The abstract is read at the publisher; the summary is JClub's.
Tryptic-peptide tokenization underperforms 3-mers for protein function prediction on 23,723 UniProt proteins because most peptide tokens are unseen in training (92.2%).
- Why it matters: Tokenization choices affect sequence-based protein function prediction, but it is unclear whether biologically grounded fragments improve transfer to unseen proteins.
- What they did: The study compared tryptic peptides, peptide topics, 3-mers, and amino-acid composition using ESM-2 embeddings and homology transfer, evaluating six Gene Ontology terms.
- The result: With tryptic peptides, 38.9% of test proteins had no seen peptide and the classifier output a single label, yielding lower overall macro-F1 (0.640 vs 0.690 for 3-mers).
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.