yHydra

Bottom-up mass spectrometry identifies proteins by matching fragmentation spectra against a database of theoretical peptides. That match becomes expensive the moment the search space grows: allowing arbitrary post-translational modifications, searching a metaproteomic or proteogenomic database, admitting unspecific cleavage, or looking for proteoforms that no reference proteome contains. Conventional search engines answer this with fragment indexing and combinatorial enumeration, which scales poorly and forces the analyst to declare in advance which modifications to consider.

yHydra reframes the problem as a retrieval task in a learned space. Two transformers — a Spectrum Transformer and a Peptide Transformer — are trained jointly so that a spectrum and its matching peptide land close together in a shared 64-dimensional Euclidean embedding, using a pairwise contrastive loss modeled on CLIP. Once trained, either encoder runs independently, and any search mode reduces to a nearest-neighbor query over vectors. The model was developed by Tom Altenburg and Bernhard Renard at the Hasso Plattner Institute with collaborators at Boston Children's Hospital, the Robert Koch Institute, and Mount Sinai, and was published in Bioinformatics after circulating as a preprint since late 2021.

It sits alongside Casanovo in the small set of transformer models for tandem mass spectra, but solves a different problem: Casanovo reads a sequence off a spectrum de novo, while yHydra provides a reusable embedding on top of which closed, open, and error-tolerant database searches are all implemented.

#Key Features

• Open search with delta masses: Every identification carries the mass difference between the precursor and the unmodified peptide, so oxidation (+16 Da), carbamidomethylation (+57 Da), and their combinations are characterized without being specified beforehand.
• Error-tolerant search by gradient descent: Because both encoders are differentiable, the gradient of a peptide embedding toward its spectrum yields a reward matrix over amino acid substitutions; a beam search over that matrix rewrites candidate peptides into proteoforms missing from the reference proteome.
• Wavelet peak encoding: Each of up to 500 peaks is encoded by a 64-dimensional sinusoidal function of its m/z and scaled by intensity, letting the model resolve peaks in close proximity without explicit peak matching.
• Multiplexed mass-bucket retrieval: Peptides are partitioned into mass buckets, each given an address vector appended to the embedding, so closed, narrow, and open searches all execute as a single GPU k-nearest-neighbor call.
• Interpretable embeddings: UMAP projections show the joint manifold organized by precursor charge and mass while modifications scatter uniformly — the property that makes open search viable.

#Technical Details

Both encoders emit 64-dimensional vectors. Spectra are truncated to their 500 most intense peaks; peptides are capped at 42 residues, with a trainable 26-by-64 amino acid lookup table multiplied into a positional encoding. Training used 19,991,263 peptide-spectrum matches drawn from 67 PRIDE repositories, deliberately excluding the ten most-studied organisms to limit proteome-specific bias, with the pairwise contrastive loss computed over 64-example mini-batches and label smoothing applied to the off-diagonal. Weights ship for Q Exactive and timsTOF instruments and for tryptic and non-tryptic peptides. Retrieval uses faiss, and peak matching and PSM scoring are GPU-accelerated.

Reported gains vary sharply by benchmark. On a cyanobacterial open search, yHydra identified 10,810 peptides at 1% FDR against MSFragger's 10,524, with better target-decoy separation but a narrow margin. The published abstract's headline of up to a 25% increase in peptide identifications traces instead to the harder proteoform settings: on a monoclonal antibody digest, yHydra found 50 peptides to MSFragger's 39 at 5% peptide FDR, rising to 55 and 43 respectively after gradient-descent post-processing. A chimpanzee plasma sample searched against the human proteome yielded 1,715 peptides versus 1,658.

#Applications

The model suits proteomics groups whose search space defeats a closed search: unrestricted PTM profiling, antibody sequencing where variable regions are absent from any reference, metaproteomics, proteogenomics, and cross-species work on organisms with only predicted proteomes. Because the error-tolerant mode consumes candidate PSMs with substantial delta masses, it can be applied as post-processing to results from an existing engine rather than replacing an established pipeline.

#Impact

yHydra shows that a single joint embedding can serve as substrate for several distinct proteomics tasks, and its gradient-based rewriting of peptide candidates is an unusual and concrete use of encoder differentiability. Inference code, search scripts, and the trained model archives are distributed under the MIT license, with training code in a separate repository; the canonical repository lives on GitLab, with a GitHub mirror. The evaluation remains narrow — a handful of datasets, with MSFragger as the sole search-engine baseline — and the largest advantages appear in proteoform-heavy cases rather than routine tryptic searches.