Tesorai Search

A modern database search engine does not really score a spectrum. It computes dozens to hundreds of hand-crafted features about one — predicted fragment intensities, predicted retention time, peptide length, charge — then trains a second classifier, usually Percolator, on the decoys of the very run being analysed to weigh those features into a score. That machine is refit for every dataset, and the same decoy sequences do double duty as training labels and as the yardstick for the false-discovery rate, which is how a search can report more identifications and a worse-than-advertised error rate at once.

Tesorai Search removes that second stage. A single pretrained network takes the complete centroided MS2 spectrum and a candidate peptide sequence and emits one calibrated score, and it is trained once — never on the dataset being searched, and never on decoys, which are left free to do only the job of estimating FDR through ordinary target-decoy competition. The model was built by Tesorai with collaborators at the Max Planck Institute of Biochemistry, and released as a preprint in August 2024 alongside a hosted search platform.

It sits alongside DDA-BERT, which rescores from a fixed checkpoint over raw MS2 peaks, and pUniFind, which unifies scoring with open de novo sequencing. What distinguishes it is less the encoder than the training objective: it learns to separate correct matches from high-scoring real alternatives, so nothing about decoy generation leaks into what it has learned.

#Key Features

• Spectrum-and-sequence input only: The network sees the full peak list, including regions outside the theoretical fragment windows, not a summary of matched ions.
• Decoy-free training objective: Negative examples are second- and third-ranked real peptides for the same scan, so decoys are reserved entirely for FDR estimation.
• No per-run refitting: One checkpoint scores every dataset, removing the tuning surface that on-the-fly classifiers introduce.
• Implicit preprocessing: Deisotoping, mass recalibration and charge correction are learned rather than configured, which is what lets one model handle unfamiliar instruments.
• Cloud-native execution: Candidate generation, GPU rescoring and quantification run as separate elastic jobs behind a browser interface.

#Technical Details

Three modules make up the model. A spectrum encoder bins and normalises the m/z and intensity vectors of a centroided scan, discarding spectra with fewer than 10 peaks and keeping the 1,100 most intense peaks otherwise. A sequence encoder one-hot encodes the peptide over 32 tokens — 20 amino acids plus 12 common modifications, including phosphorylation and TMT labels — computes theoretical a, b, y and z fragments at charges 1 and 2 with optional water and ammonia losses, and passes them through a transformer. A joint encoder fuses the two representations with a transformer and average pooling, ending in linear layers that output a single logit. Training used 289 million PSMs from roughly 16,000 Orbitrap LC-MS/MS runs across 22 public PRIDE studies, labelled by MaxQuant, Comet and MS-GF+; two epochs took 64 hours on eight V100 GPUs.

At 1% FDR across seven benchmark datasets, it identified 12% more unique peptides than FragPipe with MSBooster, 9% more than PEAKS, 21% more than Proteome Discoverer with Chimerys/Inferys and 68% more than MaxQuant, with 70% of its peptides corroborated by at least one other engine against 37–65% for the tools compared. Entrapment analysis on ISB18 against a castor-bean proteome at a 668:1 ratio kept the measured false-discovery proportion below the requested threshold from 0.1% to 10%. Although trained only on label-free Orbitrap data, it gained 13% more PSMs on TMT10-MS3, 10–30% on Bruker timsTOF Pro and 9% on a Sciex TripleTOF 6600+, 51% on a phospho-enriched sample and 43–50% on single-cell DISCO runs. Ablations put validation accuracy at 99.5%, falling to 77.6% when intensities are randomised.

#Applications

The gains are largest where identification is hardest: immunopeptidomics, where a deeper MHC-presented repertoire feeds neoantigen and vaccine work, and single-cell proteomics, where spectra are sparse. Searches run at console.tesorai.com, free for academic users — raw data and a FASTA go in, Comet, MaxQuant and MSFragger generate candidate PSMs at permissive settings, the model rescores them, and peptide and protein tables come back quantified by FFIA or IonQuant. Reprocessing 250 immunopeptidomics samples took under 45 minutes.

#Impact

The paper's broader argument is that much of the machinery bolted onto search engines over the past decade compensates for scoring functions that never used the spectrum fully, and its nine-engine benchmark is a reusable reference for testing that claim. The practical constraint is access: neither weights nor inference code are released, and the public repository holds only MIT-licensed notebooks that regenerate the manuscript's figures from outputs deposited on Mendeley Data, so results can be reproduced but the model cannot be run offline or retrained. The work covers data-dependent acquisition only, leaving DIA to future work, and the authors note that accurate PTM site localisation remains unsolved.