acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Spectral Conditioning for NMR Structure Elucidation

Abstract

Diffusion models for nuclear magnetic resonance (NMR) structure elucidation typically condition generation through a single global spectral representation, applying the same modulation to every atom and, in graph models, every bond. We study whether this uniform conditioning interface limits structure recovery and replace it with per-element spectral cross-attention, allowing each atom and candidate bond to query the spectral tokens independently. On simulated SpectraBase spectra, this raises top-1 exact recovery on NMR-DiGress from for our same-protocol FiLM reproduction to . The same conditioning change transfers to the coordinate-based ChefNMR architecture, where per-atom cross-attention improves over the released ChefNMR-S model by 7.24 percentage points under a common evaluation. Fine-tuning the cross-attention models with an auxiliary spectral reconstruction loss further reaches on NMR-DiGress and on ChefNMR. These results show that per-element spectral conditioning improves structure recovery across both graph and coordinate diffusion models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
NMR structure elucidation, molecular graph generation, diffusion models, cross-attention conditioning, spectral conditioning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_wsYhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the limitations of global spectral conditioning in diffusion-based molecular structure elucidation from NMR spectra. The authors identify that existing methods typically compress spectral information into a single global representation, which is uniformly applied to all atoms and bonds, potentially limiting the model's ability to capture local spectrum–structure relationships. To address this issue, they propose per-element spectral cross-attention, allowing individual atoms and candidate bonds to independently attend to relevant spectral tokens. The proposed conditioning mechanism is integrated into two existing diffusion architectures, NMR-DiGress and ChefNMR, and further enhanced with an auxiliary spectral reconstruction loss. Experiments on the simulated SpectraBase dataset demonstrate substantial improvements in top-1 exact structure recovery, increasing from 43.26% to 52.10% on NMR-DiGress and reaching 76.64% on ChefNMR. With additional spectral reconstruction supervision, performance further improves to 54.38% and 78.29%, respectively. The authors also provide theoretical analysis of the greater expressiveness of element-specific conditioning and evaluate transfer to experimental NMR spectra, where improvements in exact recovery are more limited. Overall, the paper highlights the importance of fine-grained spectral conditioning for molecular structure generation and demonstrates its effectiveness across different diffusion architectures.

Critical strengths

Clear motivation and meaningful problem formulation. The paper identifies a potential limitation of global spectral conditioning in diffusion-based NMR structure elucidation and investigates whether fine-grained conditioning can better capture local spectrum–structure relationships.

Critical weaknesses

Limited methodological novelty. The main contribution is replacing global FiLM/AdaLN conditioning with standard per-element cross-attention, a widely adopted mechanism in conditional generative modeling. The authors should better clarify the methodological novelty beyond adapting existing cross-attention techniques to NMR-based molecular structure elucidation.

Limited improvements on experimental NMR spectra. Although the proposed method achieves substantial gains on simulated SpectraBase data, its improvements on experimentally measured spectra are marginal. On SpecTeach, top-1 recovery improves only from 41.2% to 42.9%, while performance on NMRShiftDB2 remains approximately unchanged. This raises concerns about the practical generalization of the proposed approach.

Insufficient compute-matched comparisons. The cross-attention modules introduce additional parameters and computational overhead. More rigorous comparisons under matched parameter counts, training budgets, and inference costs would help isolate the benefits of the proposed conditioning mechanism.

Questions for authors

See Weaknesses.

1 comment
AuthorsOpenReview ↗

Reply to Reviewer wsYh: pooled conditioning limits two models on simulated spectra, the SpecTeach numbers are corrected, and after the same training time our model is about 7 points above ChefNMR-S on simulated spectra

Thank you for your review.

"The authors should better clarify the methodological novelty beyond adapting existing cross-attention techniques to NMR-based molecular structure elucidation."

We agree that cross-attention is a widely used mechanism. What we show is that pooled FiLM and adaLN conditioning limits these two models on simulated spectra, and we quantify the effect. The same change improves two models that generate molecules in very different ways, with NMR-DiGress improving from 43.3% to 52.1% and ChefNMR-S from 69.2% to 76.6%. On simulated spectra, the published ChefNMR-L reaches about 72% with three times the parameters of our ChefNMR model, CA-ChefNMR-S. On simulated spectra the improvement also holds for molecules whose ring scaffold never occurs in the training data, as described in point 2 of the summary comment. We will clarify this contribution in the revised paper, which we will upload before the discussion period ends.

