From artificial intelligence promise to translational practice in drug discovery: Reproducibility, validation and evidence-based implementation

Artificial intelligence (AI) can support drug discovery, but benchmark performance alone does not establish experimental, operational or clinical benefit. This keynote review distinguishes predictive generalization, experimental confirmation, workflow utility and clinical outcomes, separating demonstrated failures from missing validation. We critically examine quantitative case studies of benchmark bias, prospective hit finding and clinical development, alongside data quality, reproducibility, uncertainty and governance. Practical recommendations link each intended use to appropriate comparators, evaluation designs and acceptance criteria. The aim is not to presume that AI fails to translate but to specify what evidence is needed to demonstrate reliable prospective use in drug discovery.