. Number of samples, n, are indicated in parenthesis. P-values determined by two-sided Wilcoxon rank-sum test. I) Kaplan-Meier plots depicting OS difference in HCC patients stratified by median value of the average abundance of proteins in Clique 46. Number of samples, n, are indicated in the plot. P-values and HR same as in (F). Significance is indicated as *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. For boxplots, centerline indicates the median, box limits indicate upper and lower quartiles, whiskers indicate the 1.5 interquartile range, and number of samples per group indicated in parentheses.

Some CORUM complexes associated with cancer-related processes displayed robust connectivity among their members in FunMap but not in BioPlex, such as complexes involved in cell cycle and DNA replication, gene expression and regulation, signal transduction, cell motility, and innate immunity (Figure 3B). Unlike BioPlex, which used data from only two in vitro cell lines, FunMap utilized data from over 1,000 human tumor samples, making it potentially more effective in uncovering functional modules relevant to in vivo cancer biology.

To extend our analysis beyond CORUM complexes, we applied the iterative clique enumeration (ICE) algorithm²³ to FunMap (Methods). This algorithm identifies relatively independent cliques, which are fully connected subnetworks (dense modules) with limited overlap to each other. Through this approach, we identified 281 dense modules, each with five or more genes (Supplementary Table 3). Of these, 130 (46%) overlapped significantly with CORUM complexes, an additional 37 (13%) with BioPlex complexes, and another 49 (17%) Gene Ontology (GO) annotations (FDR <0.05, Fisher’s exact test followed by Benjamini-Hochberg adjustment, Figure 3C, Supplementary Table 3). These results emphasize the functional coherence of genes within these de novo identified dense modules.

To evaluate the cancer relevance of these dense modules, we compared the average standardized protein abundance between tumor and normal samples for each of the five cancer types (Supplementary Table 3). Of the 276 modules with sufficient data for statistical analysis, 273 showed significantly different abundance in tumors compared with normal samples in at least one cancer type (adjusted p < 0.01, Wilcoxon rank-sum test followed by Benjamini-Hochberg adjustment). Notably, 43 of the 273 (16%) had no significant overlap with CORUM, BioPlex, or GO annotations (adjusted p > 0.01, Hypergeometric test), and 203 (74%) had more than half of their edges unique to FunMap compared to other networks (Supplementary Table 3). These observations underscore the value of FunMap in uncovering previously unrecognized, cancer relevant dense modules.

Seventy-eight dense modules showed significant differential expression across all five cancer types, with 36 (46%) having less than 25% edge overlap with the other networks (Extended Data Fig. 3A). Many overexpressed modules were enriched in processes related to replication and proliferation. Moreover, three highly overexpressed modules (Clique 160, Clique 96, and Clique 54) were associated with extracellular matrix (ECM) organization (Figure 3D-E and Extended Data Fig. 3B-E), and higher module levels were significantly associated or trending toward worse overall survival (OS) in various cancer types (Figure 3F, Extended Data Fig. 3F-G, Supplementary Table 3). Fewer modules were underexpressed, and those related to cell adhesion (Clique 46 and Clique 17, Figures 3G-H, Extended Data Fig. 3A) may contribute to increased cell motility and tumor aggressiveness. This was supported by tumors with underexpression of Clique 46 showing worse OS in HCC (Figure 3I).

In summary, these results demonstrate the ability of FunMap to identify functionally and clinically relevant dense modules. Importantly, many of these modules were associated with cancers of diverse histological origin but had limited overlap with other networks, highlighting a unique connection of FunMap to cancer biology and disease progression.

Hierarchical modular organization linked to cancer hallmarks

The coexistence of scale-free topology (Figure 2F) and a high clustering coefficient (Figure 2G) in FunMap indicates a hierarchical modular organization, where genes form smaller modules that combine into larger ones across multiple scales²⁴. Using NetSAM²⁵, a specialized computational tool for uncovering the hierarchical organization in biological networks, we identified eight hierarchical levels and 255 modules with at least 20 genes in FunMap (Figure 4 and Supplementary Table 4). Of these, 243 (95%) significantly overlapped with at least one GO annotation (FDR<0.05, Fisher’s exact test followed by Benjamini-Hochberg adjustment) (Supplementary Table 4), indicating their functional coherence.

Figure 4: Hierarchical modular organization of FunMap statistically linked to cancer hallmarks.

Hierarchical modular organization of FunMap. Nodes represent NetSAM derived modules with node size proportional to module size. Nodes are colored based on the significance of pan-cancer tumor vs normal protein abundance difference and ordered according to the significance levels of level 2 modules for each branch. Top enriched cancer hallmark annotations for the 10 largest level 2 branches are annotated in gray text. Green text indicates biological processes highlighted in Figure 5.

We focused on the enriched GO annotations that have been previously linked to cancer hallmarks^26,27 (Supplementary Table 4). The top ten largest branches were associated with various hallmarks (Figure 4, Methods), including tumor microenvironment-related hallmarks like avoiding immune destruction and tumor-promoting inflammation, with the largest branch linked to tumor promoting inflammation (1,118 genes). These findings underscore the strength of using tumor-derived data in network construction, which can capture complex, biologically significant information that may be missed in cell line-based protein-protein interaction networks.

To assess the clinical significance of these modules, we calculated meta-p values for differential expression between tumors and normal samples across the five cancer cohorts (Supplemental Table 4). Tumor-overexpressed branches were linked to hallmarks such as enabling replicative immortality, genome instability and mutation, sustaining proliferative signaling, evading growth suppressors, avoiding immune destruction, resisting cell death, and activating invasion and metastasis (Figure 4). A detailed examination of these branches revealed their hierarchical functional organization. For example, the level 3 module L3_M55, associated with “protein folding” and “protein transport”, was divided into two level 4 modules: L4_M58 (“protein folding”) and L4_M59 (“protein transport”) (Figure 5A). The latter further split into level 5 modules for “protein targeting to ER” (L5_M51) and “ER to Golgi vesicle-mediated transport” (L5_M50). In tumor cells, ongoing replication, growth, and genetic aberrations disrupt protein homeostasis²⁸, increasing the need for protein folding and related protein transport to resist cell death and avoid immune destruction, two hallmarks linked to this branch. Overexpression of the “protein folding” module (L4_M58) was associated with worse OS in CCRCC (Figure 5B), with similar trends in HNSCC, LUSCC, and LUAD (Supplementary Table 4), supporting its pro-tumor role.

Figure 5. In-depth analysis of selected FunMap branches and their clinical associations.

