Research map / Probabilistic and scientific modeling / Molecules and proteins
Curriculum Learning for Biological Sequence Prediction: The Case of 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 RefineNovo, a non-autoregressive transformer for de novo peptide sequencing that combines CTC-path-based curriculum masking with iterative self-refinement at inference, reporting higher precision and recall than prior AR and NAR models on 9-species benchmarks.
Problem
De novo peptide sequencing from tandem mass spectra with non-autoregressive transformers is hampered by CTC training instability (loss explosion, overfitting, slow convergence) and by the lack of conditioning that prevents iterative refinement.
Approach
Uses the model's own most likely CTC alignment path as a teacher signal, masking tokens with an adaptive ratio tied to current accuracy so training moves from easy (mostly unmasked) to hard, and reuses the learned curriculum token embedding to iteratively refine predictions at inference, followed by precise mass control decoding.
Key results
- On 9-species-V2, averages 0.907 amino acid precision and 0.790 peptide recall, versus 0.891 and 0.777 for PrimeNovo.
- On 9-species-V1, averages 0.800 amino acid precision and 0.653 peptide recall, versus 0.788 and 0.638 for PrimeNovo.
- Across 20 random MassIVE-KB subsets, PrimeNovo fails in 18 training runs while the proposed method fails once (reported as more than 90% reduction in failures).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RefineNovo instead of PrimeNovo | 9-species-V1 · 9-species-V1 Average | AA Precision: 0.788 -> 0.800 | Helped |
| Use RefineNovo instead of PrimeNovo | 9-species-V1 · 9-species-V1 Average | Peptide Recall: 0.638 -> 0.653 | Helped |
| Use RefineNovo instead of PrimeNovo | 9-species-V2 · 9-species-V2 Average | AA Precision: 0.891 -> 0.907 | Helped |
| Use RefineNovo instead of PrimeNovo | 9-species-V2 · 9-species-V2 Average | Peptide Recall: 0.777 -> 0.790 | Helped |
| Use RefineNovo instead of PrimeNovo | NovoBench · NovoBench 9Species (yeast) | Score: $0.70$ -> $0.71$ | Helped |
| Use RefineNovo instead of PrimeNovo | NovoBench · NovoBench 7Species (yeast) | Score: $0.09$ -> $0.09$ | No clear effect |
| Use RefineNovo instead of PrimeNovo | NovoBench · NovoBench HC-PT | Score: $0.85$ -> $0.88$ | Helped |
| Add fixed-ratio curriculum learning with mask ratio 0.7 (ablation) | PrimeNovo · Ablation study | AA Precision: 0.788 -> 0.733 | Hurt |
| Add fixed-ratio curriculum learning with mask ratio 0.7 (ablation) | PrimeNovo · Ablation study | Peptide Recall: 0.638 -> 0.558 | Hurt |
| Add iterative refinement to fixed-ratio curriculum learning (ablation) | PrimeNovo · Ablation study | AA Precision: 0.733 -> 0.742 | Helped |
| Add difficulty annealing to curriculum learning (ablation) | PrimeNovo · Ablation study | AA Precision: 0.733 -> 0.793 | Helped |
| Use 3 rather than 1 iterative refinement step (ablation) | RefineNovo · Iterative refinement steps | Peptide Recall: 0.728 -> 0.736 | Helped |
Surprising findings
- Curriculum learning alone with a fixed mask ratio of 0.7 lowers performance below the naive NAT baseline (amino acid precision 0.733 vs 0.788), so fixed masking hurts and adaptive annealing is needed.
- On the NovoBench 7-species yeast set, all pretrained models including RefineNovo perform poorly (0.09), attributed to a distribution mismatch with the MassIVE-KB training data.
Limitations
- Performance on human peptide recall is slightly below ContraNovo (0.581 vs ContraNovo's higher value on V1), and the method does not beat all AR baselines everywhere.
- Generalization to out-of-distribution instrument data is weak (NovoBench 7-species yeast recall/precision near 0.09).
Experimental setup
Datasets: MassIVE-KB 9-species-V1 9-species-V2 NovoBench yeast
Benchmarks: 9-species-V1 9-species-V2 NovoBench (yeast)
Metrics: amino acid precision peptide recall
Baselines: Peaks Novo DeepNovo PointNovo Casanovo CasanovoV2 AdaNovo PrimeNovo ContraNovo HelixNovo SearchNovo InstaNovo
Builds on: PrimeNovo Connectionist Temporal Classification Glancing Transformer Precise Mass Control
Closest papers by meaning
- Universal Biological Sequence Reranking for Improved De Novo Peptide Sequencing (ICML 2025 · similarity 0.87)
- Bio-xLSTM: Generative modeling, representation and in-context learning of biological and chemical sequences (ICLR 2025 · similarity 0.87)
- NovoBench: Benchmarking Deep Learning-based \emph{De Novo} Sequencing Methods in Proteomics (NeurIPS 2024 · similarity 0.87)
- AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases (NeurIPS 2024 · similarity 0.86)
- Context-Aware Regularization with Markovian Integration for Attention-Based Nucleotide Analysis (NeurIPS 2025 · similarity 0.85)
- ContraNovo: A Contrastive Learning Approach to Enhance De Novo Peptide Sequencing (AAAI 2024 · 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.