Abstract C176: Cancer-specific AI identifies multi-modal biomarkers of therapeutic response for 1,951 drugs including TNG348, a highly selective USP1 inhibitor

Abstract

Understanding drug mechanism-of-action and improving clinical response rates are critical challenges to cancer therapeutic development. We designed a deep learning framework that learns molecular mechanisms of drug sensitivity and resistance from preclinical and/or clinical drug response data. The trained model can then generate biomarker strategies, suggest companion diagnostics, and propose rational drug-drug combinations that increase or restore drug sensitivity. The key innovation is the model’s architecture; the framework learns an embedding that reflects an interpretable relationship between a cellular process and drug response by emulating a tumor cell’s proteomic architecture. We applied the framework to learn genetic determinants of response for 1,951 drugs and 17,386 CRISPR gene knockouts screened in human cancer cell lines. The learned embeddings reflected known and novel mechanisms. RAF inhibitors, MEK inhibitors, and BRAF CRISPR knockout colocalized (P=1.7×10-11) due to their shared dependency on hyperactivation of RAF signaling. CDK4/6 inhibitor efficacy was strongly associated with a set of physically interacting proteins that control G1-to-S transition (P=5.6×10-12). Validating the clinical applicability of the finding, the trained model stratified patient response in a retrospective analysis of 70 ER+ metastatic breast cancer patients treated with the CDK4/6 inhibitor palbociclib (P=0.05, log-rank predicted responsive vs. non-responsive). As a case study, we examined mechanisms of sensitivity and resistance to TNG348 – a potent and highly selective USP1 inhibitor – in non-small cell lung cancer (NSCLC). By integrating genetics, gene expression, and CRISPR screens, the model identified seven genes whose expression explained >50% of the variation in response to TNG348 across 32 held-out NSCLC cell lines (Spearman R=0.74; P=8.7×10-7). The findings extend TNG348’s known relationship with homologous repair to additional protein complexes involved in cell cycle and DNA repair. The model leveraged the new associations to generate a six-biomarker inclusion-exclusion logic that captures both a large population (31.4% of TCGA lung adenocarcinoma patients meet inclusion criteria) and substantially enriches for strong response amongst preclinical models (OR=14.8, P=3.0×10-3). Finally, we asked the model to design rational combinations to overcome acquired resistance mechanisms. Proposed combinations included clinically successful combinations such as RAF+MEK inhibitors and were enriched for synergistic matches in a cell line screen of 986 drug pairings (Spearman rho P=6.6×10-4). These results highlight the promise of explainable AI to learn complex cellular mechanisms-of-action and generate non-obvious biomarker hypothesis and rational drug combination strategies for clinical development.

Citation Format: Adam Yaari, Lee McDaniel, Antoine Simoneau, Samuel Meier, Oliver Priebe, Eduardo Farias, Alan Huang, Jannik Andersen, Yi Yu, Maxwell Sherman. Cancer-specific AI identifies multi-modal biomarkers of therapeutic response for 1,951 drugs including TNG348, a highly selective USP1 inhibitor [abstract]. In: Proceedings of the AACR-NCI-EORTC Virtual International Conference on Molecular Targets and Cancer Therapeutics; 2023 Oct 11-15; Boston, MA. Philadelphia (PA): AACR; Mol Cancer Ther 2023;22(12 Suppl):Abstract nr C176.