"On SpecTeach, top-1 recovery improves only from 41.2% to 42.9%, while performance on NMRShiftDB2 remains approximately unchanged. This raises concerns about the practical generalization of the proposed approach."

You are right that we do not show an improvement on experimental spectra, and we have to correct the numbers you quote. We found a one-bin offset in the experimental 13C inputs, which we describe in point 3 of the summary comment. With the offset corrected, ChefNMR-S and our model solve about 55% of the SpecTeach molecules, and they are not distinguishable.

We also investigated why the accuracy on experimental spectra is so much lower.

  • 165 of the 213 SpecTeach molecules are also in the SpectraBase training set. ChefNMR-S and our model solve about 98% of them from simulated spectra, and 94% to 96% of the 44 others with a simulated spectrum, so little improvement could carry over.
  • Most of the drop comes from the measured 1H spectrum: replacing only the simulated 1H spectrum of a molecule by the measured one costs about 32 to 34 of the roughly 41 to 42 points lost on measured spectra, replacing only the 13C input about 1 point.

As used here, NMRShiftDB2 only contains 13C shifts, so it only tests the 13C input, which has the same one-bin offset. In the revised paper we will report the result of the NMRShiftDB2 runs with the corrected input.

"More rigorous comparisons under matched parameter counts, training budgets, and inference costs would help isolate the benefits of the proposed conditioning mechanism."

  • Training budget: after the same wall-clock training time our model is about 7 points above ChefNMR-S (76.48% against 69.15%), at the same batch size. We trained our ChefNMR model from scratch, and at about the same training FLOPs it already reaches 75.77%. Our two NMR-DiGress models were trained from scratch on the same hardware, with the same batch size, learning rate and 48-hour limit as set in our run files.
  • Parameters: a pooled model three times the size of ours does not reach this accuracy on simulated spectra. The published ChefNMR-L reaches about 72% where our ChefNMR model reaches about 77%. We do not have pooled models of exactly the size of ours. Our models have 14% more parameters than the pooled models they replace (153 M against 134 M for the ChefNMR models).
  • Inference cost: we have not measured it in a controlled way.

In the revised paper we will add a table with the parameters, optimiser steps and training FLOPs of CA-ChefNMR-S and the released ChefNMR-S. If this answers your concerns, we would be grateful if you reconsidered your score.

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

This paper studies how NMR spectra conditioning effects molecular diffusion decoder performance for NMR-based small molecule structure elucidation. It proposes a cross-attention representation mechanism that lets each atom and candidate bond query the spectral tokens separately, instead of a pooled spectral representation that is shared across atoms and bonds. The paper evaluate these two spectral conditioning methods on a widely-used graph and coordinate diffusion model architecture respectively (i.e., NMR-DiGress and ChefNMR), and further tries tine-tuning the cross-attention models with an auxiliary spectral reconstruction loss. Structure recovery performance improves on simulated SpectraBase data, while gains on experimental spectra are limited. The paper also provides a single-step expressivity analysis and examines how the models respond to spectral masking.

Critical strengths

The paper addresses a clear architectural question and tests the proposed change on both graph-based and coordinate-based diffusion models. On simulation data, the performance boost is substantial, as top-1 accuracy increases from 43.26% to 52.10% on NMR-DiGress, and on ChefNMR from 45.24% to 76.64%. The authors also include useful controls involving pooling capacity, encoder choice, and auxiliary reconstruction. They report the weaker results on experimental spectra and acknowledge the distinction between repeated sampling from one checkpoint and independent training runs.

Critical weaknesses
  1. My main concern is the extent of the contribution. Replacing pooled modulation with cross-attention is a familiar architectural change, so the paper needs convincing evidence about what this teaches us specifically about spectral conditioning. The current analysis empirically shows uniform conditioning could be the bottleneck responsible for the gains, but the theoretical result only formulate an idealized single conditioning step, while the masking experiment also does not show that cross-attention makes better use of the relevant spectral regions. Also the implementation also changes normalization on DiGress, without controls that fully separate these factors from element-specific spectral access.

