Translator

An annotation that reads "probable hydrolase, similar to a characterized esterase" is a sentence, and a sentence has to be read before it can be used. Two curators describing the same activity phrase it two ways, neither phrasing joins cleanly against a database, and a model trained to produce such text is graded on wording as much as on biology. Translator never writes that sentence. Given an amino acid sequence it emits identifiers — an EC number, a GO accession, an InterPro family, a Gene3D fold — each already meaningful in an existing database and checkable there directly.

Translator is a closed commercial model built by Synthyra, a public-benefit protein-science company, and described publicly in March 2025. Its output space comes from the Annotation Vocabulary, the framework the company's founder and collaborators at the University of Delaware introduced in 2024, which maps structured biological ontologies onto transformer token embeddings so protein properties can be modeled as a language. CAMP, the representation model from that work, uses the vocabulary to reshape an embedding space and assigns EC numbers by vector search. Translator instead predicts vocabulary tokens directly from sequence across eight aspects: enzyme commission number, the three Gene Ontology branches, InterPro domain, Gene3D fold, UniProt keywords, and cofactor.

The intended position is a first pass. Similarity search is the standard opening move and is genuinely strong, but it is weakest exactly where interest is highest — a sequence distant from anything characterized, or one carrying partial functional hints no single homolog explains. Metagenomic surveys, new genomes and designed proteins produce such candidates faster than curators can reach them.

#Key Features

• Identifiers rather than prose: Output is checkable directly against UniProt, InterPro and Gene3D, and identical for two proteins sharing a function however anyone would word it.
• Eight aspects from one pass: One query returns catalytic, ontological, domain, structural, keyword and cofactor hypotheses together, rather than one predictor per aspect.
• Annotation as retrieval: Predictions rank logits over the full vocabulary and take the top k per decode position, duplicates collapsed. The deployed default k = 3 maximizes F1, with a minimum-confidence threshold trimming weak candidates.
• Tuned deliberately for recall: Test-set recall runs 0.73 to 0.86 across all eight aspects while precision spans 0.16 to 0.82, a spread Synthyra describes as making it a triage tool rather than an oracle.
• Served, not distributed: It runs on serverless GPU infrastructure behind a REST endpoint batching up to 64 sequences of at most 2,048 residues per call.

#Technical Details

Translator is a transformer mapping protein primary sequence to Annotation Vocabulary terms. The original vocabulary assigned unique integers to EC, the three GO branches, InterPro and Gene3D members, producing 30,000 to 60,000 tokens depending on the source dataset; the version Translator decodes into adds UniProt keywords and cofactors. Parameter count and training corpus are unpublished.

Two evaluations are reported, both in the supplementary material of Synthyra's Diffusion Sequence Models manuscript rather than a paper of Translator's own. On a non-redundant test set of 1,000 high-quality UniProt entries the model never saw, per-aspect F1 runs from 0.81 for Gene3D fold and 0.77 for InterPro domain down to 0.33 for EC number and 0.27 for biological process. A prospective evaluation on 656 UniProt entries deposited after training, all with experimentally verified annotations, raises fold to 0.93, domain to 0.88 and EC to 0.55, while biological process falls to 0.23. EC recall of 0.85 at precision 0.21 means the shortlist very probably contains the right number alongside roughly four wrong ones.

#Applications

Translator suits the stage between having a sequence and deciding whether to spend real effort on it: annotating metagenomic assemblies and non-model genomes where homology search returns little, attaching functional context to de novo designs, and labeling sequence sets too large to curate. It complements similarity search rather than replacing it: a close homolog remains authoritative when one exists. Synthyra uses it internally as the annotation step for DSM, and serves it as its own endpoint alongside the annotation lane of the company's Atlas platform.

#Impact

Translator is a product rather than a publication: no preprint of its own, no code and no downloadable checkpoint, and the DSM manuscript calls it closed-source outright, so independent verification is limited to what the hosted API returns. Its interest is the design choice rather than a leaderboard position — treating annotation as retrieval over a fixed ontology vocabulary makes the output machine-checkable and the recall-precision tradeoff an exposed knob rather than an implicit one. Synthyra is candid about the cost: unfiltered output fed into a downstream training set inherits the model's precision, which for biological process means most added labels are wrong, and the stronger prospective numbers reflect a set skewed toward proteins characterized because they were interesting, not evidence of performance on novel metagenomic sequence.