acceptodds
Under review as a conference paper at ICLR 2027

dIon: Fragmentation-Based Invariance for Self-Supervised Learning of Tandem Mass Spectra

23 views

Abstract

We introduce a novel invariance for peptide tandem mass spectrometry data, unlocking self-supervised representation learning that improves de novo sequencing of peptides. This invariance exploits the physical relationship between precursor properties (mass and charge) and fragment-ion evidence, without requiring peptide sequence labels. We introduce **dIon**, which adapts the DINO framework with two latent prediction tasks, both recovering a clean teacher representation: one from a spectrum mixture, using the precursor as a selection query, and one from a partial spectrum with the precursor withheld. The first associates precursor information with fragment-ion evidence; the second prevents representational collapse onto that information alone. Mechanistic probes support both effects, and ablations show that the full objective performs best. Under identical end-to-end training, dIon initialization improves de novo peptide precision over training from scratch by 5.5 and 8.4 percentage points on the held-out MassIVE-KB and Kingdoms test sets, and by 2.3 and 4.8 percentage points with a larger supervised training corpus. The resulting models surpass fully supervised state-of-the-art de novo sequencing models on the diverse, multi-species Kingdoms corpus under the same greedy-decoding protocol. Without peptide labels, dIon learns strong native peptide-similarity geometry compared with other learned models; with limited peptide-supervised adaptation, it achieves the best retrieval and pair-discrimination performance across all representation benchmarks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Self-Supervised Learning, Representation Learning, De Novo Sequencing, Mass Spectrometry, Proteomics

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_HSb3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper introduces dIon, a self-supervised framework for representation learning from tandem mass spectra. dIon employs two complementary latent prediction tasks, both aiming to recover a clean teacher representation: a precursor-conditioned mixture task that encourages the model to associate precursor information with relevant fragment-ion evidence, and a precursor-withheld partial-view task that prevents the model from relying solely on precursor information. Experiments demonstrate that dIon improves de novo\textit{de novo} peptide sequencing performance over training from scratch and achieves better results against existing supervised methods.

Critical strengths

Domain-specific self-supervised learning objectives. The paper introduces two complementary latent prediction tasks tailored to the physical properties of tandem mass spectra, providing a well-motivated approach to learning representations without peptide sequence labels.

Critical weaknesses
  1. Comparison with InstaNovo-FM under different pretraining data. dIon and InstaNovo-FM are pretrained on different datasets, making it difficult to determine whether dIon's superior performance stems from the proposed objectives or differences in pretraining data.

  2. Limited comparisons with recent de novo\textit{de novo} sequencing methods. The paper only compares against Casanovo and InstaNovo, omitting other recent methods such as π-HelixNovo, LIPNovo and ReNovo.

Questions for authors
  1. Comparison with InstaNovo-FM's self-supervised objective. Could the authors replace dIon's proposed objective with InstaNovo-FM's masked spectrum reconstruction objective under the same training settings? This would help establish the advantages of the proposed objectives.

  2. Comparison with recent de novo\textit{de novo} sequencing methods. Could the authors provide comparisons with more recent methods mentioned in Weakness 2.

3 comments

Thank you for the review and for recognizing the motivation of the self-supervised objectives. Your feedback has helped us present the contribution and the scope of the comparisons more clearly.

1. Comparison with InstaNovo-FM's self-supervised objective. Could the authors replace dIon's proposed objective with InstaNovo-FM's masked spectrum reconstruction objective under the same training settings? This would help establish the advantages of the proposed objectives.

Thank you for highlighting this distinction. We agree that our comparison with InstaNovo-FM cannot isolate the contribution of the pretraining objective, because both the architecture and pretraining data differ. Our comparison with its released encoder is intended as an external reference for native representation quality, not as a controlled demonstration that our objective is superior to masked reconstruction.

We would also like to highlight that InstaNovo-FM appeared as a concurrent preprint on September 7, 2026, only 11 days before the ICLR abstract deadline. Given this very short window, incorporating its released model into our representation benchmarks was already a substantial additional effort before submission.

