acceptodds
Under review as a conference paper at ICLR 2027

PeakCloud: Benchmarking and Learning on Peak-Level Mass Spectra for Metabolomic Pattern Recognition

Abstract

Mass spectra acquired from non-invasive biospecimens such as urine, breath, and feces enable untargeted metabolomics to capture comprehensive molecular profiles for disease characterization. While machine learning can extract informative patterns from these spectra, current studies are largely confined to single cohorts and suffer from disparate data processing pipelines and evaluation protocols, limiting systematic assessments across multiple cohorts and data formats. To address these challenges, we establish PeakCloud, a benchmark covering nine classification tasks across three public datasets and an clinical cohort of 700 subjects. Crucially, this clinical cohort contains paired measurements across two instruments, enabling controlled evaluation of hardware shifts alongside cross-batch settings under unified protocols. Within this benchmark, although CNN-based deep learning models are competitive on several tasks, their reliance on dense spectral grids limits biological explainability. We therefore introduce PeakCloudNet, a set-based neural network that processes raw spectra while preserving individual peak identities for attribution. PeakCloudNet achieves competitive performance against grid-based models, accurately pinpoints disease-associated peaks, and retrieves candidate formulas aligned with known metabolic pathways. We hope that this benchmark and architecture will facilitate generalizable and interpretable pattern recognition in metabolomics. Upon publication, we will release the processed peak-level data for all datasets in PeakCloud, together with the PeakCloudNet implementation, to support reproducible research in this area.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
mass spectrometry, metabolomics, deep learning, spectrum learning, benchmarking, dataset

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_rNdMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces PeakCloud, a benchmark for sample-level disease classification from untargeted MS1 mass spectra, and PeakCloudNet, a set-based neural network for learning directly from peak sets.

The benchmark unifies four cohorts in a shared peak-level format under fixed, subject-disjoint splits: three public LCMS cohorts (MTBLS17 serum HCC vs. cirrhosis, n=262; MTBLS28 urine lung cancer vs. control, n=1004; ST001000 feces IBD vs. control, n=220) and a new breath PTR-TOF-MS cohort (MSP001, n=700 TB vs. healthy controls) in which every subject was measured concurrently on two instruments, with longitudinal acquisition batches and paired ambient-air references used to remove environmental confounds. Evaluation covers in-distribution performance plus three OOD regimes: cross-instrument transfer (both directions, on 140 paired held-out subjects), temporal batch shift, and external-cohort transfer.

PeakCloudNet encodes each peak with a resolution-aware damped-Fourier m/z embedding and a centered log-ratio intensity embedding, fuses them per peak, and pools with a Set Transformer. Peak-CAM attribution plus a frozen pretrained molecular-formula retriever (Appendix A.9) are used for explainability, with candidate formulas matched against previously reported disease-associated metabolites (e.g., 3-heptanone and methylnaphthalenes for TB breath; creatine riboside / NANA / bile acids on the LCMS cohorts).

Validation: 8 baselines (RF, SVM, PLS-DA, PCA+LR, MLP, DeepMSProfiler, LCMS-Net) under identical splits; AUC/BA with 2000-resample bootstrap CIs over 3 seeds. Headline results: PeakCloudNet is best in-distribution on both breath instruments (0.97 AUC) and transfers best across instruments in both directions (losing only 0.02-0.09 AUC, vs. up to 0.21 for RF and a collapse to chance for DeepMSProfiler in one direction); on batch/cohort shifts it is competitive with RF; ablations attribute most of the gain to the m/z encoder.

