Signature-aware deep learning reveals distinct driver gene programs and mutational processes in glioblastoma and colon adenocarcinoma.

The precise identification of cancer driver mutations is essential for precision oncology; however, it remains a significant challenge because of the overwhelming presence of passenger mutations and the complexity of mutational processes. In this study, we introduce ResMLP-GL, a signature-aware residual multilayer perceptron designed for variant-level cancer driver prediction, which explicitly integrates COSMIC SBS context probability vectors with > 100 functional and sequence features. The network incorporates two projection residual blocks alongside a feature-wise gating module that multiplicatively modulates hidden activations, thereby enhancing the gradient flow and facilitating process-aware representation learning. An Optuna-guided search optimizes parameters such as width, dropout, learning rate, and L2 regularization, whereas ADASYN addresses class imbalance. Trained on harmonized TCGA GBM/COAD exomes and reserved for testing, ResMLP-GL achieved an AUC of 0.949 on held-out data and an AUC of 0.921 on independent ICGC cohorts, surpassing CHASMplus, OncodriveFML, and MutSigCV. SHAP analysis indicated that functional scores (REVEL, AlphaMissense, CADD) and specific SBS probabilities collectively drive predictions, offering interpretable connections between mutational processes and driver selection. The model elucidates tissue-specific, signature-aligned programs (e.g., SBS1 and DNA-repair-related signatures for TP53/PTEN/EGFR in GBM and SBS1/MMR/POLE for APC/KRAS/PIK3CA in COAD), and a model-derived driver burden stratifies survival. The code, trained weights, and processed feature tables were made available for reproducibility. In summary, ResMLP-GL demonstrates that residual-gated MLPs with a quantitative signature context provide state-of-the-art interpretable driver prediction across cancers. Our findings underscore the significance of explicitly incorporating the context of mutational processes, which not only complements but, in certain oncological contexts, surpasses methodologies that rely exclusively on recurrence frequency or functional impact scores. This approach offers a robust framework for addressing tissue-specific challenges in predicting driver mutations. All code, trained models, and processed data are available at https://github.com/Zubair11122/ResMLP-GL to promote transparent and reproducible precision oncology research.