Attention-enhanced Multi-omics Model for Pan-cancer Drug Response Prediction and Biomarker Discovery

Abstract

Cancer drug discovery remains challenged by tumour heterogeneity and limited experimental scalability. Recently, virtual drug screening integrating machine learning algorithms has yielded numerous research results, some of which have been successfully patented and are expected to be further translated and deployed in drug discovery pipelines. Most current models for virtual drug screening fail to effectively integrate multi-omics data or capture nonlinear cross-omics interactions, restricting predictive accuracy and biomarker discovery across diverse cancers with high heterogeneity. We developed a multi-omics fusion deep learning model integrating mutation, methylation, transcriptomic, and metabolomic profiles from over 900 pan-cancer cell lines derived from the DepMap database. Our framework synergizes random forest-based feature selection to prioritize biologically relevant omics features and multi-head attention mechanisms to model nonlinear interactions between cellular multi-omics landscapes. Further biological analysis of the selected features enabled the deciphering of potential biomarkers related to drug effects. Our multi-omics fusion model attained high-performance drug response prediction across nearly 900 cancer cell lines (median Pearson r = 0.50 vs Pearson r = 0.22 for the former model for all included drugs). Validation demonstrated robust accuracy for the MEK inhibitor Trametinib (r = 0.78, MAE = 0.59) and the non-oncology agent BMOV (r = 0.73). The model identified BRAF-mutant melanoma sensitivity and PI3K/AKT bypass resistance, consistent with existing findings. Feature mining revealed TERT modulation and oxidative stress induction as BMOV's probable anticancer mechanisms, while Benzamide targeted metabolic vulnerabilities. This study introduced a deep-learning-based multi-omics fusion model to predict pan-cancer drug response. The framework achieved the expansion of the anticancer spectrum of existing anticancer drugs and explored the potential anticancer effects of non-anticancer drugs. Moreover, the integration of SHAP/MDI feature interpretation algorithms enabled mechanistic biomarker discovery. However, the prediction results that were not reported in previous research are yet to require further experimental verification. In conclusion, this work established a robust DL-driven platform for virtual drug screening and biomarker discovery, providing a computational platform that could aid virtual drug screening and biomarker discovery and facilitate the development of precision oncology.