pUniFind

A tandem mass spectrum of a peptide carrying an unexpected modification looks nothing like the fragment ladder a search engine computes from the bare sequence. Open search engines answer this by widening the precursor mass window so any mass shift is allowed, which multiplies the candidate list and hands a much harder discrimination problem to a scoring function built from a handful of manually curated features. Deep learning has mostly been bolted onto the edges of that pipeline, predicting a spectrum or a retention time that is folded back in as one more feature for the same shallow re-scorer.

pUniFind replaces the scoring function itself. Built by the pFind group at the Institute of Computing Technology, Chinese Academy of Sciences with DP Technology, it encodes a peptide sequence and an MS/MS spectrum with dedicated transformer encoders, then pretrains the two against each other: the peptide representation must predict the spectrum, each spectral peak must predict how many amino acids its fragment carries and whether it is a b- or y-ion, and the spectrum representation must predict the peptide's length. A joint modality scorer reads both embeddings and emits the peptide-spectrum match (PSM) score, with an auxiliary listwise ranking task sharpening it against competing candidates for the same spectrum.

Because that alignment is learned over open-search annotations, non-tryptic digestion and rare modifications sit in the training signal instead of being filtered out of it. The same frozen checkpoint therefore serves two jobs: rescoring database-search candidates, and decoding a peptide directly from a spectrum. Unlike Casanovo and other de novo transformers, trained against a small fixed modification vocabulary, pUniFind does open de novo sequencing zero-shot. DIA-CLIP, from an overlapping author group, applies related cross-modal pretraining to data-independent acquisition.

#Key Features

• Cross-modality pretraining: Spectrum prediction, peak-level amino-acid counting, fragment-ion-type classification and peptide-length prediction train jointly, forcing the peptide and spectrum embeddings into correspondence rather than merely comparing them.
• Open de novo sequencing: A Peptide Length Aware decoder predicts the peptide's length first, then decodes sequence and modifications conditioned on it, covering more than 1,300 modification types without prior knowledge of which are present.
• Scoring decoupled from decoding: The joint modality scorer sits apart from the sequence decoder, a design the authors credit for low entrapment rates where competing end-to-end scorers showed inflated ones.
• Deep-learning quality control: End-to-end score, predicted spectrum similarity, retention time and missing fragment-ion sites combine into a filter that removes unreliable de novo results.
• Practical search workflow: Open-pFind retrieves the top 10 candidates per spectrum (20 for non-tryptic cases) and pUniFind reranks them, so it never scores an exhaustive candidate list.

#Technical Details

The published architecture pairs a 9-layer spectrum encoder with a 9-layer peptide encoder at a hidden width of 512 and 8 attention heads, feeding a 4-layer joint-encoding scorer. Training used 100 million PSMs drawn from 6,524 MS/MS data files and annotated by Open-pFind. That annotated corpus is released rather than merely described — roughly 125 million rows of parquet on Hugging Face, archived on Zenodo, with the raw runs traceable to 26 ProteomeXchange accessions — which is unusual for a proteomics model at this scale. Across nine species datasets, pUniFind identified 2–18% more peptides than Open-pFind at the peptide level, and 42.6% more in immunopeptidomics. Entrapment testing on V. mungo with an A. mellifera trap gave 49,364 peptides at a trap ratio of 0.408, against 43,452 at 0.454 for Open-pFind: more identifications at a lower error rate. On the 21PTM benchmark it reached 63.8% average peptide-level recall, a 60% improvement over pNovo with all 21 PTMs enabled, despite a search space 300 times larger. The quality-control module recovered 38.5% additional peptides, 1,891 of which map to the human genome but are absent from reference proteomes, and lifted agreement with RNA-Seq evidence from 65.4% to 85.0%.

#Applications

Bottom-up proteomics groups can drop pUniFind in behind an existing pFind search to deepen identifications without changing sample preparation, and use its de novo mode where no adequate reference proteome exists. The clearest gains are in immunopeptidomics and neoantigen discovery, where non-tryptic HLA-bound peptides defeat conventional engines' tryptic assumptions, and in proteogenomics, metaproteomics and PTM studies where the peptide of interest carries a modification nobody thought to specify. Separate checkpoints cover Thermo and timsTOF instruments, and a Windows GUI and hosted GPU web app serve labs without local hardware.

#Impact

pUniFind shows that the PSM scoring function, long the least examined part of a proteomics pipeline, benefits from pretraining at the scale the rest of computational biology takes for granted, and that one scorer can serve database search and de novo sequencing at once. Its limits are stated plainly: ion-trap and ETD/EThcD spectra are unsupported, de novo prediction is restricted to precursor mass errors within 20 ppm and peptide lengths of 6 to 40 residues, and Astral and timsTOF data benefit from a brief fine-tuning pass. The authors also withhold internal methods they report improve half of the failure cases, so the released system is not the full one they describe operating.