A matched comparison would require substantial pretraining and hyperparameter tuning to find suitable settings for InstaNovo-FM's objective within our architecture. Those runs were not feasible within the time and compute available during the review period.

Our appendix already reports an iBOT-style masking experiment, where the student predicts the clean teacher's latent targets at masked peak positions. We added this loss with weight 0.5 to the dual objective. At the last shared evaluation step, 96,000, strict pair AUROC is 0.6323 versus 0.6758 for dual alone, while broad mAP is higher, 0.4556 versus 0.4077. This experiment uses learned teacher targets rather than reconstructing the spectrum inputs, and adds to dual rather than replacing it.

We will also clarify in the Discussion that InstaNovo-FM's reported lack of a measurable de novo gain was obtained under its own dataset and experimental protocol, which differ from ours. It does not establish that masked reconstruction is generally a poorer pretraining objective.

2. Comparison with recent sequencing methods. Could the authors provide comparisons with more recent methods mentioned in Weakness 2.

Thank you for suggesting these additional comparisons. Our main contribution is a self-supervised pretraining method and the demonstration that it improves de novo sequencing over matched training from scratch. To our knowledge, this benefit from self-supervised spectrum pretraining had not previously been clearly demonstrated. Our central comparison therefore holds the architecture, supervised data and training protocol fixed, and varies the encoder initialization. The external supervised models provide context for the resulting performance; an exhaustive comparison of supervised sequencing methods is beyond the scope of this study.

We selected InstaNovo based on its strong performance in a public de novo sequencing benchmark covering 83 datasets. The cross-dataset ranking plot places InstaNovo 1.2.2 first by median rank for peptide precision-coverage AUC, ahead of π-HelixNovo, one of the methods named in the review. We used InstaNovo 1.2.2, the latest available release for our experiments.

We included Casanovo as a well-established method that pioneered transformer-based de novo peptide sequencing. We evaluated its latest available release, 5.2.1, incorporating updates since the original publications.

We did not have time to complete additional comparisons with π-HelixNovo, LIPNovo or ReNovo during the review period.

We hope our explanations of the InstaNovo-FM comparison and the choice of supervised baselines address the questions you raised. Our responses to the other reviewers also include new mechanistic controls, an unseen-peptide analysis and repeated-seed results. If these additions and clarifications increase your confidence in the paper sufficiently, we would appreciate your considering an increase in your rating from 2 to 3.

Thank for the authors' detailed responses. I appreciate the clarifications regarding the baseline selection and the scope of the paper. Given the overall contribution and experimental evidence, I have decided to increase my rating from 2 to 3.

Reviewer_6JVTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper adapts DINO self-distillation to peptide MS/MS data using two types of student views. In the first, the anchor spectrum is merged with a random in-batch spectrum, and the anchor's precursor is supplied; the student must recover the teacher's clean-anchor assignment. In the second, a 60% peak subset is shown with the precursor replaced by a null token. A 127M-parameter encoder is pretrained on 9.2M PXD010000 spectra and fine-tuned as a Casanovo-style de novo model. Under-matched training improves peptide precision over scratch by 5.5/8.4 points (MSKB-final/Kingdoms), and by 2.3/4.8 points with the larger DNL-v1 corpus. Retrieval, pair-discrimination, and probing results are also reported.

