A billion-scale mass spectral corpus arrives: 1.54 billion spectra across 35 species give de novo sequencing models a shared starting point

Mass spectrometry proteomics now has a large-scale spectral corpus organized under one schema and directly usable by deep learning frameworks, spanning animals, plants and microorganisms across 35 species. Earlier spectral libraries were heavily human-centric, with limited coverage of other model organisms, and stored peak lists separately from peptide sequences, charge states and modification sites in formats that varied across tools. The corpus integrates 27,643 liquid chromatography-tandem mass spectrometry runs from 648 projects in the PRIDE and iProX repositories, searched with FragPipe and Sage at a 1% false discovery rate; after merging and deduplication it holds 548 million target peptide-spectrum matches, 4.61 million peptide precursors and 133,422 proteins.

The companion MSDT format stores each spectrum as schema-defined numerical arrays, with fragment m/z values and intensities kept as aligned arrays alongside target or decoy labels and provenance metadata. Conventional formats such as MGF and mzML are row-oriented text, expensive to parse and poorly suited to selective column-wise loading, which raises input/output overhead on large datasets. Seven serialization formats were benchmarked on 15 randomly selected raw files from each of the Orbitrap, timsTOF and TripleTOF platforms; Parquet offered a favorable balance of compression ratio and read/write speed and became the standard storage format, with run-level MSDT files about 82% smaller than the combined source files.

XuanjiNovo, trained on MassNet, predicts a complete peptide sequence in parallel rather than residue by residue, reducing sequential decoding overhead. Autoregressive models generate one token at a time from left to right, which carries sequential decoding cost and can be unstable on complex heterogeneous data. The model was trained on 100 million PSMs randomly sampled from MassNet, with training and validation split at the peptide-sequence level so the two sets share no peptide sequences.

A curriculum learning scheme keeps the parallel-decoding model converging on large, heterogeneous mass spectrometry data instead of collapsing during training. Non-autoregressive models predict all positions at once, expanding the search space exponentially with sequence length; in the authors' experiments up to 80% of training runs failed to converge. Early in training the decoder receives a partially masked ground-truth alignment path as conditional input, and the masking ratio increases as training proceeds, while the model is still optimized under the original CTC objective.

The corpus includes only spectra confidently identified by FragPipe and Sage at a 1% false discovery rate, leaving unidentified spectra out, and was built around three modifications: methionine oxidation, N-terminal acetylation and cysteine carbamidomethylation. On the 15-species benchmark XuanjiNovo reaches peptide recall of 0.45 on human and 0.68 on mouse, and the abstract reports a 38.35% to 143.57% improvement over state-of-the-art methods on human Astral data. Data are distributed in Parquet and converted to an LMDB memory-mapped key-value store during training to support random access.

On the nine-species benchmark, XuanjiNovo without fine-tuning yields lower peptide recall than existing methods for most species; the authors attribute this to the benchmark's earlier-generation Q Exactive instruments and older PEAKS annotation pipeline differing from the training data. Fine-tuning with 30 million PSMs from MassIVE-KB improves performance across species, though on human data it remains slightly below ContraNovo. The PSM rescoring model was trained on only about 4 million PSMs and identified fewer peptides than FragPipe and Sage, which the authors expect to improve with more data and tuning. The model's handling of modifications such as phosphorylation, ubiquitination and glycosylation remains limited.