Billion-Spectrum AI Corpus Pushes Proteomics Into the Deep Learning Era

Proteomics has long labored under a paradox: the field generates staggering volumes of data, yet its machine learning models have been starved of the kind of curated, standardized corpora that fueled the revolutions in language and vision. Now, a team led by researchers at Westlake University, Fudan University and the Shanghai Artificial Intelligence Laboratory has taken a decisive step toward closing that gap. In a study published in Nature Methods, they present MassNet, an artificial intelligence-ready corpus of 1.54 billion tandem mass spectrometry spectra drawn from 35 species, alongside a new deep learning model, XuanjiNovo, that sets a fresh benchmark for de novo peptide sequencing. The work is explicitly modeled on the datasets that transformed other branches of AI, such as ImageNet in computer vision and OpenWebText in natural language processing, and it may mark the moment proteomics acquires its own foundational data infrastructure.

The scale of the resource is considerable. MassNet aggregates data-dependent acquisition mass spectrometry data from 648 projects in the PRIDE and iProX public repositories, spanning 27,643 liquid chromatography–tandem mass spectrometry runs and roughly 30 terabytes of raw data collected across more than ten high-resolution instrument platforms. After processing through two independent search engines, FragPipe and Sage, each controlling false discovery at 1 percent, the corpus retains 548 million peptide-spectrum matches covering 4.61 million peptide precursors, 2.54 million unique peptides and 133,422 proteins. Human samples contribute approximately 1.7 million peptide precursors, while model organisms central to biomedical research, including mouse, fruit fly, nematode worm and budding yeast, are represented alongside plants and microorganisms. Every spectrum carries detailed annotations, from project identifiers to instrument parameters and search engine scores, ensuring full traceability back to the original raw files.

What distinguishes MassNet from earlier spectral libraries is not merely its size but its format. Existing resources such as the Global Proteome Machine Database, PeptideAtlas, ProteomeTools and MassIVE-KB have curated high-quality spectra for years, but they typically store peak lists separately from identification information in heterogeneous, text-based formats like MGF, mzML and .msp files. These lack unified schemas, resist efficient random access, and impose heavy parsing and harmonization burdens on anyone attempting large-scale model training. To solve this, the team benchmarked seven serialization formats across Orbitrap, timsTOF and TripleTOF platforms and settled on a columnar representation built on Apache Parquet, which they call the Mass Spectrometry DDA Tensor, or MSDT. Each spectrum is encoded as schema-defined numerical arrays of mass-to-charge ratios and intensities, linked directly to peptide annotations, target-decoy labels and rich metadata. Run-level MSDT files came out roughly 82 percent smaller than their combined source files, and the format can be optionally converted to a Lightning Memory-Mapped Database for fast shuffled access during training.

Built on this foundation, XuanjiNovo tackles one of proteomics’ hardest computational problems: inferring peptide amino acid sequences directly from fragmentation spectra without consulting a reference protein database. This de novo capability matters enormously for characterizing novel proteins, non-model organisms and unannotated corners of the proteome, where database search fails. The model extends the nonautoregressive transformer architecture of the earlier π-PrimeNovo system, predicting entire peptide sequences in parallel rather than residue by residue, which dramatically accelerates inference while enabling bidirectional context modeling. A transformer-based encoder learns representations of each spectrum, and a mass-constrained decoding module, formulated as a knapsack-style optimization solved by dynamic programming, ensures the predicted sequence matches the observed precursor mass within a defined error tolerance.

The most technically inventive element is the training strategy. Nonautoregressive transformers are notoriously unstable on large, heterogeneous datasets; in the team’s experiments, up to 80 percent of training runs failed to converge, because predicting all tokens in parallel expands the hypothesis space exponentially. The researchers countered this with a curriculum learning scheme built on connectionist temporal classification, or CTC. Early in training, the model is shown the most probable CTC alignment path for each ground-truth peptide with most tokens revealed, making the task nearly trivial. As training progresses, progressively more tokens are masked, forcing the model to reconstruct sequences from partial evidence until it can generate complete peptides unaided. During inference, an initial prediction is iteratively refined, with each pass conditioning the next, allowing the model to correct its own errors over multiple rounds before the mass-constrained module finalizes the output.