Critical strengths
  1. The scratch vs. dIon comparison is controlled (same decoder, schedule, and data, two training regimes), and the gain holds on all 70 Kingdoms species, from +3.2 points (B. subtilis) to +6.6 (Table 25). This is the result I find most convincing.
  2. Figure 2 makes the precursor shortcut visible: the pure-mixture objective gains broad mAP while its strict pair AUROC declines. It justifies the partial views with data rather than argument.
Critical weaknesses
  1. My main concern is that the state-of-the-art claim conflates the benefit of pretraining with differences in training-data coverage, and test overlap is not measured. DNL-v1 includes 863k PXD010000 bacterial spectra (Table 6). In Table 25, dIon leads InstaNovo 1.2.2 by 1.8 points on average over the nine mammals (0.0 on R. norvegicus) but by 3.9 over the 31 bacteria. Scratch trails InstaNovo by 3.7 on mammals and only 0.9 on bacteria. The gain over scratch is similar in both groups, so the matched result stands; the external margin mostly reflects data coverage. On top of that, the Kingdoms test set is run-disjoint, but the peptide overlap with PXD010000 and DNL-v1 is never reported. The abstract qualifies the claim as holding "under the same greedy-decoding protocol" (Table 28), but greedy decoding is not the recommended setting for either baseline, so the comparison does not reflect their reported performance. This matters because surpassing the fully supervised state of the art is a headline claim of the abstract.
  2. The evidence that the model learns the intended precursor-fragment relation is thin. Competing spectra are random in-batch spectra separated by >10 ppm, so presumably typical pairs differ by hundreds of Da, unlike real chimeras inside a 1-2 m/z window. On measured co-isolated spectra, conditioning on the second precursor flips the margin in only 6 of 45 pairs, none closer than 0.2 Da. Frozen dIon scores 0.704 AUROC on chimericity, below the random Transformer's 0.760 (Table 20). This may follow from the mixture objective, which trains invariance to interfering peaks, but it is not discussed. More tellingly, even fine-tuned dIon (0.799) stays below frozen Casanovo 4.0 (0.807) and InstaNovo-FM (0.827), and the Discussion's claim of exceeding every other representation leaves out chimericity without explanation.
Questions for authors
  1. What is de novo precision on Kingdoms test peptides absent from both PXD010000 and DNL-v1, split by mammals and bacteria?
  2. How do Casanovo 5.2.1 and InstaNovo 1.2.2 score under their recommended decoding (beam search; knapsack or refinement for InstaNovo)? Does dIon's lead survive?
  3. Can you give three seeds for scratch vs. dIon on MSKB-final and for Table 10?
  4. Is the low frozen chimericity AUROC a direct consequence of the mixture-invariance objective? If so, what does this imply for the proposed DIA use, where detecting co-fragmentation matters?
  5. The abstract states that dIon learns "strong native peptide-similarity geometry compared with other learned models". In Table 2, GLEAMS beats frozen dIon on all six strict metrics, Casanovo 4.0 on two of six, and the training-free binned spectral angle on all six. dIon is uniformly better than only InstaNovo-FM. Please qualify this sentence.
6 comments

Thank you for the detailed review and for recognizing the strength of the matched scratch comparisons. Your questions led to additional analyses that have strengthened the paper and clarified the scope of our conclusions.

1. What is de novo precision on Kingdoms test peptides absent from both PXD010000 and DNL-v1, split by mammals and bacteria?

The reviewer raises an important point. We considered run and project disjointness in the benchmark design, but did not require the Kingdoms test to be peptide-disjoint from our training corpora. We agree that evaluating peptide-disjoint generalization is a valuable additional check and will include this analysis in the appendix.

This also addresses the overlap and training-data concerns in Critical Weakness 1. We compared Kingdoms test peptide backbones with PXD010000 pretraining and DNL-v1 training, ignoring modifications and treating I/L as equivalent. The overlap covers 9.89% of test spectra with PXD010000, 17.65% with DNL-v1, and 23.05% with at least one of the two. The remaining 3,790,931 unseen test spectra include 361,032 mammal spectra and 1,128,385 bacterial spectra. Peptide precision at 100% coverage (dIon and scratch trained on DNL-v1):

ModelAll speciesMammalsBacteria
dIon0.58840.58760.5925
Scratch0.54100.53840.5433
Casanovo 5.2.10.48170.47200.4883
InstaNovo 1.2.20.55510.56270.5509

The gain over scratch is 4.9 percentage points in both groups. We used the saved predictions and existing matcher, pooling correct predictions and spectra within each group. The verified roster has nine mammal and 32 bacterial entries, including the cyanobacterium Synechocystis_nigrescens. The three archaeal entries are kept separate. The I/L-distinct sensitivity analysis gives the same overall ranking. These exclusions concern our two training corpora; overlap with external models' training data remains unknown.

The abstract already restricts the external-model comparison to greedy decoding. We will clarify in the experiment description and Discussion that external-model rankings compare systems trained on different corpora, while the matched dIon-versus-scratch experiments establish the initialization benefit. We will add the overlap statistics and unseen-subset results to the appendix, and clarify the run-based split in the main text.

