Integrating artificial intelligence across the cancer drug discovery pipeline using a design-test-refine workflow

Abstract

Artificial intelligence-driven drug design (AIDD) is increasingly transforming cancer drug discovery, but its applications are often considered as individual computational tasks rather than interconnected stages of the drug discovery pipeline. This mini-review addresses this gap by firstly identifying the following four interconnected stages of anticancer drug development (1) target identification and biomarker-guided prioritization (2), structure-based and generative molecular design (3), perturbational mechanism-of-action assessment, and (4) drug response, resistance, and combination prioritization. It, secondly, describes how AI can be integrated across these four stages in a design-test-refine workflow where AI-generated predictions are progressively evaluated and refined through experimental and patient-relevant evidence. We emphasize that the role of AI in predicting target dependency, molecular activity, mechanism-of-action, or drug response should be to guide rather than replace experimental discovery or clinical judgment. Key limitations to the application of AIDD in cancer are highlighted, including in training and benchmarking data, in accounting for biological heterogeneity, as well as in model generalization. Importantly, robust validation across increasingly complex cancer models is required and it is proposed that AIDD should be used to prioritize testable treatment predictions. Ultimately, translationally useful AIDD workflows should move beyond isolated predictions toward iterative, biologically informed therapeutic development, to form a holistic design-test-refine workflow.