Unleashing AI for preclinical drug development

Artificial intelligence (AI) is reshaping preclinical drug development across target discovery, therapeutic design, synthesis, functional characterization and safety assessment. In this review, we present an integrated view of AI across the preclinical pipeline, beginning with target identification, modality selection and representation learning for molecular, protein, structural, cellular and multimodal data. We then examine major generative modeling paradigms including autoregressive, latent-variable, diffusion, flow-based and guided approaches and their applications across small molecules, peptides, enzymes, antibodies, Tcell receptors and mRNA therapeutics. We discuss how these methods support de novo design, structure- and pocket-conditioned generation, sequence–structure co-design, affinity and activity optimization, developability engineering and synthesis-aware discovery. We further review AI-enabled retrosynthesis, reaction prediction and closed-loop experimental optimization, emphasizing the distinction between computational synthesizability and experimentally executable chemistry. Beyond molecular generation, we organize functional prediction from binding and kinetics through target engagement, pharmacological activity, mechanism of action, pathway response and phenotypic outcome. We also examine singlecell perturbation modeling, organoid-based drug-response prediction, toxicity and off-target assessment, and prospective validation across therapeutic modalities. Across these areas, we highlight the importance of uncertainty estimation, out-of-distribution generalization, reproducibility and experimentally grounded benchmarking. Finally, we discuss how integrating generative AI, mechanistic biology and iterative design–make–test–analyze cycles could transform AI from isolated predictive tools into decision-support systems for preclinical drug development.