2. How do Casanovo 5.2.1 and InstaNovo 1.2.2 score under their recommended decoding (beam search; knapsack or refinement for InstaNovo)? Does dIon's lead survive?

Five-beam decoding took roughly 2.5 times as long as greedy decoding for our models, so we capped the evaluation at 20,000 spectra per species. This gives 1,309,619 held-out Kingdoms spectra across all 70 species, with the same spectra used for every model and both decoding settings. All predictions are scored using the same peptide-matching criteria; missing predictions remain errors in the denominator.

For InstaNovo, we followed the decoding settings in the public de novo benchmark: five beams, without knapsack or diffusion refinement. All four models use five-beam search without these additional stages; model-native preprocessing and confidence scores are retained.

ModelGreedy peptide precision (%)Beam-five peptide precision (%)Greedy PC AUCBeam-five PC AUC
dIon, DNL-v1 trained, 200 peaks59.92562.3090.8656660.872367
Scratch, DNL-v1 trained, 200 peaks55.08757.9060.8356050.845617
Casanovo 5.2.149.19251.1150.7664020.771717
InstaNovo 1.2.256.52262.6460.8504150.858882

With five beams, InstaNovo effectively ties dIon in full-coverage precision, with a 0.337-percentage-point advantage, while dIon retains the higher precision-coverage AUC. InstaNovo gains more from beam search (+6.12 points, versus +2.38 for dIon); we have not established why.

The abstract already limits the external-model comparison to greedy decoding. We chose that simple protocol to keep additional search and postprocessing from obscuring the initialization comparison; the dIon-versus-scratch gain remains with five beams as well. We will add this comparison and its decoding settings to the manuscript. The smaller species cap changes the species weights relative to the full test set, so comparisons should stay within this table.

3. Can you give three seeds for scratch vs. dIon on MSKB-final and for Table 10?

We completed three downstream training runs per initialization, using paired seeds 0, 10 and 20. The pretrained checkpoint is fixed for each dIon/objective variant; decoder initialization and downstream training randomness vary. Scratch also varies encoder initialization. Every selected checkpoint maximizes validation peptide precision.

Canonical MSKB-final held-out test, 200,000 spectra:

InitializationSeed 0 (%)Seed 10 (%)Seed 20 (%)Mean +/- sample SD (%)
Canonical dIon79.69579.48979.60279.596 +/- 0.103
Scratch73.50473.35374.01573.624 +/- 0.347

The mean paired test improvement is 5.972 +/- 0.335 percentage points.

Table 10 development validation, 100-epoch pretrained encoders:

InitializationSeed 0 (%)Seed 10 (%)Seed 20 (%)Mean +/- sample SD (%)
Dual objective78.30578.52477.84578.225 +/- 0.346
Pure mixture77.61677.66677.04777.443 +/- 0.344
Scratch74.00274.66174.42174.361 +/- 0.334

Dual exceeds pure mixture in every paired seed, with a mean gap of 0.782 +/- 0.086 percentage points. The original fine-tuned de novo scores were close, with a small advantage for dual. Our choice of SSL objective was driven mainly by the clearer differences in strict pair AUROC (0.684 vs. 0.587), end-AA linear accuracy (90.11% vs. 62.33%) and frozen-encoder de novo validation precision (0.675 vs. 0.638).

SD is the sample SD across three downstream seeds (ddof=1). All repeats used the same validation split and evaluation settings. These experiments assess downstream stability conditional on each pretrained checkpoint. Variation across independent SSL pretraining runs and a 300-epoch objective-ranking comparison remain unmeasured.

Reviewer_25U1OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces dIon, a self-supervised representation learning method for peptide tandem mass spectra. Built on the DINO teacher–student self-distillation framework, it combines a precursor-conditioned mixture task with a precursor-withheld partial-view task to learn fragment representations associated with the target peptide from interfering and incomplete observations, while reducing reliance on precursor metadata. After pretraining on approximately 9.23 million spectra, the model is transferred to tasks including de novo peptide sequencing, spectrum retrieval, and same-peptide discrimination. Under identical supervised training settings, dIon initialization improves full-peptide prediction accuracy by 2.3–8.4 percentage points over training from scratch, demonstrating the transfer value of the proposed self-supervised pretraining approach.

