Abstract 2696: Unlocking novel gene-based biomarkers using an explainable deep learning framework for precision oncology.

Abstract Somatic genomic alterations are widely profiled in cancer and remain the primary indicator for personalized therapy, yet their clinical utility is limited to a small set of actionable targets benefiting only a fraction of patients. AI/ML offers a transformative path to precision oncology by leveraging large-scale genomic data, but poor interpretability, often reduced to single-gene insights, limits clinical adoption and overlooks cancer’s biological complexity. We address this gap with a novel framework combining PhenoMap and PhenoSurv. PhenoMap converts gene-level alterations into a phenotypic feature map using over 9,000 pan-cancer genomes and transcriptomes, trained to emulate enrichment scores for 6,000+ pathways from MSigDB using a combined LASSO-Random Forest approach. This map serves as input to PhenoSurv, a variational autoencoder designed to predict disease-free survival. PhenoSurv employs an innovative architecture that distinctly embeds eleven phenotypes and optimizes reconstruction loss, Kullback-Leibler divergence, and Cox survival loss on the latent space. This yields survival-relevant latent features, enabling discovery of multi-phenotypic vulnerabilities and prognostic biomarkers, which we tested and validated in breast, lung, and brain cancer. PhenoMap pathways that exceeded baseline performance in internal and external validation included 3,382 for lung, 3,772 for brain, and 4,188 for breast cancer. Clustering by PhenoMap scores revealed patient subsets with shared dominant signaling pathways. Notably, PI3K and KRAS signaling-enriched clusters included patients lacking PIK3CA (54%) or KRAS (67%) mutations in breast and lung cancer, respectively, indicating PhenoMap identifies broader cohorts for targeted therapy beyond single-gene biomarkers. PhenoSurv outperformed existing AI/ML models for survival prediction across cancer types while providing multilevel explainability at phenotype, pathway, and gene levels. Using DeepSHAP, we identified key pathways driving predictions across cohorts. Over 50 pathway-level biomarkers showed significant survival differences between high- and low-signaling groups, including NOTCH1 signaling in breast cancer (TCGA: p = 0.0098; MSK: p = 0.003), the CLEC7A inflammasome pathway in lung (TCGA: p = 0.0045; MSK: p = 0.041), and inositol phosphate metabolism in brain cancer (both p 0.0001). Survival declined progressively with the number of poor prognostic markers across cancer types. PhenoMap and PhenoSurv bridge a critical gap in precision oncology by delivering interpretable models based on clinical genomic data that reveal clinically actionable biomarkers. These mechanism-based insights enable biologically informed patient stratification, paving the way for more effective personalized treatment strategies. Citation Format: Sydney Grant, Aritro Nath, . Unlocking novel gene-based biomarkers using an explainable deep learning framework for precision oncology abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 2696.