acceptodds
Under review as a conference paper at ICLR 2027

FreeDIA: Search-Free Pretraining for DIA De Novo Peptide Sequencing

Abstract

De novo peptide sequencing reconstructs peptide sequences directly from tandem mass spectra without relying on a reference database search. However, sequencing from data-independent acquisition (DIA) spectra is challenging because peaks from multiple co-isolated precursors and background signals are mixed within the same spectrum. Existing DIA de novo methods often use search-dependent supervision based on database-identified peptides, limiting the use of unidentified DIA data for representation learning. We introduce FreeDIA, a search-free pretraining framework that learns precursor–fragment attribution from sequence-independent chromatographic supervision. Instead of using rule-based assignments as hard filters, FreeDIA uses them as supervision while retaining the complete observed spectrum, jointly learning denoising and deconvolution and transferring precursor-specific fragment evidence to peptide decoding. Across two datasets, FreeDIA achieves strong performance under both unseen-peptide and sample-holdout evaluations, with state-of-the-art generalization to unseen peptide sequences. Further analyses show that the learned attribution captures precursor-specific and graded spectral evidence beyond binary assignments, while ablation studies confirm the complementary roles of pretraining and reranking. These results highlight search-free spectral pretraining as an effective way to exploit DIA data without requiring peptide identifications during pretraining. Code for reproducing our experiments is provided in the supplementary material.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
de novo peptide sequencing, mass spectrometry, proteomics, representation learning, pretraining

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vLd3OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes FREEDIA, a search-free pretraining framework for data-independent acquisition (DIA) spectra data. FreeDIA leverages DIA-Umpire to generate precursor–fragment attribution pseudo-labels from chromatographic co-elution patterns, enabling encoder pretraining without peptide sequence annotations. The pretrained encoder is subsequently fine-tuned for peptide sequencing, followed by a spectrum-evidence reranker to improve candidate selection. Across twodatasets, FREEDIA achieves strong performance under both unseen-peptide and sample-holdout evaluations.

Critical strengths
  1. Search-free pretraining for DIA data. FreeDIA introduces a pretraining framework that does not require peptide sequence annotations, enabling the utilization of large-scale unidentified DIA data for representation learning.

  2. Strong experimental performance. FreeDIA achieves strong results across two DIA datasets under both unseen-peptide and sample-holdout evaluations.

Critical weaknesses
  1. Unclear peptide overlap between pretraining and downstream evaluation. Although the unseen-peptide split ensures no sequence overlap among the downstream training, validation, and test sets, the paper does not clearly report whether peptide sequences in the pretraining corpus overlap with those in the OC/UTI test sets. While pretraining does not use peptide sequence labels, potential exposure to test peptide spectra could still affect the interpretation of generalization to unseen peptides.

  2. Although FreeDIA's main contribution is its pretraining framework, it additionally employs a reranker absent from DIANovo. Without the reranker, FreeDIA achieves lower peptide recall than DIANovo on both OC and UTI in the unseen-peptide setting.

Questions for authors
  1. Could the authors report the peptide sequence overlap ratio between the pretraining corpus and the OC/UTI unseen-peptide test sets and evaluate performance after excluding overlapping test peptides?

  2. Could the authors apply the same reranker to DIANovo or incorporate the proposed pretraining into DIANovo to better isolate the contribution of the pretraining framework?

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

FreeDIA introduces search-free pretraining for DIA de novo peptide sequencing. It learns precursor–fragment relationships from DIA-Umpire assignments without requiring peptide identifications, then transfers the encoder to peptide generation and applies spectrum-based candidate reranking. Evaluations cover peptide-disjoint and sample-holdout splits on OC and UTI, modified peptides, and cross-species transfer.

Critical strengths
  1. A useful way to exploit unidentified DIA data. Using chromatographic assignments as supervision while retaining unassigned spectral evidence addresses a practical limitation of identification-dependent pretraining. This provides a meaningful route to learning from data that cannot supply peptide sequence labels.

  2. Evidence that the learned representations are useful. The pretraining and reranking ablations show complementary benefits. Figure 4(c) also shows improved recovery of peptide-matched fragments among traces unassigned by DIA-Umpire, providing evidence beyond simply reproducing the training assignments.

