bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / protein / dinovo
Protein
Chinese Academy of SciencesShandong University of TechnologyBeijing Institute of LifeomicsInstitute of Computing Technology, Chinese Academy of SciencesUniversity of Chinese Academy of SciencesChinese Academy of Medical SciencesChina Medical UniversityReleased March 2025

DiNovo

De novo peptide sequencing from mirror-protease mass spectra, reading paired complementary spectra to recover near-complete fragment ion coverage.

The short version

  • —Sequences peptides no database contains, for antibodies, neoantigens, and metaproteomes
  • —Reads a pair of complementary spectra jointly rather than merging them into one
  • —Recognizes mirror spectral pairs without pre-sequencing, under target-decoy error control
  • —Estimates FDR by mapping sequenced peptides to a decoy database, with no prior search
60Openness2Citations
4GitHub stars
LGPL-3.0License

Where to run it

No providers recorded yet. Browse all providers

De novo peptide sequencing reads an amino acid sequence directly out of a tandem mass spectrum, with no protein database to match against. It is the only route to peptides no database contains: antibodies from an immunized animal, neoantigens, venom proteins, and metaproteomes. The obstacle is that one spectrum rarely carries every fragment ion needed to read a sequence off unambiguously, and one protease rarely digests a proteome fully.

DiNovo, from Yan Fu's group at the Chinese Academy of Sciences with collaborators at Shandong University of Technology and the Beijing Institute of Lifeomics, attacks the problem from the experiment rather than the model. Mirror proteases cut on opposite sides of the same residue: trypsin and LysargiNase both act at lysine and arginine, Lys-C and Lys-N at lysine. Digesting one sample with both yields mirror peptides sharing an identical interior sequence but carrying that residue at opposite termini. Under higher-energy collisional dissociation the trypsin product yields rich y ions and the LysargiNase product rich b ions, filling each other's gaps.

DiNovo makes this usable end to end: it recognizes which spectra form mirror pairs, sequences them with either a deep network or a graph algorithm, and estimates a false discovery rate on the result. Its neural sequencer, MirrorNovo, is the first deep learning model built for mirror spectra, in the lineage of single-spectrum sequencers such as DeepNovo, PointNovo, and Casanovo.

#Key Features

  • Mirror-pair recognition without pre-sequencing: MirrorFinder pairs spectra directly from precursor masses, fragment masses, and intensities, skipping the slow, quality-sensitive step of sequencing each spectrum first, and controls pairing FDR with a target-decoy strategy.
  • Two interchangeable sequencers: MirrorNovo is a neural model needing a GPU; pNovoM2 uses spectrum graphs and dynamic programming on CPU, far faster. Under 1% FDR, over 99% of pairs sequenced by both yield identical sequences.
  • Joint reading of paired spectra: MirrorNovo builds a peak-to-ion matching matrix per spectrum and aligns the two at homologous fragmentation sites, exploiting complementarity without merging peak lists.
  • Database-free error control: Target-decoy mapping scores sequenced peptides against forward and reversed databases, giving an FDR estimate needing no prior peptide-spectrum matches.
  • Two protease pairs, not one: Adding Lys-C/Lys-N to trypsin/LysargiNase addresses missed cleavages as well as missing ions, since the second pair yields longer peptides.

#Technical Details

MirrorNovo matches observed peaks in both spectra against 18 theoretical fragment ion types: singly and doubly charged b, y, and a ions, their neutral-loss forms, and six internal ion types. A T-Net of three 1D convolutional layers extracts peak features, a GRU captures dependencies across peaks, three fully connected layers follow, and a softmax emits probabilities over 20 amino acids with leucine and isoleucine merged. Decoding is iterative, using beam search, a knapsack constraint, and bidirectional prediction, stopping when prefix and precursor masses agree within 20 ppm. Training used mirror pairs annotated at 1% FDR from Vero, MC2155, and human testis samples.

Across eight E. coli and yeast datasets, mirror spectra reached average fragment ion coverage of 98.4% and 98.0%, against 90.2% and 89.7% for single spectra. Against single-protease sequencers — PEAKS, pNovo3, PointNovo, and Casanovo, the last trained on 30 million MassIVE-KB spectra — DiNovo returned 110.2% to 195.3% more high-confidence peptides, 35.9% to 74.2% higher amino acid coverage, and 10.3% to 19.9% higher protein coverage. Against trypsin alone it sequenced 458.9% and 509.2% more high-confidence peptides. At matched 1% FDR it recovered 85.4% and 73.7% of the amino acids and 92.4% and 91% of the proteins found by the pFind search engine.

#Applications

The beneficiaries are groups sequencing proteins databases do not cover — antibody characterization, immunopeptidomics, venomics, and metaproteomics — where completeness matters more than throughput. Target-decoy mapping matters more broadly: putting de novo results on the same FDR footing as a database search lets de novo sequencing be reported as a primary identification method rather than a fallback. Adopting it does require changing the wet-lab protocol: a sample must be split and digested with both members of each protease pair.

#Impact

DiNovo is the first system to support the full mirror-protease workflow, from pairing spectra through sequencing to quality control — Lys-Sequencer was never released and pNovoM shipped without the pairing step that makes the strategy work. MirrorNovo trained on far fewer spectra than Casanovo or π-PrimeNovo because annotated mirror pairs are scarce, leaving clear headroom as such data accumulates. Pairing also scales quadratically in spectrum count, needing tighter retention-time bounds or spectral clustering on large runs. The source is archived on Zenodo under LGPL-3.0, though the README also limits DiNovo to academic use, a restriction LGPL-3.0 itself does not impose. The trained MirrorNovo weights ship only inside the 570 MB Windows standalone package, not the repository or the Zenodo archive, and the deep sequencer needs a GPU.

At a glance

Released
March 2025
Category
Protein
License
LGPL-3.0
Organizations
Chinese Academy of Sciences / Shandong University of Technology / Beijing Institute of Lifeomics / Institute of Computing Technology, Chinese Academy of Sciences / University of Chinese Academy of Sciences / Chinese Academy of Medical Sciences / China Medical University

Related models

  • XuanjiNovo

    Westlake University / Fudan University / Shanghai AI Laboratory / Zhejiang University / University of British Columbia / Westlake Omics

  • RefineNovo

    Fudan University / University of British Columbia / Shanghai AI Laboratory / Zhejiang University

  • OmniNovo

    Fudan University / Shanghai AI Laboratory / Tsinghua University / Westlake University / Tongji University / Shanghai Innovation Institute / University of British Columbia / Zhejiang University / Stony Brook University

  • Casanovo

    Noble Lab

  • PLMNovo

    Duke University

Links

GitHub RepositoryResearch PaperbioRxiv PreprintOfficial WebsiteLink

Tags

cnngrumass_spectrometryproteomicssupervised

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.