A) Hierarchical organization of five modules related to protein folding and protein transport. Node color and size of the modules are the same as in Figure 4. P-values determined by hypergeometric test. B) Kaplan-Meier plots depicting overall survival (OS) difference in CCRCC patients stratified by median value of the average abundance of proteins in module L4_M58. Number of samples, n, are indicated in the plot. Logrank p-values and hazard ratio (HR) shown with 95% confidence intervals derived from Cox-proportional hazard models. C) Hierarchical organization of modules in an angiogenesis and metastasis associated branch. Node color and size are the same as in (A). Node outline indicates enrichment of ECM genes with pro- vs anti-angiogenic roles. P-values determined by hypergeometric test. D-E) Kaplan-Meier plots depicting OS difference in HCC patients stratified by median value of the average abundance of proteins in module L3_M19 (D) or in HCC and HNSCC patients stratified by L4_M13 abundance (E). Number of samples, n, are indicated in the plots. P-values derived as in (B).

Tumor-underexpressed branches were linked to cancer hallmarks including deregulating cellular energetics, tumor promoting inflammation, inducing angiogenesis, and activating invasion and metastasis (Figure 4). Although the association with tumor-promoting hallmarks initially seemed counterintuitive, further examination provided deeper insight. For example, the branch rooted in L2_M12, associated with inducing angiogenesis and activating invasion and metastasis, was enriched in functional categories including ECM structure, cell adhesion, and angiogenesis, with modules deeper within the branch showing more specialized roles (Figure 5C). While L2_M12 was overall underexpressed, it was divided into an underexpressed module (L3_M19) tied to antitumor functions like cell adhesion and an overexpressed module (L3_M18) linked to pro-tumor functions like angiogenesis. Both overexpressed and underexpressed modules were enriched with ECM components, but anti-angiogenic ECM components were enriched in underexpressed modules, while pro-angiogenic ECM components were enriched in overexpressed modules (Supplementary Table 4). Interestingly, underexpressed dense modules related to cell adhesion (Cliques 17 and 46) were entirely covered by L3_M19, whereas the overexpressed dense modules related to ECM (Cliques 54, 96, 160) were found entirely within L3_M18. Consistent with the good-prognosis association observed for Clique 46 (Figure 3I), higher expression of L3_M19 was correlated with a longer OS in HCC (Figure 5D), with a similar trend observed for LUAD and CCRCC (Supplementary Table 4). In contrast, higher expression of the tumor overexpressed module L4_M13, which was under L3_M18 and included most components from the poor-prognosis Cliques 54, 96, and 160 (Figure 3F, Supplementary Figures 3F-G), was correlated with a shorter OS in HCC (Figure 5E) and other cancer types (Supplementary Table 4). Thus, the hierarchical module analysis not only reinforced the clique-based analysis results but also revealed the broader functional context and systematic organization of the dense modules.

In summary, network analysis revealed a hierarchical modular organization of FunMap, in which the major branches were statistically aligned to cancer hallmarks, supported by both functional analysis and the examination of clinical outcomes.

Connecting somatic mutations to protein modules

A major goal of cancer proteogenomics is to understand how somatic mutations impact the cancer proteome. Previous studies have used univariate analysis to examine the cis- and trans-effects of individual mutations^29,30. Here we used a machine learning approach to simultaneously model the impact of all significant mutations on individual functional modules in FunMap to better capture the complexity of biological systems (Methods).

We identified 77 genes that were significantly mutated (q-value < 0.1) in at least one of the 10 CPTAC cancer types. For each of the 536 modules identified by NetSAM or ICE, we trained an XGBoost model to predict average standardized protein abundance based on mutation status of the 77 genes. In a 5-fold cross validation based on data from 1,021 tumors across 10 cancer types, 32 modules showed a non-random correlation (Pearson’s correlation coefficient > 0.2, p < 0.00001) between predicted and actual abundance, suggesting a significant connection between mutation status and protein abundance of these modules. Feature importance analysis highlighted TP53 as a top predictor across all 32 modules, consistent with its role as a master regulator, while some other genes were specific to certain modules (Figure 6A, Supplementary Table 5).

Figure 6: Connecting somatic mutations to functional protein modules.

A) Heatmap depicting the most important mutant genes in predicting protein abundance of 32 modules. The modules were clustered based on membership similarity. Heatmap color corresponds to relative importance in the XGBoost model. B) Associations defining module L2_M40. Node size corresponds to node degree. C) Boxplot comparing L2_M40 protein abundance in samples with and without KEAP1 mutations in selected cancer cohorts. Number of samples, n, are indicated in parenthesis. P-values were derived from two-sided Wilcoxon rank-sum test. D) Associations defining module L3_M58. Node size corresponds to node degree. E) Boxplot comparing L3_M58 protein abundance in samples with and without TP53 mutations across cancer cohorts. Number of samples, n, are indicated in parenthesis. P-values were derived from two-sided Wilcoxon rank-sum test. F) Clique 254, a cancer/testis antigens-associated dense module. Significance is indicated as *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant. For boxplots, centerline indicates the median, box limits indicate upper and lower quartiles, whiskers indicate the 1.5 interquartile range, and number of samples per group indicated in parentheses.

Hierarchical clustering of the 32 modules based on pairwise membership overlap revealed a predominant cluster with 19 modules (highlighted by red lines in the dendrogram in Figure 6A). These modules comprised genes involved in the cell cycle or cellular division processes (Supplementary Tables 3-4). The most distinctive mutant genes defining this cluster included RB1, ACVR2A, SETD1B, and TBC1D23. Mutations or deletions of RB1 are common across various cancers and disrupt cell cycle control, leading to uncontrolled cell proliferation³¹. While the roles of ACVR2A, SETD1B, and TBC1D23 are less extensively documented, mutations in these genes have been implicated in cell proliferation and tumorigenesis^32,33.

Another cluster of three modules were dominated by KEAP1 mutations (highlighted by pink lines in the dendrogram in Figure 6A), with L2_M40, a module comprising 22 genes (Figure 6B), showing a particularly strong effect. L2_M40 exhibited increased protein abundance in tumors with KEAP1 mutations across all cancer types that had a sufficient number of KEAP1 mutant samples for statistical comparison (Figure 6C). Moreover, expression of genes in this module showed the highest degree of co-regulation at both mRNA and protein levels (average PCC > 0.5) in these four cancer types compared to the other cancer types (Extended Data Fig. 4A). Importantly, all genes in the module are known targets of NRF2^34–41, which is activated by loss-of-function mutations in KEAP1, the gene encoding an inhibitor of NRF2. Therefore, this example serves as a strong positive control for our prediction.

