Data-centric training enables meaningful interaction learning in protein–ligand binding affi nity prediction

Predicting protein–ligand binding affi nity remains a fundamental challenge in drug discovery, where machine learning models often struggle with rotational sentitivity, overfi tting, and poor generalization. Here we present a CNN–based scoring function that achieves robust and stable performance through careful attention to training methodology rather than architectural complexity. We systematically evaluate data augmentation strategies and fi nd that random rotation, unlike commonly used 90-degree rotations, eff ectively mitigates rotational variance. We introduce molecular dropout, a simple regularization technique that reduces reliance on isolated protein or ligand features and promotes learning of intermolecular interactions. The resulting model generalizes across diverse test conditions, including time-based and structurally disjoint splits, and performs competitively on standard and external benchmarks. Through systematic ablation studies, error analysis, and interpretability experiments, we provide insights into model behaviour, highlighting the importance of data-centric design in developing reliable scoring functions.