Artificial Intelligence in Computational Drug Discovery: Reproducibility, Validation, and Decision Stability

Abstract

Artificial intelligence (AI), machine learning, and molecular modeling have become integral components of modern computational drug-discovery workflows. But, as has been demonstrated in the industry, success in benchmark datasets does not always guarantee success in real screening projects, or in different experimental conditions. Given this limitation, the present review investigates the ability of AI-driven computational workflows to support a stable compound prioritization process in the face of uncertainty, but not just a higher retrospective score. Relevant literature published from 2018 to 2025 is reviewed, in addition to previous docking and virtual-screening studies, and community benchmarks, as well as reports of limitations under practical discovery conditions. The reviewed evidence indicates that unnoticed data leakage, validation approaches that do not mimic real-world utilization, sensitivity to docking protocols and preprocessing, stochastic variation across runs, and selection effects in large-library screening are some of the recurring factors that can lead to overconfidence. In real cases seemingly minor modifications, such as protonation-state assignments, definition of the docking-box, or random-seed selection, could make a significant difference in compound ranking that affects decisions in the lab. Validating the model in the deployment environment, performing replicate analyses, checking for rank agreement, validating the model through appropriate negative controls, and developing calibration methods are also important operational considerations for reliable prioritization. These findings highlight the growing importance of AI-assisted computational workflows that emphasize validation, reproducibility, and reliable decision support in modern drug-discovery research. Reproducibility is still important but the real utility of computational workflows is to help make decisions that are defensible in experimentation under realistic discovery conditions.