MassNet: A Foundational Resource for Advancing AI in Proteomics

MassNet is the largest publicly available DDA-based proteomics dataset to date, and the first specifically optimized for AI applications.

~30 TB of raw DDA-MS data

1.54 billion MS/MS spectra;

558 million peptide-spectrum matches (PSMs);

Coverage across 35 species, including animals, plants, and microorganism;

The human subset covers ~98% of annotated proteins;

The release of MassNet marks a new chapter in AI-driven proteomics:

The first foundational training dataset in proteomics, comparable to those in NLP and CV;

Enables AI-driven applications in non-model organism research, biomarker discovery, and PTM identification;

Built on a standardized format and high-performance architecture to support scalable and reproducible proteomics analysis;

MassNet

Species: 36
Data volume: ~30 TB of DDA-MS data
LC/MS runs: 27,643
MS/MS spectra: ~1.53 billion
PSMs: 558 million
Precursors: 4.7 million
Peptides: 2.69 million
Protein groups: 133,495

MSDT (Mass Spectrometry Data Tensor)

MassNet structures raw spectra into 2D tensors, preserving key features like m/z and intensity through a unified data format.

Standardized Structure: Unified spectra–peptide–protein format ensures consistency and interpretability.
Non-Autoregressive Design: Greatly improves inference speed and efficiency.
Parameter-Free Optimization: Handles complex data without manual tuning.

XuanjiNovo

XuanjiNovo is a MassNet-based decoding model integrating multiple core innovations.

Curriculum Learning: Dynamically adjusts difficulty for faster, more stable training.
Non-Autoregressive Design: Greatly improves inference speed and efficiency.
Parameter-Free Optimization: Handles complex data without manual tuning.