Deep representation learning for anticancer sensitivity prediction

Variability in tumor response to anticancer drugs remains a major challenge in oncology. In vitro assays, which assess how cancer cells respond to varying drug concentrations, provide valuable insights into drug efficacy. While the response to single-agent therapies can typically be modeled with a dose-response curve, combination therapies require modeling more complex, multidimensional response surfaces. However, experimental limitations and biological heterogeneity hinder the ability to exhaustively characterize all drug–tumor–dose combinations. This motivates the development of computational models that can generalize to unseen scenarios, accelerating therapeutic discovery and guiding personalized treatment strategies. This thesis introduces three machine learning models for dose-response prediction, each targeting a specific aspect of this problem. The first model, CANDELA, estimates IC50 values using genomic features and molecular graphs, offering a compact summary of drug efficacy. The second, ARCANet, predicts full dose-response curves for individual drugs, enabling finer-grained modeling across concentrations. The third, PanThera, leverages geometric deep learning to predict full dose-response surfaces for drug combinations, with architectural features designed to enforce biologically and geometrically meaningful inductive biases such as treatment permutation invariance and sham compliance. Empirical evaluations demonstrate that CANDELA and ARCANet outperform established baselines on their respective prediction tasks. PanThera achieves comparable or improved performance relative to specialized synergy models, particularly in out-of-distribution settings such as unseen concentrations or novel drug combinations. Notably, although trained only on single- and double-drug treatments, PanThera can generate plausible predictions for higher-order combinations, suggesting its potential utility in exploratory drug combination studies. These findings underscore the potential of deep learning to enhance drug development pipelines and contribute foundational capabilities toward AI-assisted therapy recommendation systems.