Critical strengths
  1. The self-supervised tasks are well suited to the characteristics of mass spectrometry data.
    Rather than directly adopting image cropping, the paper constructs views using the relationship between precursor information and fragment-ion evidence. The mixture task requires conditional source selection, while the partial-view task limits reliance on explicit precursor information. Both have clear motivations, and this approach to view construction offers a degree of novelty.

  2. The matched training-from-scratch comparisons provide the strongest evidence in the paper.
    The authors compare different initializations using the same backbone, decoder, and supervised training protocol, obtaining consistent improvements across two supervised training corpora and two test sets. This design demonstrates the practical value of self-supervised pretraining more convincingly than comparisons between different publicly available models alone. Improvements persist with the larger DNL-v1 training corpus, indicating that the benefits are not confined to settings with limited supervised data.

  3. The evaluation explicitly controls for precursor-based shortcuts.
    The paper recognizes that broad retrieval can achieve high scores using mass and charge alone, and therefore evaluates strict retrieval and pair discrimination among spectra with the same charge and similar precursor masses. This places greater emphasis on fragmentation patterns rather than simple metadata matching. Including the training-free binned spectral angle baseline also helps assess the practical value of learned representations objectively.

  4. The evaluation is broad and supported by extensive appendices.
    Beyond sequencing, the paper evaluates frozen representations, metric learning with limited labels, auxiliary tasks, objective ablations, and real mixture spectra. The Kingdoms sequencing evaluation covers 70 species and includes per-species results. Data provenance, splits, peak preprocessing, decoding, and computational costs are also described in considerable detail.

  5. The authors are transparent about several negative results and limitations.
    They explicitly report that frozen dIon representations underperform some simple methods, that performance on auxiliary tasks is not uniformly superior, and that the real DIA evaluation relies on oracle precursor information. These disclosures help prevent the results from being interpreted as demonstrating universal superiority across mass spectrometry tasks.

Critical weaknesses
  1. The current mechanistic probes do not adequately rule out precursor-information shortcuts.

    The source-selection probe in Table 3 and Appendix C.2 compares the mixture representation conditioned on precursor A with a clean reference representation that also receives precursor A. Under this evaluation, a representation primarily encoding precursor mass and charge could still achieve high source-selection accuracy. Near-perfect performance therefore does not, by itself, establish that the model identifies the corresponding fragment-ion evidence.

    The partial-view probe shows that the model can match sources without explicit precursor input. However, the partial view shares many peaks with the clean reference spectrum, leaving simple peak-overlap matching or inference of precursor properties from the spectrum as possible explanations. Metadata-only controls, simple spectral-similarity baselines, and more stringent counterfactual controls are needed.

  2. The objective ablations change multiple factors simultaneously and do not cleanly isolate the contribution of the mixture task.

    The mixture-free control in Appendix C.1 does more than remove the interfering spectrum: it also reverses the allocation of precursor information, withholding the precursor from the 95% global views while retaining it in the 60% partial views. Both mixing and precursor conditioning therefore differ from the full model. Consequently, the full model’s advantage cannot be attributed specifically to the mixture task. A more direct control would preserve the peak-retention fractions, precursor allocation, and loss structure while removing only the competing spectrum. Moreover, the end-to-end validation peptide precisions of the full and pure-mixture objectives in Table 10 are 0.782 and 0.778, respectively—a difference of only 0.4 percentage points. Without repeated runs, the stability of this difference is difficult to assess.

