Foundation model enables interpretable open and error-tolerant searching for mass spectrometry-based proteomics
TL;DR
The foundation model yHydra jointly embeds mass spectrometry spectra and peptide sequences into a shared space, enabling open and error-tolerant searches that increase peptide identification by up to 25% compared to existing methods.
Problem / question
Identifying peptide sequences from tandem mass spectrometry data is computationally difficult because real spectra contain missing peaks, noise, and complex isotopic patterns, making it especially hard to detect proteoforms with unexpected mutations or post-translational modifications.
Methods
The researchers developed yHydra, a foundation model comprising a Spectrum Transformer and a Peptide Transformer, pre-trained contrastively on nearly 20 million peptide spectrum matches from 67 repositories. They used wavelet encoding for spectral peaks and implemented a GPU-accelerated k-nearest neighbor open search, alongside a gradient descent-based beam search for error-tolerant identification of modified peptides.
Key findings
In a cyanobacteria dataset, yHydra's open search identified 10,810 peptides at a 1% false discovery rate (FDR), compared to 10,524 by MSFragger. When applied to a monoclonal antibody digest, yHydra's error-tolerant search increased peptide identifications from 50 to 55 at a 5% FDR. In a cross-species search of chimpanzee plasma against a human reference proteome, yHydra identified 1,717 peptides at a 1% FDR, outperforming MSFragger's baseline of 1,658. Overall, the model showed up to a 25% increase in peptide identification at a constant FDR.
Why it matters
By mapping both spectra and peptides into a shared Euclidean space, yHydra allows researchers to rapidly and unrestrictedly discover novel protein variants, genetic mutations, and post-translational modifications without relying on exhaustive, predefined search databases.
Limitations
The authors note that in the k-nearest neighbor search, selecting a larger neighborhood size (k) can introduce low-scoring matches that reduce the final number of identified peptides, and the error-tolerant beam search requires balancing sensitivity against computational runtime.
Takeaway
yHydra uses deep learning to translate mass spectrometry data and peptide sequences into a shared mathematical space, making it significantly faster and more accurate to discover modified or unmapped proteins.