Recursive Attention-Based Prediction of Theoretical MS/MS Spectra From Peptide Sequences

In tandem mass spectrometry (MS/MS)-based proteomics, precise prediction of peptide spectra is key to improving peptide identification accuracy and the overall reliability of proteomic studies. However, existing theoretical MS/MS prediction methods are limited by their focus on predicting only b and y backbone fragment ions and their inability to effectively capture the complex long-range dependencies among distant residues within peptide sequences.

To address these issues, we propose DeepMultiIon, a novel prediction framework based on a mathematically grounded recursive attention mechanism, designed to enhance MS/MS spectral fitting and improve peptide detection sensitivity. By combining local chunking with recursive attention, the model enables deep modeling of long-range residue interactions and the corresponding fragment intensity dependencies. In addition to conventional b and y backbone fragment ions, DeepMultiIon further extends its predictive coverage to a ions and precursor ions, while also encompassing their neutral loss ions (-H₂O, -NH₃) and isotopic peaks, offering comprehensive coverage of all major spectral peak types.

Performance is evaluated using Pearson correlation coefficient (PCC) and entropy similarity (ES), with comparisons against b/y-only models such as Prosit, pDeep, and AlphaPeptDeep. Experimental results show that DeepMultiIon achieves notable improvements, with average increases of 0.18 in PCC and 0.22 in ES.

These findings demonstrate that incorporating recursive attention and multi-ion prediction significantly improve both the peak coverage and accuracy of theoretical MS/MS spectrum prediction, thereby enhancing peptide-spectrum fitting and contributing to improved peptide detection sensitivity, making DeepMultiIon a powerful tool for advanced proteomics analysis.