Research map / Probabilistic and scientific modeling / Molecules and proteins
Latent Imputation before Prediction: A New Computational Paradigm for 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 LIPNovo, which imputes latent representations of theoretical b/y-ion peaks from observed spectra via learnable peak queries and bipartite matching before autoregressive peptide decoding, improving de novo peptide sequencing on three benchmarks.
Problem
De novo peptide sequencing from tandem mass spectra suffers performance loss when fragmentation peaks are missing from the observed spectrum.
Approach
Impute missing fragment information in latent space rather than raw peaks: a set-prediction imputation module with learnable peak queries generates latent representations of theoretical b/y-ions, trained with Hungarian bipartite matching against encoded theoretical spectra, then combined with observed spectrum features for peptide decoding.
Key results
- On Nine-species, LIPNovo reaches 0.797 amino acid precision and 0.582 peptide precision, versus 0.741 and 0.529 for the retrained CasaNovo baseline.
- On Seven-species, amino acid precision improves to 0.557 from 0.357 for the retrained baseline (+20.0%).
- Parameter-matched baseline extension (12 layers plus FFN) gives only marginal gains (0.750 amino acid precision), while LIPNovo reaches 0.797.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add latent imputation before peptide prediction (ablation) | CasaNovo baseline · Nine-species | amino acid precision: 0.741 -> 0.797 | Helped |
| add latent imputation before peptide prediction | CasaNovo baseline · Seven-species | amino acid precision: 0.357 -> 0.557 | Helped |
| add latent imputation before peptide prediction | CasaNovo baseline · HC-PT | amino acid precision: 0.525 -> 0.637 | Helped |
| add complementary spectrum (ablation) | CasaNovo baseline · Nine-species | amino acid precision: 0.741 -> 0.755 | Helped |
| add latent imputation to baseline with complementary spectrum (ablation) | CasaNovo baseline · Nine-species | amino acid precision: 0.755 -> 0.766 | Helped |
| add theoretical-spectrum CE supervision (ablation) | LIPNovo without theoretical-spectrum CE supervision · Nine-species | amino acid precision: 0.766 -> 0.797 | Helped |
| add complementary spectrum (ablation) | LIPNovo without complementary spectrum · Nine-species | amino acid precision: 0.782 -> 0.797 | Helped |
| increase peak queries from 50 to 100 (ablation) | LIPNovo · Nine-species | amino acid precision: 0.787 -> 0.797 | Helped |
| increase peak queries from 100 to 150 (ablation) | LIPNovo · Nine-species | amino acid precision: 0.797 -> 0.792 | Hurt |
| use 1 imputation decoder layer instead of 3 (ablation) | LIPNovo · Nine-species | amino acid precision: 0.797 -> 0.796 | Hurt |
| use 6 imputation decoder layers instead of 3 (ablation) | LIPNovo · Nine-species | amino acid precision: 0.797 -> 0.793 | Hurt |
| set filtering threshold to 0.6 instead of 0.8 (ablation) | LIPNovo · Nine-species | amino acid precision: 0.797 -> 0.790 | Hurt |
Surprising findings
- Upper-bound analysis using ground-truth theoretical spectrum representations reaches 0.939 amino acid precision, indicating large headroom from better imputation.
Limitations
- Performance gains diminish at very high missing-fragmentation ratios (e.g., +0.8% on HC-PT at ratios 0.7-0.8).
- Imputation quality is limited when signal peaks are heavily lost, requiring cross-sample information.
Experimental setup
Datasets: Nine-species Seven-species HC-PT GraphNovo dataset
Benchmarks: Nine-species Seven-species HC-PT
Metrics: amino acid precision amino acid recall peptide precision peptide AUC PTM precision PTM recall
Baselines: PEAKS DeepNovo PointNovo InstaNovo CasaNovo AdaNovo pi-HelixNovo GraphNovo
Builds on: CasaNovo DETR AdaNovo pi-HelixNovo
Closest papers by meaning
- NovoBench: Benchmarking Deep Learning-based \emph{De Novo} Sequencing Methods in Proteomics (NeurIPS 2024 · similarity 0.90)
- AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases (NeurIPS 2024 · similarity 0.90)
- Universal Biological Sequence Reranking for Improved De Novo Peptide Sequencing (ICML 2025 · similarity 0.89)
- ContraNovo: A Contrastive Learning Approach to Enhance De Novo Peptide Sequencing (AAAI 2024 · similarity 0.88)
- PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction (NeurIPS 2026 · similarity 0.87)
- 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.