Research map / Label-efficient and robust learning / Recognition and long-tail
ContraNovo: A Contrastive Learning Approach to Enhance De Novo Peptide Sequencing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ContraNovo, a transformer-based de novo peptide sequencing model that adds a CLIP-style spectrum-peptide contrastive loss and prefix/suffix mass information to the decoder, reaching state-of-the-art precision on 9-species benchmarks.
Problem
De novo peptide sequencing from tandem mass spectra is treated as a plain sequence-to-sequence translation task, which overlooks the relationships between spectra and peptides and underuses amino acid mass information.
Approach
Combine a contrastive loss that aligns global spectrum and peptide embeddings with a transformer decoder that is fed prefix and suffix mass sum embeddings, and use an amino acid lookup table built from mass and index embeddings.
Key results
- On 9-species-V1, ContraNovo reaches average amino acid precision 0.785 and peptide precision 0.621, versus CasaNovoV2's 0.739 and 0.539.
- On 9-species-V2, ContraNovo averages 0.851 amino acid precision and 0.677 peptide precision, above CasaNovoV2's 0.818 and 0.640.
- Ablation on bacillus shows removing the contrastive loss drops peptide precision from 0.688 to 0.660 under beam search.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ContraNovo instead of CasaNovoV2 | de novo peptide sequencing · 9-species-V1 average | amino acid precision: 0.739 -> 0.785 | Helped |
| use ContraNovo instead of CasaNovoV2 | de novo peptide sequencing · 9-species-V1 average | peptide precision: 0.539 -> 0.621 | Helped |
| use ContraNovo instead of CasaNovoV2 | de novo peptide sequencing · 9-species-V2 average | amino acid precision: 0.818 -> 0.851 | Helped |
| use ContraNovo instead of CasaNovoV2 | de novo peptide sequencing · 9-species-V2 average | peptide precision: 0.640 -> 0.677 | Helped |
| add beam search decoding (ablation) | ContraNovo · Bacillus | amino acid precision: +3.1% | Helped |
| add beam search decoding (ablation) | ContraNovo · Bacillus | peptide precision: +2.8% | Helped |
| add contrastive learning loss (ablation) | ContraNovo · Bacillus | amino acid precision: 2.2% | Helped |
| add contrastive learning loss (ablation) | ContraNovo · Bacillus | peptide precision: 2.7% | Helped |
| add amino acid lookup-table embedding (ablation) | ContraNovo · Bacillus | amino acid precision: 2.1% | Helped |
| add amino acid lookup-table embedding (ablation) | ContraNovo · Bacillus | peptide precision: 2.7% | Helped |
| add prefix and suffix mass information (ablation) | ContraNovo · Bacillus | amino acid precision: 0.1% | Helped |
| add prefix and suffix mass information (ablation) | ContraNovo · Bacillus | peptide precision: 3.1% | Helped |
Surprising findings
- Removing the prefix/suffix mass information barely changes amino acid precision (0.1% drop) but reduces peptide precision by 3.1%.
- On 9-species-V2, ContraNovo does not beat CasaNovoV2 on peptide precision for yeast and bacillus, where CasaNovoV2 is equal or higher.
Limitations
- Gains on 9-species-V2 are not uniform; yeast and bacillus peptide precision do not improve over CasaNovoV2.
- Contrastive loss weighting uses an ad hoc dynamic lambda rule (0.1 if CSLoss above 0.1, else 0).
- Evaluated only on the two nine-species benchmarks; generalization beyond them is not shown.
Experimental setup
Datasets: MassIVE-KB 9-species-V1 9-species-V2
Benchmarks: 9-species-V1 9-species-V2
Metrics: amino acid precision peptide precision precision-coverage curve
Baselines: Peaks DeepNovo PointNovo CasaNovo CasaNovoV2
Builds on: CLIP CasaNovo Transformer DeepNovo
Closest papers by meaning
- AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases (NeurIPS 2024 · similarity 0.93)
- NovoBench: Benchmarking Deep Learning-based \emph{De Novo} Sequencing Methods in Proteomics (NeurIPS 2024 · similarity 0.90)
- Universal Biological Sequence Reranking for Improved De Novo Peptide Sequencing (ICML 2025 · similarity 0.89)
- Latent Imputation before Prediction: A New Computational Paradigm for De Novo Peptide Sequencing (ICML 2025 · similarity 0.88)
- PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction (NeurIPS 2026 · similarity 0.85)
- Curriculum Learning for Biological Sequence Prediction: The Case of De Novo Peptide Sequencing (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.