Despite its broad importance, TP53 mutations showed the strongest importance for Module C253 and Module L3_M58 (Figure 6A). Module L3_M58, comprising 51 genes including highly interconnected COP9 signalosome subunits (Figure 6D), showed decreased protein abundance in TP53-mutant tumors across nine out of the 10 cancer types, with a statistically significant decrease in five (Figure 6E). Notably, gene expression in this module was more co-regulated at the protein level than at the RNA level in most of the cancer types (Extended Data Fig. 4B). The COP9 signalosome is known to promote p53 degradation by targeting it for ubiquitination⁴². Our data suggest a negative feedback loop in which wild-type p53 activates the signalosome to suppress p53 levels, and the process is disrupted by TP53 mutations, leading to increased mutant p53 accumulation. This is consistent with the elevated p53 levels observed in TP53-mutant tumors (Extended Data Fig. 4C).

Some modules, like C254, lacked a dominant predictor (Figure 6A). This module, comprising four MAGE family cancer/testis (C/T) antigens and a testis-specific protein DCAF4L2⁴³ (Figure 6F), showed no significant associations with any top-ranked mutant genes in univariate analysis. However, several top predictors, such as PBRM1, ATRX, TP53, and KDM5C, have been linked to immunosuppression and immunotherapy response^44–48, aligning with the role of C/T antigens in triggering immune responses.

In summary, our machine learning approach effectively connected somatic mutations with protein abundance across various functional modules. While some modules had clear dominant predictors and others did not, our models consistently identified key mutant genes whose functions aligned with the overarching function of the modules, demonstrating a clear functional basis for our predictions.

Illuminating understudied cancer proteins

Despite the massive disparity in our knowledge of individual genes—ranging from 9,282 publications in the Gene Reference Into Function (GeneRIF) database for TP53 to 0 publications for 700 “dark” genes—protein degrees in FunMap (i.e., the number of edges) were comparable across the entire spectrum of knowledge depth (Figure 7A), offering a significant opportunity to illuminate understudied genes. Notably, while known cancer driver genes were concentrated among well-studied genes, proteins differentially expressed between tumor and normal samples, based on a meta-analysis of five cancer types, were evenly distributed across the proteome, including the 700 dark genes with no publications (Figure 7A, Supplementary Table 6). Specifically, 125 of these dark genes were highly significantly overexpressed in tumors, whereas 92 were highly significantly underexpressed (meta p value < 1.0e-16, Figure 7B).

Figure 7: FunMap predicts functions of understudied proteins.

A) Heatmap of the adjacency matrix of FunMap with genes sorted based on Gene Reference Into Function (GeneRIF) counts. Genes with a 0 GeneRIF count are defined as “dark genes”. Edge count depicts the log2 count of the number of edges per gene. Cancer driver annotation indicates whether a gene is annotated as a cancer gene in the Cancer Gene Census database. Tumor vs normal annotation plots the signed -log10 meta-p-value comparing protein abundance in tumor vs normal across cancer cohorts. Positive sign indicates higher in tumor and negative sign indicates lower in tumor compared to normal. B) Heatmap depicting signed -log₁₀ meta p-values (p < 1.0e-16) computed as described in (A). Yellow text indicates genes analyzed in subsequent panels. C) Proportions of the dark genes with significantly enriched GO terms in enrichment analysis of the network neighborhood. P-values derived from hypergeometric test and FDR adjusted p-values derived by Benjiamini Hoochberg method. D) Network neighborhood of CXorf38 with genes associated with the enriched GO terms highlighted. E-F) Relationship between protein abundance of CXorf38 and RNA-seq inferred ESTIMATE ImmunoScore in HNSCC (E) and LSCC (F) tumors. Number of samples, n, are indicated in the plots. P-values were derived from two-sided Spearman’s rank correlation. Shaded area depicts the 95% confidence interval. G) Network neighborhood of MAB21L4 with genes associated with the enriched GO term highlighted. H-I) Protein abundance (log2 MS1 intensity) of MAB21L4 by histological tumor grade in HNSCC (H) and LUAD (I) tumors. Number of samples, n, are indicated in parenthesis. P-values were derived from two-sided Jonckheere-Terpstra test. For boxplots, centerline indicates the median, box limits indicate upper and lower quartiles, whiskers indicate the 1.5 interquartile range, and number of samples per group indicated in parentheses.

To gain functional insights into the 700 dark genes, we used the network topology analysis algorithm in WebGestalt⁴⁹ to establish a neighborhood of 50 genes for each dark gene and performed GO enrichment analysis (Method). We found significant enrichment in biological process for 76.2% of the genes, in molecular function for 74.5%, and in cellular component for 65.5% (FDR<0.05, Fisher’s exact test followed by Benjamini-Hochberg adjustment) (Figure 7C). This analysis connected 496 of the 700 dark genes, including 200 shown in Figure 7B, to at least one GO annotation. Although these genes lack publication records in GeneRIF, 315 have existing GO annotations. Of these, 183 (58%) had their top 10 predicted GO terms overlap with one or more existing annotations. This high overlap, compared to just 0.63 from random gene sets, represents a 290-fold increase, underscoring the effectiveness of our approach in predicting gene function.

The dark genes RBM34 and RBM12B were among the most significantly overexpressed genes in tumors (meta p < 1.0e-100, Figure 7B, Extended Data Fig. 5A), consistent with their frequent somatic amplification across various cancers (Extended Data Fig. 5B). Both genes encode RNA-binding motif (RBM) proteins, though their functions have not been experimentally characterized. The network neighborhood of RBM34 was enriched for genes involved in rRNA processing (Extended Data Fig. 5C), whereas that of RBM12B was enriched for genes associated with RNA splicing (Extended Data Fig. 5D). This analysis connected their amplification and overexpression to distinct functional roles, supported by computational inference from the GO consortium based on an orthogonal phylogenetic approach⁵⁰.

The dark gene CXorf38 was significantly overexpressed in tumors compared to normal samples in four out of the five cancer types (meta p=8.6e-31, Extended Data Fig. 6A). Its network neighborhood was enriched for genes associated with cytokine-mediated signaling pathway, major histocompatibility complex (MHC) protein binding, and proteasome complex (Figure 7D), suggesting an immune function. As supporting evidence, CXorf38 protein levels correlated significantly with the immune infiltration scores computed based on RNASeq data in most CPTAC cancer types (Figure 7E-F, Extended Data Fig. 6B). Moreover, single cell data from the Human Protein Atlas show that CXorf38 is highly expressed in immune cells (Extended Data Fig. 6C), reinforcing its inferred immune role.

