Transformer (NAR)
8 methods · 2024–2026
Transformer (NAR): Non-autoregressive transformers, which predict every residue position in one forward pass instead of left to right. Faster, and they need an explicit mechanism to keep the residues consistent with the precursor mass.
Non-autoregressive transformers, which predict every residue position in one forward pass instead of left to right. Faster, and they need an explicit mechanism to keep the residues consistent with the precursor mass.
The earliest of its 8 methods is π-PrimeNovo (2024); 7 more have followed.
| Methods | 8 |
| Papers describing them | 10 |
| Authors | 45 |
| Active | 2024-05-17 to 2026-08-03 |
| Deep learning | 8 of 8 |
| Kinds | algorithm (8) |
| Acquisition | DDA (7) |
Methods (8)
Oldest first, by the paper that describes each one.
- π-PrimeNovo (2024): NAR Transformer (CTC)
- CrossNovo (2025): AR + NAR hybrid
- RefineNovo (2025): Curriculum learning
- XuanjiNovo (2025): Billion-scale pretraining
- DiffNovo (2025): Transformer-diffusion model
- OmniNovo (2025): PTM-aware NAR
- Prime-DiffNovo (2026): Non-autoregressive Transformer de novo peptide sequencer with a diffusion-based refinement step at inference time: the NAR pass drafts a full sequence in parallel, and the diffusion loop iteratively corrects residue predictions to sharpen accuracy.
- FoxNovo (2026): Non-autoregressive de novo sequencer specialised for HLA-I immunopeptides. Uses dual-token m/z encoding (integer + decimal vocabularies, ~4k tokens instead of ~3M fine-grained bins) in the spectrum encoder, then dynamic-programming top-K decoding over the NAR probability matrix to enforce exact precursor-mass constraints. Reaches >90% peptide accuracy at ~2,800 spectra/s, over 100x faster than the beam-search baseline, and was used to re-analyse 168M spectra from 4,423 raw files in 18 h on one GPU, recovering 41 of 42 targeted-MS-validated non-canonical peptides.
How they score
1 of the 8 has been run on denovo_benchmarks, which ranks 17 tools over 84 datasets. The family’s best median rank is 5.
- π-PrimeNovo: median peptide-level average precision 0.808, median rank 5 of 17
Read these next to the rest of the field, not on their own: what the numbers mean.
Papers describing them (10)
- π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing (2024, bioRxiv, preprint)
- π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing (2025, Nature Communications, peer-reviewed)
- Distilling Non-Autoregressive Model Knowledge for Autoregressive De Novo Peptide Sequencing (2025, ICLR 2025, preprint)
- Curriculum Learning for Biological Sequence Prediction: The Case of De Novo Peptide Sequencing (2025, ICML 2025, preprint)
- MassNet: billion-scale AI-friendly mass spectral corpus enables robust de novo peptide sequencing (2025, bioRxiv, preprint)
- DiffNovo: A Transformer-Diffusion Model for De Novo Peptide Sequencing (2025, Bioinformatics and Computational Biology (BICOB 2025), peer-reviewed)
- Bidirectional Representations Augmented Autoregressive Biological Sequence Generation (2025, NeurIPS 2025, preprint)
- Accurate de novo sequencing of the modified proteome with OmniNovo (2025, arXiv, preprint)
- Prime-DiffNovo: Accurate Peptide De Novo Sequencing via Non-autoregressive Generation and Diffusion Refinement (2026, Lecture Notes in Computer Science (ISBRA 2026), peer-reviewed)
- Accurate and ultra-fast de novo HLA-I immunopeptide sequencing with FoxNovo (2026, LangTaoSha (LTS) Preprint, preprint)