AI and machine learning for predicting in vitro/preclinical drug toxicities and adverse events in clinical patients

Clinical toxicity accounts for approximately 70% of drug development failures, as 40% of failures attributed to poor efficacy may often be due to dose-limiting toxicities that prevent adequate dose escalation. Conventional drug toxicity studies from in vitro assays and preclinical animal models often fail to predict adverse events (AEs) in humans at clinically relevant doses. To address this gap, the U.S. Food and Drug Administration has incentivized New Approach Methodologies to supplement or replace traditional animal-based toxicity testing. Recent advances in artificial intelligence (AI) and machine learning (ML) have enabled the development of predictive models for drug-induced toxicity in vitro , in animal models, and in clinical settings. These models leverage molecular structure to forecast outcomes such as hERG channel inhibition, Ames mutagenicity, hepatotoxicity, and LD 50 values in animals. AI/ML approaches have also been employed to predict clinical AEs based on drug-specific pharmacogenomic targets or across drug classes in patient populations; other models aim to create AE predictions using the chemical structure and target profiles of hundreds or thousands of drugs. Despite rapid and ongoing progress, current AI/ML frameworks remain inadequate for accurately predicting clinical AEs at therapeutic doses. Future models must incorporate drug structure, on- and off-target interactions, tissue- and cell-specific selectivity, target expressions, and clinical doses to enhance predictive performance.