  3. The value of the “general-purpose foundation representation” is demonstrated primarily after adaptation rather than in the frozen setting. In Table 2, frozen dIon underperforms binned spectral angle on strict retrieval and pair discrimination across all three datasets. It also substantially trails other frozen encoders on some auxiliary tasks. For example, Table 20 reports a chimericity AUROC of 0.704 for frozen dIon, compared with 0.827 for InstaNovo-FM. This does not diminish dIon’s value as a pretrained initialization, but claims of strong native geometry and broad generality require qualification. The paper should distinguish more clearly among:

    • Frozen representations usable directly without labels;
    • Representations adapted using peptide supervision;
    • Task performance obtained through large-scale end-to-end fine-tuning.
Questions for authors
  1. How do the mechanistic probes rule out precursor-based shortcuts?
    Could the source-selection probe in Table 3 still achieve near-perfect accuracy if the representation were solely a function of precursor mass and charge? Please report metadata-only and simple spectral-similarity controls, and consider withholding precursor information from the clean reference spectra to test whether the mixture representation actually selects the corresponding fragment-ion evidence.

  2. Can you provide an ablation that changes only the mixing operation?
    Please preserve the full model’s peak-retention fractions, precursor allocation, and loss structure while removing only the competing spectrum. Why does the current mixture-free control also reverse precursor conditioning across the two view types?

  3. Are the gains from the dual objective stable?
    The end-to-end difference between the full and pure-mixture objectives in Table 10 is 0.4 percentage points. Does this difference persist across fine-tuning seeds? Does the ranking obtained from the 100-epoch ablations remain representative at the full training horizon?

  4. Can different pretraining objectives be compared under matched conditions?
    Please provide baselines using standard DINO-style views, masked-peak reconstruction, or masked latent prediction with the same backbone, pretraining corpus, and comparable computational budgets. If such comparisons are not feasible, should the claim that reconstruction objectives are unsuitable for this modality be qualified?

  5. How can “physical invariance” be distinguished from generic denoising or metadata inference?
    Does source selection remain effective when the interfering and target spectra have similar precursor masses, their relative intensities vary, or the competing spectrum’s peak count is no longer capped by that of the target? Could controls that disrupt specific mass relationships while preserving the overall spectral distribution identify which physical information the model uses?

  6. What explains the relationship between frozen representations and fine-tuning gains?
    Why does frozen dIon underperform binned spectral angle on strict retrieval and perform relatively poorly on some auxiliary tasks, yet provide a strong initialization for end-to-end training? Which experiments could distinguish directly transferable task information from benefits arising primarily through improved optimization?

  7. How should the real-mixture experiments be interpreted?
    Please explain the performance gap between synthetic probes and real co-isolated spectra. Also clarify whether the requirement in Appendix C.2 that each peptide have at least two reference spectra is consistent with the statement that some successful cases rely on a single reference spectrum. For the DIA pilot, please explicitly limit the conclusion to conditional decoding given the correct precursor.

7 comments

Thank you for the careful and detailed review. Your constructive feedback has made the paper stronger and led to more informative experiments and clearer interpretations.

1. How do the mechanistic probes rule out precursor-based shortcuts?

We agree with the reviewer that these probes do not rule out a precursor shortcut. This was not their intended purpose, and we will update the manuscript to make their purpose and limitations clearer. The source-selection probe was intended to check whether changing the queried precursor shifts the mixture embedding toward the corresponding clean source. We wanted to check that the model had learned this behavior, beyond seeing its training loss decrease. Since the queried precursor matches the correct clean reference's metadata by construction, metadata alone can solve this task.

To address this concern more directly, we have added a control where the mixture still receives the queried precursor, while both clean references receive no precursor input (denoted NULL below). dIon achieves 88.03% mean source accuracy in this setting. We will include this control and its result in the revised manuscript. This removes the direct match between the supplied precursor metadata of the mixture and targets.

Both comparisons below use the same 4,488 directed source pairs from the NineSpecies development-validation set. We report mean accuracy, averaging the two precursor-query or partial-view decisions per pair.

Representation / reference inputsMixture source selection (%)
dIon, precursor-conditioned references92.12
dIon, NULL-precursor references88.03
Binned spectral angle50.00
Supplied precursor metadata only100.00

We also found a reproducibility issue affecting only the reported cosine-similarity mixture probe. We will update the manuscript with the verified rerun values above; the partial-view result is unchanged.

