Researchers have compiled 1.54 billion tandem mass spectra into a standardized dataset for developing artificial intelligence models in proteomics. The MassNet resource, described in Nature Methods, includes data from 35 animal, plant, and microbial species.
Mass spectrometry is widely used to identify and measure proteins in biological samples. Public repositories contain large quantities of proteomic data, but differences in file formats, metadata, and analysis methods can make these records difficult to combine for machine learning.
Jun A and colleagues of Westlake University in Hangzhou, China, created MassNet from 648 projects in the PRIDE and iProX repositories. The dataset contains 27,643 liquid chromatography-tandem mass spectrometry runs generated using more than 10 high-resolution instrument platforms. The original data occupied approximately 30 terabytes.
The researchers reanalyzed the files using two established database-search tools, FragPipe and Sage. After combining the results and removing duplicate identifications, the dataset contained 548 million peptide-spectrum matches, 4.61 million peptide precursors, 2.54 million unique peptides, and 133,422 proteins. Human and mouse samples contributed most of the data.
To make the resource easier to use for artificial intelligence, the team developed a format called Mass Spectrometry DDA Tensor, or MSDT. It stores spectral measurements alongside peptide identifications, retention times, instrument settings, search scores, and information about the source data.
MSDT files were approximately 82 percent smaller than the combined spectral and search-result files used to create them. The format can also be converted for faster, randomized access during deep-learning model training.
The researchers tested the resource by developing XuanjiNovo, an artificial intelligence model for de novo peptide sequencing. This process determines a peptide’s amino acid sequence directly from its tandem mass spectrum without relying on a reference protein database. It can therefore help analyze novel proteins, organisms with incomplete reference data, and previously unannotated protein regions.
XuanjiNovo was trained on 100 million peptide-spectrum matches. In a benchmark covering 15 species, it performed better than five other de novo sequencing models. For human data, its peptide recall was 0.45, compared with 0.32 for the next-best model in that comparison. Recall for mouse data was 0.68, compared with 0.60.
Increasing the training data from 30 million to 100 million peptide-spectrum matches improved accuracy and performance across species. Results still varied between datasets and mass spectrometry platforms, and some older instrument data required additional model training.
MassNet also supported retention-time prediction and peptide-spectrum match rescoring. Predicted and observed retention times had a correlation of 0.98. However, the experimental rescoring model identified fewer matches than established FragPipe and Sage workflows.
For laboratory and medical researchers, MassNet provides a common structure for training and comparing proteomics models. Such models could support protein identification and biomarker research, but the study did not assess clinical diagnostic performance.
The resource includes only spectra confidently identified by FragPipe or Sage, which may introduce search-related bias and excludes unidentified spectra. It also covers only a limited set of common protein modifications. Phosphorylation, glycosylation, ubiquitination, and other clinically relevant modifications were not systematically included.
MassNet is available through PRIDE under accession PXD071590. The code, documentation and trained XuanjiNovo models are also publicly available.
Newsletters
Receive the latest analytical science news, personalities, education, and career development – weekly to your inbox.
