acceptodds
Under review as a conference paper at ICLR 2027

Spectrum-Guided Label Refinement for Robust Peptide Sequencing

Abstract

Neural *de novo* peptide sequencing models learn from database-search labels that contain both false matches and systematic errors. We introduce Physics-Consistent Self-Correction (PCSC), which uses observed fragment peaks to refine these labels. A model predicts a peptide for each spectrum. The prediction replaces the current label when it receives a higher score from a fixed spectrum–peptide matcher. Models are then retrained on the updated labels. With 60% global corruption, five label updates recover 63.3% of corrupted labels, while four retraining rounds raise peptide recall on an independent nine-species benchmark from 0.3142 to 0.3970. On a separate corpus of over 35 million PSMs, one update improves recall in seven of nine species. The procedure also improves an autoregressive model, with gains of 6.88 percentage points under greedy decoding and 0.52 under mass-filtered beam search. These results show that fragment-based label refinement can improve sequencing accuracy across different models and training corpora.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
de novo peptide sequencing, learning with noisy labels, label refinement, self-training, tandem mass spectrometry

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QMFHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Physics-Consistent Self-Correction (PCSC) to address noisy training labels in de novo peptide sequencing. The method combines a neural sequencer with a fixed spectrum peptide matcher, iteratively generating candidate peptides, replacing current labels with higher scoring predictions, and retraining models from scratch. Experiments show improved reference sequence recovery and peptide recall on an independent nine-species benchmark under synthetic label corruption, with further performance improvements observed using a large-scale corpus of real-world PSMs and an autoregressive model.

Critical strengths
  1. The authors address the practically important problem of noisy training labels in de novo peptide sequencing. Using fragment evidence from experimental spectra to assist label correction is a reasonable idea that fits the task.

  2. The authors propose a simple and intuitive method that combines neural candidate generation with fixed spectrum–peptide matching. The method naturally incorporates mass spectrometry domain knowledge and can be applied to different sequencing models.

  3. The authors report higher final peptide recall than the evaluated robust learning and confidence-based label update baselines under synthetic label corruption, alongside improvements in label recovery and fragment matching. Preliminary experiments help explain model behavior under noisy supervision, while tests on real-world data and an autoregressive model provide additional support.

Critical weaknesses
  1. The methodological novelty is limited. PCSC mainly combines neural candidate generation, fixed spectrum–peptide scoring, label replacement, and iterative retraining. The authors do not clearly establish what substantive new mechanisms or generalizable insights this combination offers beyond existing self-training and label correction frameworks.

  2. The authors do not sufficiently justify the method’s key design choices or discuss its relationship to existing label correction and candidate rescoring methods. Higher matching scores do not necessarily imply more accurate labels. It remains unclear why the method uses only one candidate per spectrum, accepts any positive score difference, and repeatedly retrains from scratch.

  3. The baseline comparisons are insufficient. The robust learning baselines mainly come from older work, and their adaptation and tuning are not adequately explained. PCSC achieves 0.3970 peptide recall versus GCE’s 0.3790, but uses initial training plus four retraining rounds, whereas GCE is trained once. Although confidence-based updates are compared over the same number of rounds, recent robust learning methods, budget-matched robust learning baselines, and simple score-based filtering, weighting, or candidate reranking alternatives are missing.

  4. The practical benefits are not sufficiently validated. The authors do not establish that the synthetic perturbation types and proportions reflect real database search errors. On real-world annotated data, the unweighted mean peptide recall across nine species improves by only about 0.13 percentage points, with declines in some species. The PTM experiments report improved matching scores and matched peak counts without independent validation of sequence or modification-site accuracy.

Questions for authors
  1. How well does PCSC recover labels for peptides without any correctly labeled PSMs in the training set, compared with peptides that have other clean PSMs?

  2. Is there evidence independent of the matching score that label updates yield net corrections on real world data, accounting for originally correct labels that are incorrectly replaced?

  3. Does PCSC retain its advantage over GCE and other robust learning baselines under comparable total training budgets?

  4. What additional gains does multi-round retraining provide over score-based filtering or weighting, or single round label refinement using multiple candidates?

  5. How does PCSC compare with representative noisy label learning, label correction, or sequence self-training methods from the past two years? If these methods are inapplicable, please explain why.

