Research map / Label-efficient and robust learning / Recognition and long-tail
AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases
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 AdaNovo, a Transformer de novo peptide sequencing model that uses conditional mutual information between spectrum and amino acids/peptides to reweight training losses against PTM scarcity and spectral noise.
Problem
De novo peptide sequencing from tandem mass spectra struggles with post-translational modified (PTM) amino acids, which are rare in training data, and with noisy or missing peaks that make peptide-spectrum match labels unreliable.
Approach
Computes amino acid-level conditional mutual information between the spectrum and each target amino acid (via two decoders, one with and one without spectrum input) and aggregates it into PSM-level mutual information; both are used to adaptively reweight the cross-entropy loss during training.
Key results
- AdaNovo averages 0.499 peptide-level precision vs 0.459 for Casanovo (rep.) across nine species, and 0.716 amino acid-level precision vs 0.672.
- AdaNovo improves PTM identification precision by 3.6%-9.4% over competitors across datasets.
- On synthetic-noise Clam bacteria data, AdaNovo reaches 0.397 peptide precision vs 0.297 for Casanovo.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add amino acid-level adaptive training weights based on conditional mutual information (ablation) | AdaNovo · Human | PTM precision: 0.314 -> 0.483 | Helped |
| Add PSM-level adaptive training weights based on mutual information (ablation) | AdaNovo · Clam bacteria with synthetic noise | peptide precision: 0.336 -> 0.397 | Helped |
| Add amino acid-level adaptive training weights based on conditional mutual information (ablation) | AdaNovo · Yeast | PTM precision: 0.616 -> 0.665 | Helped |
| Use inverse-frequency amino acid re-weighting | Casanovo · Yeast | PTM precision: 0.552 -> 0.564 | Helped |
| Replace cross-entropy with focal loss | Casanovo · Yeast | PTM precision: 0.552 -> 0.550 | Hurt |
| Use AdaNovo instead of previous de novo sequencing methods | 9-species benchmark · 9-species benchmark (Average) | peptide-level precision: 0.459 -> 0.499 | Helped |
Surprising findings
- Simple upsampling re-weighting by inverse class frequency transfers poorly across species because PTM distributions differ between train and test species.
Limitations
- Cannot identify never-before-seen PTM types.
- Adds a second decoder, increasing parameter count by 40%.
- Hyperparameters s1 and s2 are sensitive and must be tuned per dataset; no error bars for most baselines.
Experimental setup
Datasets: 9-species benchmark (DeepNovo)
Benchmarks: 9-species benchmark
Metrics: peptide-level precision amino acid-level precision PTM identification precision
Baselines: DeepNovo PointNovo Casanovo PEAKS Re-weight upsampling Focal loss
Builds on: Casanovo DeepNovo PointNovo focal loss
Closest papers by meaning
- ContraNovo: A Contrastive Learning Approach to Enhance De Novo Peptide Sequencing (AAAI 2024 · similarity 0.93)
- NovoBench: Benchmarking Deep Learning-based \emph{De Novo} Sequencing Methods in Proteomics (NeurIPS 2024 · similarity 0.91)
- Universal Biological Sequence Reranking for Improved De Novo Peptide Sequencing (ICML 2025 · similarity 0.90)
- Latent Imputation before Prediction: A New Computational Paradigm for De Novo Peptide Sequencing (ICML 2025 · similarity 0.90)
- PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction (NeurIPS 2026 · similarity 0.87)
- PROSPECT PTMs: Rich Labeled Tandem Mass Spectrometry Dataset of Modified Peptides for Machine Learning in Proteomics (NeurIPS 2024 · similarity 0.86)
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.