Loading…
Sequence-only interaction model scoring protein-protein and protein-ligand pairs plus functional annotation across whole proteomes in seconds.
No providers recorded yet. Browse all providers
A proteome of 20,000 proteins holds roughly two hundred million possible pairs, and a co-folding model that spends seconds on each turns an all-versus-all screen into a compute budget rather than an experiment. Atlas, built by Synthyra, takes the pair out of the expensive step: each sequence is encoded once into a dense vector by its own encoder, and a pair score is a comparison between two vectors computed without reference to each other. A screen of N proteins against M therefore costs N+M encoder passes and then arithmetic: a 20,000 by 20,000 matrix returns in under a second, at roughly 375 million pair scores per second after encoding.
The same input feeds three lanes. Atlas-PPI scores whether two proteins interact, on a single pair, a neighborhood around a query, or an entire proteome. Atlas-PLI scores whether a protein and a small molecule associate, with compounds supplied as SELFIES strings so the ligand side runs through the same embed-then-compare machinery. Atlas-CAMP places a sequence in structured functional annotation space, so an interaction or ligand hypothesis arrives with functional context attached.
Synthyra first described Atlas publicly in October 2025 and launched it as a product in August 2026. Atlas-PPI is the productized successor to the company's Synteract line, retrained on a larger encoder, and Atlas-CAMP grows out of its Annotation Vocabulary work. Atlas reaches users as a hosted REST API and web explorer rather than downloadable weights, which distinguishes it from DSM, Synthyra's masked-diffusion protein language model.
On the Bernett gold-standard split, where no sequence above 40 percent identity crosses between train and test, Atlas reports a Matthews correlation coefficient of 0.34 against 0.30 for the next best published method, while running about five orders of magnitude faster than the quickest comparator publishing a rate. The three lanes are not equally hard: Atlas-CAMP reaches a test MCC of 0.963, reflecting how strongly function tracks sequence family, while Atlas-PLI is the weakest at 0.503 on test against 0.682 on validation, a gap the authors describe as a generalization warning.
Cross-species interaction is the harder surface. A 2026 Nature Machine Intelligence audit scored eight published methods on human-SARS-CoV-2 labels between 0.444 and 0.515 ROC AUC, indistinguishable from chance. On a homology-controlled version, an Atlas checkpoint trained on human data alone reaches 0.645 and the production checkpoint 0.790. On the human intra-actome, that checkpoint exceeds ProteomeLM's published ROC AUC by 0.123 absolute on the strict zero-shot subset. Full-surface precision-recall remains low, because the real class balance is severe.
Atlas is aimed at the stage before an experiment is designed: narrowing a proteome, a compound library, or a host-pathogen interface to the corner worth testing first. The host-pathogen report works three cases. A checkpoint with all SARS-CoV-2 sequences held out names one uncharacterized viral protein predicted to engage 409 of 433 human olfactory receptors. A 33-second screen of three Andes hantavirus proteins against 2,638 FDA-approved small molecules and 626 biologics nominates mupirocin and dornase alfa, two long-approved drugs, above every antiviral with documented hantavirus activity. A blind recall test on Bundibugyo ebolavirus places all four approved monoclonal antibodies on the correct antigen out of nine viral proteins, with a 3.7-fold margin and no score overlap.
Atlas is a commercial platform rather than a published model: there is no preprint, no code release and no downloadable checkpoint, so independent replication is limited to what the API returns. It demonstrates that a sequence-only interaction model with honest leakage controls can clear the random floor on a surface where every prior method sat at chance, fast enough that the screen stops being the rate-limiting step. Synthyra reports the failures alongside the successes: the ebolavirus run misses the VP24-karyopherin interaction entirely despite a published co-crystal structure, ranks the canonical entry receptor NPC1 at 361 of 19,933 human proteins but below the operating threshold, and is not corroborated by an independent folding model, which inverts the antibody result. Nothing here establishes binding, efficacy or safety; the nominated candidates are interesting because they are cheap to test.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.