Zero-Shot Protein-Ligand Binding Site Prediction

This zero-shot protein-ligand binding site predictor identifies, at residue resolution, which positions in a protein contact a given small molecule — using only the protein sequence and the ligand's SMILES string, with no structural input. Binding site prediction accuracy has long varied widely across ligand families and data regimes: common cofactors have thousands of annotated examples, while most ligands have only a handful or none at all. The work reframes this as a stratified problem, evaluating performance separately on overrepresented, underrepresented (few-shot), and zero-shot (entirely unseen) ligands.

Developed by Mahdi Pourmirzaei, Salhuldin Alqarghuli, Kai Chen, Mohammadreza Pourmirzaei, and Dong Xu at the University of Missouri, with a collaborator at Politecnico di Milano, and posted to bioRxiv in September 2025, the method is a three-stage modeling suite that progressively adds ligand conditioning and, ultimately, zero-shot capability. Stage 1 trains a separate predictor per ligand on top of a frozen protein language model; Stage 2 merges these into a single multi-ligand model via a learned ligand embedding table; Stage 3 replaces that table with a pretrained chemical language model over SMILES, so the model can generalize to ligands never seen during training.

The suite follows the established pattern of thin supervised heads over frozen pretrained biological language models, extending it into cross-modal, zero-shot territory by pairing a protein encoder with a chemical encoder.

#Key Features

• Sequence-and-SMILES input: Predicts binding residues from protein sequence and a ligand SMILES string alone, requiring no experimental or predicted 3D structure.
• Three-stage design: Progresses from per-ligand predictors (Stage 1) to a single conditioned multi-ligand model (Stage 2) to a fully zero-shot model (Stage 3).
• Genuine zero-shot generalization: Stage 3 reaches F1 = 0.31 on 5,612 previously unseen ligands, a true unseen-ligand result rather than a held-out split of the training library.
• PLM scale is the dominant lever: Ablations show that enlarging the protein encoder consistently improves accuracy across all regimes, while scaling the chemical encoder gives only modest or inconsistent gains.

#Technical Details

The protein encoder is ESM-2, kept frozen, with the headline Stage-3 model using the 650M-parameter variant; ablations sweep ESM-2 from 8M to 3B parameters. Ligands are represented by a chemical language model over SMILES — primarily MolFormer, with UniMol-2 also evaluated. A lightweight classification head maps the combined protein and ligand embeddings to per-residue binding predictions. Stage 2 raises Macro F1 on the overrepresented test set from 0.4769 (Stage 1) to 0.5832, outperforming sequence- and structure-based baselines, while Stage 3 attains a zero-shot F1 of 0.3109 on 5,612 unseen ligands and remains competitive on represented ligands. Across five ESM-2 scales and multiple chemical models, larger protein backbones consistently increase Macro F1. The code is released open source on GitHub, without a stated license.

#Applications

The suite supports mechanistic biology and early-stage drug discovery, where knowing which residues a candidate small molecule engages helps prioritize targets, interpret activity, and guide medicinal chemistry. Because Stage 3 requires only a sequence and a SMILES string, it can screen novel or poorly characterized ligands against proteins that lack an experimental structure — settings where docking-based tools are costly or unreliable.

#Impact

The work demonstrates that residue-level binding site prediction from sequence and SMILES is feasible even for ligands unseen at training, and identifies protein-language-model scale as the primary axis for future improvement. The absolute zero-shot F1 of 0.31 is modest, reflecting the difficulty of true unseen-ligand generalization rather than a limitation of the framing. As an open-source preprint whose repository carries no stated license, its influence will grow as the community builds on the stratified evaluation protocol and tests larger protein backbones.