Objective:

To create a standardized, large-scale mass spectrometry dataset for developing and evaluating artificial intelligence models in proteomics.

Approach:

• Dataset assembly: Researchers combined 648 projects from PRIDE and iProX, covering 27,643 LC–MS/MS runs, more than 10 high-resolution instrument platforms, and 35 species.
• Reanalysis and format: They reanalyzed spectra with FragPipe and Sage, removed duplicate identifications, and developed the Mass Spectrometry DDA Tensor (MSDT) format to store spectra, identifications, metadata, and search results.
• Model evaluation: The team trained XuanjiNovo on 100 million peptide-spectrum matches and benchmarked it against five other de novo peptide-sequencing models across 15 species.

Key Findings:

• MassNet contains 1.54 billion tandem mass spectra and 548 million peptide-spectrum matches, representing 2.54 million unique peptides and 133,422 proteins.
• MSDT files were approximately 82% smaller than the combined spectral and search-result files used to create them.
• In the reported human benchmark, XuanjiNovo peptide recall was 0.45 versus 0.32 for the next-best model; for mouse, recall was 0.68 versus 0.60.
• Increasing training data from 30 million to 100 million peptide-spectrum matches improved accuracy and performance across species.
• Predicted and observed retention times correlated at 0.98. The experimental rescoring model identified fewer matches than established FragPipe and Sage workflows.

Interpretation:

MassNet provides a standardized resource for training and comparing proteomics AI models. The study did not evaluate clinical diagnostic performance.

Limitations:

• The dataset includes spectra confidently identified by FragPipe or Sage, potentially introducing search-related bias and excluding unidentified spectra.
• Coverage of protein modifications is limited; phosphorylation, glycosylation, ubiquitination, and other modifications were not systematically included.
• Performance varied across datasets and instrument platforms, and some older-instrument data required additional model training.
• Human and mouse samples contributed most of the data.

Conclusion:

MassNet is available through PRIDE under accession PXD071590; code, documentation, and trained XuanjiNovo models are also publicly available.

Sources:

• PRIDE dataset, accession PXD071590