Ligand-aware protein language model that cross-attends SaProt embeddings to ligand SMILES, beating SaProt across six downstream benchmarks.
No providers recorded yet. Browse all providers
Protein language models learn from sequence and, more recently, from predicted structure — but not from the small molecules that proteins actually work with. Cofactors, metabolites, and metal ions determine which conformations a protein adopts and often decide its function outright, yet a model trained only on residues and backbone geometry cannot distinguish paralogues that share 80% sequence identity while binding different metabolites. ProtLigand, from Dan Kalifa and Kira Radinsky at the Technion with Eric Horvitz at Microsoft, addresses that gap by letting ligand chemistry shape the protein representation during pre-training itself.
The construction is straightforward. ProtLigand takes SaProt — a structure-aware protein language model whose tokens pair each residue with a Foldseek 3Di structural state — and adds a cross-attention pathway to a small-molecule encoder. During masked language modelling, ligand-derived keys and values inform the protein's residue queries, so the representation used to reconstruct masked residues has already been conditioned on the cognate binder. No docking coordinates are required: the inputs are a sequence, an AlphaFold 2 backbone, and a SMILES string.
Existing protein–ligand methods take the opposite approach. Docking pipelines such as DiffDock and Uni-Mol train separate protein and molecule encoders and fuse them only at the pose-prediction head, leaving the protein representation itself ligand-blind. The work was posted to bioRxiv in May 2025 and published in Bioinformatics in August 2025.
Pre-training uses PDBbind v2020, from which 17,393 protein–ligand pairs (14,331 training, 3,062 validation) covering 3,347 unique proteins across 12 protein families were assembled. Structures come from AlphaFold DB by UniProt ID with pLDDT-below-70 regions filtered out, following SaProt's protocol. The protein encoder is the 650M-parameter SaProt AF2 checkpoint; the ligand encoder is ChemBERTa-77M-MLM.
Evaluation covers six tasks drawn from the SaProt benchmark suite: HumanPPI from PEER, thermostability regression on the FLIP human-cell split, metal ion binding, both the binary and 10-class DeepLoc subcellular localization variants, and Enzyme Commission number prediction from DeepFRI. ProtLigand leads on all six against six baselines, with Cohen's d effect sizes above 0.8 throughout and results averaged over three fine-tuning seeds with Holm–Bonferroni-corrected paired t-tests. The largest gains fall where ligand annotation is densest: HumanPPI accuracy rises from 86.67% to 90.00% and thermostability Spearman ρ from 0.710 to 0.731. Gains on DeepLoc and EC are smaller, consistent with ligand modulation being less central to localization and enzyme classification.
The model is a drop-in replacement for a general protein encoder wherever binding chemistry is likely to matter — thermostability engineering, cofactor-dependent enzyme characterization, and interaction prediction for proteins whose partnerships are mediated by shared ligands. The ligand generator and decoder extend it to hypothesis generation for orphan targets, proposing plausible binder scaffolds from sequence and structure alone. The authors also frame the model as a research instrument rather than a leaderboard entry: cases where adding ligand context produces an outsized prediction gain flag interactions worth experimental follow-up, an argument they illustrate with heme-dependent proteins such as heme oxygenase-2 and the autophagy pair ATG7/ATG10, which SaProt misclassifies.
ProtLigand demonstrates that a third modality can be folded into an existing protein language model cheaply — no retraining from scratch, no docked poses — and still yield consistent, statistically controlled gains over a strong sequence–structure baseline. The evaluation is careful by the standards of this literature, with homology-controlled splits, multiple seeds, multiple-comparison correction, and calibration analysis alongside raw accuracy. The main constraint is data scale: pre-training draws on 17,393 PDBbind pairs, orders of magnitude smaller than the sequence corpora behind the underlying language models, which bounds the chemical diversity the cross-attention can learn. Checkpoints for both the encoder and the ligand generator are released, though hosted on Google Drive rather than a model hub, and the code is MIT-licensed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.