Biologically Interpretable Representation Learning for Mechanistic Insights into Cancer Immunotherapy Resistance

Abstract

Resistance to immune checkpoint inhibitors (ICIs) remains a critical barrier to effective cancer therapy, driven by complex, multi-scale interactions that current biomarkers often fail to capture. This dissertation introduces the Biologically Disentangled Variational Autoencoder (BDVAE)—an interpretable deep learning framework designed to uncover mechanistic drivers of ICI resistance through multi-omic data integration. Using RNA-seq and wholeexome sequencing data from 366 patients across melanoma, renal cell, urothelial, and gastric cancers, BDVAE learns low-dimensional latent representations that are both predictive of response and biologically meaningful. The model reveals distinct latent dimensions aligned with immune regulation, tumorintrinsic signaling, metabolism, and neuroimmune interactions. SHAP-based interpretation and pathway analysis highlight key resistance-associated programs, including immunosuppressive cytokine signaling, metabolic signaling, and neuroactive pathways such as calcium and cAMP signaling. Unsupervised clustering identifies three tumor subtypes—responder-dominant, non-responder-dominant, and an intermediate group—suggesting plastic or transitional immune states. Survival analyses confirm the clinical relevance of these clusters and expose heterogeneity within standard RECIST categories. Overall, this work presents a novel, interpretable framework for modeling ICI response, offering insights into resistance mechanisms and actionable paths for biomarker discovery, patient stratification, and therapeutic innovation in precision immuno-oncology.