From Artificial Intelligence Outputs to Medicinal Chemistry Decisions in Kinase Inhibitor Discovery: Evidence Boundaries, Resistance-Aware Design, and DMTA Learning

Abstract Kinase inhibitor discovery increasingly requires integrated decision-making that extends beyond biochemical potency. Selecting the next analogue, defining the assay package, evaluating structural hypotheses, anticipating resistance, and assessing developability collectively require evidence from distinct experimental layers. In this Perspective, we argue that the practical value of artificial intelligence-assisted drug discovery (AIDD) in kinase inhibitor programs should be assessed not by predictive performance alone but by whether computational outputs improve medicinal chemistry decisions that can be implemented and experimentally tested. We treat AI-supported outputs as testable hypotheses whose value is determined by their effect on analogue selection, assay prioritization, resistance panel design, and information gain across Design−Make−Test−Analyze (DMTA) cycles. Using resistance-focused examples from EGFR, ALK/ROS1/TRK, BCR−ABL1, and BTK, we illustrate how explicit evidence boundaries, validation gates, and falsifying experiments can guide redesign decisions and learning updates.