Ptarmigan-1

Virtual screening asks which molecules in a vast chemical space are worth synthesizing and testing. Nearly every modern answer routes through an explicit three-dimensional pose — docking places a ligand in a pocket, and co-folding models such as Boltz-2, Protenix, and AlphaFold 3 build the complex outright. That presumes a well-defined pocket exists. The non-orthosteric, cryptic, and intrinsically disordered sites where much unexplored ligandability sits offer none, and pose construction is slow enough to put billion-compound screens out of reach.

Ptarmigan-1, from Talus Bioscience in Seattle, removes the pose entirely. It co-embeds each residue of a protein and each candidate compound into a shared 256-dimensional latent space from amino acid sequence and two-dimensional chemistry alone. Engagement is the cosine similarity between a residue and a compound, so screening reduces to nearest-neighbor retrieval against a library embedded once. The work was posted to bioRxiv on 30 July 2026 and has not yet been peer reviewed.

That separates Ptarmigan-1 from earlier contrastive protein-ligand screeners. LigUnity and DrugCLIP also score by proximity in a joint space, but both encode the protein as a resolved three-dimensional pocket and so inherit docking's structural prerequisite. Ptarmigan-1 needs no structure at any stage, and its per-residue predictions report where a compound engages, not only whether it binds.

#Key Features

• Pose-free, residue-resolved scoring: Each prediction is a per-residue engagement map from sequence and SMILES, nominating the site a compound engages without constructing a three-dimensional pose.
• Mixed-resolution training: Paired contrastive objectives, one over residues and one over whole proteins, let a single model learn from residue-resolved chemoproteomics and from assays recording only a binary binding outcome.
• Screening as retrieval: Libraries are embedded once into an approximate nearest-neighbor index, so each additional target becomes a lookup rather than a fresh screen.
• Zero-shot generalization: The model enriched actives above random for 32 of 33 held-out targets whose sequences never appear in training, at a median adjusted logAUC of 0.31.
• Emergent cryptic-pocket detection: Marginalizing engagement over a diverse library yields a ligandability score that finds cryptic sites the model was never trained to predict.

#Technical Details

Ptarmigan-1 couples two frozen backbones: ESM Cambrian at 600M parameters for protein sequences and ChemBERTa for compound SMILES. Only rank-32 LoRA adapters and two linear projection heads are trained, mapping both encoders into an L2-normalized 256-dimensional space; a temperature-scaled softmax over residue similarities gives a calibrated protein-level engagement probability. Training ran 25,000 steps at an effective batch of 1,024 pairs on Protein Data Bank complexes (residues within 5 Å of a ligand labeled engaged), BindingDB, KIBA and LCIdb at a 10 µM threshold, four published activity-based protein profiling studies, and a proprietary internal chemoproteomic corpus supplying most residue-level labels.

On the five LIT-PCBA target sets with published structure-based baselines, Ptarmigan-1 ranked second by adjusted logAUC behind Boltz-2 and above the docking and co-embedding methods scored alongside. Across 874 COValid covalent actives the engaged cysteine fell in the top 1% of residues for eight of nine targets, and on 337 drug-like PoseBusters complexes the median pocket-versus-rest AUROC was 0.99, with the top residue within 5 Å of the ligand for 92% of complexes. On CryptoBench the per-protein AUPRC was 0.27 against 0.22 for P2Rank given the apo structure, though P2Rank led on COACH420's well-folded pockets (0.63 versus 0.58). Against 40 patent-disclosed STAT6 inhibitors and property-matched decoys, Ptarmigan-1 reached an AUROC of 0.94 while site-informed docking and blind Boltz-2 both scored 0.58. Scoring took 10 ms per ligand against 54 s for Boltz-2 on one H100, and screening all 20,431 human proteins against the 3.4-billion-compound OnePot CORE library took 20 H100 GPU-hours.

#Applications

The model targets hit finding where structure-based pipelines struggle: covalent cysteines away from canonical sites, allosteric grooves, and disordered proteins that offer a co-folding model no pocket. Because the embedding space supports the inverse query, a chemist can also rank the whole proteome against one compound to estimate selectivity and off-target liability. Its speed suits upstream triage, narrowing billions of compounds to a sub-library a pose-resolving engine can refine.

#Impact

Ptarmigan-1 is evidence that molecular recognition can be predicted, localized, and screened at proteome scale without building a structure, and that training on assay data rather than crystallography reaches sites the Protein Data Bank underrepresents. The caveats are substantial. All results are retrospective and in silico, with prospective experimental validation being pursued. Code and weights have not been released — only a public web server is planned — so the numbers cannot yet be independently reproduced. The internal chemoproteomic corpus and its disordered-target benchmark will not be released, though public-data ablations show localization survives their removal. The non-orthosteric class holds only two targets, and on one — the KRAS switch-II pocket — Boltz-2 was stronger. It reports no binding pose, only the residues a compound is predicted to engage.