The dark gene MAB21L4 was significantly underexpressed in tumors in three cancer types (meta p=9.9e-56) (Extended Data Fig. 6D). Its network neighborhood was enriched for genes associated with epithelial cell differentiation (Figure 7G), the suppression of which plays a critical role in tumorigenesis. Remarkably, MAB21L4 protein abundance was lower in poorly differentiated tumors (G3) compared to well differentiated (G1) and moderately differentiated (G2) tumors in both HNSCC and LUAD (Figure 7H-I). These findings, consistent with a recent study showing that loss of MAB21L4 blocks differentiation to drive the development of squamous cell carcinoma⁵¹, provided strong evidence to support a tumor suppressor role of MAB21L4.

Together, our systematic evaluation using existing GO annotation and the specific examples illustrate the utility of FunMap as a systematic framework to illuminate understudied genes, including many understudied cancer-associated proteins.

Discovery of drivers with low mutation frequency

Leveraging advancements in graph neural network-based deep learning, we developed a positive-unlabeled learning algorithm that integrates the FunMap network, gene mutation significance scores from 10 CPTAC cohorts, and known cancer genes to train a graph attention network (GAT) model for classifying unlabeled genes as cancer or non-cancer genes (Extended Data Fig. 7, Methods).

For performance evaluation, we used 274 cancer genes from the original Cancer Gene Census (CGC)⁵² as the positive set and 449 genes added later as hidden positives (Supplementary Table 7). The FunMap GAT model outperformed a random forest classifier trained without using network data, with a 6.5% improvement in area under the receiver operating characteristic (AUROC), 27.8% in area under precision-recall curve (AUPRC), and 35.7% in Average Precision at k (AP@K) (Methods). We also trained alternative GAT models using other networks including BioGrid²⁰, BioPlex¹⁸, HI-union¹⁹ and STRING²¹. The FunMap GAT model outperformed all alternative models for all three evaluation metrics (Figure 8A).

Figure 8: Discovery of cancer drivers with low mutation frequency using FunMap.

A) Performance comparison between models trained with various networks and that without network information, using Area Under the Receiver Operating Characteristic Curve (AUROC), Area Under the Precision-Recall Curve (AUPRC), or Average Precision @k (AP@k) as evaluation metrics. B) Percentages of hidden positive genes among the top-20 predictions generated by various models. C) Mutation frequencies across various cancer types for the top 15 newly predicted cancer drivers by the FunMap-based model. D) Number of manually confirmed publications with direct experimental evidence implicating a causal role for a given predicted cancer driver. E) Oncoplot depicting copy number alterations in the newly predicted cancer drivers with > 1% alteration frequencies in TCGA PanCancer Atlas in cBioPortal. F) Boxplots comparing LGI3 RNA expression and protein abundance in tumor vs normal samples demonstrating tumor-underexpression in the cohorts shown. Number of samples, n, are indicated in parenthesis. P-values derived from two-sided Wilcoxon rank-sum test. G) Violin plots depicting dependency scores after LGI3 or FAT1 CRISPR KO in cell lines from annotated lineages downloaded from the DepMap resource. Number of samples, n, are indicated in parenthesis. P-values derived from one-sample, one-tailed t-test. For each cancer type, the first and second p-values correspond to significance of LIG3 KO and FAT1 KO, respectively. For boxplots, centerline indicates the median, box limits indicate upper and lower quartiles, whiskers indicate the 1.5 interquartile range, and number of samples per group indicated in parentheses.

Among the top FunMap GAT predictions, 60.0% of the top 5, 40% of the top 10, and 25% of the top 20 were hidden positives, far exceeding the expected 4.3% by random chance (p<0.01, Fisher’s exact test). In this analysis, models incorporating network data clearly outperformed the one that did not (Figure 8B), and there was minimal overlap among the top 20 predictions when different networks were used or when network data was not used (Supplementary Table 7). These results underscore the significant impact of network information on prediction outcomes.

Despite low mutation frequencies (Figure 8C), 12 of the top 15 (80%) putative driver genes predicted by FunMap and not covered by CGC had at least one publication that supports a causal role in cancer through genetic and/or pharmacologic perturbation in model systems (Figure 8D, Supplementary Table 7, Methods). Moreover, nine genes showed frequent copy number alterations in TCGA data (Figure 8E), providing independent support to our predictions because copy number data was not used in the FunMap GAT model. Notably, LGI3, though lacking causal evidence in the literature (Figure 8D), was recurrently deleted in 3% of the 5,656 TCGA samples and significantly downregulated at both RNA and protein levels in tumors from CPTAC cancer cohorts where LGI3 was quantified in both tumor and normal samples (Figure 8F). Furthermore, an analysis of CRISPR knockout (KO) dependency scores for cancer cell lines available through DepMap revealed a significant increase in cell fitness across various lineages following LGI3 KO (p<0.05, one-sample t-test), and the effect was on par with that observed for well-known tumor suppressor genes listed in CGC, such as FAT1⁵³ (Figure 8G). These results collectively suggest LGI3 as a putative tumor suppressor gene.

Taken together, our data highlight the effectiveness of FunMap in uncovering genes with a low mutation frequency as putative cancer genes, presenting them as promising candidates for further experimental validation.

DISCUSSION

Large-scale omics profiling has massively expanded the landscape of somatic mutations and cancer-associated proteins, but the difficulty in functional interpretation hinders their prioritization and translation into clinical practice. By utilizing machine learning techniques on pan-cancer proteogenomics data, FunMap provides a systematic framework to tackle this challenge.

With 196,800 associations among 10,525 proteins and an LR of 50, FunMap provides both a comprehensive and unbiased proteomic coverage and a high level of functional relevance. The key differences between our approach and previous studies on gene co-expression networks include the use of protein profiling data obtained from over 1,000 human tumor samples spanning 11 cancer types and a supervised machine learning approach for functional network construction. Consistent with previous reports, protein co-expression is a much more reliable predictor of gene co-functionality than mRNA co-expression^10,12, however, combining both protein and mRNA co-expression provides the highest level of predictive power. One unexpected observation is that our co-expression based functional network outperforms protein-protein interaction networks in discriminating between functionally relevant and irrelevant gene pairs. Thus, functional networks constructed from proteomic and proteogenomic data offer a complementary approach to protein-protein interaction networks, thereby expanding systems biology frameworks for functional genomics research. Indeed, analyses from our study clearly demonstrate utilities of FunMap in providing functional annotation of understudied cancer proteins, obtaining functional insights into somatic mutations, and shedding global insights into cancer proteome organization and function.

