Loading…
Peptide tandem mass spectrum prediction from sequence, giving fragment-ion intensities and indexed retention time for any organism or protease.
No providers recorded yet. Browse all providers
A database search engine decides whether a tandem mass spectrum belongs to a candidate peptide largely by counting which theoretical b- and y-ions it can match. It cannot ask whether those ions appear in the right proportions, because fragment intensity depends on the peptide's sequence in ways that rule-based models never captured. A decoy of similar composition therefore matches nearly as many peaks as the correct peptide, and false discovery rate control degrades as the search space grows.
Prosit removes that blind spot by predicting the spectrum in advance. Built at the Chair of Proteomics and Bioanalytics of the Technical University of Munich with collaborators at SAP, JPT Peptide Technologies and Thermo Fisher Scientific, and published in Nature Methods in 2019, it maps a peptide sequence, its precursor charge and the instrument's normalized collision energy to a complete fragment-ion intensity pattern, and the sequence alone to an indexed retention time (iRT). Both come from a fixed trained checkpoint; nothing about a new sample is measured first.
The training resource is ground truth rather than inference: the team extended the ProteomeTools synthetic peptide library to 550,000 tryptic peptides and 21 million high-quality tandem mass spectra, adding 219,089 peptides to broaden coverage of human proteins and isoforms. Because every peptide was chemically synthesized, its sequence is known independently of any search engine. Prosit is the base model of a family the catalog also carries: Prosit-PTM, Prosit-XL and Prosit-Cit extend the same design to modified, cross-linked and citrullinated peptides.
Prosit is a recurrent encoder–decoder. The peptide sequence passes through an embedding layer into two bidirectional GRU networks and an attention layer, while precursor charge and collision energy each pass through a dense layer; the three encoded representations are multiplied element-wise into a fixed-size latent space. For fragment intensity, a bidirectional GRU decoder emits six values — b and y ions at charges 1 to 3 — for each of up to 29 fragmentation positions, covering peptides of at most 30 residues. The iRT model reuses the same encoder with dense layers as its decoder. Training minimizes a masked spectral-angle loss, so physically impossible fragments contribute nothing.
On 217,377 peptides held out of training, predicted and measured iRT agree at Pearson R = 1.00. Spectral angles between predicted and measured spectra peak near 0.9 (R ≈ 0.99) on held-out data, comparable to the agreement between replicate measurements. Supplied as Percolator input, Prosit's scores yielded more true-positive peptide-spectrum matches at every false discovery rate tested, matching the identification yield of a conventional Andromeda search at a threshold more than ten times stricter.
Prosit enters bottom-up proteomics at two points. Groups running data-dependent acquisition rescore MaxQuant or other search output with predicted intensities and recover peptides that ion-counting scores discard — valuable in immunopeptidomics, where non-tryptic peptides and small sample amounts make identification fragile, and in metaproteomics, where the database spans thousands of organisms. Groups running data-independent acquisition can build spectral libraries computationally for organisms that have no measured library at all. A hosted interface in ProteomicsDB takes a peptide list and returns a library or rescoring input, requiring neither a GPU nor model expertise.
Prosit made predicted fragment intensity a routine component of peptide identification rather than a research curiosity, and the pattern it established — train once on synthetic ground truth, then apply the fixed checkpoint to whatever proteome, protease or instrument a lab has — carried into a family of successors covering post-translational modifications, cross-links and citrullination. The trained weights and their training data are archived on figshare under Apache 2.0 alongside source code on GitHub, and current checkpoints are served through the Koina inference API; the authors have deprecated the original 2019 repository in favour of that tooling. The released model handles peptides of at most 30 residues and only two modifications, carbamidomethylated cysteine and oxidized methionine, which is precisely the gap the later family members were built to close.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.