acceptodds
Under review as a conference paper at ICLR 2027

Persistent Homology with Vineyards for Robust Spectral Representation

Abstract

Automated molecular structure elucidation from experimental spectra is fundamentally challenged by noise, baseline distortions, overlapping signals, and modality-specific artifacts. Existing neural encoders typically represent spectra either as dense intensity grids, entangling experimental artifacts with chemical information, or as sparse peak lists that depend on heuristic detection thresholds. We introduce VINEA, a modality-general framework that represents spectra through their multiscale topological structure, capturing persistent spectral features while suppressing transient noise and preserving chemically meaningful relationships. VINEA provides a unified representation across 1D NMR, HSQC, IR, and Raman spectra. With approximately 1M parameters, VINEA provides the strongest overall robustness on H NMR and HSQC retrieval, leads Raman robustness when trained on clean spectra, and leads IR functional-group prediction under corruption-aware training. After fine-tuning on experimental spectra, VINEA matches or exceeds the strongest specialized architectures on NMR, HSQC, and Raman while remaining competitive on IR. These results establish spectral topology as a compact, robust, and modality-agnostic representation for automated structure elucidation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
NMR representation, TDA, IR, Raman, Spectroscopy

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_P3vKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper proposes VINEA, a spectral representation for structure elucidation that replaces dense intensity grids or thresholded peak lists by a set of H0 persistent branches. For 1D spectra (¹H NMR, Raman, IR) each local maximum of the signed, shifted signal becomes a branch with birth, death and persistence (equal to topographic prominence in 1D), decorated with position, area, widths at three relative heights, asymmetry and neighbour spacings; for HSQC, connected cross-peak regions play the same role. Each branch is followed across a Gaussian scale-space ladder (a persistence vineyard) to obtain a relative-persistence profile and a scale-death statistic. Up to 128 branches with the highest persistence are encoded by a two-block set transformer with Fourier position features, persistence-weighted embeddings and a translation-invariant relative-position attention bias (about 0.6M parameters). The paper also introduces a synthetic corruption benchmark with five severities for four modalities and compares VINEA with 15 retrained published encoders over ten seeds, in clean-trained and corruption-aware (S1–S3 ramp) regimes, plus zero-shot and fine-tuned transfer to measured spectra, field-strength transfer on GISSMO, and architecture ablations.

Critical strengths
  1. A broad, carefully controlled multi-modality robustness benchmark. Four modalities, five calibrated severity levels, 15 retrained published encoders plus internal controls, ten seeds, and corruptions that are byte-identical across models for each item (Appendix H.1) make the comparisons paired and reproducible. Under corruption-aware training, VINEA has clear and statistically convincing margins on several tasks: ¹H NMR S3 Recall@1 0.583 ± 0.022 vs 0.446 for NMRViT, HSQC 0.109 vs 0.023, and IR S5 macro-AUROC 0.780 vs 0.712 (Table 12), with an order of magnitude fewer parameters than most baselines. The benchmark itself is a useful resource.

  2. Thorough and fairly candid reporting. The appendix reports results that do not favour the method — zero-shot transfer to measured spectra (Tables 13–14), single-artifact retention (Table 10) and latency (Table 15) — and includes an informative negative control: shuffling the association between branch position and the other descriptors reduces MRR-AUDC from 0.533 to 0.126 (Table 16), showing that the model relies on which morphology occurs at which chemical shift.

Critical weaknesses

1. The central claim that the topological representation is robust to real acquisition artifacts is not supported outside the authors' own synthetic corruptions.

  • Zero-shot transfer to measured ¹H NMR (Table 13, ramp pretraining). VINEA obtains 0.068 / 0.047 / 0.201 / 0.221 Recall@1 on Chemotion / BMRB / GISSMO / NMRXiv, versus 0.173 / 0.106 / 0.278 / 0.327 for NMR2Struct and 0.144 / 0.125 / 0.283 / 0.303 for NMRViT; it beats only SECS. With clean pretraining it is also below NMR2Struct (e.g., 0.021 vs 0.037 on Chemotion). On measured HSQC (Table 14) VINEA-HSQC is near chance (0.020 vs 0.010 random), while SPECTRE reaches 0.097; on measured IR it is below ACT and UltraIR (0.684 vs 0.692). Only Raman favours VINEA (0.095 vs 0.090, s.d. 0.031). This contradicts the contribution "compact robustness that transfers beyond simulation" (l. 103–107), and the abstract mentions only fine-tuned results.
  • The representation alone does not suppress noise. In Table 10, clean-trained VINEA retains only 1.6% (¹H) and 1.2% (HSQC) of its clean performance under thermal noise at S3, 6.1% under ¹H shimming and 5.6% under Raman shot noise, despite the abstract's claim of suppressing transient noise (l. 017–018). Robustness appears mainly after corruption-aware training, which improves all architectures (l. 424). The clean-trained "7–16×" advantage on ¹H (l. 104–105, 390–391) is a ratio between near-chance values (chance 5.4×10⁻⁵).
  • Co-designed benchmark, unprocessed baselines. The signed filtration, the persistence normalization and the translation-invariant attention target the phase, baseline and referencing corruptions of the same benchmark, whereas dense baselines receive corrupted inputs before any model-specific preprocessing (l. 373–374) and without standard baseline correction, phasing or denoising.
  • Fine-tuning claims lack statistics. The differences in §5.2 (0.382 vs 0.375 on ¹H, 0.336 vs 0.332 on HSQC) are shown only in Fig. 5a. l. 442 refers to Tables 13 and 14 for the fine-tuning result, but those tables contain only zero-shot numbers.

2. The experiments do not isolate what persistent homology and vineyards contribute, and the ablation study is incomplete and inconsistent with the main results.

  • No peak-list baseline. The paper frames its contribution against dense grids and peak lists (l. 071–078) and discusses NMRTrans / NMR-Text (l. 118–124), yet all ¹H baselines are dense or patch encoders (Table 8); NMRTrans, NMR-Text and UltraNMR appear only in the learning-rate list (l. 2705–2707). Since 1D persistence equals topographic prominence (l. 161–162) and VINEA keeps the top-128 branches by persistence (Algorithm 1), the decisive control is the same set encoder on peaks from a standard prominence/width peak picker (optionally at several smoothing scales) with the same geometric descriptors. Without it, the gains cannot be attributed to the topological construction rather than to set encoding with relative-position attention.
  • Interim three-seed ablations from a different protocol. Table 16 uses three seeds and is described as an "interim aggregate" to be treated as a trend until the remaining seeds are complete (l. 2594, 2611–2612, 2696–2698). In it, full VINEA has clean Recall@1 0.276 but S3 Recall@1 0.619 — better under corruption than clean — whereas Table 12 reports 0.778 (clean) and 0.583 (S3) for ramp-trained VINEA on the same pool, and Table 11 reports 0.859 for clean-trained VINEA. The ablation uses a different curriculum and checkpoint rule (l. 2416–2418), so its conclusions do not transfer to the reported model. The vineyard ablation ("− blur axis") changes MRR-AUDC by only −0.020 with standard deviations of about 0.015–0.018, so the title component is not shown to matter for robustness.
  • Inconsistencies that affect fairness. Appendix G states that batch size (the InfoNCE negative pool) is 32 for every encoder and dataset (l. 2700–2701) but then lists VINEA batch sizes of 32/4/64/64 for ¹H/HSQC/Raman/IR (l. 2714–2716). Table 11's caption describes the Raman task as 5,056-candidate retrieval, Table 12's caption says Table 11 reports a pharmaceutical classification task, and Appendix E.10 says balanced accuracy is reported for Raman. The main text gives 0.588 for ¹H S3 (l. 425) while Table 12 gives 0.583. SPECTRE appears in the HSQC tables (0.381 clean, stronger than DeepSAT's 0.326) but is not among the described baselines, and l. 393–394 compares VINEA only with DeepSAT. Leftover text ("unequal n ... at the freeze date" although all n = 10, l. 2534–2535; "historical result key topo scalespace", l. 2355) suggests the appendix was not finalised. Finally, VINEA-HSQC needs 455 ms per spectrum versus 0.8–6 ms for the baselines (Table 15), so "compact" holds for parameters only.
Questions for authors