A limitation of this study is that data from only 11 cancer types were included in the pan-cancer FunMap construction. We expect that proteomic and proteogenomic profiling will be applied to more cancer types in the future, and a more comprehensive analysis can be performed as more cancer types are included in future studies. Moreover, the CPTAC cohorts used in the study have limited follow-up duration, with the incidence of death events varying significantly among different cancer types. Therefore, the statistical power to detect associations with survival is generally low and varies considerably across cohorts, which constrains the scope of our prognostic analysis. To mitigate this limitation, it would be beneficial to seek out cancer cohorts that have been followed for a longer period. For some cancer types such as breast cancer and lung cancer, there are already multiple independent proteomic and proteogenomic studies. In this scenario, our approach can also be used to integrate independent datasets from a single cancer type to build cancer type-specific FunMaps. Additionally, this study focused on assessing the value of proteogenomic profiling data in mapping the functional network of human cancer, but the approach can be easily expanded to integrate expression data with other types of data, such as protein-protein interaction data, to generate a more comprehensive functional network. Although FunMap GAT outperformed other models to some extent in distinguishing between driver and passenger mutations, the accuracy was far from satisfactory for all models, highlighting the difficulty of this persistent challenge. Further improvements may be made in both FunMap construction and network-based driver gene prediction. Lastly, the associations identified in our analysis represent pairs of genes that work in coordination within the complex tumor tissue system, which includes not only cancer cells but also the surrounding microenvironment. Because the data we used originated from bulk tissues, it is impossible to determine associations within specific cell types. The emerging single-cell proteomics technology would be ideal for addressing this limitation⁵⁴.

In conclusion, this study highlights the significant potential of integrating machine learning and proteogenomic profiling to gain a deeper understanding of complex cancer systems. By generating a comprehensive functional network, this approach provides a robust framework for cancer functional genomics research, offering valuable insights into somatic mutations and cancer-associated proteins. These findings can greatly aid in prioritizing targets for clinical translation, ultimately contributing to the development of more effective cancer therapies.

METHODS

Data acquisition

CPTAC data for 10 cancer cohorts were harmonized by the CPTAC pan-cancer working group as previously described¹⁵. HCC data were downloaded from the original publication⁵⁵. In total, we collected mRNA and proteomics data for 11 cancer cohorts, where 5 cohorts also included data for matched normal samples for both mRNA and protein. For each of the 32 mRNA or proteomics dataset, we required that each gene or protein had at least 20 valid data points to be included in the analysis. The union set of all valid genes was denoted as $g_{valid}$.

Network construction

A machine learning model using XGBoost⁵⁶ was trained to predict the probability of co-functionality for a gene pair. For each gene pair $(A,B)$, the Pearson’s correlation coefficient ${PCC}_{AB}$ was computed between their mRNA expression vectors or protein expression vectors in each of the 32 datasets. We further calculated the mutual rank (MR) of a gene pair in each dataset based on a modified version of a previously published definition¹⁷: $MR(A,B)=\frac{1}{n-1}\sqrt{r_{AB}{r}_{BA}}$, where $r_{AB}$ is the rank of ${PCC}_{AB}$ among all PCCs between gene A and its partners. Rank starts with 0 and a larger PCC results in higher ranks. The total number of genes is denoted as $n$. The mutual rank values are in the range of [0,1]. In the case of ${PCC}_{AB}$ missing in a cohort, we will treat $r_{AB}$ as a missing value. The 32 MRs for a gene pair were used as input features for training the XGBoost model.

To prepare the data for training and validating the XGBoost model, we downloaded a gold standard set that was previously constructed using the Reactome pathway database¹². In brief, functionally associated protein pairs (labeled as positive) are defined as pairs that are found in the same ‘detailed’ pathway. Here each protein is annotated to a subset of the lowest level pathways. Only pathways that contain ≤ 200 proteins were included to make sure that only closely related protein pairs were positively labeled. Protein pairs that are not included in the same pathway at any annotation level are labeled negative. We included only those pairs where both proteins are in $g_{valid}$ as the final dataset $D$ for training the classification model. We partitioned the data into training ($D_{train}$) and test ($D_{test}$) sets, with a 50–50 split. The ratio of positive and negative labels were kept the same in training and test sets using stratified splitting technique. However, it’s worth noting that the original dataset exhibited a significant class imbalance issue, with a considerably larger number of instances in the negative class compared to the positive class. To tackle this challenge, we applied undersampling specifically to the negative class within the training dataset. This step involved reducing the number of negative class instances, aligning them with the number of positive class instances. The goal was to create a balanced training dataset that allows the machine learning model to learn from both classes more equitably. We then performed hyperparameter tuning by applying grid search with 5-fold cross-validation. The parameter grid was defined as follows: {“n_estimators”: [50, 150, 250], “max_features”: [0.2, 0.4, 0.6, 0.8], “min_samples_split”: [2, 4, 6]}. We used AUROC as the performance metric for hyperparameter tuning. After the model was trained, we predicted the labels for all possible pairs of proteins in $g_{valid}$. We required that the MR of a pair must be larger than 0.95 (i.e., top 5% among all gene pairs) in at least one data cohort. The final prediction performance was measured with log likelihood ratio (LLR) using the gold standard subset $D_{test}$. Here, LLR is defined as:

$$LLR=ln\left(\frac{C(PP\&P)/C(PP\&N)}{C\left(P\right)/C\left(N\right)}\right)$$

where $PP$ is the set of predicted positive protein pairs, while $P$ and $N$ are sets of positive and negative pairs in $D_{test}$, respectively. Set intersection is denoted as $\&$, while function $C(\cdot )$ returns the size of a set. To determine the number of pairs to be included in the final network, we first sorted the pairs in descending order of being positive (based on the predicted probability). We then computed the LLRs while designating more top pairs with a step size of 100 as $PP$. The LLR drops with the inclusion of less confidently predicted pairs. We stopped the process as soon as LLR dropped below 3.912 (likelihood ratio 50). All protein pairs selected with this procedure were included as edges in a functional association network named FunMap.

