Machine Learning for Protein-Ligand Binding Prediction

Abstract

Protein-ligand binding prediction is a cornerstone of computational drug discovery, determining whether a candidate molecule will bind to a therapeutic target with sufficient affinity and selectivity, and machine learning models that predict binding affinity from molecular features now complement physics-based docking and free-energy methods, yet their predictive accuracy and generalisability vary widely across model architectures and training paradigms. We evaluated 212 ML-based protein-ligand binding prediction programmes across centres affiliated with the Swiss Institute of Machine Intelligence in Zurich between 2017 and 2021, spanning five model categories: descriptor-based regression models, structure-based convolutional scoring functions, graph neural network approaches, sequence-based binding predictors, and physics-informed hybrid models. A Protein-Ligand Binding Prediction Index (PLBPI) was constructed from five sub-scores -- binding affinity prediction accuracy, pose discrimination capability, scaffold-hopping generalisation, virtual screening enrichment, and prospective hit-rate validation -- with weights from regression against sustained adoption into drug discovery pipelines. PLBPI correlated with adoption at r = +0.83 and discriminated adopted from non-adopted models with an AUC of 0.880. Structure-based CNN scoring functions scored highest (mean PLBPI 0.820), while sequence-based predictors trailed at 0.594. Only 34.4 percent exceeded the 0.75 threshold. Affinity prediction accuracy carried the largest weight (beta = +0.280), followed by virtual screening enrichment (beta = +0.228).