  2. The broader benefit remains uncertain. Improvements are mainly seen on simulated spectra under a random molecular split, lacking a more difficult MCES or scaffold setup to evaluate generalization; the experimental-spectrum evaluations show little improvement in recovery. Independent training repeats are also missing, making the smaller auxiliary-loss gains difficult to assess. The comparison with a larger ChefNMR model is encouraging, but neither a matched-compute advantage nor improved data efficiency is established. On DiGress, the recovery gain additionally comes with reduced chemical validity. I therefore see a useful architectural improvement, but the demonstrated scope and explanatory depth are not yet sufficiently satisfying.

Questions for authors
  1. Can the authors separate the benefit of element-specific spectral access from the additional parameters and normalization changes? A pooled-conditioning control with comparable capacity and the same normalization would help interpret the main result.

  2. Given that the normalized masking analysis does not distinguish the two interfaces, what evidence supports the stronger claim that uniform conditioning is the limiting bottleneck, beyond showing empirically that the replacement module performs better?

  3. Can the authors explain why the simulated-data gains do not translate into clear improvements on experimental spectra? An analysis of which errors the new model fixes or introduces on those datasets would help establish its practical value.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer xCLd: for our ChefNMR model the improvement of about 7 points on simulated spectra is obtained with the same kind of normalisation, after the same training time and on unseen scaffolds, and why experimental spectra give no gain

Thank you for the detailed review.

"A pooled-conditioning control with comparable capacity and the same normalization would help interpret the main result."

ChefNMR-S and our ChefNMR model (CA-ChefNMR-S) use the same kind of adaptive layer normalisation, and only NMR-DiGress has the normalisation change you point out. So for our ChefNMR model the improvement of about 7 points is obtained with the same kind of normalisation. For the parameters two results are close to the control you ask for, although we do not have a pooled control of exactly our size. For NMR-DiGress, three pooled vectors instead of one give about 1.5 points, and cross-attention gives about 9. Compared with ChefNMR-S, ChefNMR-L adds 328 M parameters for about 3 points. Our model adds 19 M for about 7.

"what evidence supports the stronger claim that uniform conditioning is the limiting bottleneck"

The first contribution of our paper names pooled conditioning "a bottleneck of spectral conditioning", not the only one. A new check shows that our models give each atom its own view of the spectrum. If at inference every atom (for NMR-DiGress also every atom pair) gets the average of these views, top-1 of our ChefNMR model falls from about 77% to about 1% (3,000 random SpectraBase test molecules) and top-1 of our NMR-DiGress model from about 51% to below 1% (1,500 of these molecules). This does not show that a pooled model must fail. We agree with the limits you name: the lemmas only describe one idealised conditioning step, and the masking analysis does not separate ChefNMR-S from our ChefNMR model. In the revised paper, which we will upload before the discussion period ends, we will state these limits next to the lemmas.

"Can the authors explain why the simulated-data gains do not translate into clear improvements on experimental spectra?"

For most SpecTeach molecules we can, see point 3 of the summary comment. ChefNMR-S and our model solve about 98% of the 165 SpecTeach molecules that are also in the training set from simulated spectra, so little improvement could carry over. And most of the drop on measured data comes from the measured 1H spectrum: about 32 to 34 of the roughly 41 to 42 points lost on measured spectra.

"lacking a more difficult MCES or scaffold setup to evaluate generalization"

On simulated spectra the improvement also holds for unseen scaffolds. On the unseen-scaffold list of Xiong et al. (3,770 test molecules, point 2 of the summary comment) the improvement is about 7 points. For its 1,788 molecules with rings, our model solves about 54% and ChefNMR-S about 45% in our evaluation. We do not have an MCES split, because this would need retraining. Within our test set, at an MCES distance of 5 to below 10 from the nearest training molecule (an upper bound on the true distance), the improvement is about 10 to 12 points. At most 4 of the 3,000 and 12 of the 3,770 molecules are at a distance of 10 or more, the threshold MassSpecGym (Bushuiev et al., 2024) uses to separate its folds. So we cannot show it beyond the MassSpecGym threshold.

"Independent training repeats are also missing"

You are right. We do not have repeated training runs, because we lack the compute for them, so in the revised paper we will present the improvements from the auxiliary loss (about 2 points, obtained by fine-tuning) as a trend. With the loss, our ChefNMR model gains 1.65 points over a run trained at least as long without it.

"neither a matched-compute advantage nor improved data efficiency is established."

After the same wall-clock training time our model is about 7 points above ChefNMR-S (76.48% against 69.15%), and at about the same training FLOPs it reaches 75.77%. You are right that we did not state this comparison in the paper: Table 11 already reports these two values of our model, and we will state the comparison in the main text of the revised paper. We have not measured the inference cost in a controlled way and did not study data efficiency.

