Abstract 549: Biologically informed machine learning for discovery of protein assemblies underlying cancer therapeutic resistance

Abstract Despite substantial advances in cancer therapeutics, the near inevitable emergence of treatment resistance limits overall survival of advanced cancer patients. Genomic drivers of intrinsic and acquired resistance are highly multifaceted. Extensive profiling of tumors has demonstrated that the disparate resistance driving alterations often converge onto shared cellular processes. Motivated by this observation, we developed an intrinsically interpretable, gradient descent-based algorithm to learn assemblies of physically and functionally interacting proteins that are jointly predictive of tumor sensitivity to chemical and genomic perturbations. Importantly – and unlike existing solutions – the model learns protein assemblies de novo during training rather than being limited to rigid gene sets from public databases or literature. By co-embedding experimentally derived PPI maps with sample-specific multi-omics, the model to identifies interacting proteins that predict perturbation response. Once identified, protein assemblies can be used as compound biomarkers, for new target identification, for drug-target deconvolution, and for combination design amongst other therapeutic applications. We benchmarked the predictive accuracy of our method using a panel of 23 clinical oncology drugs and related genetic perturbations targeting multiple areas of cell biology, including Ras/Raf signaling, G1/S checkpoint control, and DNA damage response. Our single model, trained on all 23 perturbations, outperformed state-of-the-art methods optimized for each individual perturbation in a held-out response prediction task (Mean accuracy: Random Forest, 0.41; Elastic Net, 0.37; FCNN, 0.33; this work, 0.43, highest accuracy for 15/23 perturbations). Our method achieved increased predictive accuracy while providing interpretable importance statistics for the members of each protein assembly. To predict CDK4/6 inhibitor response, the model learned a G1/S checkpoint assembly including CDK4 (P=1.4×10-8), CDK6 (P=4.5×10-8), CCNE1 (P=7.0×10-15), and RB1 (P=4.7×10-17). To demonstrate the application of our method to target discovery, we performed a synthetic lethality screen in cell lines with high and low activation of a G2/M protein assembly predictive of Nutlin-3A response (P=3.1×10-34). As expected given p53’s central role in the protein assembly (P=3.6×10-56), top hits included known TP53-driven lethalities including TP53BP1 (P=2.8×10-34) and USP28 (P=2.8×10-27). Moreover, we identified a synthetic dependency on FBXO42 (P=5.2×10-8) associated not only with TP53 loss-of-function but also with two other assembly members involved in apoptosis regulation. These results demonstrate the potential for biologically informed machine learning models to discover protein assemblies that explain drug resistance and enable new therapeutic avenues. Citation Format: Oliver Priebe, Eric Ho, Fransisco Guedes, Chandni Khandwala, Eduardo Farias, Trey Ideker, Adam Yaari, Maxwell Sherman. Biologically informed machine learning for discovery of protein assemblies underlying cancer therapeutic resistance [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 549.