A mass spectrometer does not output a protein name. It breaks peptide ions into fragments and records peaks: mass-to-charge values paired with intensities. Identifying the peptide means learning which sequence of amino acids could have produced that pattern under the instrument’s conditions.[1]
MassNet turns a vast collection of those patterns into an AI-ready corpus. The authors assembled 1.54 billion tandem mass spectra from 27,643 liquid chromatography–mass spectrometry runs across 648 projects and 35 species. The source data occupied about 30 terabytes before the processing and standardisation pipeline.[1]
From raw instrument files to training records
How MassNet changes the data
| Stage | What is present | What the model gains |
|---|---|---|
| Raw acquisition | Instrument-specific spectra and run files | The original physical measurements, but in heterogeneous formatsSource 1 |
| Database search | FragPipe and Sage propose peptide-spectrum matches with confidence estimates | Labels and target–decoy information for supervised learningSource 1 |
| MSDT conversion | Spectra, precursor data, peptide annotations and metadata in structured Parquet records | A consistent tensor-like representation that modern frameworks can streamSource 1 |
| Optional LMDB | A memory-mapped form for rapid random access | Faster input and output during deep-learning trainingSource 1 |
After combining the two search engines and removing overlapping identifications, the resource contains 548 million target peptide-spectrum matches, 4.61 million peptide precursors, 2.54 million unique peptides and 133,422 proteins. The standardised corpus occupies 4.45 terabytes. Comparing the reported 30-terabyte raw collection with that corpus gives an approximate 6.7-fold size reduction, although the two stores do not contain identical information.[1]
MassNet at a glance
| Measure | Value | Context |
|---|---|---|
| Projects | 648 | Public PRIDE and iProX projects plus grouped in-house dataSource 1 |
| LC–MS/MS runs | 27,643 | Runs across animal, plant and microbial samplesSource 1 |
| Tandem spectra | 1.54 billion | The broad measurement corpus before limiting to identified matchesSource 1 |
| Target PSMs | 548 million | Combined high-confidence peptide-spectrum matches after overlap removalSource 1 |
| Unique peptides | 2.54 million | Distinct peptide sequences represented after processingSource 1 |
| Model-ready storage | 4.45 TB | Structured MSDT corpus based on Apache ParquetSource 1 |
The first model learns to read peptides directly from spectra
The team trained XuanjiNovo on 100 million spectra from MassNet. De novo peptide sequencing differs from database search because the model must infer the amino-acid sequence directly rather than select among entries in a reference database. The system uses a non-autoregressive transformer, curriculum learning and iterative refinement constrained by the precursor mass.[1]
On human Astral data, the authors report improvements ranging from 38.35% to 143.57% over the compared state-of-the-art methods, depending on the evaluation. That wide range should remain attached to the specific datasets and metrics. It is not one universal accuracy gain for all instruments, species or peptide modifications.[1]
A large corpus also scales its biases
MassNet retains spectra confidently identified by FragPipe or Sage. That improves label quality and tractability, but it excludes unidentified spectra that may contain rare, difficult or previously unknown peptides. Differences between search engines can also influence which records survive, embedding parts of their scoring behaviour in the training set.[1]
The current processing focuses on unmodified peptides and a small set of common modifications: methionine oxidation, N-terminal acetylation and cysteine carbamidomethylation. Many biologically important post-translational modifications—including phosphorylation, glycosylation, ubiquitination and methylation—are not systematically represented. Models trained only on the present corpus may therefore be strongest where the corpus is already easiest to identify.[1]
Why the format may matter as much as the model
MassNet can support more than de novo sequencing. The paper demonstrates retention-time prediction and peptide-spectrum-match rescoring, and its traceable metadata allow models to condition on instrument, fragmentation method, species and project. A shared representation also makes it easier to compare architectures without rebuilding the input pipeline for every laboratory.[1]
The lasting contribution may be infrastructural. In language and vision, large standardised corpora created a common substrate for repeated experiments. MassNet attempts the same move for proteomics: preserve billions of individual spectra in a form that can be trained on, audited and extended. The next test is whether future versions add the hard spectra and uncommon modifications that the first corpus necessarily leaves behind.[1]
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