BOLD-GPCRs

G protein-coupled receptors are the target of roughly a third of marketed drugs, yet only about 134 of the ~800 human receptors have an approved drug against them. The gap is largely a data problem: ligand-based and structure-based screening both depend on prior chemical or structural knowledge, so the receptors that most need new chemistry — orphan and understudied subtypes — are exactly the ones those methods serve worst. BOLD-GPCRs (BERT-Optimized Ligand Discovery for GPCRs) attacks that gap by transferring what a protein language model already knows about receptor sequence and function onto the bioactivity prediction task.

Developed by Davide Provasi, Kirill Konovalov and colleagues in Marta Filizola's lab at the Icahn School of Medicine at Mount Sinai and posted to bioRxiv in August 2025, the framework represents each receptor with ProteinBERT embeddings and each ligand with a dense network over RDKit molecular descriptors, concatenates the two into a simple multimodal vector, and classifies the pair as active or inactive. Because receptor identity enters as a residue-level sequence rather than a one-hot label, the same trained model also scores point mutations — a capability that per-receptor classifiers cannot offer at all.

The published version appeared in the Journal of Chemical Information and Modeling in January 2026. A hosted web application makes the pretrained model usable without any local setup.

#Key Features

• Family-wide single model: One checkpoint covers all non-olfactory class A GPCRs rather than a separate classifier per receptor, so understudied subtypes borrow signal from data-rich relatives.
• Mutation-effect prediction: Residue-level sequence encoding lets the model infer how signaling-relevant point mutations change ligand activity, supporting mutagenesis design and work on disease-associated receptor variants.
• Attention-based interpretability: Global attention scores concentrate on conserved functional motifs including DRY and NPxxY, the sodium-binding pocket, ICL2, and helix 8, indicating the model attends to residues implicated in receptor activation and transducer coupling.
• Hosted screening interface: The web app accepts FASTA receptor sequences and SMILES ligands, handles up to 1,000 ligands per submission, and can score every receptor-ligand combination to profile selectivity and off-target risk.

#Technical Details

ProteinBERT version 1.0.1 supplies the receptor track: six transformer blocks running paired local (sequence) and global (function) representations, pretrained self-supervised on 106 million UniRef90 sequences plus 8,943 Gene Ontology annotations. Ligands are encoded as 43 RDKit descriptors covering physicochemical properties, atom counts, and topological indices, passed through parallel dense-and-dropout blocks. The concatenated embedding feeds a final dense layer with softmax output. Supervised fine-tuning used all non-olfactory class A GPCR records from ChEMBL — 383,638 receptor-ligand pairs (214,816 active, 168,822 inactive) thresholded at 1 μM potency — in two stages, ten epochs with ProteinBERT frozen followed by ten epochs unfrozen. Mutation data came from GPCRdb: 10,920 datapoints spanning 592 ligands, 2,459 mutations, and 45 targets.

On random 20% held-out splits the model reaches average precision 0.72 and recall 0.85, versus 0.66/0.86 for a conventional DNN and 0.64/0.82 for a random forest trained on the same descriptors with one-hot receptor labels. Holding out entire targets, precision stays at 0.66 against 0.64 and 0.62 for the baselines. Withholding the whole opioid subfamily drops average precision from 0.78 to 0.68 and recall from 0.89 to 0.67. On mutation data, one extra epoch of few-shot fine-tuning lifts recall from 0.43 to 0.96 while holding precision at 0.85. The gain over baselines is concentrated in precision; recall is essentially unchanged. Processed training data is on Zenodo under CC BY 4.0, but neither source code nor trained weights are distributed — the model is reachable only through the hosted app.

#Applications

Medicinal chemists and pharmacologists can use BOLD-GPCRs to triage compound libraries against receptors with little or no annotated chemistry, including orphan GPCRs, and to profile a candidate across the class A family for selectivity before committing to assays. The mutation channel supports precision-pharmacology questions — whether a patient-derived receptor variant will alter a drug's activity — and helps prioritize which residues to mutate in mechanistic studies.

#Impact

BOLD-GPCRs is a concrete demonstration that protein language model embeddings carry transferable pharmacological signal: swapping a one-hot receptor label for a ProteinBERT representation improves precision across the family and, more importantly, degrades gracefully on receptors never seen in training. The attention analysis adds supporting evidence that these representations encode activation-relevant structure from sequence alone. The main limits are honest ones the authors state: recall does not improve over simpler baselines, reliability varies for receptors with sparse or atypical pharmacology, high-attention residues outside known functional motifs remain experimentally unverified, and the absence of released code or weights means independent reproduction depends on the hosted server.