Zero-Shot Protein-Ligand Binding Site Prediction from Protein Sequence and SMILES

Stage 3 enables zero-shot residue-level protein-ligand binding-site prediction for previously unseen ligands, reaching pooled residue-level F1 0.3109 and ligand-macro F1 0.2338.

• Why it matters: Predicting binding sites is crucial for biology and drug discovery, but performance varies with ligand family and how much ligand-specific data is available, including the zero-shot case.
• What they did: Using a BioLiP2-derived benchmark with ligand-specific sequence clustering, a three-stage sequence-to-binding-site framework was trained, with Stage 3 conditioning on SMILES via a pretrained chemical language model.
• The result: Control tests with constant, mismatched, and scrambled SMILES inputs reduced performance across all regimes, supporting ligand-input dependence, while scaling PLM size generally improved results.

AI summary of the abstract · may contain errors · not medical advice · read the paper.