Clarifications that could change my assessment

  1. Please add the same set encoder trained on conventional peak lists (prominence/width-based peak picking, optionally at several smoothing scales, with the same geometric descriptors), and an NMRTrans-style peak-set baseline, under the same corruption protocol and seeds.
  2. Why does full VINEA in Table 16 reach only 0.276 clean Recall@1 while scoring 0.619 at S3, compared with 0.778/0.583 in Table 12? Please re-run the ablations under the main training protocol with all ten seeds.
  3. How do the dense baselines perform when given standard preprocessing (baseline correction, automatic phasing, smoothing) before encoding, both in the clean-trained setting and on measured spectra?
  4. Please provide a table of the fine-tuning results in Fig. 5a with standard deviations, candidate-pool sizes and paired tests.
  5. Which batch size (number of InfoNCE negatives) did each model use on each modality, and what hyper-parameter search budget did baselines receive compared with VINEA?
  6. What does Table 11 report for Raman (retrieval R@1 or classification BAcc)? What is SPECTRE, how was it trained, and why is it not in Table 8?

Suggestions (not required for my assessment)

  1. Analyse which real-data factors make VINEA worse than dense encoders in zero-shot ¹H transfer (e.g., solvent or impurity peaks, multiplet structure, the M = 128 token cap under noise), and report performance as a function of M under thermal noise.
  2. Report end-to-end throughput alongside parameter counts, and qualify "compact" accordingly.
11 comments
AuthorsOpenReview ↗

Response to Reviewer P3vK (Part 1)

We thank Reviewer P3vK for the detailed assessment and for recognizing the benchmark’s breadth, reproducibility, and reporting of negative results. The review identifies several places where our claims exceed the evidence or where the experimental description is inconsistent. We address these points below.

Please add the same set encoder trained on conventional peak lists (prominence/width-based peak picking, optionally at several smoothing scales, with the same geometric descriptors), and an NMRTrans-style peak-set baseline, under the same corruption protocol and seeds.

We thank Reviewer P3vK for this suggestion. We evaluated a conventional peak-set control and NMRTrans and UltraNMR retrieval adaptations on the same 1^1H NMR benchmark, using the same corruption conditions and ten matched training seeds. The conventional control uses peak positions, prominence, widths, areas and neighboring-peak gaps with the VINEA Set Transformer design and the same training budget. It removes the lifetime and blur-profile descriptors while retaining multiscale extraction. NMRTrans receives comparable peak geometry measured directly from the spectrum through its own encoder. The adaptation uses automatically detected peaks from scipy.signal.find_peaks and was trained from scratch without externally supplied chemical annotations, with a learning rate of 310-5 selected through learning-rate screening.

Peak-based retrieval comparisons. Spectrum-to-molecule R@1 over 18,493 candidates, reported as mean sample SD across ten seeds.

ModelClean R@1 (%)Combined S3 R@1 (%)
VINEA77.93$\pm$0.5658.76$\pm$2.05
Conventional peaks + VINEA encoder76.01$\pm$0.4755.73$\pm$2.32
NMRTrans, retrieval adaptation77.18$\pm$0.5750.21$\pm$3.34
UltraNMR, 1^1H-only adaptation47.22$\pm$1.005.30$\pm$0.79

VINEA achieves the highest mean R@1 in both conditions, exceeding the conventional-peak control by 1.93 percentage points on clean spectra and 3.03 points at S3, calculated before rounding. NMRTrans approaches VINEA’s clean performance, but its gap increases to approximately 8.55 points under combined S3 corruption. These results support the robustness of the evaluated VINEA configuration relative to the tested peak-based alternatives.

UltraNMR is included as a supplementary comparison. It receives only automatically detected 1^1H chemical shifts, retains its Fourier embedding, and was trained from scratch without the 13^{13}C inputs or pretrained weights of its published setup. Its results therefore describe a restricted adaptation.

Why does full VINEA in Table 16 reach only 0.276 clean Recall@1 while scoring 0.619 at S3, compared with 0.778/0.583 in Table 12? Please re-run the ablations under the main training protocol with all ten seeds.

Thank you for identifying this inconsistency. Table 16 reported a three-seed ablation study using a different training protocol from the main benchmark. The main benchmark used 20 clean-training epochs followed by an 80-epoch S1-S3 corruption ramp. The original ablation instead began the ramp immediately, reached S3 after the first 70% of training, and selected checkpoints by clean validation MRR only from epoch 75 onward. Its full-model reference was therefore not directly comparable to the main-benchmark model.

Separately, Table 12 contained a stale seven-seed aggregate despite being labelled as ten seeds. This was caused by a bug in our aggregation pipeline. We apologize for this reporting error. Following this, we subsequently verified all numerical tables against the retained per-seed results and controlled the fine-tuning. Recomputing the complete ten-seed gives the following results:

ExperimentSeeds representedClean R@1S3 R@1
Original Table 16, separate ablation protocol30.276$\pm$0.0190.619$\pm$0.030
Original Table 12, stale aggregate70.778$\pm$0.0060.583$\pm$0.022
Finalized main benchmark100.779$\pm$0.0060.588$\pm$0.021

We retrained all NMR ablation variants with ten matched seeds using the main training schedule and compared them against the existing full VINEA checkpoints from Figure 4, and we will provide attached to the extended tables of ablations for our method. The no-blur control achieves 0.760$\pm$0.005 clean R@1 and 0.557$\pm$0.023 S3 R@1.

AuthorsOpenReview ↗

Response to Reviewer P3vK (Part 2)

How do the dense baselines perform when given standard preprocessing (baseline correction, automatic phasing, smoothing) before encoding, both in the clean-trained setting and on measured spectra?

Our experimental NMR/HSQC evaluations are made of processed spectra and not raw FIDs, and dense baselines and VINEA receive the same experimental inputs. We also retain preprocessing built into baseline implementations, including ACT’s baseline correction and UltraIR’s smoothing. Thus, the reported experimental results already include these existing processing steps. However, this does not replace a controlled preprocessing ablation: we have not separately quantified the benefit of adding baseline correction, automatic phasing, and smoothing to each encoder in either the clean-trained corruption benchmark or experimental evaluation. We will clarify this distinction and avoid attributing the observed performance differences solely to the representations.

