Mass-spec sequencing stops relying on spectra alone: PepPr adds a proteome prior and lifts median peptide recall by 15.8 points across nine species

De novo peptide sequencing can now draw on reference-proteome knowledge during decoding instead of committing to a closed candidate list. Database search identifies only peptides listed in advance, while spectrum-only de novo sequencing is limited by incomplete fragmentation and trails database search in recall. On the nine-species dataset, each species used a prior trained on its own reference proteome and combined with Casanovo; median peptide recall rose by 15.8 percentage points, all nine species improved, and the largest gain was 33 percentage points on T. endoloripes.

The prior enters through a small fusion head that supplies sequence context when spectral evidence is ambiguous and yields to the spectrum when the sequence diverges from the reference. Earlier hybrid methods either corrected final predictions after the fact or retrieved general-purpose peptide-spectrum match libraries, so reference information could not shape which sequence hypotheses the network explored. The prior is a transformer language model trained on reference peptides, single-amino-acid variants, and reference prefixes joined to random suffixes; the fusion head was trained on 200,000 reference-peptide PSMs from MassIVE-KB, with top-1 and top-2 prior scores swapped during training to simulate an incorrect prior.

The prior's benefit tracks how closely the sample relates to the reference, with the same direction across reference peptides, variants and antibodies. Homology-driven database search decays steeply as sequence divergence grows, while prior-free de novo sequencing substantially outperformed that search for every species pair considered. With a human prior on 14 non-human animal species and an E. coli prior on ten bacterial species, PepPr stayed within 1.8 percentage points of prior-free Casanovo on the most distant species and improved recall by up to 10.5 percentage points on closer ones, with gains rising as exact tryptic overlap increased.

Evaluation centers on relatively high-quality spectra from data-dependent acquisition on high-resolution Orbitrap instruments, covering reference peptides, single-amino-acid variants, homology-driven proteomics and antibody sequencing; settings with more ambiguous spectra, such as data-independent acquisition or single-cell proteomics, remain untested.

The prior explicitly models only reference peptides, single-residue variants and divergent sequences; splicing, insertions and deletions are not included, and the proportions of the three training components were not tuned. The authors used a single all-rounder prior rather than application-specific ones, and note that de novo sequencing still lacks confidence calibration independent of database-search ground truth.