Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Universal Biological Sequence Reranking for Improved 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 RankNovo, a list-wise deep reranker that combines candidate peptides from multiple de novo sequencing models using axial attention over an MSA-like candidate grid, trained with two mass-deviation regression targets (PMD and RMD).
Problem
De novo peptide sequencing from tandem mass spectra is limited by noise and data-specific biases of individual deep learning models, so no single model consistently yields the correct peptide.
Approach
Treat de novo sequencing as list-wise reranking of candidate peptides from multiple base models, modeling candidates as an MSA and applying row and column axial attention plus spectrum cross-attention, supervised by a peptide-level alignment-based mass deviation (PMD) and a residue-level prefix-mass deviation (RMD) instead of binary correctness labels.
Key results
- On 9-species-V1, RankNovo reaches average peptide recall 0.660, versus 0.647 for the strongest base model ByNovo's 0.623 and ContraNovo's 0.618.
- On 9-species-V2, RankNovo reaches average peptide recall 0.781 and amino acid precision 0.906, above ByNovo's 0.758 and 0.897.
- Zero-shot on unseen base models, average peptide recall rises from 0.586 with 2 models to 0.649 with 6 models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| rerank candidates from multiple base models with RankNovo | de novo peptide sequencing · 9-species-V1 | peptide recall: 0.623 -> 0.660 | Helped |
| rerank candidates from multiple base models with RankNovo | de novo peptide sequencing · 9-species-V2 | peptide recall: 0.758 -> 0.781 | Helped |
| add RMD loss (ablation) | RankNovo · 9-species-V1 | peptide recall: 0.652 -> 0.660 | Helped |
| add PMD loss (ablation) | RankNovo · 9-species-V1 | peptide recall: 0.650 -> 0.660 | Helped |
| add column-wise attention (ablation) | RankNovo · 9-species-V1 | peptide recall: 0.653 -> 0.660 | Helped |
| use PMD and RMD losses instead of point-wise reranking objective (ablation) | RankNovo · 9-species-V1 | peptide recall: 0.647 -> 0.660 | Helped |
| use PMD and RMD losses instead of pair-wise reranking objective (ablation) | RankNovo · 9-species-V1 | peptide recall: 0.648 -> 0.660 | Helped |
| use PMD and RMD losses instead of list-wise reranking objective (ablation) | RankNovo · 9-species-V1 | peptide recall: 0.646 -> 0.660 | Helped |
| increase the number of base models (ablation) | RankNovo · 9-species-V1 | peptide recall | Helped |
| add unseen base-model predictions at inference | RankNovo trained on two base models · 9-species-V1 | peptide recall: 0.586 -> 0.649 | Helped |
Surprising findings
- Point-wise, pair-wise and list-wise RankT5-style rerankers all reach about 0.647 recall, nearly identical, and the proposed mass-deviation losses are needed to exceed them.
Limitations
- Inference is slower because it requires running multiple base models to collect candidates.
- Performance on longer peptides declines and the advantage narrows with length.
- Relies on the candidate pool; base model selection bounds the upper performance.
Experimental setup
Datasets: MassIVE-KB 9-species-V1 9-species-V2
Benchmarks: 9-species-V1 9-species-V2
Metrics: peptide recall amino acid precision
Baselines: PEAKS DeepNovo PointNovo Casanovo Casanovo V2 ContraNovo ByNovo RankT5-style point-wise, pair-wise, list-wise rerankers
Builds on: ContraNovo Casanovo RankT5 MSA Transformer
Closest papers by meaning
- AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases (NeurIPS 2024 · similarity 0.90)
- NovoBench: Benchmarking Deep Learning-based \emph{De Novo} Sequencing Methods in Proteomics (NeurIPS 2024 · similarity 0.90)
- ContraNovo: A Contrastive Learning Approach to Enhance De Novo Peptide Sequencing (AAAI 2024 · similarity 0.89)
- Latent Imputation before Prediction: A New Computational Paradigm for De Novo Peptide Sequencing (ICML 2025 · similarity 0.89)
- Curriculum Learning for Biological Sequence Prediction: The Case of De Novo Peptide Sequencing (ICML 2025 · similarity 0.87)
- PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction (NeurIPS 2026 · similarity 0.84)
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.