Detection of network modules

We used two complementary algorithms to identify modules from FunMap. First, we applied the Iterative Clique Enumeration (ICE) algorithm²³ to identify relatively independent maximal cliques in the network as functional modules. Overlap between the modules are allowed but are restrained due to the inherent design of the algorithm. The stringent requirement imposed by the module definition in the algorithm ensures high-level of co-functionality among all proteins in a module. The input to the software, available at http://ice.zhang-lab.org, is the network edge list file and the only required parameter is the minimal module size c. In this study, we set c to 5.

In contrast to the bottom-up approach taken in ICE, the top-down hierarchical modular organization of FunMap was uncovered based on the Network Seriation and Modularization (NetSAM) algorithm²⁵ implemented in the R package NetSAM available at https://bioconductor.org/packages/release/bioc/html/NetSAM.html. The main function of the package takes as input an network edge list file and outputs an “nsm” file that describes all detected modules organized in a hierarchical fashion. The most important parameters to the function include “minModule” and “modularityThr”. The parameter “minModule” specifies the ratio between the size of the smallest module and the total number of nodes in the network. If the size of a module identified by the function is less than the minimum size, the module will not be further partitioned into sub-modules. We set “minModule” such that the minimum size of a module is 20. To test whether a network under consideration has a non-random internal modular organization, we set the parameter “modularityThr” to 0.2 such that the network will be considered to have internal organization and will be further partitioned when its modularity⁵⁷ is above this threshold value. This parameter reflects the stringency of splitting a module into sub-modules. A higher threshold value tends to split the modules less frequently.

Connecting hierarchical modules to cancer hallmarks

Overlap between FunMap’s hierarchically organized modules and cancer hallmarks were evaluated according to 146 literature curated GO terms^{26,27,58–60}. These terms are categorized into 10 themes that map to 10 cancer hallmarks⁶¹. For each FunMap module we performed over-representation analysis (ORA) and obtained top 10 enriched terms for that module. In order to annotate each branch of the tree structure that is rooted on a second level module with the most relevant hallmark, we designed a voting scheme that works as follows: for each branch, we first designate the most overlapped hallmark as the one has the largest sum of associated negative logarithm of p-values for that hallmark over all modules in that branch. In essence, each module can vote for a representative hallmark for its residing branch using its significance of overlap with that hallmark. The designated hallmark for each branch represents the consensus annotation for the whole branch. The top associated consensus annotation for the 10 largest branches are shown in Figure 4. For selected branches, a second consensus hallmark annotation was also shown that was both closely related to the top annotation and had a sufficiently significant p-value.

Connecting hierarchical modules to ECM and angiogenesis

ECM genes encoding proteins with documented exclusive pro- or anti-angiogenic activity⁶² along with collagen type VI were first used to calculate the proportion of pro- and anti-angiogenic genes within nodes downstream of the FunMap branch rooted hierarchical module L2_M12. A final enrichment ratio for angiogenic impact was then computed by taking the previous pro-angiogenic ratio over the anti-angiogenic ratio. Values > 1 indicate a higher proportion of pro-angiogenic ECM genes in a module while values < 1 indicate a higher proportion of anti-angiogenic ECM genes. Some modules did not contain any anti-angiogenic genes and were annotated as pro-angiogenic exclusive (Supplementary Table 4).

Connecting network modules to somatic mutations

We trained an XGBoost model to evaluate the importance of gene mutation in predicting network module abundance. A total of 536 modules are considered including those revealed by NetSAM (255) and ICE (281). To compute module abundance, we first transformed the raw protein expression data in each cohort into z-scores by performing feature-wise standardization. The module abundance of a sample is defined as the average z-score of all genes in the module for that sample.

We used mutation data from 10 CPTAC tumor cohorts in this part of the study due to the lack of mutation data from the HCC study. First we selected genes that are significantly mutated in at least one cohort (q-value < 0.1). We then retrieved the actual binary mutation data of the selected genes from each cohort and merged them into a final feature data set. The resulting mutation dataset is composed of 1,021 samples and 77 genes.

For each module, we trained a regressor with XGBoost to predict module abundance based on the 77 significantly mutated genes. We applied 5-fold cross validation for hyperparameter tuning using the grid search technique. The parameter grid was defined as {“learning_rate”: [0.1, 0.2, 0.3, 0.4, 0.5], “n_estimators”: [20, 50], “max_depth”:[2, 3, 4]}. We used the Pearson correlation coefficient (PCC) between the predicted and actual abundance scores as the scoring metric for model assessment. Best parameters were used to fit a final model with the whole training data. We only included those modules that can be predicted with PCC > 0.25 in downstream analyses. This results in a total of 17 modules. The built-in feature importance scores of the trained model were used to estimate the contribution of each mutated gene in predicting the module abundance. Specifically, we used the “gain” type importance which implies the relative contribution of the corresponding feature to the model calculated by taking each feature’s contribution for each tree in the model. A higher value of this metric when compared to another feature implies it is more important for generating a prediction. This allows features to be ranked and compared with each other.

Function prediction of understudied genes

Based on the assumption that genes with similar functions are located in proximity to each other in the functional association network, we made function prediction of the “dark” genes in FunMap. We used the network topology analysis algorithm in WebGestalt⁴⁹ to establish a neighborhood of 50 genes for each dark gene and then performed GO enrichment analysis. Specifically, the algorithm lets the random walker start from each dark gene. It repeatedly moves to its neighboring nodes with an equal likelihood. At each step, it also has some probability (p=0.5) to return to the starting point. The restart probability controls how far the random walker moves away from the dark gene. The final score of a gene is defined as the steady-state probability that the walker will stay at the gene in the long run. For each dark gene, we chose the top 50 genes with the highest scores as its network neighbors, and then performed overrepresentation analysis (ORA) against GO terms for these network neighbors.

Cancer driver gene prediction

To predict cancer driver genes, we trained graph attention network (GAT)⁶³ based neural network models on FunMap and compared the performance with models trained with other publicly available networks, including BioPlex¹⁸, HI-union,¹⁹ BioGrid ²⁰, and STRING ²¹. For the STRING network, we only kept interactions with a combined score higher than 700. As a baseline, we also trained a random forest classifier without using network data.

We used mutation data from the 10 CPTAC tumor cohorts in this part of the study. First we selected genes that are significantly mutated in at least one cohort (q-value < 0.1). We then performed minus log10 transformation to the raw p-values. Each gene is characterized by a 10-dimensional vector as its features, representing mutation significance in 10 cancer cohorts.

