ChemGLaM

Predicting whether a small molecule binds a given protein is the workhorse computation of early drug discovery, and the practical bar is higher than benchmark accuracy alone. A model trusted to triage a screening library has to generalize to compounds and targets it has never seen, say how confident it is so that low-confidence calls can be discarded, and expose some structural rationale for its predictions. ChemGLaM, the chemical genomics language model from Takuto Koyama and colleagues at Kyoto University, was designed around those three requirements rather than around a leaderboard.

The architecture is deliberately modular: a chemical language model reads the compound as SMILES, a protein language model reads the target sequence, and an interaction block joins them with cross-attention, so the compound representation queries the residues of the protein directly. Because both encoders are pretrained on far more chemistry and sequence data than any compound-protein interaction (CPI) dataset contains, the fusion layer is the only part that must learn interaction-specific structure — which is what lets the model extrapolate to unseen drug-target pairs.

First posted to bioRxiv in February 2024 and published in the Journal of Cheminformatics after successive revisions, the work is notable less for its benchmark margins than for what the authors did with one trained model: they fine-tuned it once on a large ChEMBL-derived dataset, then deployed that single checkpoint across the entire drug-by-human-proteome space and into a patient-derived cell screen.

#Key Features

• Cross-attention fusion of two pretrained language models: The interaction block computes attention from compound tokens to protein residues, replacing fixed kernel similarities with a learned, context-dependent notion of which parts of a target matter for which parts of a ligand.
• Calibrated uncertainty via evidential deep learning: Epistemic uncertainty is estimated from a single forward pass, avoiding the cost of deep ensembles or MC dropout and making confidence filtering practical at library scale.
• Attention maps as structural rationale: Residues receiving high attention concentrate around known ligand-binding sites even though the model never sees a structure, illustrated on the glutathione transporter–glutathione disulfide complex (PDB 4C0R).
• Whole-proteome deployment: One trained checkpoint scored every pairing of 11,455 DrugBank drugs with 20,434 human proteins in roughly 72 hours on 12 NVIDIA T4 GPUs, at most 13 ms per pair, and the resulting predictions are published as a public database.

#Technical Details

The compound encoder is MoLFormer — specifically the publicly released checkpoint pretrained on 10% of the MoLFormer-XL corpus, since the full model is not distributed — and it is fine-tuned along with the interaction block. The protein encoder is ESM-2 t36-3B-UR50D, 36 transformer blocks and 3B parameters, held fixed on the grounds that it dwarfs the available CPI data. Training uses AdamW. Evaluation is a five-fold cross-validation in which test sets contain only compounds and proteins absent from training. On BindingDB classification, ChemGLaM ranked first in AUROC (0.829 ± 0.042), F1 (0.756 ± 0.021), MCC (0.507 ± 0.077), and accuracy (0.754 ± 0.039); on Davis it ranked first in AUPRC (0.561 ± 0.119) and MCC (0.331 ± 0.095). On PDBbind affinity regression it reached RMSE 1.521 ± 0.047 and Spearman 0.596 ± 0.024, ahead of the Vina scoring function at 0.511 ± 0.037. The authors report candidly that on some settings more recent methods such as PSICHIC and BarlowDTI outperform it. The large-scale predictions come from a model fine-tuned on 1,392,072 CPI records derived from ChEMBL v34.

#Applications

The clearest use is hit triage: rank a screening library against a target, then use the uncertainty estimate to drop out-of-domain compounds before committing assay resources. The prediction database also supports the reverse, polypharmacology query — given a drug, which human proteins might it engage. The authors demonstrated both on an amyotrophic lateral sclerosis screen in patient iPSC-derived motor neurons, scoring 1,607 compounds against 66 reported ALS targets and recovering experimentally validated hits with an enrichment factor of 3.04 at 10% and AUC 0.693, with the top predicted targets converging on the Src/c-Abl pathway that the original phenotypic screen had implicated.

#Impact

ChemGLaM demonstrates that composing two frozen or lightly tuned foundation models through a thin learned interface can reach competitive CPI accuracy at modest compute, and that the result is usable prospectively rather than only on held-out folds. Publishing whole-proteome predictions as a queryable resource is the more durable contribution, since it puts the model's output in reach of researchers who will never run it. Availability is uneven, though: the Apache-2.0 repository provides training and inference code, a Docker image, and per-benchmark cross-validation configs, but the ChEMBL-trained production checkpoint behind the database and the ALS analysis is not distributed, and the README's finetuned-model inference section is still marked as forthcoming. Reproducing the paper's benchmarks is straightforward; reusing its production model is not.