A critical appraisal of structure-based virtual screening, binding free energy methods, and AI-driven molecular design across the evolving computational landscape of drug discovery.

Structure-based drug design rests on the physical chemistry of molecular recognition, and every computational advance represents a successive attempt to capture that reality with greater fidelity and at greater scale. This review examines that progression through three lenses: the empirical performance of prospective structure-based virtual screening across 419 documented case studies, the accuracy of binding free energy methods, and the capability of generative artificial intelligence to design molecules de novo. Across the surveyed campaigns, the best hit was better than 1 micromolar in 113 of 419 studies. Ensitrelvir, discovered by virtual screening of a proprietary library followed by structure-based optimisation, received emergency regulatory approval in Japan in November 2022 and standard approval in March 2024. Scaling a docking library from 99 million to 1.7 billion compounds against the same target raises the confirmed active rate, but the reported enhancement factor ranges from 1.09-fold to 4.18-fold depending on the activity threshold and selection rule applied, so the direction of the effect is more robust than its magnitude. Extended-sampling free-energy perturbation reduces prediction RMSE to 0.78 kcal/mol, approaching the median reproducibility of the experiments against which such methods are judged. Rentosertib, the first molecule whose target and structure were both generated computationally to reach a randomised trial, met a safety primary endpoint in a 71-patient Phase IIa study; its lung-function findings were secondary and exploratory. The review introduces the Computational-to-Clinical Translation Index as a proof-of-concept reporting framework with an explicit operational specification, and articulates a mechanistically grounded research agenda.