Given the uncertainty regarding the role of an unlabeled gene as a driver or non-driver gene, the standard supervised machine learning approach is not well-suited for our task. This is due to the fact that typical supervised learning algorithms necessitate the presence of both positive and negative examples for training purposes. Therefore, we formulated our prediction task as a positive unlabeled (PU) learning problem⁶⁴ where genes in the network are divided into positive genes (known drivers) and unlabeled genes, which can contain both hidden driver genes (to be predicted positives) and non-driver genes (negatives). The goal is to train a model that uses known drivers to identify hidden drivers in the network. For known drivers, we downloaded a list of cancer drivers from the original Cancer Gene Census (CGC) publication⁵², which includes 274 genes. To test our trained model, we downloaded the 449 driver genes that were included in the CGC database after the original publication (Supplementary Table 7). Only known and hidden driver genes presented in the respective networks were used in training and performance evaluation.

We used the bagging based PU learning approach⁶⁵ to tackle the driver gene prediction task. The approach can be broken down into 4 steps: (1) Create a training set by combining all positive data points with a random bootstrapped sample set $B$ of the same size from the unlabeled samples. (2) Train a classifier with the newly assembled sample set, treating positive and unlabeled data points as positives and negatives, respectively. (3) Apply the classifier to those unlabeled samples that were not included in $B$, the out of bag (OOB) sample set, and record their predicted scores. (4) Repeat the previous three steps $T$ times ($T=10$ in this study) and assign to each sample the average of the OOB scores it has received.

To train a node classifier in the abovementioned step 2 using graph neural network (GNN), we used the GAT architecture. The learning of a GAT attention layer involves four key steps: (1) In order to obtain sufficient expressive power, a linear transformation is applied to the feature vectors of the nodes. (2) Attention coefficients determining the relative importance of neighboring features to each other were computed. To obtain the attention score between two neighbors, it first concatenates the embeddings $z$ of the two nodes obtained from the previous step, then takes a dot product of it with a learnable weight vector $a$, and applies a LeakyReLU in the end. This step can be formulated as:

$$e_{ij}=LeakyReLU\left(a^T\right(z_i\left|\right|z_j\left)\right)$$

where || denotes concatenation. (3) To make the scores easily comparable, the attention coefficients are normalized across all neighborhoods using the softmax function. (4) The final step works similar to GCN. The embeddings from neighbors are aggregated together, weighted by the attention coefficients and then are transformed by a nonlinear activation function. Similar to multiple channels in a convolutional neural network, GAT uses multi-head attention to enhance the model capacity and to stabilize the learning process. Specifically, $K$ independent attention mechanisms apply the transformations of step (1) - (3). During the last step, embeddings from different heads are averaged before applying the nonlinear transformation. In this study, we trained a model consisting of 2 GAT layers each with 8 attention heads.

For performance evaluation, in addition to the standard metrics such as AUROC and AUPRC that treat all unlabeled samples as negative, we also included the more appropriate Average Precision at k (AP@k) metric, which is widely used in the areas of information retrieval and recommendation systems. Essentially we are treating our task as a ranking problem where we aim to assign the test positive samples with higher scores (likelihood of being a driver gene) so that they rank higher in the list of sorted prediction scores ⁶⁶. After the samples are sorted by their predicted scores, AP@k is computed as: $AP@k=\frac{1}{min(m,k)}{\sum }_{i=1}^k\frac{TP\left(i\right)}{i}$, where $m$ is the total number of positive samples in the test dataset. $TP\left(i\right)$ is set to 0 if the ith sample is not a positive test sample. Otherwise, it is set to the number of positive test samples seen up to the ith position in the ranked list. AP@k is a measure that combines recall and precision for ranked results. It is considered a reasonable evaluation metric for emphasizing returning more highly likely positive samples at the top of the ranked list⁶⁷.

We trained our GAT models using the Pytorch Geometric framework⁶⁸. The inputs to the model include a feature matrix $X\in R^{N\times p}$ and network edge list (Extended Data Fig. 7). In this study, $p$ is 10, representing the significance of gene mutation in 10 cancer cohorts. Cross entropy loss was computed as $L=-\left(ylog\right(h)+(1-y\left)log\right(1-h\left)\right)$ where $h$ is the output of the network after sigmoidal activation and $y$ the node label (0 or 1). The ADAM optimizer⁶⁹ was used for training with an exponentially decaying learning rate ($\gamma =0.99$) starting at 0.001. We employed early stopping to prevent overfitting. For the baseline random forest model, only the feature matrix was needed. Default parameters provided in scikit-learn package⁷⁰ were used.

Published causal evidence supporting predicted cancer drivers

Each of the predicted cancer drivers described above were used to search PubMed with the following term on December 20, 2023: “Gene (CRISPR OR knockout OR shRNA OR siRNA knockdown OR silencing OR overexpression OR over-expression) cancer”, where “Gene” is replaced by the predicted cancer driver. Search results were sorted in descending order with respect to published date. Abstracts or manuscript texts were then manually vetted for causal evidence that genetic and/or pharmacologic perturbation of the predicted cancer driver functionally impacted cancer phenotypes (proliferation, migration, invasion, etc.) or augmented drug responses in model systems. This continued for each gene until all search records were verified or until 10 publications by recent publication date were found with casual evidence impacting cancer phenotypes and/or drug response (Supplementary Table 7).

Genetic dependency in cancer cell lines

Cancer cell line annotations (sample_info.csv) and gene effect dependency scores derived from integration of CRISPR KO screens published by Broad’s Achilles and Sanger’s SCORE projects were retrieved from DepMap Public 22Q2 (CRISPR_gene_effect_.csv)^71,72. Cancer cell lines were matched to tumor cancer types by using the following filters: BRCA: primary_disease = “Breast Cancer” and lineage = “breast”; GBM: primary_disease = “Brain Cancer” and lineage = “central_nervous_system”; LUAD: primary_disease = “Lung Cancer”, lineage = “lung”, lineage_sub_subtype = “NSCLC_adenocarcinoma”; PDAC: primary_disease = “Pancreatic Cancer”, lineage = “pancreas”. For each cancer cell lineage, a one-sample, one-tailed T test was used to identify LGI3 and FAT1 associated with significantly higher cell growth following gene knockout.

Statistics and reproducibility

All data used for machine learning and gene dependency analysis are from publicly available resources^15,71,72 with detailed methodologies for data collection, blinding, randomization, and protection. Sample sizes were from the original publications, and they were sufficient for all statistical tests performed. Non-parametric statistical tests were used whenever possible. For parametric tests, normality of data distributions was assumed, though this was not formally tested. No data were excluded from analyses. The experiments were not randomized. The Investigators were not blinded to allocation during experiments and outcome assessment.