Reviewer_Ldp8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Authors propose a physics-consistent, iterative self-correction (PCSC) label refinement method for de novo peptide sequencing that uses a fixed fragment-ion matching score. This method is used to select between a model’s predicted peptide and the current training label per spectrum. For a higher physical prediction score, it replaces the label. After that, the sequencer is retrained from scratch on the updated labels, and the process can be repeated. The results show that under controlled synthetic label corruption, PCSC fixes a large part of corrupted labels and improves peptide recall on a nine-species benchmark. Authors additionally demonstrate that the method is robust across different models and training corpora.

Critical strengths
  1. The proposed method is simple, yet a new and valuable direction for the label refinement of the peptide sequence. The method potentially offers a practical way to reduce noisy levels without manual annotations.

  2. The authors evaluated the proposed method across diverse experimental settings, which makes the method robust and model agnostic. Additionally, the method shows transferability across different decoders.

Critical weaknesses
  1. The paper used synthetic corruption, which does not represent real natural errors, and the synthetic corruptions are relatively easier to correct. The authors do not provide any further theoretical or empirical analysis in real-world settings. For that reason, the effectiveness of this method cannot be justified. Additionally, in this setup, many peptides may appear in clean as well as in corrupted forms. This may leak data for the model. As a result, the improvement in results can not be attributed solely to the proposed method.

  2. The method sets higher-score winning rules without any theoretical explanation. This does not explain why the higher score wins for a marginal score gap. It is not clear whether the method truly recovers the label or just selects peptides that better fit the scoring function. The authors did not provide any justification for this. Without the justification, the recovery method is a uncertain system.

Questions for authors
  1. Can the author establish the score-based margin for the acceptance criterion?
  2. Can the author provide a sensitivity analysis on the ion series considered, fragment-matching tolerances, and explain the results?
  3. The paper does not include a complete computational cost analysis. The method should require a large GPU training hours; how is the improvement in results justified by computational cost?
  4. No explicit False Discovery Rate analysis is given in the paper. Only increases observed in physical scores do not clearly guarantee correctness at the proteome level. Can the authors provide a setwise error analysis combining PCSC with False Discovery Rate (FDR) ?
  5. Can the author breakdown recovery by if a peptide appears only with corrupted labels to assess the method’s efficacy absent exposure to clean counterparts?
1 comment

There is no rebuttal submitted for this paper. I will maintain my original rating.

Reviewer_ZYSFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Physics-Consistent Self-Correction (PCSC), which uses model predictions and a fixed HyperScore to iteratively refine PSM labels. Experiments cover synthetic label corruption, large-scale real-world PSMs, PTM categories, and different model architectures.

Critical strengths
  1. The paper addresses an important problem in de novo peptide sequencing. The controlled experiments provide useful insights into correction and memorization under noisy supervision.
  2. PCSC is simple and well motivated. The paper includes robust-learning and acceptance-rule baselines, as well as evaluations on large-scale real-world PSMs and different model architectures.
Critical weaknesses
  1. The synthetic corruption setting may not adequately represent real database-search errors. Appendix C.1 shows that reference sequences have higher spectral-angle scores than corrupted labels in 98.62% of sampled PSMs. It remains unclear whether the reported recovery rates generalize to cases where incorrect labels have stronger spectral evidence.

  2. In the 60% global-noise dataset, 266,983 reference peptides occur with both clean and corrupted PSMs (Section 3.1). The model may benefit from clean PSMs of the same peptides during training, making it difficult to isolate its ability to recover sequences without such examples.

  3. The improvements on real-world PSMs are relatively limited. Table 2 shows an average peptide-recall improvement of approximately 0.13 percentage points across nine species, with decreases on B. subtilis and S. lycopersicum. Although amino-acid precision and recall improve for both species, the peptide-recall results suggest that the benefits are not consistent across species.

  4. The controlled comparisons and autoregressive experiments use a single training seed (42). In the autoregressive experiment, the reported 95% intervals are based on resampling evaluation spectra rather than repeated training runs. Furthermore, mass-filtered beam search yields only a 0.52-percentage-point improvement, with peptide recall decreasing in three species.

  5. The PTM experiments do not directly establish the correctness of the refined labels. HyperScore is used for label selection, so its increase is expected. Although matched peaks and TIC fraction also improve, exact sequence and modification-site accuracy are not measured, as acknowledged in Section 5.3.

Questions for authors
  1. Can the authors evaluate label recovery when no clean PSMs of the same reference peptide are available during training?

  2. Can the authors evaluate corrupted labels that have spectral matching scores closer to their reference sequences?

  3. Can the authors report results across multiple training seeds for the controlled corruption and autoregressive experiments?

  4. Can the authors provide independent evidence of exact sequence and modification-site accuracy for the refined real-world PSM labels?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.