Answers are grounded in this paper and cite their sources.
Generative artificial intelligence (AI) is increasingly used for de novo small-molecule design, but the practical and translational utility of generated candidates depends on the configuration of the entire discovery workflow rather than on generative model architecture alone. This review adopts a workflow-based perspective on generative molecular design, tracing the process from existing molecular data sets, preprocessing and representation, to molecule generation, conditioning, candidate prioritization, evaluation, and iterative refinement. It examines how generative AI model families, constraints, and predictive oracles interact within this pipeline, and evaluates generated candidates using intrinsic metrics, computational assessments, and experimental and translational evidence. Active learning, computational and experimental feedback, and human expert input are further discussed as mechanisms for closed-loop refinement. Across these stages, we identify data construction and representation choices, oracle quality, evaluation protocols, and the availability of informative feedback as major determinants of how model performance should be interpreted and whether generated structures progress toward biologically and translationally meaningful candidates, with selected examples. In conclusion, reliable generative drug discovery requires coordinated design across the entire workflow rather than isolated advances in individual components.
No takes yet. Share an insight, caveat, or question.
Fan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: