acceptodds
Under review as a conference paper at ICLR 2027

Non-Tryptic Terminus Simulation for Generalizable De Novo Peptide Sequencing

Abstract

Tandem mass spectrometry (MS/MS)-based de novo peptide sequencing provides a high-throughput approach for identifying peptide sequences without relying on a predefined protein sequence database, making it valuable for discovering novel protein sequences. The performance of de novo peptide sequencing AI model is fundamentally dependent on the scale and diversity of training data. However, existing training datasets are dominated by tryptic peptides, whose C termini are strongly enriched for lysine (K) or arginine (R) due to the cleavage specificity of trypsin. This constrained cleavage pattern substantially limits the diversity of peptide sequences and fragmentation patterns represented in training data. Here, we propose Non-Tryptic Terminus Simulation (NTS), an interpretable and general data augmentation method that generates MS/MS spectra with non-tryptic C termini directly from existing tryptic peptide-spectrum matches. NTS removes the C-terminal residue from a tryptic peptide and transforms its corresponding spectrum by selectively removing or mass-shifting fragment ions dependent on the removed residue, while preserving their original intensities. This targeted and interpretable transformation substantially expands the diversity of training data without requiring additional wet-lab experiments or spectrum-generation models. Across multiple widely used datasets and three representative de novo sequencing models, Casanovo, π-HelixNovo, and π-HelixNovo2, NTS significantly improves model performance and generalization, yielding average relative peptide recall gains of 50.9%, 12.4%, and 17.2%, respectively. These results demonstrate that NTS provides an interpretable and broadly applicable solution to the training-data bottleneck in de novo peptide sequencing.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Tandem mass spectrometry, de novo peptide sequencing, data augmentation, spectrum simulation, non-tryptic peptides, proteomics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_84UqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a data augmentation method called Non-Tryptic Terminus Simulation (NTS), which generates simulated MS/MS spectra with non-tryptic C termini from existing tryptic PSMs. The data augmentation represents an alternative for generating a diverse set of non-tryptic C-termini PSMs without requiring the need for generative models or lab experiments. Experiments conducted on three models and 11 test datasets show that data augmented with NTS substantially improves amino-acid precision, amino-acid recall, and peptide recall across both trypsin and non-trypsin digested-peptide chains. Additionally, the data augmentation broadens the frequency distribution of amino acids beyond lysine (K) and arginine (R).

Critical strengths
  • S1. The paper introduces NTS, which generates simulated MS/MS spectra corresponding to peptides with non-tryptic C termini from existing tryptic PSMs. The paper takes b, a, and y fragment ions into consideration and either retains peaks if they remain unaffected by residue removal or shifts/deletes peaks if they are affected by the removal process. The relative peak-intensities are retained in the process.

  • S2. Experiments on three models and 11 test datasets show that data augmented with NTS improves metrics for both trypsin-digested and non-tryptic peptides. Additionally, NTS augmentation increases the C-terminal amino-acid distribution beyond lysine and arginine.

Critical weaknesses
  • W1. My main concern is that the augmented data is never cross-referenced with actual data from a non-tryptic counterpart of an original tryptic PSM. Given the relative complexities involved in data generation, whether the new peaks actually match a genuinely removed version could provide evidence for the effectiveness of the method.

  • W2. In Section 3.3 the b_(n-1) ion was removed from the simulated spectrum without an explanation. Suppose the original sequence contained ‘NHGMHFR’, the b_(n-1) fragment sequence is ‘NHGMHF’ which does not contain the arginine.

Questions for authors
  • Q1. Could the authors clarify why the b_(n-1) ion was removed from the simulated spectrum, given that this fragment does not contain the C-terminal residue being removed?

  • Q2. Could the authors provide evidence that the NTS- augmented data match experimentally observed spectra of the corresponding non-tryptic peptides?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 84Uq

Thank you for your comments and raising the constructive questions.

Reply to Q1:

Although the original bn−1b_{n-1} ion does not contain the removed C-terminal residue, it no longer represents a valid b-type fragment after truncation. For example, in the original peptide NHGMHFR, the b6b_6 ion (NHGMHF) is generated by cleavage of the peptide bond between F and R. After removing R, NHGMHF becomes the complete truncated peptide, and there is no longer a peptide bond after F that can be cleaved to generate this b ion. Therefore, NTS retains b1b_1 through bn−2b_{n-2}, while removing bn−1b_{n-1}.

Reply to Q2:

To evaluate how well the augmented spectra approximate real non-tryptic spectra, we directly compared the NTS-augmented spectra with experimental non-tryptic peptide spectra from MSV000094014. In total, 382 peptide sequences were shared between the two datasets, of which 226 had matching modifications and precursor charge states, yielding 16,848 experimental–simulated spectrum pairs. Using one-to-one assignments, we computed the square-root-intensity cosine similarity across all theoretical b/y-ion channels, with unobserved ions assigned zero intensity. With equal weighting across peptides, the b/y-ion similarity was 0.710±0.1710.710 \pm 0.171. As a control, we randomly reassigned the intensities of the ions in the NTS-augmented spectra among the theoretical b/y-ion channels, which reduced the similarity to 0.211±0.0530.211 \pm 0.053.

We further compared full-spectrum similarity, including non-b/y peaks, between simulated spectra and their corresponding experimental spectra versus randomly paired spectra. Specifically, each simulated spectrum was paired with 20 randomly selected experimental peptides of the same precursor charge (precursor m/z within ±5\pm 5 Da) that did not overlap with the corresponding peptide. Using the same simulated spectra, the full-spectrum similarity was 0.382±0.1290.382 \pm 0.129 for corresponding peptides versus 0.063±0.0390.063 \pm 0.039 for randomly paired peptides. Mirror plots of representative examples, annotated with detected b/y ions and similarity scores, are shown in Figure 4.

Together, these results indicate that NTS-augmented spectra reasonably approximate real non-tryptic spectra (4.5 Comparing NTS-augmented spectra with experimental non-tryptic spectra).

I thank the authors for their effort. The rebuttal addressed my concern. I maintain my original positive score.

Reviewer_2C8vOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary
  1. The augmentation method is simple and directly applicable to existing models.
    NTS constructs training examples through explicit mass calculations and peak operations, without requiring an additional spectrum-generation network or new experimental data. The algorithm is clearly described and accounts for multiple charge states and isotope peaks, facilitating implementation and adoption.

  2. The improvements are consistent across models and evaluation conditions.
    The authors evaluate three sequencing models rather than a single architecture, covering different species, tissues, and digestion protocols. Improvements are observed under the vast majority of conditions in Table 1, supporting the approach’s applicability across multiple settings.

  3. The fixed-step training comparisons reduce confounding from additional optimization budgets.
    Both the original-data and augmented-data conditions use 600k training steps with the same batch size. The results therefore cannot be explained simply by the augmented models receiving more gradient updates. This design helps establish the contribution of changes to the training-data distribution.

Critical strengths
  1. The augmentation method is simple and directly applicable to existing models.
    NTS constructs training examples through explicit mass calculations and peak operations, without requiring an additional spectrum-generation network or new experimental data. The algorithm is clearly described and accounts for multiple charge states and isotope peaks, facilitating implementation and adoption.

  2. The improvements are consistent across models and evaluation conditions.
    The authors evaluate three sequencing models rather than a single architecture, covering different species, tissues, and digestion protocols. Improvements are observed under the vast majority of conditions in Table 1, supporting the approach’s applicability across multiple settings.

  3. The fixed-step training comparisons reduce confounding from additional optimization budgets.
    Both the original-data and augmented-data conditions use 600k training steps with the same batch size. The results therefore cannot be explained simply by the augmented models receiving more gradient updates. This design helps establish the contribution of changes to the training-data distribution.

Critical weaknesses
  1. The actual behavior of the peak-matching tolerance is inconsistent with its explanation in the text.
    Under Equation (13), the tolerance is always 0.2 Da when m/z is below 10,000. Thus, the tolerance is effectively fixed at 0.2 Da throughout the typical fragment-ion range. This contradicts the explanation that the relative tolerance varies with mass and avoids the limitations of a fixed Da window.

  2. Preserving experimental intensities and noncanonical peaks does not establish sufficient chemical realism.
    Removing a terminal K/R residue changes more than the theoretical masses of certain ions: it may also alter proton distribution, fragment charge states, and relative b/y-ion intensities. Previous experimental studies have shown that the positions and numbers of basic residues influence fragment-ion intensities.

  3. Key ablations are missing, leaving it unclear whether the detailed peak transformations drive the gains.

Questions for authors
  1. How is the chemical realism of the simulated spectra validated?
    Are paired experimental spectra available for original peptides and their actual truncated counterparts, allowing comparisons of peak coverage, mass consistency, and intensity relationships? How are neutral-loss peaks, residual precursor peaks, and fragment charge-state distributions affected by terminal-residue removal handled?

  2. Does the full spectral transformation outperform simpler controls?
    Could you compare full NTS with controls that update only the sequence and precursor information, delete only selected peaks, or apply generic spectral perturbations matched for the number of modified peaks? These comparisons would help determine whether the gains arise from chemically appropriate mappings or from changes in terminal-residue priors and regularization.

  3. Was train–test overlap checked again after augmentation?
    Please report exact sequence and peptidoform overlap between the original and truncated training sequences and each test set, together with performance after excluding overlaps. Please also provide the numbers of spectra and unique peptides in each split, as well as the validation and checkpoint-selection procedures.

  4. What is the precise denominator of peptide recall?
    Does the model produce a prediction for every test spectrum? If predictions can be absent or filtered, please report full-coverage accuracy, precision among emitted predictions, and coverage, and provide the complete sequence-matching criteria.

  5. Are the performance gains stable across training runs?
    Please provide results across multiple random seeds for the key experiments and confidence intervals based on grouping by peptide or acquisition run. Please clarify how the average relative improvements in the abstract are aggregated, and report absolute improvements and sample sizes for each evaluation condition.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 2C8v

Thank you for your constructive comments. We have run additional experiments and revised the manuscript. Main results are summarized below, and details are in the sections indicated.

Reply to Q1

We compared NTS-augmented spectra with experimental non-tryptic spectra from MSV000094014. The b/y-ion similarity was 0.710±0.1710.710 \pm 0.171, compared with 0.211±0.0530.211 \pm 0.053 when intensities were randomly reassigned among b/y-ion channels. Full-spectrum similarity was 0.382±0.1290.382 \pm 0.129 for corresponding experimental–simulated pairs, compared with 0.063±0.0390.063 \pm 0.039 for random pairs. These results indicate that NTS-augmented spectra reasonably approximate experimental non-tryptic spectra. Details are provided in Section 4.5, and Figure 4 shows representative mirror plots annotated with b/y ions and similarity scores.

NTS accounts for fragment charges up to 3+ and isotope peaks from M+0 to M+2, with charge-dependent m/zm/z shifts. Residual precursor peaks are removed during MS/MS preprocessing, and precursor m/zm/z is recalculated after truncation while its charge remains unchanged. Neutral-loss peaks are retained without modification. NTS is intended as a lightweight, fragmentation-aware augmentation method rather than a full simulation of peptide fragmentation.

Reply to Q2

We conducted ablation experiments using Casanovo, keeping the training budget, the amount of augmented data, and the random seed identical across all augmentation variants. The results show that neither updating only the peptide sequence and precursor information nor applying peak removal without peak shifting improved peptide recall over the baseline. In contrast, incorporating peak shifting into the sequence and precursor updates substantially increased peptide recall from 12.1% to 17.9%, while the full NTS, which combines peak shifting and removal, further increased recall to 19.0%. Furthermore, removing unmatched peaks from NTS-generated spectra reduced recall from 19.0% to 15.0%, highlighting the benefit of retaining these peaks. Replacing the fragmentation-aware transformations with chemistry-agnostic perturbations (random peak removal, shifting, or intensity perturbation), while keeping the number of affected peaks the same, resulted in a substantially lower recall of 7.2%. These results demonstrate that the performance gains stem from the fragmentation-aware transformations and that retaining unmatched peaks in NTS-generated spectra is beneficial. Details of the ablation variants and their results are provided in Appendix A.5 and Table 7.

Reply to Q3

After excluding test PSMs whose peptide sequences appeared in either the original or NTS-augmented training set, NTS continued to improve peptide identification on unseen peptides across different models and random seeds. The results are reported in Tables 4-6.

Across the deduplicated test datasets, 5,639 of 745,786 unique peptide sequences (0.7561%) and 5,995 of 835,807 unique peptidoforms (0.7173%) overlapped with the original training set. For the NTS-augmented training set, these numbers increased slightly to 6,012 (0.8061%) and 6,377 (0.7630%), respectively. Details of the overlap analysis, together with per-dataset spectrum and unique peptidoform counts, are provided in Appendix A.4.

For checkpoint selection, all models used the final checkpoint at 600,000 training steps, without any additional selection or tuning. This procedure was applied consistently across all models.

Reply to Q4

Every test spectrum receives exactly one prediction from each model, so the number of predicted peptides equals the number of ground-truth peptides. To ensure a full-coverage evaluation, even if a prediction were empty, it would still be compared with the ground-truth peptide.

Reply to Q5

To further assess the robustness of NTS, we repeated the key experiments with two additional random seeds (three independent runs in total) and evaluated all models on the test sets. The results show that NTS consistently improves peptide identification across random seeds and datasets. We also report the mean ± standard deviation (SD), the relative and absolute improvements, the numbers of spectra and unique peptidoforms, and confidence intervals calculated by grouping spectra according to their acquisition runs.

For the average relative improvements reported in the abstract, we first calculate the relative improvement for each paired run, average the results across three runs within each dataset condition, and then take the equally weighted mean across all dataset conditions.

Response to the comment on Eq. (13)

To ensure consistency with the actual implementation, we revised the description of Eq. (13) by replacing “relative tolerance” with “a fixed tolerance of 0.2 Da” and removing the previous claim regarding the advantages of a mass-dependent tolerance.

Thank you for your reply. It has cleared up most of my questions.

Reviewer_5Ku6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

NTS augments peptide-spectrum pairs by removing the C-terminal residue and deleting or mass-shifting affected fragment peaks while preserving retained peak intensities. Experiments with three de novo sequencing models on 11 test datasets report broad improvements across species and digestion conditions without requiring additional experimental data or a learned spectrum generator.

Critical strengths

NTS offers a simple, model-agnostic augmentation that diversifies peptide termini without additional wet-lab data or a learned generator. Figure 3 shows that the combined K/R terminal frequency falls from 95.8% to 49.8%, directly addressing the dominant training-data bias.

Table 1 reports broad gains across three models and 11 test datasets under matched 600k-step training budgets. For example, π-HelixNovo2’s reported peptide recall on the non-enzymatic dataset rises from 23.8% to 32.6%, supporting practical cross-condition benefits.

Critical weaknesses

The experiments compare NTS with no augmentation but do not include simpler augmentation baselines or ablations of its spectral transformations. Consequently, the contribution of terminal diversification versus generic regularization remains unclear.

Preserving original intensities and leaving unmatched peaks unchanged is an approximation whose spectral fidelity is not quantified in the main text. The results support NTS as useful augmentation more directly than they establish faithful simulation of truncated-peptide spectra.

Questions for authors

On one backbone, how does NTS compare with simple peak dropout or intensity perturbation under the same training budget? Which spectral transformations are essential to the improvement?

Using existing experimental spectra or targeted ablations, can you assess the effect of preserving intensities and retaining unmatched peaks on simulation quality and downstream performance?

Equation 18 divides matched peptides by predicted peptides: does every test spectrum receive exactly one prediction from every model? If not, please clarify how missing predictions are counted and report recall over all evaluated spectra.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 5Ku6

Thank you for the constructive comments. We have performed additional experiments and revised the manuscript accordingly (A.5 Ablation study).

  1. Under the same training budget, we compared NTS with generic regularization such as peak dropout and intensity perturbation to generate the same amount of augmented data as NTS. The results in Table 7 show that peak dropout and intensity perturbation reduced performance relative to the no-augmentation baseline, which is expected because random perturbations corrupt key b/y signals.

  2. We performed further ablation experiments to evaluate the effectiveness of each NTS component:

    • Sequence/precursor-only: update only the peptide sequence and precursor information after removal of the C-terminal residue.
    • Sequence/precursor + peak removal: update the peptide sequence and precursor information and remove the bn−1b_{n-1}, an−1a_{n-1}, and y1y_1 ions after removal of the C-terminal residue.
    • Sequence/precursor + peak shift: update the peptide sequence and precursor information and shift the y2,…,yn−1y_2,\ldots,y_{n-1} ions after removal of the C-terminal residue.

    The results in Table 7 show that peak shifting contributed the largest gain because it affects the largest number of b/y ions. Peak removal alone gave no clear improvement, but combining it with peak shifting yielded a larger gain than shifting alone.

  3. We evaluated the effect of preserving intensities and retaining unmatched peaks on simulation quality and downstream performance.

    (a) Intensity perturbation randomly alters b/y signals and increases the deviation from real spectra. Its performance drop in the intensity perturbation experiment described in Point 1 supports the necessity of preserving original intensities.

    (b) Removing all unmatched peaks from NTS-augmented spectra lowered downstream peptide recall from 19.0% (NTS) to 15.0% (Table 7), because the resulting spectra contain only b/y signals and lack the non-b/y peaks present in real spectra. To quantify simulation quality, we compared augmented spectra with experimental non-tryptic spectra from MSV000094014 (382 shared peptide sequences, 226 of which had matching modifications and precursor charges, yielding 16,848 experimental–simulated pairs). Using one-to-one assignments, the full-spectrum similarity (square-root-intensity cosine) decreased from 0.382±0.1290.382 \pm 0.129 to 0.215±0.1110.215 \pm 0.111 after the unmatched peaks were removed.

    Together, these results support preserving intensities and retaining unmatched peaks.

  4. Every test spectrum receives exactly one prediction from each model, so the number of predicted peptides equals the number of ground-truth peptides. We checked all predictions and found no empty outputs. To ensure a full-coverage evaluation, even if a prediction were empty, it would still be compared with the ground-truth peptide.

  5. Because preserving original intensities and leaving unmatched peaks unchanged is an approximation, we revised the statement “NTS produces simulated spectra that remain faithful to the complexity of real experimental measurements” to: “NTS produces augmented spectra that retain the complexity of the original experimental spectra.” (page [6], lines [281-282]).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.