Please provide a table of the fine-tuning results in Fig. 5a with standard deviations, candidate-pool sizes and paired tests.

We now provide the updated transfer results following calibrated-ramp training, as suggested by Reviewer KbXD, including zero-shot (ZS) and fine-tuned (FT) evaluation on identical held-out test sets. HSQC retains the original augmentation coefficients. Values are mean ± SD over 10 seeds (%): top-1 retrieval accuracy, and macro-AUROC for IR.

Experimental test setCandidatesVINEA ZSBest dense ZSVINEA FTBest dense FTPaired pp, ZS / FT
¹H NMR — BMRB2,01212.83 ± 0.91NMR2Struct: 12.73 ± 0.4515.83 ± 2.31NMRViT: 15.65 ± 0.691.000 / 1.000
¹H NMR - Chemotion3,55624.77 ± 0.75NMR2Struct: 20.10 ± 0.5223.19 ± 1.91NMR2Struct: 20.04 ± 1.050.0469 / 0.0508
¹H NMR - nmrXiv31649.05 ± 0.99NMR2Struct: 43.45 ± 2.0747.97 ± 1.78NMR2Struct: 43.99 ± 1.880.0469 / 0.0508
HSQC - nmrXiv1914.74 ± 4.15DeepSAT: 24.21 ± 5.6620.00 ± 7.77DeepSAT: 35.26 ± 4.990.0508 / 0.0703
IR - NIST FTIRN/A71.54 ± 1.23IRCNN: 71.06 ± 1.0181.90 ± 1.29UltraIR: 82.59 ± 0.881.000 / 0.6738
Raman - SOPRANO356.86 ± 3.86DeepCNN: 7.43 ± 2.419.43 ± 3.58RamanNet: 6.29 ± 1.811.000 / 0.3164

The reported pp-values compare VINEA with the displayed dense baseline using two-sided exact paired sign-flip permutation tests across ten matched training seeds. Holm correction covers all 24 VINEA/baseline comparisons across datasets, separately for zero-shot and fine-tuned evaluation, including baselines omitted from the table. After Holm correction, VINEA significantly outperforms the strongest dense baseline in zero-shot transfer on Chemotion and nmrXiv 1^{1}H NMR (both (p=0.0469p=0.0469)).

Which batch size (number of InfoNCE negatives) did each model use on each modality, and what hyper-parameter search budget did baselines receive compared with VINEA

We thank Reviewer P3vK for requesting this clarification. Our statement that every model used batch size 32 was incorrect. The main synthetic benchmarks used:

ModalityModelsBatch sizeInfoNCE negatives per query per direction
1^1H NMRNMR2Struct, CHEFNMR, NMRViT, SECS, VINEA3231
HSQCDeepSAT, HSQC-CNN, NMRViT-HSQC, VINEA-HSQC43
RamanDeepCNN, RamanNet, SANet, RamanFormer, ReZeroNet, ResNet-1D, Spectral Transformer, VINEA6463
IRACT, IRCNN, SpectraML-CNN, UltraIR, VINEA64Not applicable

For symmetric InfoNCE, a batch of BB matched pairs provides B−1B-1 negatives per query in each direction, or 2(B−1)2(B-1) negative comparisons per matched pair across both directions. Thus, the negative pools were consistent within each retrieval modality. IR uses supervised multilabel classification.

The historical SPECTRE comparison used batches of 32, four gradient-accumulation steps and 300 training epochs, compared with 30 epochs for the other HSQC models. Accumulation did not expand its InfoNCE pool: each query still had 31 negatives.

Regarding hyperparameter-search budgets, the documented NMR learning-rate screen was shared by NMR2Struct, CHEFNMR, NMRViT, SECS and VINEA. Each tested (10−410^{-4},3⋅10−43\cdot 10^{-4},10−310^{-3}) using seed 17 and 30 epochs per trial: three attempted fits per model. VINEA received additional exploration: its NMR architecture sweep and initial HSQC pilot each evaluated four configurations with three seeds (12 fits each), followed by nine HSQC configurations with three seeds (27 fits). We do not claim equal overall tuning budgets. We will correct the batch-size statement and clarify these budgets in the updated manuscript.

AuthorsOpenReview ↗

Response to Reviewer P3vK (Part 3)

What does Table 11 report for Raman (retrieval R@1 or classification BAcc)? What is SPECTRE, how was it trained, and why is it not in Table 8?

We thank the reviewer for identifying these inconsistencies. Tables 11 and 12 report Raman spectrum-to-molecule Recall@1 over 5,056 ChEMBL candidates. The reference to pharmaceutical classification in Table 12’s caption and the BAcc description in Appendix E.10 incorrectly describe these results. We will correct these descriptions. However, this task-label correction does not change the underlying retrieval scores.

SPECTRE (Xu et al., 2026) is a Transformer for NMR-based molecular fingerprint prediction. We explored several training configurations for an HSQC retrieval adaptation using our shared contrastive head. The exploratory configuration reported in the SI used ten seeds, a 300-epoch budget, microbatches of 32, four gradient-accumulation steps, and a Noam schedule with 8,000 warmup steps. Despite these attempts and the extended training budget, our adaptation showed unsatisfactory learning progress and convergence. We therefore excluded it from the final main comparison. Its preliminary results were inadvertently retained in Supplementary Table 11 during manuscript preparation, creating the discrepancy with Table 8. We will remove the obsolete SI entry and clarify the status of this exploratory adaptation.

Analyse which real-data factors make VINEA worse than dense encoders in zero-shot ¹H transfer (e.g., solvent or impurity peaks, multiplet structure, the M = 128 token cap under noise), and report performance as a function of M under thermal noise. Report end-to-end throughput alongside parameter counts, and qualify "compact" accordingly.

We thank Reviewer P3vK for this question. As requested, we further investigated the experimental transfer gap through different ablations, such as token-budget ablations under thermal noise, isolated noised experiments, and diagnostics on measured spectra.

The token-budget results show that retaining more tokens helps VINEA under moderate thermal noise. Increasing M from 128 to 256 improves S3 R@1 by 1.03 percentage points (95%95\% CI [0.46,1.78]). At S5, accuracy stays around 1%1\% for all three budgets, so increasing the token limit provides little benefit under severe noise.

We additionally evaluated individual noise artifacts across S1-S5 using ten paired training seeds. Each artifact was applied alone. We compared full VINEA with the no-blur control to measure how their performance difference depends on the noise type and severity. In 1^1H NMR, full VINEA achieves higher R@1 in all 25 conditions, with an average gain of 2.602.60 percentage points (95%95\% CI [2.31,2.86][2.31,2.86]). At S5, the gains reach 7.107.10 points under phase error and 4.054.05 points under baseline drift. Shimming shows a gain of 6.366.36 points at S4 and 0.970.97 points at S5. Raman and HSQC show mixed effects. For example, the S5 differences are +3.38+3.38 points for Raman fluorescence and −0.92-0.92 for cosmic rays. For HSQC, they are +1.52+1.52 points for shift jitter and −1.92-1.92 for baseline drift.

We hypothesized that solvent or impurity peaks in experimental spectra could contribute to VINEA's transfer gap, as experimental spectra of products might contain impurities. We tested this possibility by adding solvent-reference signals to held-out MSD spectra and comparing VINEA with the dense baselines. We collected a list of most common solvents added each separately at peak heights of 0%0\%, 1%1\%, 5%5\%, 10%10\%, 20%20\% and 50%50\% of the original spectrum's maximum absolute intensity. All models received the same perturbed spectra. The results showed that at 10%10\% addition, VINEA achieves the highest mean R@1 for ten of the twelve solvent references. For example, it reaches 75.28%75.28\% under acetonitrile interference and 76.12%76.12\% under cyclohexane interference, compared with NMR2Struct's 59.63%59.63\% and 62.47%62.47\%, respectively. Averaging across all twelve solvents gives R@1 of 61.57%61.57\% for VINEA and 54.55%54.55\% for NMR2Struct. Ethyl acetate and DMF are the two exceptions, highlighting specific chemical-interference patterns that may contribute to the experimental transfer gap. We summarize all results in the next Table.

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

The paper proposes VINEA, which represents spectra using H0H_0 persistence branches augmented with physical positions, local geometry, and Gaussian scale-space profiles. A set encoder supports molecular retrieval or functional-group prediction. Experiments cover NMR, HSQC, Raman, and IR, focusing on robustness to synthetic corruptions and transfer to experimental spectra.

Critical strengths
  • The method is physically motivated. Combining peak prominence, local shape, and scale-space evolution within a common representation and encoder template across four modalities is valuable.
  • Broad and careful experiments. The paper covers four modalities, retrains 15 published baselines on identical partitions and applies deterministic corruptions shared across all models before any model-specific preprocessing.
  • Strong robustness on IR with a compact model. VINEA is robust on IR under both clean-only and corruption-aware training with about 1M parameters, and the corruption suite could become a useful community resource.
Critical weaknesses
  • Insufficient isolation of the vineyard contribution Removing the blur axis reduces MRR-AUDC from 0.533 to 0.513, although its effect on clean retrieval is larger. These three-seed, NMR-only ablations do not sufficiently establish the vineyard’s contribution to the broader robustness gains. Since finite one-dimensional H0H_0 persistence corresponds to peak prominence, a conventional peak-set baseline with comparable features, the same encoder, and a matched training budget is particularly important.
  • Limited robustness to noise and experimental-domain shifts Under S3 thermal noise, clean-trained VINEA retains only 1.6%/1.2% of its clean NMR/HSQC performance; corruption-aware training improves this substantially, indicating strong dependence on augmentation. Moreover, ramp-pretrained VINEA trails several dense encoders in zero-shot measured NMR and HSQC evaluations. These limitations do not negate the synthetic benchmark gains, but weaken their connection to robustness under experimental conditions, particularly without quantitative severity calibration against measured data.
Questions for authors
  1. Can you compare VINEA with a conventional peak-set baseline using comparable geometric features, the same set encoder, and an identical training budget, reporting both clean and corrupted performance?
  2. Can you repeat the full-versus-no-vineyard ablation under the main experimental protocol on NMR and at least one additional modality, reporting paired confidence intervals for the robustness improvement?
  3. Can you calibrate the synthetic noise levels against measured spectra and test whether the resulting augmentation improves zero-shot NMR/HSQC transfer for VINEA and the strongest dense baseline, compared with the current augmentation?
4 comments
AuthorsOpenReview ↗

Response to Reviewer KbXD (Part 1)

We deeply thank Reviewer KbXD for their constructive and detailed feedback. We have carefully considered your critical weaknesses and questions. We address them point-by-point below:

Can you compare VINEA with a conventional peak-set baseline using comparable geometric features, the same set encoder, and an identical training budget, reporting both clean and corrupted performance?

Thank you for this suggestion. To address it, we evaluated a conventional peak-set control using standard peak picking to extract positions, prominence, widths, areas and neighboring-peak gaps. These features are passed to the same Set Transformer design without the vineyard lifetime and blur-profile descriptors, using the same training budget and ten random seeds (Conventional peaks + VINEA encoder).

We also included NMRTrans and UltraNMR retrieval adaptations using a peak detection algorithm from scipy.signal.find_peaks. NMRTrans receives the same measured peak geometry through its own encoder, providing a complementary comparison. The conventional-peak control retains the VINEA set-encoder design. The table below reports 1^1H NMR spectrum-to-molecule R@1 over 18,493 candidates, as mean sample SD across ten seeds.

ModelClean R@1 (%)Combined S3 R@1 (%)
VINEA77.93$\pm$0.5658.76$\pm$2.05
Conventional peaks + VINEA encoder76.01$\pm$0.4755.73$\pm$2.32
NMRTrans, retrieval adaptation77.18$\pm$0.5750.21$\pm$3.34
UltraNMR, 1^1H-only adaptation47.22$\pm$1.005.30$\pm$0.79

VINEA achieves the highest mean R@1 in both conditions, exceeding the conventional-peak control by 1.93 percentage points on clean spectra and 3.03 points at S3, calculated before rounding. NMRTrans is competitive on clean spectra (77.18% versus 77.93%), but the gap widens to approximately 8.55 points under combined S3 corruptions (50.21% versus 58.76%). Its clean-to-S3 decrease is approximately 26.97 points, compared with 19.17 points for VINEA.

The restricted UltraNMR adaptation obtains 47.22%47.22\% clean and 5.30%5.30\% S3 R@1. These results describe a proton-shift-only model trained from scratch and should not be interpreted as the performance of the published pretrained UltraNMR model with its full inputs. Together, these comparisons support the robustness of the evaluated VINEA configuration relative to the tested peak-based alternatives, especially in noisier environments.

Can you repeat the full-versus-no-vineyard ablation under the main experimental protocol on NMR and at least one additional modality, reporting paired confidence intervals for the robustness improvement?

We thank Reviewer KbXD for this request. We repeated the full-versus-no-blur comparison under the main training protocol on 1^1H NMR, Raman and HSQC, using ten matched training seeds per modality. We evaluated combined corruptions and individual instrumental artifacts across S1-S5 to characterize artifact-specific contributions.

We additionally extended the paired 1^1H NMR evaluation to chemical noises, in which we perturb the original spectra with common solvent impurities at different percentages. We added main-band solvent references individually to held-out MSD spectra, with S0-S5 corresponding to added peak heights of 0%0\%, 1%1\%, 5%5\%, 10%10\%, 20%20\% and 50%50\% relative to each original spectrum's maximum absolute intensity. The full and no-blur models receive identical perturbed spectra, using the same trained checkpoint pairs as the instrumental-artifact comparison. The no-blur control masks the lifetime and blur-profile descriptors in NMR and Raman, and the lifetime descriptor in HSQC, while retaining multiscale extraction.

The instrumental-corruption benchmarks and solvent perturbations both use spectrum-to-molecule retrieval with fixed candidate pools of 18,493 molecules for NMR, whereas Raman contains 5056 and 500 for HSQC. For combined instrumental corruptions, we compute MRR on clean spectra and at S1-S5. MRR-AUDC is the trapezoidal area across these six conditions, divided by five. We also report clean R@1 for reference. For solvent interference, we report R@1 separately at S0-S5, with S0 providing the clean reference.

