Cancer is rarely caused by a single defective gene. Instead, it arises from the coordinated dysfunction of many genes acting together as gene modules - sets of co-regulated genes that respond to cellular conditions in concert. Understanding which gene modules drive tumor growth is essential for designing treatments that correct the underlying network rather than targeting isolated downstream effects.
Existing methods for detecting cancer gene modules have two critical limitations. First, they largely ignore the three-dimensional (3D) structure of chromosomes inside the cell nucleus. Recent technologies like Hi-C, which measure how distant regions of DNA physically contact each other in space, have revealed that this 3D genome architecture is fundamental to gene regulation - but few existing methods incorporate this information.
Second, existing approaches cannot assign importance to specific biological features for individual genes within a module. Cancer genes can be disrupted through genetic mutations, epigenetic silencing, altered chromatin structure, or protein interaction changes - but most co-expression clustering methods treat all these mechanisms equally and cannot reveal which mechanism matters most for each specific gene.
This study introduces CGMega, a graph attention network framework that integrates genomic, epigenomic, 3D chromatin, and protein interaction data to predict cancer genes, detect the modules they form, and explain which biological features drive each prediction - with direct application to breast cancer and acute myeloid leukemia (AML).
CGMega builds a biological network graph where each node represents a gene and edges represent protein-protein interactions (PPIs) between genes. Each gene node carries a 15-dimensional feature vector that combines epigenomic data (ATAC-seq chromatin accessibility, CTCF binding, H3K4me3 and H3K27ac histone modifications), genetic variation data (single nucleotide variant frequencies, copy number variant frequencies), and condensed Hi-C features capturing the gene's 3D spatial position in the nucleus.
Incorporating Hi-C data presented a key technical challenge because Hi-C contact maps are noisy, sparse, and very high-dimensional. The team systematically compared different strategies for encoding Hi-C data. Adding Hi-C contacts as additional graph edges provided only marginal improvement. But treating Hi-C data as condensed node features using singular value decomposition (SVD) produced a major performance boost, achieving the best results with just 5 dimensions of condensed Hi-C information.
On top of this multi-omics graph, CGMega applies a graph attention network (GAT) with transformer architecture to predict which genes are cancer drivers. The attention mechanism allows the model to weight different neighboring genes differently, capturing the unequal influence that genes exert on each other in real biological networks. The model is trained in a semi-supervised manner using labeled known cancer genes.
To make predictions interpretable, CGMega uses GNNExplainer, a model-agnostic neural network interpretation tool. For any predicted cancer gene, GNNExplainer identifies the compact subgraph of neighboring genes (up to two hops away, with up to 20 edges) and the specific feature dimensions that most strongly influenced the prediction. This reveals both the gene module structure and the biological mechanisms driving each cancer gene's role.
CGMega was benchmarked on the MCF7 human breast cancer cell line, a well-characterized model system with high-quality multi-omics data. On the primary performance metric of area under the precision-recall curve, CGMega achieved an AUPRC of 0.9140 - outperforming all competing methods across all four evaluation metrics (AUPRC, AUROC, accuracy, and F1 score).
The comparison included both general machine learning approaches (graph convolutional networks, standard graph attention networks, multilayer perceptrons, and support vector machines) and methods specifically designed for cancer gene prediction: MTGCN, EMOGI, and MODIG. CGMega outperformed all of these, demonstrating that integrating 3D chromatin structure with standard multi-omics data provides substantial new predictive power.
A critical practical advantage emerged in transfer learning scenarios. When labeled training data is scarce - as is common in rare cancers - non-pretrained models lose accuracy rapidly. CGMega pretrained on MCF7 and then fine-tuned on new cancers maintained high performance even with very few labeled genes (fewer than 200), while competing models collapsed. Hi-C features were especially important in this low-data regime.
The model was robust across a wide range of conditions: different Hi-C data resolutions (5 kilobase to 25 kilobase), reduced sequencing depths, extreme ratios of positive to negative training examples, and multiple different protein-protein interaction databases. This robustness indicates that CGMega can be applied to a wide range of cancer types even when data quality or quantity is imperfect.
