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

Stage 3 zero-shot protein-ligand binding-site prediction using SMILES achieves pooled residue-level F1 of 0.3109 and ligand-macro F1 of 0.2338 on previously unseen ligands.

• Why it matters: Binding-site prediction performance varies with ligand family and data availability, and it remains unclear how well sequence-only models generalize to ligands absent from training.
• What they did: They built a BioLiP2-derived, ligand-aware benchmark with ligand-specific sequence clustering and trained a three-stage sequence-based suite, replacing ligand embeddings with a pretrained chemical language model over SMILES for zero-shot.
• The result: Zero-shot performance is supported by controls where constant, mismatched, or scrambled SMILES reduce results across test regimes, and scaling PLM size generally helps while CLM scaling yields smaller, less consistent gains.

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