The performance gains are striking. Trained on 100 million spectra from MassNet, XuanjiNovo achieved improvements of 38.35 to 143.57 percent over state-of-the-art methods on human data acquired with the latest Orbitrap Astral platform, reaching a peptide recall of 0.45 on human and 0.68 on mouse, compared with 0.32 and 0.60 for π-PrimeNovo and 0.26 and 0.52 for InstaNovo+. On the notoriously difficult tomato dataset, where competing models managed recalls between 0.01 and 0.06, XuanjiNovo still achieved 0.25, demonstrating robustness under low-coverage, noisy conditions. The model also showed superior discrimination of near-isobaric amino acid pairs, a classic pitfall in the field, distinguishing lysine from glutamine, whose masses differ by less than 0.04 daltons, with accuracies of 0.78 and 0.71 respectively, and handling phenylalanine versus oxidized methionine with similarly improved precision.

Equally important is the evidence that scale itself drives performance. The team trained versions of XuanjiNovo on 30 million, 65 million and 100 million peptide-spectrum matches and evaluated them on a rigorous 15-species benchmark containing 30.78 million spectra, assembled with strict separation between training and evaluation peptide sequences. The largest model converged faster, maintained the lowest validation loss and achieved amino acid recall of approximately 0.82, more than 20 percent above π-PrimeNovo’s roughly 0.60. Cross-species generalization improved consistently with data volume: recall in Escherichia coli rose from 0.35 to 0.48, in fruit fly from 0.37 to 0.44, and in human from 0.36 to 0.45. Even species where the 30-million model already exceeded 0.72 recall, such as Bacteroides thetaiotaomicron and maize, climbed further to 0.79 and 0.80, indicating that benefits from scale have not yet plateaued.

The study is candid about the challenges of cross-platform generalization. On the widely used nine-species benchmark, acquired on older Q Exactive instruments and annotated with a legacy search engine, XuanjiNovo initially underperformed, a shortfall the authors attribute to differences in mass resolution, scan speed, chromatographic configurations and annotation pipelines that measurably shift spectral characteristics. A single fine-tuning step using 30 million peptide-spectrum matches from MassIVE-KB closed much of the gap, boosting recall by 64.62 percent in yeast and 60.23 percent in Bacillus subtilis, and lifting human performance to 0.60. On independent datasets from Orbitrap Fusion and Fusion Lumos instruments, including hepatocellular carcinoma tissue and synthetic ProteomeTools peptides, the base model outperformed rivals by margins of roughly 77 percent and 3 to 32 percent respectively, all without fine-tuning.

MassNet’s utility extends beyond sequencing. A retention time prediction model trained on Sage-derived MSDT data achieved a Pearson correlation of 0.98 between predicted and observed retention times, with a mean absolute error of 2.97 and 95 percent of deviations within 16.50, matching the established DeepLC tool. A peptide-spectrum match rescoring model built on the DDA-BERT architecture converged stably, reaching accuracy above 0.88 and area under the curve above 0.95 on validation data, and its identifications overlapped heavily with those of FragPipe and Sage on independent yeast and human datasets, confirming consistency in practical workflows.

The authors acknowledge real limitations. MassNet includes only spectra confidently identified by the two search engines, discarding unidentified spectra that may carry informative signals, and it was built with just two variable modifications, methionine oxidation and N-terminal acetylation, plus fixed cysteine carbamidomethylation, leaving phosphorylation, glycosylation, ubiquitination and other biologically critical modifications largely unmodeled. Future versions, the team suggests, could integrate additional search engines, retain curated unidentified spectra, incorporate data-independent acquisition modes and support specialized tasks such as immunopeptidomics. The entire dataset has been deposited in PRIDE under accession PXD071590, with code and pretrained weights released under an Apache 2.0 license on GitHub, alongside converter tools and Docker images. If the analogy to ImageNet holds, the release of a billion-scale, model-ready corpus could prove the inflection point at which AI-driven proteomics stops being a collection of bespoke models and becomes a scalable, cumulative discipline.

Subject of Research: A billion-scale AI-ready mass spectrometry corpus and a nonautoregressive transformer for de novo peptide sequencing in proteomics

Article Title: MassNet: billion-scale AI-ready mass spectral corpus empowering scalable deep learning in proteomics

Article References: MassNet: billion-scale AI-ready mass spectral corpus empowering scalable deep learning in proteomics. (n.d.). https://doi.org/10.1038/s41592-026-03247-y

Image Credits: AI Generated

DOI: 10.1038/s41592-026-03247-y

Keywords: proteomics, mass spectrometry, deep learning, de novo peptide sequencing, MassNet, XuanjiNovo, transformer, curriculum learning, spectral library, MSDT, retention time prediction, PSM rescoring