Extended Data

Extended Data Fig. 1. Quantification of inter-sample heterogeneity through gene-wise standard deviation.

A) Distributions of gene-wise standard deviations across individual datasets (n = 17,733 to 19,113 mRNAs and n = 7,961 to 11,815 proteins). For boxplots, centerline indicates the median, box limits indicate upper and lower quartiles, whiskers indicate the 1.5 interquartile range. B) Median values of the median standard deviations across various dataset groups. T: Tumor; N: Normal.

Extended Data Fig. 2. Breakdown of feature importance in the XGBoost model.

A) Barplot showing importance of individual features. B) Pie chart depicting aggregated importance by data and sample type pairs.

Extended Data Fig. 3. Characterization of dense modules.

A) Heatmap depicting log2 fold change (log2FC) of average protein abundance of dense modules (cliques) in tumor vs normal for each of the five cancer cohorts shown. All 78 cliques have concordant tumor over- or under-expression in all five cohorts (FDR < 0.01 in each cohort). Table shows the number and maximum number of overlapping edges with other networks as indicated. Gene ontology biological processes (GO_BP) indicates the top enriched term of a given clique (GO_BP_FDR). B-C) Tumor overexpressed, ECM-associated dense modules, Clique 96 (B) and Clique 54 (C). Edge color indicates lack of overlap in BioGRID, BioPlex, HI-union, STRING, and CORUM (pink) or overlap in any of these resources (gray). D-E) Boxplots comparing average protein abundance of Clique 96 (D) and Clique 54 (E) in tumor and normal samples demonstrating tumor overexpression in five cancer cohorts. Number of samples, n, are indicated in parenthesis. P-values determined by two-sided Wilcoxon rank-sum test. F-G) Kaplan-Meier plots depicting overall survival (OS) difference in patients from indicated cohorts stratified by median value of the average abundance of proteins in Clique 96 (F) and Clique 54 (G). Logrank p-values and hazard ratio (HR) shown with 95% confidence intervals derived from Cox-proportional hazard models. Significance is indicated as ****p < 0.0001. For boxplots, centerline indicates the median, box limits indicate upper and lower quartiles, whiskers indicate the 1.5 interquartile range, and number of samples per group indicated in parentheses.

Extended Data Fig. 4. Connecting somatic mutations to functional protein modules.

A) Average pairwise Pearson’s correlation coefficient for genes in L2_M40 based on mRNA or protein data in different cancer types. B) Average pairwise Pearson’s correlation coefficient for genes in L3_M58 based on mRNA or protein data in different cancer types. C) Comparison of TP53 protein abundance (log2 MS1 intensity) in TP53 wildtype (wt) and mutant (mut) samples across 10 cancer types. Number of samples, n, are indicated in parenthesis. P-values were derived from two-sided Wilcoxon rank-sum test. Significance is indicated as *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant. For boxplots, centerline indicates the median, box limits indicate upper and lower quartiles, whiskers indicate the 1.5 interquartile range, and number of samples per group indicated in parentheses.

Extended Data Fig. 5. Illuminating understudied cancer proteins RBM34 and RBM12B.

A) Boxplots comparing protein abundance of RBM34 and RBM12B in tumor and normal samples demonstrating tumor over-expression in five cancer cohorts. Number of samples, n, are indicated in parenthesis. P-values determined by two-sided Wilcoxon rank-sum test. For boxplots, centerline indicates the median, box limits indicate upper and lower quartiles, whiskers indicate the 1.5 interquartile range, and number of samples per group indicated in parentheses. B) Barplots depicting frequency of somatic copy number and mutations in RBM34 and RBM12B from TCGA PanCancer Atlas Studies in cBioPortal. C-D) Network neighborhood of RBM34 (C) or RBM12B (D) with genes associated with the enriched GO terms highlighted.

Extended Data Fig. 6. Illuminating understudied cancer proteins CXorf38 and MAB21L4.

A) Boxplots comparing protein abundance of CXorf38 in tumor and normal samples demonstrating tumor over-expression in five cancer cohorts. Number of samples, n, are indicated in parenthesis. P-values determined by two-sided Wilcoxon rank-sum test. B) Relationship between protein abundance of CXorf38 and RNA-seq inferred ESTIMATE ImmunoScore in eight cancer types. P-values were derived from two-sided Spearman’s rank correlation. Shaded area depicts the 95% confidence interval. C) Single cell data from the Human Protein Atlas showing that CXorf38 is expressed across all cell types, but the highest expression occurs in immune cells. D) Boxplots comparing protein abundance of MAB21L4 in tumor and normal samples in five cancer cohorts. Number of samples, n, are indicated in parenthesis. P-values determined by two-sided Wilcoxon rank-sum test. Significance is indicated as *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant. For boxplots, centerline indicates the median, box limits indicate upper and lower quartiles, whiskers indicate the 1.5 interquartile range, and number of samples per group indicated in parentheses.

Extended Data Fig. 7. Graph neural network architecture for predicting cancer driver genes based on network topology and mutation data.

DATA AVAILABILITY

Proteomics and RNASeq data for the 10 CPTAC cancer types were derived from the CPTAC pan-cancer study¹⁵: https://proteomic.datacommons.cancer.gov/pdc/cptac-pancancer. Proteomics and RNASeq data for HCC were downloaded from the original publication55. The data tables derived from these resources and used as input for FunMap construction are available at https://zenodo.org/record/7948944. Derived feature data for XGBoost model training are available at https://zenodo.org/records/7949375. XGBoost prediction scores for all gene pairs are available at https://zenodo.org/records/10080764. FunMap edge list, dense modules, and hierarchical modules can be downloaded at: https://funmap.linkedomics.org/. FunMap edge list, dense modules, and hierarchical modules can be downloaded at: https://funmap.linkedomics.org/. The same web site also provides visualization tools to explore gene neighborhoods, dense modules, and hierarchical organization of FunMap. Additionally, FunMap network and modules have been integrated into WebGestalt⁷³ for enrichment analysis of user provided gene lists. Cell line annotations and CRISPR KO dependency scores can be retrieved from the DepMap website: https://www.depmap.org. Other datasets used in the study included gene co-functionality “gold standard” derived from the Reactome pathway database¹², ProHD¹², BioPlex¹⁸, HuRI¹⁹, HI-Union¹⁹, and BioGRID²⁰.