A mass spectrometer does not hand a scientist a protein name. It breaks molecules into charged fragments and records a spectrum: peaks whose positions and intensities carry clues about the peptide that produced them. MassNet turns 1.54 billion of those tandem spectra into a corpus intended for large-scale machine learning.[1]
The data came from 27,643 liquid-chromatography mass-spectrometry runs in 648 public projects, spanning 35 animal, plant and microbial species and about 30 terabytes of raw files. Approximately 1.7 million peptide precursors in the resulting resource come from human samples.[1]
MassNet at a glance
| Measure | Value | Context |
|---|---|---|
| Tandem spectra | 1.54 billion | Raw spectral observations collected from public DDA proteomics experiments. |
| Projects | 648 | Studies drawn from the PRIDE and iProX repositories. |
| Instrument runs | 27,643 LC–MS/MS runs | Individual acquisition runs across the contributing projects. |
| Species | 35 | Fourteen animal, five plant and sixteen microbial species. |
A spectrum becomes training data only after several decisions
Raw spectra are abundant but not automatically labelled. Database-search engines compare observed fragment patterns with candidate peptide sequences and produce peptide-spectrum matches. MassNet used two search pipelines, combined their results and removed overlaps. The run-level corpus contained 548 million target matches before further global filtering and deduplication.[1]
From somebody else’s experiment to an AI corpus
| Stage | What is preserved or added | What can go wrong |
|---|---|---|
| Original experiment | Sample, species, instrument and acquisition settings | Public repositories overrepresent well-funded organisms, instruments and protocols. |
| Database search | A proposed peptide label for each spectrum | Reference databases and scoring thresholds can produce misses or false identifications. |
| Global curation | Deduplication, filtering and common metadata | Standardisation can hide useful variation if provenance is discarded. |
| Model training | Repeated exposure to spectra and peptide labels | Train–test leakage can occur when near-duplicates cross evaluation boundaries. |
The format is part of the contribution
The researchers created Mass Spectrometry DDA Tensor, or MSDT, as a compact standard representation for storage, curation and reuse. It can also be converted into a Lightning Memory-Mapped Database so training systems can read data more quickly. This is less glamorous than a new model, but foundation-scale science often fails first at input consistency and I/O.[1]
A common tensor format lets researchers ask broader questions than one laboratory’s pipeline normally supports. Models can learn fragment patterns across species, instruments and sample types; they can be pretrained once and adapted to peptide sequencing, spectrum prediction, quality control or retrieval. Provenance remains essential because the same peak pattern can have a different interpretation under different acquisition conditions.
The proof of concept was de novo sequencing
The paper introduces XuanjiNovo, a de novo peptide-sequencing model trained with the resource, and reports substantially better performance than previous approaches. De novo sequencing attempts to infer peptide sequences without restricting the answer to a known protein database, which makes broad spectral pretraining especially relevant. The result is an early demonstration of what corpus scale can enable, not proof that one model has solved proteomics.[1]
MassNet’s importance is therefore both scientific and archival. Proteomics accumulated an enormous public record in fragments, each project designed for a different question. The new corpus makes that record computationally legible. Its value will depend on whether future models learn transferable chemistry rather than the recurring fingerprints of the repositories that assembled it.
Sources & further reading
- MassNet: billion-scale AI-ready mass spectral corpus empowering scalable deep learning in proteomics Nature Methods · 14 Oct 2026 · Accessed 16 Oct 2026