FullMS2Former: A Transformer Model for Near-Complete Prediction of Peptide Tandem Mass Spectra

Abstract

Accurate prediction of peptide tandem mass spectra is essential for confident peptide identification. Existing deep learning models achieve high similarity to experimental spectra but remain constrained by fixed fragment dictionaries of common fragment ions, limiting their ability to capture uncommon or previously unreported ions. Here, we present FullMS2Former, a transformer-based model that predicts a comprehensive fragment-ion feature vector encompassing canonical fragments and their integer mass offsets. By annotating canonical fragments as well as their integer mass offsets, the model can learn both common and uncommon fragment ions directly from peptide CID (collision-induced dissociation) spectra. Trained on large-scale NIST HCD (higher-energy collisional dissociation) data sets, FullMS2Former performs comparably to UniSpec on validation data and outperforms UniSpec on most testing data. Beyond overall similarity scores, FullMS2Former uniquely predicts rare sequence-dependent fragment ions that appear in experimental spectra but fall outside traditional annotation schemes.