Ligand Conformational Variability Enhances Machine Learning Prediction of Protein-Ligand Binding Affinity

Rapid and accurate screening of protein-ligand binding affinities using machine learning (ML) remains a challenging yet critical task in drug discovery. In this work, two closely related challenges are addressed: the inherent sensitivity of the binding affinity to the ligands' geometry within the protein-ligand complex and the absence of prior knowledge of this geometry for previously unseen compounds. Three representative ML methods of varying complexity─Kernel Ridge Regression (KRR), SchNet, and Polarizable Atom Interaction Neural Network (PaiNN)─were evaluated on prediction of the binding affinities of compounds against the main protease SARS-CoV-2 (M^pro). Employing a multi-instance learning framework, models were trained on multiple ligand conformers per compound and tested on sets of unseen compounds represented by multiple conformers unrelated to the binding geometry. Comprehensive error analysis reveals that the incorporation of multiple conformers of unseen compounds significantly improves the prediction accuracy, exceeding the gains achieved by refining individual models alone.