Proteomics gets a billion-scale spectral corpus: MassNet holds 548 million PSMs and leads de novo sequencing across 15 species
Synopsis
Proteomics has plenty of public mass spectra, but mixed formats, missing metadata and a human-heavy species mix make them hard to train deep models on. MassNet gathers DDA data from 648 projects across 35 species, holds 548 million PSMs, and stores spectra in the tensor-based MSDT format with traceable metadata. On a benchmark spanning 15 species and 9.66 million PSMs, the XuanjiNovo model trained on this corpus reaches peptide recall of 0.45 on human and 0.68 on mouse data, and 0.25 on tomato, above the listed de novo sequencing models.
Interpretation
Proteomics now has a large standardized corpus that can feed deep-learning training directly, turning identified spectra scattered across public repositories into one uniformly loadable data resource. Earlier public spectral libraries used varied formats, mostly line-based text that is hard to load on demand as tensor batches, lacked metadata traceable to raw files and search parameters, and concentrated heavily on human samples. MassNet draws on 648 projects from PRIDE and iProX across 35 species; FragPipe and Sage searched independently with 1% FDR at the PSM, peptide precursor, peptide and protein levels, and the merged, deduplicated result holds 548 million target PSMs, 4.61 million peptide precursors, 2.54 million unique peptides and 133,422 proteins.
The companion MSDT format keeps only what AI modeling needs, stores spectra as tensor structures, and relies on separate metadata files for provenance. Seven data formats were compared on three types of mass spectrometers, and Parquet was chosen as the MSDT carrier on combined read speed and compression. MSDT can be converted to LMDB for random access, easing the I/O bottleneck in model training; the format and the corpus together form the input base for downstream tasks.
With peptide de novo sequencing as the validation task, XuanjiNovo trained on MassNet leads on a cross-species benchmark and keeps improving as training data grows. That benchmark covers 15 species, 307 raw mass spectrometry files, 30.78 million MS/MS spectra and 9.66 million PSMs, larger than the previously published nine-species benchmark. XuanjiNovo outperforms π-HelixNovo, Casanovo V2, InstaNovo+, ContraNovo and π-PrimeNovo across all species and spectral coverage levels; XuanjiNovo-100M, trained on 100 million PSMs, reaches about 0.82 amino-acid-level and about 0.68 peptide-level recall, above smaller-scale models and π-PrimeNovo.
Perspective
The demonstrated scope is DDA mass spectrometry data covering 35 species and more than ten high-resolution platforms, validated on peptide de novo sequencing with retention-time prediction and PSM rescoring as two downstream tasks; the corpus and MSDT format are public through PRIDE PXD071590 and a GitHub repository.
Native XuanjiNovo-100M shows limits on the nine-species benchmark, and only XuanjiNovo-130M, fine-tuned with 30 million PSMs from MassIVE-KB, improves broadly across species; the corpus remains dominated by human and mouse data, with plant samples clearly fewer than animal samples.