Improving Stereochemical Limitations in Protein–Ligand Complex Structure Prediction

AlphaFold3 has revolutionized biology by enabling the prediction of protein complexes with various biomolecules, including small molecular ligands. However, the ligand structures predicted by the model often contain stereochemical errors. In this study, our comprehensive evaluation of AlphaFold3 and its clone model, Boltz-1, demonstrated significant limitations in the ligand stereochemistry of their predicted structures, including chirality, bond lengths, and bond angles. To address this issue, we developed the restraint-guided inference method that applies stereochemical restraints during the reverse diffusion process. This approach perfectly reproduces the chirality specified in the input chemical structure and improves bond and angle geometries of ligands, while maintaining comparable performance in binding pose prediction. Our method provides a practical solution to the stereochemical errors in predicting protein-ligand complexes, thereby enhancing structure prediction applications in structural biology and drug discovery.