Clean accuracy and robustness across severities. Values are mean sample SD across ten seeds. MRR-AUDC denotes full minus no-blur, with a paired 95%95\% confidence interval.

ModalityFull VINEA clean R@1Full VINEA MRR-AUDCNo-blur MRR-AUDCMRR-AUDC [95% CI]
1^1H NMR0.779$\pm$0.0060.613$\pm$0.0110.596$\pm$0.010+0.0175[+0.0096,+0.0245]
Raman0.746$\pm$0.0060.608$\pm$0.0070.601$\pm$0.004+0.0069[+0.0021,+0.0110]
HSQC0.423$\pm$0.0140.370$\pm$0.0160.370$\pm$0.012-0.0002[-0.0084,+0.0079]
AuthorsOpenReview ↗

Response to Reviewer KbXD (Part 2)

To characterize artifact-specific behavior, we evaluated 80 artifact-severity conditions across 1^1H NMR, Raman and HSQC. These comprise five artifact types for NMR, five for Raman and six for HSQC, each evaluated individually at S1-S5 using ten seed pairs per modality. The table reports all sixteen modality-specific artifact conditions at S5, including positive and negative effects, as full-minus-control R@1 in percentage points. Confidence intervals use 10,000 paired training-seed bootstrap resamples. They are pointwise, unadjusted and conditional on the fixed test split and corruption bank.

ModalityIndividual artifact at S5R@1, percentage points [95%95\% CI]
1^1H NMRThermal noise+0.34[+0.13,+0.53]
1^1H NMRBaseline drift+4.05[+3.60,+4.42]
1^1H NMRPhase error+7.10[+6.10,+8.00]
1^1H NMRShift jitter+2.12[+1.86,+2.37]
1^1H NMRShimming+0.97[+0.19,+1.79]
RamanFluorescence+3.38[+2.72,+3.95]
RamanCosmic rays-0.92[-1.56,-0.28]
RamanShot noise-0.80[-2.39,+0.64]
RamanWavenumber shift-0.23[-0.61,+0.17]
RamanResolution broadening+0.63[+0.02,+1.28]
HSQCThermal noise+0.16[-0.18,+0.52]
HSQCBaseline drift-1.92[-3.74,-0.16]
HSQCPhase error-1.32[-2.80,+0.10]
HSQCt1t_1 noise-1.86[-3.12,-0.22]
HSQCShift jitter+1.52[+0.32,+2.88]
HSQCLine broadening+0.14[-1.34,+1.64]

Chemical interference from solvent references. We summarize R@1 at S3, corresponding to 10% added solvent peak height. Each solvent is added individually. Values are mean sample SD across ten seeds. Differences are full minus no-blur in percentage points, with paired 95%95\% confidence intervals.

SolventFull VINEA R@1 (%)No-blur R@1 (%)R@1, percentage points [95%95\% CI]
Clean S077.94$\pm$0.5676.03$\pm$0.46+1.91[+1.68,+2.16]
Ethanol70.04$\pm$1.0267.71$\pm$0.98+2.33[+1.52,+3.13]
Methanol73.77$\pm$0.6772.22$\pm$0.99+1.55[+0.85,+2.17]
Acetonitrile75.28$\pm$0.6473.74$\pm$0.93+1.54[+0.94,+2.17]
1-Propanol65.73$\pm$1.2663.00$\pm$1.63+2.73[+1.47,+4.12]
1-Butanol61.46$\pm$1.5358.17$\pm$1.85+3.29[+1.85,+4.87]
Ethyl acetate32.19$\pm$4.6531.08$\pm$6.62+1.11[-3.37,+5.13]
Cyclohexane76.12$\pm$0.6974.58$\pm$0.75+1.54[+0.97,+2.14]
DMF53.341.5748.54$\pm$3.15+4.80[+2.61,+7.10]
Acetic acid74.52$\pm$0.7972.82$\pm$0.88+1.70[+1.21,+2.19]
Glycerol50.92$\pm$1.9547.78$\pm$2.36+3.14[+1.59,+4.57]
Pyridine33.77$\pm$2.8730.27$\pm$3.85+3.49[+0.98,+5.89]
2-Butanone71.65$\pm$1.0569.77$\pm$1.12+1.88[+0.74,+3.01]

The new results show that at 10% addition, full VINEA achieves higher mean R@1 across all twelve solvent references, with positive pointwise confidence intervals for eleven. The largest difference occurs for DMF, at +4.80 points [+2.61,+7.10]. Averaging equally across all twelve solvents and five positive addition levels gives an absolute R@1 advantage of +2.19 points [+1.53,+2.81]. The corresponding reduction in degradation relative to each model's own clean S0 is +0.28 points [-0.24,+0.84]. Thus, the comparison supports higher absolute retrieval accuracy under solvent interference, while the additional clean-relative robustness effect remains uncertain.

For the noise-type ablations across S1–S5, full VINEA achieves higher NMR R@1 in all 25 conditions, with every pointwise, unadjusted paired 95%95\% interval above zero. The mean gain is +2.60 percentage points [+2.31,+2.86], reaching +7.10 [+6.10,+8.00] under S5 phase error. Raman and HSQC show positive and negative artifact-specific effects, with mean differences of +0.06 [-0.24,+0.38] and +0.20 [-0.80,+1.17] points, respectively. Their largest S5 gains occur under Raman fluorescence (+3.38 [+2.72,+3.95]) and HSQC shift jitter (+1.52 [+0.32,+2.88]).

In summary, VINEA follows a shared design across modalities, combining multiscale peak representations with set encoding. The full configuration achieves higher combined-corruption MRR-AUDC in 1^1H NMR and Raman, together with higher absolute NMR retrieval accuracy under the tested solvent additions. Artifact-specific results reveal modality-dependent benefits and tradeoffs, particularly in Raman and HSQC. These findings motivate retaining the full configuration as a practical common default, accompanied by explicit reporting of its modality- and artifact-specific effects.

AuthorsOpenReview ↗

Response to Reviewer KbXD (Part 3)

Can you calibrate the synthetic noise levels against measured spectra and test whether the resulting augmentation improves zero-shot NMR/HSQC transfer for VINEA and the strongest dense baseline, compared with the current augmentation?

We added an experimentally informed calibration study using Spinach simulations of assigned 1^1H NMR spin systems and ORCA harmonic calculations for IR/Raman. For each calibration molecule, we applied our original corruption pipeline to its simulated spectrum and fitted three strengths: broadening, baseline/background amplitude, and noise amplitude. We matched three measured descriptors, median peak width, broad background variation, and fluctuations in quiet spectral regions, using bounded, robust nonlinear least squares:

θi⋆=arg⁡min⁡θmin⁡≤θ≤θmax⁡[∑k=13ρ ⁣(wklog⁡d^ik(θ)dikexp)+R(θ,θS3)].\theta_i^\star = \underset{\theta_{\min}\leq\theta\leq\theta_{\max}}{\arg\min} \left[ \sum_{k=1}^{3} \rho\!\left( w_k\log\frac{\widehat d_{ik}(\theta)}{d_{ik}^{\mathrm{exp}}} \right) + R(\theta,\theta_{\mathrm{S3}}) \right].

Here, dikexpd_{ik}^{\mathrm{exp}} is the measured descriptor, aggregated across repeated measurements of molecule ii, and d^ik(θ)\widehat{d}_{ik}(\theta) is the corresponding descriptor after corrupting its simulated reference, taking the median over three stochastic realizations. The weights wkw_k balance the descriptors, ρ\rho is a robust loss, and RR weakly regularizes the coefficients toward the original S3 settings. Other corruption parameters remain unchanged. The fitted vectors form an empirical distribution that we sample jointly during training. We froze this distribution before evaluating it on separate validation molecules against the original S3 coefficients, using the same operator-activation probability as training. The resulting coefficients describe experimentally informed augmentation strengths, not an exact transformation from simulated to experimental spectra.

We then retrained VINEA and all retained baselines over 10 seeds, keeping the original curriculum, 20% clean training followed by progressively stronger corruption, but sampling experimentally calibrated coefficient vectors. We also added SOPRANO [1] organic-pigment Raman spectra, contributing 65 eligible molecules without overlap with synthetic training/validation, 35 form the held-out test set below. HSQC could not be calibrated because we did not have a dataset of assigned peaks to simulate against experimental spectra.

The following results use zero-shot evaluation without experimental fine-tuning. Values are mean ± SD across seeds, in percent: top-1 retrieval accuracy for NMR/HSQC/Raman and macro-AUROC for IR. The dense comparator is the baseline with the highest mean on each test set.

Experimental test setVINEA zero-shotBest dense zero-shotVINEA fine-tunedBest dense fine-tuned
¹H NMR - BMRB (2,012 molecules)12.83 ± 0.91NMR2Struct: 12.73 ± 0.4515.83 ± 2.31NMRViT: 15.65 ± 0.69
¹H NMR - Chemotion (3,556)24.77 ± 0.75NMR2Struct: 20.10 ± 0.5223.19 ± 1.91NMR2Struct: 20.04 ± 1.05
¹H NMR - nmrXiv (316)49.05 ± 0.99NMR2Struct: 43.45 ± 2.0747.97 ± 1.78NMR2Struct: 43.99 ± 1.88
HSQC - nmrXiv (19)14.74 ± 4.15DeepSAT: 24.21 ± 5.6620.00 ± 7.77DeepSAT: 35.26 ± 4.99
IR - NIST FTIR (848)71.54 ± 1.23IRCNN: 71.06 ± 1.0181.90 ± 1.29UltraIR: 82.59 ± 0.88
Raman - SOPRANO (35)6.86 ± 3.86DeepCNN: 7.43 ± 2.419.43 ± 3.58RamanNet: 6.29 ± 1.81

Using a unified architecture across modalities, VINEA achieves the highest mean zero-shot performance on all three 1^1H NMR datasets and IR, and the highest mean fine-tuned performance on all three 1^1H NMR datasets and Raman, demonstrating strong transfer from synthetic training with calibrated noise to experimental spectra against modality-specific baselines.

[1] W. Fremout, S. Saverwyns, Identification of synthetic organic pigments: the role of a comprehensive digital Raman spectral library, J. Raman Spectrosc. 43 (2012) 1536-1544, DOI:10.1002/jrs.4054

AuthorsOpenReview ↗

Response to Reviewer KbXD (Part 4)

Insufficient isolation of the vineyard contribution Removing the blur axis reduces MRR-AUDC from 0.533 to 0.513, although its effect on clean retrieval is larger. These three-seed, NMR-only ablations do not sufficiently establish the vineyard’s contribution to the broader robustness gains. Since finite one-dimensional H0 persistence corresponds to peak prominence, a conventional peak-set baseline with comparable features, the same encoder, and a matched training budget is particularly important.

We agree that the original ablation on three random seeds and only NMR was insufficient. The conventional peak-set control reported in Question 1 already includes peak prominence and comparable geometric features. The additional vineyard descriptors encode how peaks persist and change across Gaussian smoothing widths, so the comparison directly addresses the value of this additional information. As detailed in Question 2, we expanded the evaluation to ten seeds across NMR, Raman and HSQC, with paired confidence intervals and artifact-specific analyses. The full configuration has higher combined-corruption MRR-AUDC in NMR and Raman, whereas the HSQC experiment does not establish an overall lifetime benefit. We therefore qualify the contribution as modality-dependent. The remaining implementation and configuration differences described above also limit how exclusively these gains can be attributed to the vineyard descriptors.

We emphasize that VINEA follows a shared architectural template across modalities, combining peak geometry, multiscale information and set encoding with modality-specific input adaptations. The consistent NMR gains and benefits under selected Raman and HSQC artifacts motivate retaining the full configuration as a common default.

Limited robustness to noise and experimental-domain shifts Under S3 thermal noise, clean-trained VINEA retains only 1.6%/1.2% of its clean NMR/HSQC performance; corruption-aware training improves this substantially, indicating strong dependence on augmentation. Moreover, ramp-pretrained VINEA trails several dense encoders in zero-shot measured NMR and HSQC evaluations. These limitations do not negate the synthetic benchmark gains, but weaken their connection to robustness under experimental conditions, particularly without quantitative severity calibration against measured data.

We agree that the severe degradation under S3 thermal noise reveals a limitation of clean-trained VINEA, and that robustness on synthetic corruptions alone does not establish robustness to experimental domain shifts. Our claims should explicitly acknowledge the benefit of corruption-aware training.

As detailed above, we now provide quantitative calibration against measured spectra and evaluation over 10 seeds. In the updated evaluation, VINEA achieves the highest mean zero-shot performance on all three 1^1H NMR datasets, strengthening the evidence for experimental transfer. HSQC remains a limitation: its coefficients were not recalibrated, and DeepSAT performs better both zero-shot and after fine-tuning. These findings support experimentally informed augmentation and competitive transfer, but do not establish augmentation-independent robustness or isolate an improvement over the original ramp. We will qualify the claims accordingly.

Reviewer_sAV6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces VINEA, a multi-scale topological representation method for molecular spectroscopy. The approach begins by performing zero-dimensional persistent homology analysis on 1D NMR, Raman, and IR spectra, as well as 2D HSQC spectra, representing local maxima or connected regions as persistent branches. These branches are then tracked within a Gaussian scale space to construct a "persistence vineyard". By integrating peak positions, persistence, local line shapes, inter-peak spacing, and scale-based survival, the method generates an unordered set of tokens, which are subsequently processed via a Set Transformer to derive a final spectral representation. The study establishes a synthetic, degraded benchmark across four spectroscopic modalities and evaluates the method’s performance in tasks including molecular retrieval, functional group classification, and the transfer from simulated to experimental spectra.

Critical strengths
  1. The research motivation is significant. The paper addresses challenges, such as noise, baseline drift, phase errors, coordinate shifts, and resolution variations, that arise when transferring simulated spectral models to experimental spectra; this carries clear practical relevance.

  2. The methodological design is sound. By integrating persistent homology, physical coordinates, local line shapes, and multi-scale persistence information, VINEA reduces reliance on fixed intensity thresholds and predefined sampling grids.

  3. The representation structure is well-aligned with the problem. Since persistent branches naturally form unordered sets, the use of a permutation-invariant set encoder is appropriate. Furthermore, the combination of absolute position and relative displacement channels is well-suited to the characteristics of spectroscopic signals.

