Protein function annotation model that parses sequences into residue clusters via community detection on ESM-2 attention, then maps them to GO terms.
No providers recorded yet. Browse all providers
Sequence-to-function tools mostly work at the domain or motif level, and both units come from human curation: PROSITE encodes hand-written residue patterns, Pfam and InterPro encode families assembled from alignments. Neither exists for a protein nobody has characterised, which is precisely where an annotation tool is needed. Word2Function, from Boxue Tian's group at Tsinghua University, replaces the curated unit with one derived automatically from a protein language model's own internal structure.
The pipeline has two halves. An unsupervised tool, Protein Wordwise, feeds a sequence through ESM-2 and treats each of its attention matrices as a weighted graph over residue pairs; Louvain community detection over those graphs yields tightly attending residue clusters, which the authors call "protein words." A word is 5 to 20 residues and may span up to two gaps — the shape a catalytic triad or binding pocket actually takes in sequence space, and one a contiguous motif regex cannot express. Word2Function is the supervised half: it embeds each word as the mean of its residue embeddings, predicts Gene Ontology terms for the whole protein from those embeddings, and uses Integrated Gradients to attribute the prediction back to individual words — annotation at residue-set resolution from sequence alone, with no structure and no alignment.
Two notes on naming. "Protein words" is a crowded phrase: the catalog's Protein Words Language Model is an unrelated VQ-VAE tokenizer and generator from a different Tsinghua lab, developed independently and answering a different question. And the toolkit ships under two names — Protein Wordwise extracts the words, Word2Function predicts function from them. It was published in Advanced Science after a January 2025 preprint.
ESM-2 contributes 33 layers of 20 attention heads, giving 660 attention matrices per sequence. The general dictionary was built from one million UniRef50 sequences — 50 sampled from each of 20,000 Pfam families, to avoid over-weighting large ones — producing 56,423,445 raw words, reduced to 2,581,748 entries under the 12-residue-type degeneracy scheme; family-specific dictionaries exist for 20,762 Pfams. On a deep mutational scanning benchmark of 198 sequences, in which 37,033 of 56,606 residues are experimentally functional, Protein Wordwise achieves median and mean functional residue coverage of 0.900 and 0.843 across the 182 proteins of usable length, against 0.045 and 0.393 for PROSITE, which produces no output at all for 35% of them.
The authors also assemble PWNet, an evaluation resource spanning 10 functional-residue tasks — binding residues for seven classes of partner, plus catalysis and ion-channel activity, from 95 to 8,238 sequences each. Protein Wordwise reaches a median coverage of 1.000 on five of the ten tasks, with mean coverage between 0.709 and 0.934, where PROSITE ranges from 0.000 to 0.406. Word2Function is trained on ExpGO65, 56,299 sequences annotated with 65 experimentally supported GO terms, using a transformer layer over word embeddings and a linear head. It reaches a mean functional Matthews correlation of 0.424 there, with 28 terms above 0.500, against 0.147 for PROSITE, and 0.284 on PWNet as an external test set against 0.150.
The immediate use is hypothesis generation for uncharacterised proteins: given a sequence with no informative domain hit, the toolkit proposes which residues form functional units and which GO terms they support, pointing mutagenesis at specific positions. Case studies recover known catalytic triads, DNA-contacting residues, and ATP-binding Walker motifs directly from sequence. A second use is immunological: the UniRef50 dictionary contains words matching 2,832 known MHC class I and 2,622 class II peptides, against 169 and 0 for a random dictionary, so presented peptides surface without the supervised training that TransHLA requires.
The contribution is a definition — a functional unit a language model finds on its own, rather than one a curator wrote down — and a demonstration that it outperforms PROSITE patterns on every dataset examined. Absolute performance keeps the claim in proportion: mean functional MCC of 0.284 on an external test set is a useful signal, not a solved problem, and word-level precision stays low, at 0.202 mean word accuracy on PWNet. Access is partial: the repository declares an Apache 2.0 license and ships extraction code with pretrained downstream weights, but the full Pfam and UniRef50 dictionaries — what makes inference on a new sequence work well — are available on request, with an open download site under construction.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.