In this website, we adopt the term protein word to refer to semantic units within protein sequences, derived systematically and automatically from the attention matrices of a protein language model (PLM) using a community detection algorithm. As an alternative to the traditional concept of “motifs” for functional prediction, protein words are defined as sets of 5-20 amino acid residues. Importantly, a protein word may be either contiguous or discontiguous within the sequence, and can contain up to two gaps—each of which may span any number of residues and is denoted by an underline in visual representations.
To leverage these semantic units for functional inference, we developed Word2Function, a supervised model based on ESM-2. This model predicts protein function at the full-length sequence level and enables the mapping of individual protein words to specific Gene Ontology (GO) terms through feature attribution analysis. The curated GO term subset used in this site reflects functional annotations supported by this mapping process.
For more details, please refer to our paper.
The Search interface enables flexible querying of the protein word-GO term database. Users can perform individual or combined searches by entering a protein word and/or GO term. Two matching modes are available: exact match and contain match, allowing precise or partial string matching based on the query.
Results are presented in a structured table, with each row representing a protein word-GO term pair. The interface supports sorting and pagination for easier browsing. Additionally, selected results can be exported as a CSV file for downstream analysis.
The Match interface is designed to identify known “protein words” within user-provided protein sequences. During matching, an underscore in a protein word represents only a single residue.
To use the Match function, paste a full-length protein sequence in FASTA format into the input field. The system will scan the sequence and highlight all matching protein words, along with their associated GO terms and sequence positions. For each match, a visual alignment is provided to show the local sequence context and matched region.
All match results are displayed in a table with checkboxes for selection. Users can export selected matches to a CSV file for further analysis.