π-SPECFormer learns from 291 million mass spectra and flags six mutation peptides missing from databases in lung adenocarcinoma

Much of the mass-spectrometry proteomics signal cannot be explained by conventional database searching, and π-SPECFormer shifts learning from annotated peptide labels to the spectra themselves, so unassigned spectra also yield transferable representations. Earlier deep-learning models were mostly trained with supervision on peptide–spectrum matches produced by database searches, so their labels were confined to the sequence and modification space those searches covered and could not exploit the much larger pool of unannotated spectra. Re-analyzing the ten largest datasets in π-MSNet, conventional methods gave identification rates consistently between about 20% and 31%, showing that unassigned spectra are a general phenomenon rather than an artifact of one dataset.

The model first learns contrastively on unlabeled spectra and is then refined with match information, and one pre-trained encoder transfers to de novo sequencing, spectrum clustering and modification detection. In the first stage each spectrum is randomly peak-masked into two views, encoded by an online encoder and a momentum-updated encoder, and optimized with an InfoNCE objective against other spectra held in a dynamic queue; the second stage treats spectra assigned to the same precursor as positive pairs and differently assigned spectra as negatives, using peptide identities only to build pairs and never as model input. On the DeepNovo nine-species benchmark, Stage 2 reached the highest peptide precision in all nine datasets, with its largest relative gain over the strongest non-π-SPECFormer baseline at 12.1%, from 33.0% to 37.0% in C. endoloripes; Stage 1 exceeded all four reference methods in eight of the nine datasets and matched the best baseline in the remaining one.

Applied to a clinical cohort, the pre-trained model reads single-amino-acid mutation peptides out of spectra that routine searching misses and turns them into experimentally testable candidates. On paired tumor and adjacent non-tumor tissue from 103 individuals with lung adenocarcinoma, a de novo sequencing workflow combined with reference-proteome alignment, spectral quality control and quantification produced 5,103 candidate single-amino-acid mutations across 1,712 proteins. Cross-referencing dbSNP, GDC, COSMIC and PubMed left 3,946 of the 5,103 candidates (77.3%) absent from the resources examined; after synthetic-peptide spectral similarity checks, nine mutations were called high-confidence, six unreported in the databases examined and three independently corroborated by external resources.

The demonstrated scope is pre-training on 291 million tandem mass spectra and evaluation on the nine-species de novo sequencing benchmark, the HeLa, cerebellum and nine-species clustering datasets, synthetic phosphopeptide and LUAD phosphoproteomic detection data, and the 103-individual CNHPP-LUAD cohort; the mutation–drug-target interface resource is explicitly framed as a hypothesis-generation tool rather than a catalogue of experimentally confirmed interactions.

Training-data diversity is limited, and extending to more species and to data-independent acquisition data is named as future work; validation of the many novel mutations and unannotated peptides currently rests mainly on computational false-discovery-rate control and structural simulation, and the authors call for co-immunoprecipitation and in vivo cellular models to confirm binding and function. The interface analysis also states plainly that AF3 confidence scores, FoldX energy estimates and contact changes do not show that the predicted complexes form in cells, that the substitutions alter binding affinity, or that the mutation-centered regions are pharmacologically tractable.