Applying CGMega to 358 known cancer genes in the MCF7 breast cancer cell line revealed that cancer genes are not randomly distributed across gene modules - they cluster together in the same modules, consistent with the biological principle that disease-associated genes form localized network neighborhoods. TP53 was the most connected hub, participating in 139 distinct cancer gene modules, followed by ESR1 and AKT1.
CGMega assigned each gene a set of feature importance scores across all 15 input dimensions, revealing which biological mechanism drives each cancer gene. Clustering genes by their importance profiles identified five distinct classes: some genes are driven primarily by genetic mutations (SNVs), others by epigenetic regulation, and others by 3D chromatin structural features - demonstrating that cancer genes are mechanistically heterogeneous even within the same tumor.
A particularly striking discovery was a high-order gene module connecting all four members of the ErbB family (EGFR, HER2, HER3, and HER4). These receptor tyrosine kinases are individually well-studied cancer targets, but CGMega revealed they are interconnected through three shared hub genes: NRG1 (a ligand that activates ErbB signaling), PPM1A (a phosphatase controlling cell cycle progression), and DLG2 (a scaffolding protein modulating phosphorylation). This module suggests a coordinated phosphorylation homeostasis network that may govern resistance to single-target ErbB inhibitors.
CGMega also revealed mechanistic differences between BRCA1 and BRCA2 gene modules. BRCA1 connected broadly to 20 other genes, reflecting its broad role in DNA damage response. BRCA2 connected through ROCK2, and laboratory experiments confirmed that combining BRCA2 and ROCK2 inhibitors was significantly more effective than either alone after 24 hours of treatment in MCF7 cells - validating a CGMega-predicted combination therapy strategy.
The team applied CGMega to multi-omics data from eight AML patients spanning seven different mutation subtypes, achieving an average AUPRC of 0.8528. AML is particularly challenging to study because its gene networks are highly heterogeneous between patients - the same disease can arise from very different molecular configurations in different individuals.
CGMega identified 396 candidate AML genes predicted to be cancer drivers across all eight patients. These candidate genes were significantly enriched for known pan-cancer genes (P = 1.32e-22) and, when combined with previously known AML genes, revealed 15 hematopoietic and blood disease biological processes - including leukocyte migration and T-cell receptor signaling - that could not be identified using known AML genes alone.
A subset of genes emerged as hub genes appearing in more than 20 cancer gene modules across patients. Five new candidate hub genes were identified: ESR1, HDAC1, FYN, LYN, and GRB2. These genes may serve as novel AML gene panel candidates for clinical genetic testing, alongside well-established hubs like EGFR, MYC, TP53, and STAT3.
Beyond shared modules, CGMega identified patient-specific candidate AML genes for each individual patient. This patient-level resolution is clinically meaningful because AML heterogeneity explains why patients with similar genetic profiles can have drastically different responses to the same treatment. Identifying the unique gene module driving each patient's leukemia could eventually guide personalized therapy selection.
CGMega advances cancer gene discovery in two important ways simultaneously. It predicts which genes are cancer drivers with state-of-the-art accuracy, and it explains why by revealing the specific gene modules and omics features that determine each prediction. Most existing deep learning methods for cancer gene prediction are black boxes - CGMega's explainability turns its predictions into biological hypotheses.
The transfer learning capability makes CGMega especially relevant for rare cancers where labeled training data is scarce. By pretraining on well-characterized breast cancer data and fine-tuning on a new cancer type, CGMega can extract meaningful gene modules even when only a small number of known cancer genes exist as training examples. This extends the reach of the framework far beyond the cancers where training data is abundant.
The experimentally validated combination therapy prediction for BRCA2 and ROCK2 inhibitors demonstrates that CGMega's gene module outputs are actionable, not merely descriptive. High-order modules connecting multiple cancer genes can nominate new drug combinations that target the network structure driving a cancer rather than individual nodes within it.
Future applications of CGMega could include mapping patient-specific gene modules in clinical samples to predict drug sensitivity, identifying modules that become activated during treatment resistance, and extending the multi-omics framework to incorporate single-cell data as it becomes available for AML and other blood cancers. The complete framework and predicted gene lists were made publicly available to support these future studies.