Abstract
Predicting protein–ligand binding affinity remains a fundamental challenge in drug discovery, where machine learning models often struggle with rotational sensitivity, overfitting, and poor generalization. Here, we present a convolutional neural network (CNN)–based scoring function that emphasizes the careful design of the training strategy over architectural complexity. Systematic evaluation of data augmentation strategies shows that random rotation, unlike commonly used 90-degree rotations, effectively mitigates rotational variance. We also introduce a simple regularization technique that reduces reliance on isolated protein or ligand features and, together with random rotation, promotes learning of meaningful intermolecular interaction patterns. The resulting model, built on a simple architecture, generalizes across diverse test conditions, including structurally disjoint splits, and matches or exceeds the performance of state-of-the-art methods. Through systematic ablation studies and interpretability experiments, we provide insights into model behaviour, highlighting the importance of data-centric design in developing reliable scoring functions.

This publication has 56 references indexed in Scilit: