Pith. sign in

REVIEW 3 major objections 6 minor 16 references

Diffusion Decoding for Peptide De Novo Sequencing

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A discrete diffusion decoder trained with the DINOISER loss lifts amino acid recall in peptide de novo sequencing from 0.081 to 0.454 over the autoregressive Casanovo baseline.

desk verdict The recall gain is real but an artifact of length-inflated predictions; the authors' own qualitative examples show it. read the letter →

arxiv 2507.10955 v1 pith:E5ELSXEO submitted 2025-07-15 cs.LG

classification cs.LG
keywords peptidedenovosequencingdiffusiondecodingdiscreteDINOISERlosstandemmassspectrometryautoregressiveaminoacidrecallknapsackbeamsearch
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Peptide de novo sequencing reconstructs amino acid sequences directly from tandem mass spectrometry data, without a database to look up, which matters for discovering novel peptides from unsequenced organisms. The paper asks whether the standard left-to-right autoregressive decoder can be replaced by a discrete diffusion decoder that can start generating from any part of the peptide, avoiding errors that cascade down the whole sequence. It finds a naive decoder swap worsens every metric, but pairing one diffusion design with the DINOISER noise-manipulation loss lifts amino acid recall from 0.081 to 0.454, a statistically significant gain of 0.373 over the autoregressive Casanovo baseline. Exact peptide matches stay at zero, and the paper itself notes predictions tend to be much longer than the ground truth and end in noise, so the recall gain is partly a property of the metric's normalization.

What carries the argument

The carrying object is the discrete diffusion decoder: a decoder that treats the peptide as a sequence of discrete amino-acid tokens and iteratively denoises a randomly initialized sequence, so generation can begin at any position instead of strictly left-to-right. It is trained with the DINOISER loss, a discrete-diffusion objective that corrupts predicted logits and clips small noise scales to improve generalization. The paper also tests knapsack beam search, a candidate-generation strategy that prunes sequences using precursor-mass constraints; it is not part of the best configuration, since it worsened all metrics and raised training time from about 1.5 hours to 35 hours.

What would settle it

Truncate every DINOISER prediction to the true peptide's length, or require the predicted mass to fall within precursor tolerance, and recompute amino acid recall; if the 0.373 advantage over Casanovo collapses, the headline gain is an artifact of unpenalized trailing tokens.

Watch

Extended reading notes

Core claim

The paper's central claim is that a discrete diffusion decoder, trained with the DINOISER loss and placed inside Casanovo's encoder, recovers a substantially larger fraction of true amino acids than Casanovo's original autoregressive transformer decoder. On the combined tryptic and non-tryptic test set, the best configuration, Casanovo-DM2 with DINOISER, achieves amino acid recall 0.454 versus 0.081 for Casanovo (p<0.001, Wilcoxon signed-rank test), while amino acid precision is slightly lower (0.070 versus 0.080) and peptide precision and peptide coverage stay at 0. The authors interpret this as evidence that diffusion decoding can enhance sensitivity by decoding high-confidence regions first, and that the choice of a diffusion-specific loss, rather than the decoder swap alone, is what unlocks the improvement.

Load-bearing premise

The load-bearing premise is that amino acid recall, the fraction of true amino acids recovered, is a trustworthy measure of sequencing quality even though predicted peptides can be much longer than the real one and end in noise, which the paper itself acknowledges after reporting the gain.

Editorial extensions

If this is right

  • With the DINOISER loss, the best diffusion decoder raises amino acid recall from 0.081 to 0.454, so diffusion decoding can substantially increase sensitivity relative to autoregressive decoding.
  • Because a decoder swap without the diffusion-specific loss lowers every metric, the loss function is the ingredient that makes discrete diffusion useful here.
  • Since peptide precision and peptide coverage remain 0, the gain does not yet translate into exact whole-peptide identification, and practical use depends on removing the trailing noise from predictions.
  • Knapsack beam search is not a useful decoding upgrade in this setting: it reduces all metrics and multiplies training time, so future effort should go instead into truncation or noise-removal strategies.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural test of the paper's account would be to truncate each DINOISER prediction at the true peptide's length, or at the precursor mass, and recompute metrics; if the recall gain then becomes peptide-level matches, the diffusion decoder is a practical front-end rather than a sensitivity amplifier with unusable outputs.
  • The reported 0.373 gain is tied to the recall definition, which divides by the true sequence length and does not penalize appended tokens; a length-normalized or mass-constrained metric would likely shrink the gap.
  • The same decoder-plus-DINOISER combination could be run on non-tryptic or cross-species spectra to see whether the sensitivity gain generalizes beyond the Casanovo dataset.
  • Discrete diffusion's infilling-from-any-position property is not specific to peptides, so the design could transfer to other sequence reconstruction problems with noisy flanks, such as antibody sequencing or protein assembly from partial spectra.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper investigates replacing the autoregressive transformer decoder in the Casanovo peptide de novo sequencing model with discrete diffusion decoders. Three diffusion decoder designs (Casanovo-DS, Casanovo-DM1, Casanovo-DM2) are integrated into the Casanovo framework, and the authors evaluate knapsack beam search and alternative loss functions (weighted entropy and DINOISER). The main reported result is that Casanovo-DM2 trained with the DINOISER loss achieves an amino acid recall of 0.454 versus 0.081 for the Casanovo baseline, a statistically significant increase of 0.373. However, the paper also reports that peptide precision and peptide coverage remain 0 for all diffusion models, amino acid precision for the best model (0.070) is below the Casanovo baseline (0.080), and the predicted sequences are typically much longer than the ground truth with trailing noise. The authors acknowledge this limitation explicitly in Section 4.3 and Section 5.

Significance. If the reported amino acid recall gain represented a genuine improvement in sequencing accuracy, the work would be a useful contribution to the growing literature on diffusion-based peptide sequencing, complementing methods such as InstaNovo+. The paper also provides useful negative results: simply replacing the transformer decoder with a diffusion decoder degrades all metrics, and knapsack beam search worsens performance while increasing training time from roughly 1.5 hours to 35 hours. The study is transparent about its limitations and reports standard metrics on a public dataset with the same train/validation/test split as Casanovo v4.2. The main significance is therefore conditional: the central positive claim rests entirely on a single metric, amino acid recall, whose interpretation is compromised by the length inflation described in the paper itself.

major comments (3)
  1. [Section 4.3, Table 4, Figure 9] The headline claim that DINOISER improves amino acid recall by 0.373 over Casanovo is not evidence of better peptide sequencing, because the paper itself states that most outputs are much longer than the ground truth and match the true prefix followed by noise. Since amino acid recall is normalized by the number of true amino acids, a prediction that matches the true prefix and then appends arbitrary amino acids can achieve near-perfect recall while remaining biologically unusable. The fact that the best model's amino acid precision (0.070) is lower than Casanovo's (0.080) and that peptide precision and coverage are 0 reinforces this concern. The authors should report length-controlled or truncation-aware metrics (e.g., recall after predicted-length truncation, prefix match rates, or precision-recall curves across prediction lengths) before claiming that the recall gain reflects enhanced sequencing capability.
  2. [Abstract and Section 5] The abstract's conclusion that the findings 'highlight the potential of diffusion decoders to ... drive significant advancements in peptide de novo sequencing' is not supported by the reported metrics: the best configuration has zero peptide-level precision and coverage, amino acid precision below the baseline, and predictions that exceed the desired mass capacity with trailing noise. This is a load-bearing interpretive overreach. The authors should either temper the claims to a narrow statement about amino acid recall under the current metric, or provide additional evidence that the recalled prefix is usable in practice, for example by showing that truncation recovers correct full-length peptides at a non-trivial rate.
  3. [Section 3.5 and Table 3] The comparison between Casanovo and the diffusion models' amino acid recall is asymmetric: Casanovo's predictions are length-constrained by its autoregressive decoder, while the diffusion models are free to generate arbitrarily long sequences. The paper's own qualitative analysis in Section 4.3 shows that DINOISER training encourages length inflation. To make the recall comparison meaningful, the authors should report the distribution of predicted sequence lengths for each model, control for length in the evaluation, or use a metric that penalizes extraneous predictions, such as amino acid precision or a combined F1 score. Without such an analysis, the 0.373 recall gain cannot be attributed to better sensitivity rather than to a decoding bias.
minor comments (6)
  1. [Section 2.1] There is a typo in 'long short-tappearingerm memory'; it should read 'long short-term memory'.
  2. [Section 5] The word 'incured' should be 'incurred'.
  3. [Tables 3 and 4] Tables 3 and 4 present overlapping content with an inconsistency: Table 3 reports Casanovo-DM1 weighted-entropy AA recall as 0.423, while Table 4 reports 0.0423. Please correct this and consider merging the tables or clearly indicating what each adds.
  4. [Section 3.1] The paper states that the complete code is attached as a zip folder, but no repository link or DOI is provided. For reproducibility, please provide a persistent link to a public code repository.
  5. [Section 3.3 and Figure 3] The three diffusion decoder architectures are described only through figures. Please provide the key hyperparameters in text, including the number of diffusion steps, noise schedule, embedding dimension, and inference-time denoising steps, as these are essential for reproducing the results.
  6. [Section 4.3] The statistical significance test is described as a Wilcoxon signed-rank test, but the paper does not state what unit was used for the paired observations (e.g., per-spectrum recall) or how tied values and zeros were handled. Please clarify the test setup and report the effect size or confidence interval.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the diffusion-decoder comparison is empirical, and the acknowledged recall-metric caveat is a validity concern, not a derivation loop.

full rationale

This paper makes no derivation claim that is forced by its own inputs. The headline result — Casanovo-DM2 with the DINOISER loss raising amino-acid recall from 0.081 to 0.454 (p < 0.001) — is an empirical comparison against an external Casanovo baseline on a fixed public test split, not a quantity computed from a fitted parameter, from an equation, or from a self-citation. The diffusion decoder designs and loss functions are adopted from prior work (SEDD, DINOISER), but the variant-specific results are measured and reported, not entailed by those citations. The paper itself explicitly flags the main limitation in Section 4.3 and again in Section 5: most predicted sequences are much longer than the ground truth, match the true prefix, and then append noise, while amino-acid recall remains high because its denominator is the true sequence length. That is a metric-validity caveat, not a circular step: the model was not constructed or fitted to maximize this recall by definition, and the paper honestly reports that peptide precision and peptide coverage remain zero. No fitted input is renamed as a prediction, no uniqueness theorem is imported from the authors, and no load-bearing citation is self-referential. The recall-gain interpretation may be scientifically fragile, but it is not circular under the definition used here, so no significant circularity is found.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper does not introduce new mathematics or entities. The load-bearing assumption is that the AA recall metric is informative despite unbounded predicted-sequence length; the paper's own qualitative results directly challenge this. All other assumptions concern fidelity to existing diffusion methods and fairness of the baseline comparison.

assumptions (4)
  • domain assumption Amino acid recall remains a valid measure of sequencing performance when predicted sequences are allowed to be much longer than the true peptide.
    Section 4.3: predicted sequences are longer than ground truth with noise at the end, and recall is normalized by true length; the paper's headline result depends on this metric being meaningful.
  • domain assumption The three diffusion decoder designs faithfully implement the discrete diffusion methods they are adapted from (SEDD, DINOISER).
    Sections 3.3 and 3.5 describe the designs but provide no independent verification beyond internal metrics.
  • domain assumption The Casanovo baseline is a fair comparison under equal training conditions.
    Section 3.3 calls the baseline 'original Casanovo' and says same training settings were used, but it is not explicit whether the baseline was retrained on the same split.
  • standard math Wilcoxon signed-rank test assumptions are satisfied by the per-spectrum amino acid recall values.
    Section 4 applies scipy.stats.wilcoxon without reporting the distribution or number of paired values.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diffusion Decoding for Peptide De Novo Sequencing." pith.science (2026). https://pith.science/paper/E5ELSXEO

@misc{pith2026250710955,
  author       = {Pith},
  title        = {Pith review of: Diffusion Decoding for Peptide De Novo Sequencing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5ELSXEO}},
  note         = {Machine review of arXiv:2507.10955}
}
read the original abstract

Peptide de novo sequencing is a method used to reconstruct amino acid sequences from tandem mass spectrometry data without relying on existing protein sequence databases. Traditional deep learning approaches, such as Casanovo, mainly utilize autoregressive decoders and predict amino acids sequentially. Subsequently, they encounter cascading errors and fail to leverage high-confidence regions effectively. To address these issues, this paper investigates using diffusion decoders adapted for the discrete data domain. These decoders provide a different approach, allowing sequence generation to start from any peptide segment, thereby enhancing prediction accuracy. We experiment with three different diffusion decoder designs, knapsack beam search, and various loss functions. We find knapsack beam search did not improve performance metrics and simply replacing the transformer decoder with a diffusion decoder lowered performance. Although peptide precision and recall were still 0, the best diffusion decoder design with the DINOISER loss function obtained a statistically significant improvement in amino acid recall by 0.373 compared to the baseline autoregressive decoder-based Casanovo model. These findings highlight the potential of diffusion decoders to not only enhance model sensitivity but also drive significant advancements in peptide de novo sequencing.

Figures

Figures reproduced from arXiv: 2507.10955 by the authors.

Figure 1
Figure 1. Overview of the peptide de novo sequencing process, illustrating the flow from inputting an observed spectrum to generating the corresponding peptide sequence, copied from Casanovo (Yilmaz et al., 2022). fragmentation of precursor peptides and noise present in MS/MS spectra (seen as black peaks in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Sample deep learning model architecture, copied from Casanovo (Yilmaz et al., 2022). This paper addresses the challenge of using diffusion de￾coding for peptide de novo sequencing and explores how to better integrate diffusion in the deep learning model archi￾tecture. The objectives of the project include incorporating different diffusion decoder designs into peptide sequencing workflows to replace traditional autor… view at source ↗
Figure 3
Figure 3. Overview of our methodology, with Part 1 exploring the replacement with different diffusion decoders in Casanovo, Part 2 trying knapsack with beam search, and Part 3 evaluating two of the diffusion model designs with different loss functions. Zenodo link on Casanovo’s website (Melendez & Noble, 2024). The combined dataset was created by obtaining data from MassIVE-KB v.2018-06-15 (tryptic) and MassIVE￾KB v2.0.15 (no… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: , [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Casanovo-DM2. Lastly, amino acid recall evaluates the model’s sensitiv￾ity by measuring the proportion of true amino acids that are correctly predicted, calculated as the ratio of matched amino acids to the total number of actual amino acids in the true sequences. In a…
Figure 7
Figure 7. Figure 7: Comparison of sample predicted sequences from models in Part 1 highlighting the impact of using three different decoder designs (Casanovo-DS, Casanovo-DM1, and Casanovo-DM2) versus an autoregressive transformer decoder (Casanovo) [PITH_FULL_IMAGE:figures/full_fig_p007…
Figure 8
Figure 8. Figure 8: Comparison of sample predicted sequences from models in Part 2 showcasing the impact of knapsack beam search decoding [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Sample predicted sequences produced by Casanovo-DM1 and Casanovo-DM2 using the three loss functions where Weighted refers to the weighted entropy loss function. Although our work showed decreased performance when using the knapsack algorithm during beam search, there m…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 11 canonical work pages

  1. [1]

    P., Skwark, M

    Eloff, K., Kalogeropoulos, K., Mabona, A., Morell, O., Catzel, R., Rivera-de Torre, E., Berg Jespersen, J., Williams, W., van Beljouw, S. P., Skwark, M. J., et al. Instanovo enables diffusion-powered de novo peptide sequencing in large-scale proteomics experiments. Nature Machine Intelligence, pp.\ 1--15, 2025

  2. [2]

    Github copilot, 2025

    GitHub . Github copilot, 2025. URL https://github.com/features/copilot

  3. [3]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

  4. [4]

    mgf - read and write ms/ms data in mascot generic format, 2020

    Levitsky, L., Goloborodko, A., and Gorshkov, M. mgf - read and write ms/ms data in mascot generic format, 2020. URL https://pyteomics.readthedocs.io/en/latest/api/mgf.html

  5. [5]

    S., and Hashimoto, T

    Li, X., Thickstun, J., Gulrajani, I., Liang, P. S., and Hashimoto, T. B. Diffusion-lm improves controllable text generation. Advances in neural information processing systems, 35: 0 4328--4343, 2022

  6. [6]

    Discrete diffusion modeling by estimating the ratios of the data distribution

    Lou, A., Meng, C., and Ermon, S. Discrete diffusion modeling by estimating the ratios of the data distribution. arXiv preprint arXiv:2310.16834, 2023

  7. [7]

    Mitigating the missing-fragmentation problem in de novo peptide sequencing with a two-stage graph-based deep learning model

    Mao, Z., Zhang, R., Xin, L., and Li, M. Mitigating the missing-fragmentation problem in de novo peptide sequencing with a two-stage graph-based deep learning model. Nature Machine Intelligence, 5 0 (11): 0 1250--1260, 2023

  8. [8]

    and Noble, W

    Melendez, C. and Noble, W. Data for 'accounting for digestion enzyme bias in casanovo' (melendez et al., 2024), 2024. URL https://zenodo.org/records/12587317

Show all 16 references
  1. [9]

    E., Oh, S., and Noble, W

    Melendez, C., Sanders, J., Yilmaz, M., Bittremieux, W., Fondrie, W. E., Oh, S., and Noble, W. S. Accounting for digestion enzyme bias in casanovo. Journal of Proteome Research, 23 0 (10): 0 4761--4769, 2024

  2. [10]

    Evaluation metrics and statistical tests for machine learning

    Rainio, O., Teuho, J., and Kl \'e n, R. Evaluation metrics and statistical tests for machine learning. Scientific Reports, 14 0 (1): 0 6086, 2024

  3. [11]

    Statistical functions (scipy.stats), 2025

    SciPy . Statistical functions (scipy.stats), 2025. URL https://docs.scipy.org/doc/scipy/reference/stats.html

  4. [12]

    H., Zhang, X., Xin, L., Shan, B., and Li, M

    Tran, N. H., Zhang, X., Xin, L., Shan, B., and Li, M. De novo peptide sequencing by deep learning. Proceedings of the National Academy of Sciences, 114 0 (31): 0 8247--8252, 2017

  5. [13]

    S., Cha, S

    Wang, M., Wang, J., Carver, J., Pullman, B. S., Cha, S. W., and Bandeira, N. Assembling the community-scale discoverable human proteome. Cell systems, 7 0 (4): 0 412--421, 2018

  6. [14]

    Dinoiser: Diffused conditional sequence learning by manipulating noises

    Ye, J., Zheng, Z., Bao, Y., Qian, L., and Wang, M. Dinoiser: Diffused conditional sequence learning by manipulating noises. arXiv preprint arXiv:2302.10025, 2023

  7. [15]

    Yilmaz, M., Fondrie, W., Bittremieux, W., Oh, S., and Noble, W. S. De novo mass spectrometry peptide sequencing with a transformer model. In International Conference on Machine Learning, pp.\ 25514--25522. PMLR, 2022

  8. [16]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.