Critical weaknesses
  1. The extreme differences in baseline generalization remain insufficiently explained. In Table 1, DiffNovo’s OC peptide recall changes from 79.50% under sample holdout to 0.06% under unseen-peptide evaluation, and Transformer-DIA also deteriorates sharply, while DIANovo remains substantially stronger.

  2. The connection between the proposed attribution task and improved sequence generalization is not sufficiently established. DIANovo already learns from coeluting spectral evidence through pretraining. FreeDIA’s sequence-independent supervision is a meaningful distinction, but the current ablations compare pretraining against no pretraining and do not separate the supervision choice from additional data exposure.

Questions for authors
  1. What do the existing training, validation, and decoding records indicate about the severe unseen-peptide failures of Transformer-DIA and DiffNovo? Can they distinguish sequence overfitting from invalid, empty, or mass-inconsistent predictions?

  2. Relative to DIANovo’s coelution-aware pretraining, what existing evidence supports attributing FreeDIA’s generalization advantage specifically to its attribution supervision, rather than differences in pretraining data, spectral representation, or candidate selection?

  3. Appendix D states that attribution scores are not directly used in decoding. Does the decoder receive the target precursor’s query representation or chromatographic profile, or only precursor mass and shared spectral embeddings? Please clarify how target-specific ownership information influences generation.

3 comments
AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer duZE,

Thank you again for your thoughtful review and constructive feedback. We have carefully addressed your concerns in our rebuttal, providing additional experimental results and clarifications.

As the discussion period is coming to a close, we would greatly appreciate it if you could review our responses. If you feel that your concerns have been sufficiently addressed, we would be grateful if you would consider updating your assessment.

Thank you for your time and consideration.

Thank you for the detailed responses and the additional analyses, which address my main questions.

The failure analysis of Transformer-DIA and DiffNovo clearly separates sequence memorization and mass inconsistency from invalid predictions, and the experiment directly tests the role of the supervision choice. I note that the Ident performs close to the DIANovo-style, so much of the gain over scratch appears to come from pretraining itself, with attribution supervision adding roughly 3 points. This is still a meaningful result.

One point I would like to discuss: since the decoder receives only the precursor mass and shared MS2 embeddings, and removing explicit attribution scores leaves performance unchanged, the description that the precursor-conditioned representation “directly guides peptide decoding” seems stronger than the actual mechanism. How does the decoder distinguish the target precursor from co-isolated precursors of similar mass within the same trace group? I suggest revising this claim accordingly.

Given the responses, I am willing to raising my score to 3.

We sincerely thank the reviewer for this valuable observation and apologize for any confusion caused by our previous response.

In general, precursors with similar masses (m/z and charge) and retention times are often treated as the same precursor feature. Therefore, we initially expected this issue to have a limited impact on performance. However, to carefully address the reviewer's concern and quantitatively assess how frequently such cases occur in practice, we conducted an additional analysis. We defined similar precursors using the 20 ppm mass tolerance specified in our paper. We found that 8.7% of cases involved precursors with similar masses within the same trace group, of which 8.6 percentage points corresponded to the same peptide. Thus, only approximately 0.1% of cases were potentially ambiguous, suggesting a limited impact on overall performance in our datasets.

In fact, we had initially implemented an option to incorporate explicit attribution scores into decoding. However, since we observed virtually no performance difference (0.4354 vs. 0.4355), we disabled this option in the final model. One possible explanation is that, as shown in our analysis above, cases where explicit attribution scores would be particularly beneficial are very rare in our current datasets. Nevertheless, the reviewer's observation has encouraged us to reconsider the potential importance of this module, particularly for datasets containing more distinct precursors with similar masses and retention times.

Therefore, while precursor mass conditioning and learned MS2 representations appear sufficient for effective decoding on our current datasets, incorporating explicit attribution scores or target precursor query embeddings could further improve robustness across more diverse datasets.

We also agree with the reviewer's suggestion that the phrase "directly guides peptide decoding" should be revised to more accurately reflect the actual mechanism. We will clarify this description and discuss the potential limitation and future directions in the revised manuscript.

