Foundation model enables interpretable open and error-tolerant searching for mass spectrometry-based proteomics

Area of Science:

• Proteomics
• Computational Biology
• Bioinformatics

Background:

• Mass spectrometry-based proteomics is crucial for molecular-level protein analysis.
• Noisy mass spectra and complex protein landscapes (PTMs, isoforms, large databases) challenge algorithmic identification.
• Existing methods struggle with comprehensive peptide and protein identification.

Purpose of the Study:

• To introduce yHydra, a foundation model for joint spectral and peptide embedding in proteomics.
• To enable advanced search modes like open search and error-tolerant identification of novel proteoforms.
• To improve the accuracy and scope of peptide and protein identification from mass spectrometry data.

Main Methods:

• Developed a foundation model (yHydra) for joint embedding of mass spectra and peptide sequences.
• Implemented an open search mode for querying large spectral and peptide datasets.
• Implemented an error-tolerant search for identifying uncharacterized proteoforms.

Main Results:

• yHydra provides meaningful embeddings, enabling interpretation of peptide properties.
• The open search facilitates unrestricted characterization of post-translational modifications.
• Evaluations show up to a 25% increase in peptide identification at a constant false discovery rate compared to state-of-the-art methods.

Conclusions:

• yHydra significantly advances peptide and proteoform identification in mass spectrometry-based proteomics.
• The model's open and error-tolerant search capabilities address limitations in current proteomic analyses.
• yHydra offers a powerful tool for both standalone use and post-processing of existing search engine results.