XuanjiNovo

De novo peptide sequencing reads an amino acid sequence directly off a tandem mass spectrum without matching it against a reference database, which is the only route to identifying peptides from non-model organisms, novel open reading frames, and unexpected post-translational modifications. Progress here has been limited less by architecture than by data: the community's standard nine-species benchmark contains 175 files and about 1.53 million peptide-spectrum matches (PSMs), orders of magnitude smaller than the corpora that drove progress in language and vision.

XuanjiNovo is the sequencing model built to exploit a corpus at that missing scale. It ships alongside MassNet, a curated data-dependent acquisition (DDA) mass spectrometry resource derived from roughly 30 TB of raw files and comprising 1.54 billion MS/MS spectra and 558 million PSMs across 35 species of animals, plants and microbes. Within the human subset alone MassNet covers more than 1.7 million precursors and 19,966 proteins, or 98% of annotated human proteins. Both were produced by Tiannan Guo's group at Westlake University with Siqi Sun at Fudan University and Zhiqiang Gao at Shanghai AI Laboratory, and collaborators at Zhejiang University, the University of British Columbia and Westlake Omics. The preprint was posted to bioRxiv in June 2025.

The corpus is distributed in a companion format, the Mass Spectrometry Data Tensor (MSDT), a Parquet-based schema storing one spectrum per row with precursor metadata, peak arrays and optional identification annotations, so training pipelines can stream batches to GPUs without bespoke parsers.

#Key Features

• Non-autoregressive decoding: The peptide decoder predicts token probabilities at all positions simultaneously with bidirectional attention rather than left-to-right, accelerating inference and letting the model use downstream context when resolving ambiguous residues.
• Curriculum learning without hand-tuning: A dynamic masking strategy raises task difficulty as performance improves, giving smooth convergence on the multi-distributional MassNet corpus without manual hyperparameter schedules.
• Precise mass control at decoding: A PMC module constrains the decoded sequence to match the measured precursor mass within a configurable tolerance, filtering candidates that are chemically inconsistent with the spectrum.
• Iterative self-refinement: Coarse peptide predictions are refined over several passes at inference time, with refinement iterations and CTC beam width exposed as user-facing parameters.
• Near-isotopic residue discrimination: The model resolves mass-degenerate residue pairs more reliably than prior methods, reaching recognition accuracies of 0.78 for lysine and 0.71 for glutamine, which differ by only 0.036 Da.

#Technical Details

The architecture is a Transformer spectrum encoder feeding a non-autoregressive Transformer peptide decoder, trained with a connectionist temporal classification (CTC) loss over alignment paths — the decoding paradigm of π-PrimeNovo, which the authors cite as a direct antecedent. Two checkpoints are released: XuanjiNovo-100M, trained on 100 million MassNet PSMs, and XuanjiNovo-130M, further fine-tuned on 30 million MassIVE-KB PSMs. Inference runs at roughly 280 spectra per second per GPU on an NVIDIA H200.

Evaluation used a purpose-built 15-species benchmark with 325 files, roughly five times the spectra and six times the PSMs of the nine-species set. Against π-HelixNovo, Casanovo V2, InstaNovo, ContraNovo and PrimeNovo, XuanjiNovo led across all species: peptide recall of 0.45 on human data acquired on an Orbitrap Astral (versus 0.32 for PrimeNovo and 0.26 for InstaNovo), 0.68 on mouse, 0.79 on Bacteroides thetaiotaomicron, 0.80 on Zea mays, and 0.25 on the low-coverage Solanum lycopersicum set where competing methods scored 0.001 to 0.06. Zero-shot transfer to unseen platforms held up: recall of 0.68 on an Orbitrap Fusion hepatocellular carcinoma dataset against PrimeNovo's 0.38. Training-scale ablations show amino-acid accuracy rising from 0.75 at 30M PSMs to 0.82 at 100M.

#Applications

The model targets any proteomics workflow where a reference database is unavailable or incomplete: metaproteomics, proteogenomics in crops and other non-model organisms, antibody and neoantigen sequencing, and PTM discovery. MassNet and MSDT are the more broadly reusable deliverables — the corpus is deposited in PRIDE under accession PXD071590, and a converter tool translates mzML, MGF and Bruker .d files into MSDT, letting other groups pretrain their own spectrum models without repeating the 30 TB curation effort.

#Impact

The core claim demonstrated here is that de novo sequencing accuracy scales with pretraining corpus size much as language modelling does, reframing an area that had largely optimized architectures against a fixed small benchmark. The authors are candid about the corresponding weakness: on the legacy nine-species benchmark, XuanjiNovo-100M scores below existing models on most species because that data came from older instruments than MassNet's, and only after fine-tuning on 30 million MassIVE-KB PSMs does XuanjiNovo-130M recover, with recall gains of 64.6% on S. cerevisiae and 60.2% on B. subtilis and a human recall of 0.60 that still trails ContraNovo's 0.62. Instrument-distribution shift, not model capacity, is the binding constraint. The preprint has not been peer reviewed, and checkpoints and inference code are public under a short-form Apache-2.0 notice that GitHub does not auto-detect, the Hugging Face checkpoint repository declares no licence at all, and the MassNet corpus is cited to a PRIDE accession that does not yet resolve.