New guidelines aim to standardize evaluation of AI methods for protein–ligand binding prediction by addressing data leakage, sampling bias, and run-to-run variability. The framework focuses on benchmarking discipline, reflecting recurring failures when models appear strong in offline settings but degrade in realistic use. For the biopharma pipeline, improved evaluation practices matter because protein–ligand prediction models are frequently used downstream in virtual screening and hit prioritization. Industry teams can treat the guidance as a requirements checklist for future model comparison studies, with emphasis on reproducibility and contamination control.