  4. The experimental scope is comprehensive. The study evaluates the approach across 1H NMR, HSQC, Raman, and IR/FTIR data, encompassing scenarios such as clean-only training, corruption-aware learning, extrapolation to varying severity levels, transfer to experimental spectra, and adaptation across different magnetic field strengths.

  5. The writing and figures are generally well-executed.

Critical weaknesses
  1. There is an internal inconsistency in the evaluation protocol.
  • Appendix G initially states that a batch size of 32 is used for all models and datasets to keep the InfoNCE negative sample pool constant.
  • It later reports batch sizes of 32, 4, 64, and 64 for 1H NMR, HSQC, Raman, and IR, respectively.
  1. Several key results are inconsistent across the paper.
  • Ramp-trained VINEA obtains a clean R@1 of 0.778 in Supplementary Table 12 but only 0.276 in Table 16, despite both experiments apparently using the same MSD 1H NMR benchmark, 18,493-candidate pool, and S1-S3 corruption ramp.

  • Differences in seed count or checkpoint selection do not appear sufficient to explain this gap.

  • The S3 NMR R@1 reported in the discussion of Figure 4 (0.588) also differs from Table 12 (0.583).

    The authors should reconcile these values and specify the exact configuration used for each result.

  1. The target-domain fine-tuning protocol is insufficiently documented. For each measured-spectrum dataset, the paper should report:
  • training, validation, and test set sizes;

  • scaffold-splitting and duplicate-handling procedures;

  • fine-tuning and optimization settings;

  • results for each individual dataset.

    Pooling BMRB, Chemotion, GISSMO, and NMRXiv in Figure 5 may conceal substantial dataset-specific variation.

  1. The ablation study is too limited. No corresponding ablations are reported for HSQC, Raman, or IR. Sensitivity analyses are missing for the token budget, Gaussian scale ladder, and HSQC downsampling resolution. Broader ablations are needed to establish whether the proposed components and hyperparameter choices generalize across modalities.
Questions for authors
  1. What are the actual batch sizes used for each model and modality? Is the number of negative samples for InfoNCE consistent across the different models?
  2. Why is there a difference of approximately 0.50 in the "clean R@1" metric for the "Full VINEA" model on the MSD NMR task when comparing Table 12 and Table 16?
  3. Which value for the S3 metric is correct: 0.588 (as shown in Figure 4) or 0.583 (as shown in Table 12)?
  4. Please provide the precise data splits, sample counts, deduplication rules, and checkpoint selection methods used for the fine-tuning of all experimental spectra.
  5. Why are NMRTrans, NMR-Text, and UltraNMR included in the training configurations, yet no results are reported for them?
7 comments
AuthorsOpenReview ↗

Response to Reviewer sAV6 (Part 1)

We sincerely thank Reviewer sAV6 for their constructive and detailed feedback. We have carefully considered your critical weaknesses and questions. We address them point-by-point below:

What are the actual batch sizes used for each model and modality? Is the number of negative samples for InfoNCE consistent across the different models?

Thank you for identifying this inconsistency. Indeed, the statement that all models used a batch size of 32 was incorrect. Training batch sizes were 32 for 1^1H NMR, 4 for HSQC, and 64 for Raman and IR, consistently across the reported models within each modality. Knowing that we used the symmetric InfoNCE uses B−1B-1 negatives per query per direction, giving 2(B−1)2(B-1) negative comparisons per matched pair. Negative pools were similar within each retrieval modality, but differed across modalities. IR uses supervised binary cross-entropy, so negative elements of the symmetric InfoNCE do not apply. We have further developed this in the manuscript accordingly.

Why is there a difference of approximately 0.50 in the "clean R@1" metric for the "Full VINEA" model on the MSD NMR task when comparing Table 12 and Table 16?

Thank you for highlighting this. Table 12 and 16 actually used different VINEA checkpoints under different training protocols, which we did not distinguish clearly. The main benchmark used 20 clean-training epochs followed by an 80-epoch S1 to S3 corruption ramp, with thermal noise applied probabilistically and checkpoint selection throughout training. The original ablation instead started the corruption ramp immediately, always applied thermal noise during augmentation, and restricted checkpoint selection to epoch 75 onward. Consequently, its full-model reference was not directly comparable to Table 12.

We have now rerun the ablation variants under the main protocol using ten matched seeds. The full-model reference achieves 0.7790.006 clean R@1, consistent with the finalized main-benchmark mean. We will replace the original ablation table and clarify the protocols.

ExperimentTraining protocolSeedsClean R@1
Main benchmark, finalizedMain protocol100.779
Original ablation, Table 16Separate ablation protocol30.276
Updated ablationMain protocol100.779

Which value for the S3 metric is correct: 0.588 (as shown in Figure 4) or 0.583 (as shown in Table 12)?

Thank you for also pointing out this difference. The correct S3 R@1 for the finalized ten-seed main benchmark cohort is 0.588±0.0210.588 \pm 0.021, as reported in Figure 4. Table 12 accidentally retained 0.583±0.0220.583 \pm 0.022 from the seven-seed aggregate, while its seed count was updated to ten. The same issue also affected clean R@1, which changes from 0.7780.006 to 0.7790.006 in the finalized ten-seed cohort. This was caused by a bug in our aggregation pipeline. We apologize for this reporting error.

Following this, we subsequently verified all numerical tables against the retained per-seed results and controlled the fine-tuning and field-transfer figures. Across all measured means, the largest absolute correction was of 0.045 R@1, for VINEA-HSQC zero-shot transfer (0.020 to 0.065). All corrections made are minor numerical changes that do not change the paper’s central findings on synthetic robustness, fine-tuning and field transfer. They do require revising one observation: VINEA-HSQC has now a higher mean zero-shot score than the reported dense baselines. We updated a consolidated correction for all tables and updated the affected text accordingly.

AuthorsOpenReview ↗

Response to Reviewer sAV6 (Part 2)

Please provide the precise data splits, sample counts, deduplication rules, and checkpoint selection methods used for the fine-tuning of all experimental spectra.

We fine-tuned each model from its corresponding S1–S3 pretrained checkpoint using fixed experimental splits shared across models and training seeds. All counts below are training/validation/test.

  • 1^1H NMR: Each dataset was independently split by Bemis–Murcko scaffold, targeting 70%/10%/20% with a fixed random seed: Chemotion 2,798/401/800, BMRB 1,453/262/839, GISSMO 378/76/113, and nmrXiv 298/42/85.
  • nmrXiv HSQC: Target proportions of 65%/15%/20% yielded 62/18/25 samples.
  • RamanBench: A class-stratified 70%/15%/15% spectrum-level split yielded 2,458/526/526 spectra. Training batches contained one spectrum per molecule; validation and test retrieval used 31 unique molecules.
  • Experimental IR: We preserved released fold assignments after removing 390 rows overlapping simulated training molecules and 523 additional molecular duplicates, leaving 5,576/1,276/897 spectra. Canonical-SMILES deduplication prioritized test, then validation, then training. For each model and seed, checkpoints maximized spectrum-to-molecule MRR on uncorrupted validation data, evaluated every five epochs, with caps of 100 epochs for NMR/HSQC and 60 for Raman, and early stopping after six checks without improvement. IR checkpoints maximized validation macro-average precision, evaluated every epoch over 40 epochs. Selected checkpoints were restored before testing. Test scores never informed selection. We will clarify these settings in the updated manuscript.

