MassNet releases a billion-scale, AI-ready mass spectrometry corpus for deep-learning proteomics
A team led by Westlake University in China, with collaborators at Zhejiang University, Fudan University, Soochow University, the University of British Columbia and several Shanghai research institutes, reported in Nature Methods, published online on 14 October, MassNet, an open artificial-intelligence-ready data corpus for mass-spectrometry-based proteomics.
Summary. A team led by Westlake University in China, with collaborators at Zhejiang University, Fudan University, Soochow University, the University of British Columbia and several Shanghai research institutes, reported in Nature Methods, published online on 14 October, MassNet, an open artificial-intelligence-ready data corpus for mass-spectrometry-based proteomics. The resource gathers data-dependent acquisition mass spectrometry runs from 648 public projects in the PRIDE and iProX repositories, covering 35 animal, plant and microbial species. In total it holds 27,643 liquid-chromatography-mass-spectrometry runs and about 1.5 billion tandem mass spectra, roughly 30 terabytes of raw data, searched with the FragPipe and Sage engines.
Research context (RUO). To make the data usable for machine learning, the authors define a compact tensor file format, the Mass Spectrometry DDA Tensor, that stores spectra, metadata and identification results in a single standardized structure and can be converted to a memory-mapped database for faster input and output during training. Using the corpus, the team trained XuanjiNovo, a non-autoregressive transformer model for de novo peptide sequencing, and reported large gains in de novo sequencing performance over earlier methods on data from a recent high-resolution analyzer, as well as support for retention-time prediction and peptide-spectrum-match rescoring. The full dataset is deposited in the PRIDE repository under accession PXD071590, and the model code and pretrained weights are released on GitHub under the Apache 2.0 licence. The work is a data-resource and software report with no clinical component. For proteomics and computational laboratories, it offers a large standardized training set and file format for building and benchmarking peptide-sequencing models.
