Abstract B044: Deep learning-based integration of tumor omics and functional drug screening for precision treatment selection in high-risk and metastatic breast tumors

Abstract Precision oncology centers on the selection of targeted anticancer agents in tumors harboring actionable molecular alterations. However, most tumors lack actionable alterations, and individual biomarkers are often poor predictors of response. These challenges are particularly acute in metastatic tumors with their vast inter-tumoral heterogeneity and diverse resistance mechanisms. As a result, cytotoxic chemotherapy remains a mainstay of treatment, yet few data-driven tools exist to inform chemotherapy selection. These limitations inform the need for rational treatment selection strategies for drugs lacking validated predictive biomarkers in tumors with or without actionable molecular alterations. To this end, we report a unified computational/experimental approach, combining tumor omic and functional profiling with ScreenDL, a deep learning (DL)-based cancer drug response prediction framework capable of integrating these complementary data modalities to provide treatment recommendations for high-risk/metastatic breast tumors. ScreenDL leverages a phased training schema, incorporating: (1) general-purpose pretraining on large-scale cell line pharmaco-omic screens; and (2) follow-up fine-tuning on a newly generated pharmaco-omic dataset spanning >100 patient-derived xenograft (PDX)-derived organoid (PDxO) models primarily from high-risk, endocrine resistant, treatment-refractory, and/or metastatic breast tumors representing the greatest unmet clinical need. As new patients enter our precision oncology pipeline, tumor omic profiling and functional drug screening in patient-derived organoids (PDOs) enables patient-specific fine-tuning with our ScreenAhead module, integrating a tumor’s transcriptomic and functional characteristics with prior knowledge learned from training samples to generate personalized response predictions. At baseline, ScreenDL outperforms existing DL models in never-before-seen PDxOs, achieving a median Pearson correlation (PCC) between observed and predicted response across drugs of 0.32 compared to 0.20 for the next-best tested model. After incorporating each PDxO’s response to just 12 drugs through PDxO-specific fine-tuning with ScreenAhead, ScreenDL provides high-confidence predictions (PCC > 0.5) for 49 out of 95 drugs (52%), with a median PCC across all 95 drugs of 0.59. These high-confidence drugs included both chemotherapies and targeted agents with diverse biological mechanisms, showcasing the power of personalization with a small pre-screening panel to improve predictions for the broader space of therapies. ScreenDL also outperformed biomarker-only predictors of response to talazoparib and capivasertib, two targeted agents approved for breast tumors harboring mutations in BRCA1/2 or PI3KCA/AKT/PTEN, respectively. In a retrospective preclinical pilot of 11 PDX models, our end-to-end treatment selection strategy achieved a 100% clinical benefit rate and a 65% objective response rate, highlighting the vast clinical potential of our unified computational/experimental treatment selection strategy. Citation Format: Casey Sederman, Chieh-Hsiang Yang, Emilio Cortes-Sanchez, Tony DiSera, Xiaomeng Huang, Yi Qiao, Sandra Scherer, Bryan Welm, Alana Welm, Gabor Marth. Deep learning-based integration of tumor omics and functional drug screening for precision treatment selection in high-risk and metastatic breast tumors [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Functional and Genomic Precision Medicine in Cancer: Different Perspectives, Common Goals; 2025 Mar 11-13; Boston, MA. Philadelphia (PA): AACR; Cancer Res 2025;85(5 Suppl):Abstract nr B044.