Why are NMRTrans, NMR-Text, and UltraNMR included in the training configurations, yet no results are reported for them?

Thank you for highlighting this ambiguity. Our configuration list retained experimental peak-based models beyond the comparisons reported in the manuscript. We should have distinguished these entries more clearly or have removed them. Our benchmark focused on extracting informative representations directly from numerical spectra, without externally supplied peak lists or chemical annotations. In short, NMRTrans and NMR-Text use processed spectral descriptions (chemical shifts, integration values, multiplicities and, for NMRTrans, J-couplings, obtained through spectral-processing tools such as the closed-source MestreNova software), while UltraNMR takes chemical-shift lists, with paired 1^1H/13^{13}C inputs in its published setup. Comparing these methods from the same numerical spectra therefore requires an explicit input-extraction pipeline.

Nevertheless, we thought it would still be insightful to compare those baseline, so we adapted NMRTrans and UltraNMR to use peaks detected automatically with scipy.signal.find_peaks. NMRTrans receives peak positions and geometry measured directly from the spectrum, including intensities, areas, widths and neighbouring peak distances. UltraNMR receives only the detected 1^1H chemical shifts, retaining its original Fourier embedding. Both adaptations were trained from scratch and evaluated over ten seeds, without externally supplied annotations. In the following table, we report 1^1H NMR spectrum-to-molecule retrieval over 18,493 candidates. R@1 (%) as mean ±\pm SD across ten training seeds.

ModelClean R@1 (%)Combined S3 R@1 (%)
VINEA77.93±0.56\mathbf{77.93 \pm 0.56}58.76±2.05\mathbf{58.76 \pm 2.05}
Conventional peaks + VINEA encoder76.01±0.4776.01 \pm 0.4755.73±2.3255.73 \pm 2.32
NMRTrans retrieval adaptation, selected LR = 3e-577.18±0.5777.18 \pm 0.5750.21±3.3450.21 \pm 3.34
UltraNMR H-only retrieval adaptation47.22±1.0047.22 \pm 1.005.30±0.795.30 \pm 0.79

The conventional baseline uses standard peak picking to measure positions, heights, widths and areas, then feeds these features into the VINEA set encoder without vineyard descriptors. VINEA exceeds this baseline by 1.93 points on clean spectra and 3.03 points at S3. NMRTrans uses measured peak geometry with a different encoder. After hyperparameter screening of the learning rate, the baseline reaches 77.18%77.18\% clean R@1 versus VINEA’s 77.93%77.93\%, but falls to 50.21%50.21\% at S3 versus 58.76%58.76\%, supporting VINEA’s stronger robustness under this evaluated corruption setting. UltraNMR reaches 47.22%47.22\% clean and 5.30%5.30\% at S3. This is a restricted, proton-position-only adaptation without carbon inputs or pretraining, so it remains a supplemental comparison.

AuthorsOpenReview ↗

Response to Reviewer sAV6 (Part 3)

There is an internal inconsistency in the evaluation protocol.Appendix G initially states that a batch size of 32 is used for all models and datasets to keep the InfoNCE negative sample pool constant. It later reports batch sizes of 32, 4, 64, and 64 for 1H NMR, HSQC, Raman, and IR, respectively. Several key results are inconsistent across the paper. Ramp-trained VINEA obtains a clean R@1 of 0.778 in Supplementary Table 12 but only 0.276 in Table 16, despite both experiments apparently using the same MSD 1H NMR benchmark, 18,493-candidate pool, and S1-S3 corruption ramp. Differences in seed count or checkpoint selection do not appear sufficient to explain this gap. The S3 NMR R@1 reported in the discussion of Figure 4 (0.588) also differs from Table 12 (0.583). The authors should reconcile these values and specify the exact configuration used for each result.

Answered in question 3.

The target-domain fine-tuning protocol is insufficiently documented. For each measured-spectrum dataset, the paper should report: training, validation, and test set sizes; scaffold-splitting and duplicate-handling procedures; fine-tuning and optimization settings; results for each individual dataset.

We agree that the target-domain fine-tuning protocol needs clearer documentation. In the revised manuscript, we will report the exact training, validation, and test counts for every measured-spectrum dataset, explain the dataset-specific splitting and duplicate-handling procedures, and specify the initialization checkpoints, trainable parameters, optimizer, learning rate, batch size, epoch budget, early-stopping criteria, and validation-based checkpoint selection. We will also clarify that NMR and HSQC use scaffold-based splits, Raman uses a spectrum-level stratified split with repeated molecular identities across folds, and IR retains the released fold assignments after overlap removal and deduplication. Finally, we will separate the pooled experimental results in Figure 5 and report performance for each dataset individually, also addressing reviewer KbXD’s comments on experimental transfer.

Pooling BMRB, Chemotion, GISSMO, and NMRXiv in Figure 5 may conceal substantial dataset-specific variation.

We agree that pooling BMRB, Chemotion, GISSMO, and nmrXiv in Figure 5 can obscure dataset-specific variation. We combined these results to save space, but will present each dataset separately in the revised manuscript so that differences in experimental transfer performance are explicit. This revision will also incorporate reviewer KbXD’s comments on experimental transfer.

The ablation study is too limited. No corresponding ablations are reported for HSQC, Raman, or IR. Sensitivity analyses are missing for the token budget, Gaussian scale ladder, and HSQC downsampling resolution. Broader ablations are needed to establish whether the proposed components and hyperparameter choices generalize across modalities.

We expanded the ablation experiments across 1^1H NMR, Raman and HSQC, using ten seeds per setting. We additionally report sensitivity to token budget, Gaussian scale count and range, and HSQC downsampling resolution, alongside vineyard-descriptor ablations for individual corruption types and severities.

The sensitivity analyses reveal modality-dependent choices. Using 64 rather than 128 tokens improves robustness in Raman and HSQC, while reducing the HSQC grid from 512×512 to 256×256 or 128×128 lowers MRR-AUDC from 0.370 to 0.339 and 0.303. Extending its Gaussian ladder to 100-1000 Hz improves MRR-AUDC to 0.379, while seven scales show no consistent advantage over fewer scales. VINEA achieves higher MRR-AUDC than the no-blur control in NMR (0.613 versus 0.596) and Raman (0.608 versus 0.601), with paired 95%95\% intervals excluding zero. Across isolated artifacts and severities, average NMR R@1 is higher by 2.60 percentage points (95%95\% CI [2.31, 2.86]), reaching +7.10 percentage point under S5 phase error, +6.36 under S4 shimming and +4.05 under S5 baseline drift. Raman differences depend on the artifact, with +3.38 points under S5 fluorescence and -0.80 under S5 shot noise. HSQC also shows mixed differences, including -1.92 points under S5 baseline drift.

For more details, we provide the updated ablation results alongside this response and will incorporate them into the manuscript. All results are spectrum-to-molecule retrieval over ten seeds. Candidate pools contain 18,493 NMR, 5,056 Raman and 500 HSQC molecules.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.