Critical strengths
  1. The paired cross-instrument evaluation protocol is a real methodological contribution. Measuring the same 140 subjects on two instruments isolates hardware drift from cohort composition — something no public benchmark allows (the closest precedent, Metabo-Ring, is a QC study without ML evaluation). This design is what makes the paper's central finding — that ID accuracy does not predict cross-instrument robustness, and that a grid-based CNN can collapse to chance (0.50 AUC) in one transfer direction — credible and actionable for multi-site deployment. It is the main reason this paper matters beyond another benchmark table.
  2. A novel, well-motivated architecture with honest reporting of its own weaknesses. Set-based learning for whole untargeted MS1 profiles (vs. spectrum-level or feature-detection uses of point-cloud methods in prior work) is, to my knowledge, new. tThe resolution-damped Fourier m/z encoding and CLR intensity encoding are physically motivated rather than ad hoc. The authors also report awkward results candidly (RF competitive on batch/cohort shifts; the ablation inversion), which increases confidence in the rest.
Critical weaknesses
  1. Nothing is accessible at review time, and one key component has no identifiable provenance. The paper contains zero URLs: no code link, no data link, no anonymous repository. For a benchmark paper — whose value is reproducibility and reuse by others — all artifacts are promised "upon publication," with MSP001 additionally contingent on institutional authorization. This also makes the headline dataset claim ("first public breath PTR-TOF-MS cohort with paired cross-instrument measurements") currently unverifiable. The molecular-formula retriever used for the formula-retrieval explainability results is described only as "a pretrained spectral matching model" from an unnamed, uncited "release," with results attributed to a checkpoint whose identity and evaluation sample count are unspecified. A reviewer cannot verify any part of this component, yet it underwrites the abstract's claim that the method "retrieves candidate formulas aligned with known metabolic pathways."
  2. The evidence does not statistically establish several headline claims, and the abstract overstates them. Multiple key comparisons have overlapping 95% bootstrap CIs with no significance tests or seed-level analysis (MSP001-1 ID vs RF; both cross-batch comparisons; ST001000, where PeakCloudNet is actually below SVM and RF). On MTBLS28, LCMS-Net (0.80 +/- 0.06) beats PeakCloudNet (0.75 +/- 0.06) with non-overlapping CIs, so the abstract's "competitive performance against grid-based models" is generous. The abstract also cites "nine classification tasks," which is never defined or enumerated in the paper, and there is no limitations section discussing that all biological interpretations are putative mass matches with no wet-lab or prospective validation.
Questions for authors
  1. What is the molecular-formula retriever in Appendix A.9 — its name, citation or link, and training provenance? Can you report its top-1 / top-5 accuracy measured on the attribution-selected peaks actually used in Section 5.4, rather than the "single-peak reference" result from the release?
  2. How do you count "nine classification tasks" (abstract)? Please enumerate them, since Table 1 lists four datasets with ID/OOD settings.
  3. For the comparisons where bootstrap CIs overlap (MSP001-1 ID vs RF; cross-batch on both instruments; ST001000), can you provide seed-level results or a significance test? Is the cross-instrument advantage of PeakCloudNet over RF statistically distinguishable in either direction?
  4. Why does removing both encoders (0.65) hurt more than removing only the m/z encoder (0.72) on MSP001-2, while the opposite ordering holds on MSP001-1? What does this imply about which component drives cross-instrument robustness?
  5. Will an anonymous code and data link (including processed peak-level MSP001) be available during the review period, or only after acceptance? If release is contingent on institutional authorization, what is the expected timeline? In my opinion, full evidence of reproducibility at the review time is an absolute must for venues like ICLR.
  6. Are you planning to add wet-lab or targeted confirmation of the top attributed peaks (e.g., 3-heptanone, methylnaphthalenes)? That would convert the explainability story from putative mass matching to validated evidence.
Reviewer_fUEnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces PeakCloud, a benchmark for predicting disease from MS peak data. It combines three public datasets with a new breath dataset of 700 people (TB patients and healthy controls), each measured on two instruments. It also proposed PeakCloudNet, a neural network that treats the spectrum as a set of peaks. PeakCloudNet matches or beats other models on breath data and performs best on second instrument, and it's competitive but not consistently the best on public dataset.

Critical strengths

Measuring the same people on two instruments is a good experimental design, because it rules out the effect of changing the instrument. A fair benchmarking with the same splits. The model encodes m/z and intensities is physically motivated, and the ablation shows these encoding works well.

Critical weaknesses

TB patients were tested at a hospital and the healthy controls from health check center. The room-air filter only partly addresses this; the authors should check, e.g., whether the model can tell the groups apart using only the room-air samples. So, there might leave a trace that the model can identify the testing facility instead of the chemical related to TB.

The claim that the model's design makes it robust to instrument changes is not tested. The ablations only tested on the original instrument. Also, SVM loses less AUC than PeakCloudNet when switching instruments.

The breath dataset is the main contribution, but the release depends on hospital approves.

Questions for authors

Can the model distinguish the case/control groups using only the room-air samples? The ablation study should be performed on data from both instruments.

Reviewer_2E8dOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses an important topic in mass spec metabolomics research. It introduces the PeakCloud benchmark for disease classification based on untargeted metabiolomics mass spectra. It focuses on nine tasks and the evaluation is based on three publicly available LC-MS data sets together with a new clinical cohort. A methodological advance of the paper is the PeakCloudNet tool, which is a set based neural network that processes unordered MS1 peaks, and avoids the traditional route of peak alignment and spectral binning. The experiments illustrate competitive performance in distribution and improved robustness to distribution shift and cross instrument shifts, compared to competing methodologies.

Critical strengths
  • Despite extensive prior work on benchmarking competing methods for processing metabolomics MS1 spectra, there is still a need for such benchmarks. In particular, having paired cross-instrument measurements from the same subjects is particularly valuable, since it enables a more controlled assessment of platform induced distribution shifts.

  • The design choices of resolution aware Fourier embeddings, centered log-ratio transformation are well motivated by the characteristics of MS1 spectral data.

  • The strong cross-instrument generalization results achieved by PeakCloudNet is a substantial strength. Further, it preserves individual peak identities, which can prove useful for biological interpretation for downstream analysis tasks.

Critical weaknesses
  • The proposed methodology primarily combines established components. Although this adaptation is appropriate for MS1 spectral data, it is unclear which of these components constitute a substantial methodological advance.

To that end, both the benchmark and experimental validation are informative for the computational mass spec community and metabolomics communities. Hence, the work is better suited for a more focused venue specializing in bioinformatics, computational proteomics/metabolomics.

  • The proposed method shows strengths on selective data sets, bit not across all data sets under consideration. Further, the peak removal experiments identify features that influence model predictions; however, they do not demonstrate biological relevance, as meaningful disease biomarkers.

  • The authors should take a look at the 2025 Nature Communications paper by Fiehn's group and do a direct comparison.

Questions for authors
  • How does PeakCloudNet compare with both deep learning (e.g. grid based CNNs) and signal processing methods in terms of computational cost, memory usage, and inference time? How does the computational complexity scale with the number of peaks per spectrum?

  • Which of the improved performance results showed in the paper are statistically significant?

  • On the issue of the peak removal experiments: which ones are truly biologically meaningful, given that for some of them there is large discrepancy with molecular formulas?

  • Since tuberculosis patients and healthy controls were recruited from different clinical facilities, have the authors done any analysis that controls for demographic of clinical covariates? Otherwise, it is fair to argue that the method does not learn disease related patterns, but population specific or clinic specific patterns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.