For the partial-view probe, our new binned spectral-angle baseline achieves 99.99% mean accuracy, compared with 98.55% for dIon. We agree that shared peaks make this task easy. Since DINO loss is hard to interpret, we wanted a direct check that the model had learned to match partial views to their parent spectra. Outperforming spectral angle on this task was not the aim. We will add the baseline and clarify this purpose in the manuscript.

Representation / reference inputsPartial-view source matching (%)
dIon, precursor-conditioned references98.55
dIon, NULL-precursor references99.97
Binned spectral angle99.99
Supplied precursor metadata onlyAll decisions tie

See also our answer to Question 5: fragment-mass scrambling reduces mixture source-selection accuracy to chance while retaining peak overlap with the targets, supporting the use of fragment-mass patterns beyond overlap alone.

2. Can you provide an ablation that changes only the mixing operation?

We agree that the original mixture-free baseline changed several factors and did not cleanly isolate the contribution of the mixture task. We have now trained a simpler mixture-free control for 100 epochs, removing the mixture branch entirely and keeping only partial-to-clean prediction:

student(60% A, NULL) -> teacher(95% A, NULL)

Neither network receives precursor information during pretraining. We had time to complete two end-to-end downstream seeds. MSKB-final validation peptide precision at 100% coverage is:

InitializationSeed 0 (%)Seed 10 (%)
Dual objective, 100-epoch encoder78.30578.524
Pure mixture, 100-epoch encoder77.61677.666
Fragment-only, 100-epoch encoder76.92776.587
Scratch74.00274.661

All four use the same end-to-end training/finetuning protocol. Fragment-only exceeds matched scratch and falls below dual and pure-mixture initialization in both seeds. This control removes both mixtures and precursor conditioning; a comparison retaining both original view families and changing only the competing peaks remains unrun.

The original mixture-free objective was a genuine candidate alternative. We thought it would be at a disadvantage without the implicit task of inferring precursor properties from fragments, so we kept both precursor-conditioned and precursor-withheld student views. We agree with the reviewer that a more natural ablation was possible, and will include both in the revised manuscript.

The online validation probes support the same distinction between broad retrieval and strict discrimination seen in Figure 2. The last logged values from the 100-epoch runs, averaged over the same compact validation callback subsets, are:

Pretraining objectiveStrict pair AUROCBroad retrieval mAP
Dual objective0.68370.3191
Pure mixture0.58730.4286
Original mixture-free control0.64090.0967
Fragment-only control0.62900.0876

Fragment-only exceeds pure mixture on strict pairs despite much lower broad retrieval mAP, while dual remains strongest on strict pairs. The new control uses NULL precursor inputs during these probes, whereas the original curves use precursor-conditioned inputs. We will update Figure 2 with the fragment-only run's online validation curves, aligned by pretraining epoch and with this input difference clearly labelled.

These last logged values are not a fully matched comparison: the latest probes were logged at 100,000 optimizer steps for dual and 108,000 for the other three objectives. We will show the full epoch-aligned curves in the updated report. For now, the numbers support the same general picture, with dual performing best on strict discrimination and downstream transfer.

3. Are the gains from the dual objective stable?

Yes, the dual-versus-pure-mixture gain persists across three paired downstream seeds. The original fine-tuned de novo scores were close, with a small advantage for dual. Our choice of objective was driven mainly by the clearer differences in strict pair AUROC (0.684 vs. 0.587), end-AA linear accuracy (90.11% vs. 62.33%) and frozen-encoder de novo validation precision (0.675 vs. 0.638). We ran the additional seeds at the reviewer's request to check the stability of the smaller fine-tuning difference. MSKB-final validation peptide precision is:

InitializationSeed 0 (%)Seed 10 (%)Seed 20 (%)Mean +/- sample SD (%)
Dual objective, 100-epoch encoder78.30578.52477.84578.225 +/- 0.346
Pure mixture, 100-epoch encoder77.61677.66677.04777.443 +/- 0.344

Dual wins in all three paired seeds, averaging 0.782 +/- 0.086 percentage points higher. Both remain well above matched scratch (74.361% mean). We held each SSL checkpoint fixed and varied downstream training with paired decoder/data seeds, fine-tuning the encoder and a fresh decoder end-to-end. We will add these validation results to the appendix.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.