We sincerely appreciate the reviewer's careful consideration of our work and for highlighting an aspect that we had not sufficiently addressed. This discussion has been very helpful in refining our explanation and identifying meaningful directions for future improvement.

Reviewer_EJ3ROpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FreeDIA, a search-free pretraining framework for DIA de novo peptide sequencing. It uses DIA-Umpire-derived chromatographic supervision to learn precursor–fragment attribution without peptide identifications, retaining complete spectral information for denoising and deconvolution. The pretrained encoder is subsequently fine-tuned with an autoregressive peptide decoder and a spectrum-evidence reranker. Experiments on OC and UTI datasets demonstrate strong unseen-peptide generalization, with additional cross-species and PTM evaluations supporting transferability. Ablation studies further validate the complementary benefits of pretraining and reranking.

Critical strengths
  1. The proposed search-free pretraining strategy provides a meaningful alternative to database-dependent spectral supervision. By learning precursor–fragment relationships from chromatographic evidence while preserving the complete observed spectrum, FreeDIA can exploit unidentified DIA data and addresses an important limitation of existing approaches.
  2. The experimental results provide convincing evidence of effectiveness and generalization. FreeDIA achieves the best unseen-peptide performance across all reported metrics on both datasets, while the ablation studies demonstrate substantial and complementary contributions from pretraining and reranking. The cross-species and PTM evaluations further strengthen empirical findings.
Critical weaknesses
  1. The framework remains dependent on DIA-Umpire for precursor detection and pseudo-label generation, potentially inheriting its coverage limitations and assignment errors. However, the paper does not sufficiently quantify precursor detection coverage or investigate robustness to inaccurate assignments, leaving uncertainty about performance on challenging low-abundance or heavily co-eluting precursors.
  2. The comparisons with existing methods involve substantially different inference configurations. FreeDIA employs bidirectional beam search with beam width 50 per direction and an ensemble reranker, whereas several baselines use smaller beams or greedy decoding without equivalent mass constraints. These differences make it difficult to determine how much of the performance improvement comes from the proposed pretraining strategy rather than a stronger inference pipeline.
  3. The reproducibility of the reranking component is insufficiently supported. In Appendix C, the definition of the candidate similarity function used in Eq. (14) is left as an unresolved placeholder rather than being explicitly specified. Since reranking contributes substantially to peptide-level accuracy, this missing methodological detail makes it difficult to reproduce the reported results and independently assess the contribution of this component.
Questions for authors
  1. How sensitive is FreeDIA to DIA-Umpire assignment quality? Could the authors report precursor detection coverage and evaluate performance under noisy or incomplete precursor–fragment assignments?
  2. Could the authors provide a more controlled comparison using comparable decoding budgets, mass constraints, and reranking strategies where applicable, to better isolate the improvements attributable to the proposed pretraining method?
2 comments
AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer EJ3R,

Thank you again for your thoughtful review and constructive feedback. We have carefully addressed your concerns in our rebuttal, providing additional experimental results and clarifications.

As the discussion period is coming to a close, we would greatly appreciate it if you could review our responses. If you feel that your concerns have been sufficiently addressed, we would be grateful if you would consider updating your assessment.

Thank you for your time and consideration.

Dear Reviewer EJ3R,

As the discussion period approaches its end, we would like to briefly highlight the additional analyses and clarifications provided in response to your review:

  • DIA-Umpire dependency: Since complete ground truth is unavailable, absolute detection coverage cannot be reliably measured. We therefore compared DIA-Umpire with MSFragger as an alternative reference, without assuming either method is superior. More importantly, our additional analyses showed no downstream performance degradation despite incomplete precursor detection or fragment assignments.
  • Inference fairness: We conducted additional baseline comparisons under more closely aligned inference settings, confirming FreeDIA’s superior performance in the unseen-peptide setting.
  • Reranker reproducibility: We clarified the exact candidate similarity function used in Eq. (14).

We hope these additional analyses and clarifications address your concerns. We would be happy to clarify any remaining questions before the discussion closes.

Thank you again for your thoughtful review and time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.