"the recovery gain additionally comes with reduced chemical validity"

The improvement remains when invalid structures count as misses: on simulated spectra, NMR-DiGress with cross-attention is still about 7 points ahead (1,500 SpectraBase test molecules), against about 8 points when they are skipped. The validity goes down from about 94% to about 90%, which is a drawback. For the ChefNMR models the validity does not go down (96% for both).

One remark on your summary: the 45.24% you give as the ChefNMR baseline is the model of Hu et al. in Table 1 of the paper. The baseline for our model is ChefNMR-S at about 69%, and we will split the table by backbone. If this answers your concerns, we would be grateful if you reconsidered your score.

We thank the author for the explanations upon the questions we raised in the last round. The author's clarification does reassure much of my previous concerns:

  • Comparable capacity and normalization: I can now see that you controlled normalization on ChefNMR model, and that cross-attention representation delivers better performance under much comparable model capacity is convincing.
  • Evidence of uniform conditioning being the limiting bottleneck: I'm glad to see the author promise to add statements of possible limitations next to the lemmas, though I feel like the stress test you construct only at inference-time when models learned to leverage atom-wise attention during training, the performance drop is too big to let one suspect whether this is due to the "bottleneck" we talked about here, or it's basically some severe train-inference mismatch (please correct me if I'm wrong).
  • Generalization to real spectra and new scaffolds, consistent gains across necessary training repeats: the author gave clear clarifications on both of my concerns, thanks for providing/promising to provide the new experimental results.
  • Matched compute and w/o validity filtering comparisons: I think it will be great to state the comparable-compute comparison on the main text, even though the numbers already exist at scattered places in the manuscript. The newly added w/o validity filtering experiments are also convincing.

Given these, I think this work, by introducing a new way of conditioning NMR spectrum that applied to different decoder architectures, brings consistent and substantial performance gains to the de novo task of structure elucidation from NMR. However my concerns in "Critical Weaknesses" still exists: this whole "tweak" for spectral conditioning is practically useful from ML engineering perspectives, delivering a strong applied design choice for the specific task, but will ICLR be the best venue for this work considering it might not be technically novel enough? Therefore I'm leaning towards keeping my original scores, while I would also love to hear how the author might consider this concern too.

AuthorsOpenReview ↗

Reply to xCLd: what the conditioning comparisons establish, and new experimental results

Thank you for clarifying that the capacity, normalisation and validity-filtering concerns have been addressed. We have brought the nominal-budget comparison into the main text and revised the paper around the question in its new title, "What Limits Diffusion Models for NMR Structure Elucidation?"

1. A trained query control addresses the inference-only objection

We added a separately trained shared-query control in NMR-DiGress. It uses one common spectral query for all nodes and another for all candidate bonds within each layer during both training and evaluation.

NMR-DiGress conditioningTop-1 recovery
FiLM43.26%
Shared-query cross-attention39.44%
Element-specific cross-attention52.10%

This supports a contribution from element-specific queries in these graph-model runs. The coordinate-model averaging/permutation tests remain inference-only: they show reliance on trained per-atom contexts but cannot by themselves distinguish that reliance from a training/testing mismatch. The graph control does not establish every coordinate-model mechanism.

2. Task-specific understanding across two backbone families

Cross-attention itself is established, including in NMR. Our contribution is task-specific empirical knowledge about the interface linking unassigned spectral tokens to atom representations and candidate bonds. Atom types are known and fixed: one diffusion model denoises only bond categories, the other denoises 3D atom positions. We use existing backbone families to examine modular changes.

The first row uses a cross-attention checkpoint at 97% of the baseline's nominal matrix-multiply training budget.

ChefNMR evaluationComparator top-1Cross-attention top-1
Simulated SpectraBasePublished ChefNMR-S: 69.15%75.77%
Molecules with unseen ring frameworksChefNMR-S: 45.25%54.42%

The NMR-DiGress result is in the first table. Each backbone retains its own evaluation protocol. On NMR-DiGress, the replacement also changes normalisation, initialisation and parameter count. Both ChefNMR variants use adaLN-Zero. Nominal budgets do not measure all compute consumed.

For scale context, the published 462M-parameter ChefNMR-L reaches 72.04% top-1 recovery. Our 153M-parameter cross-attention model reaches 76.64% at its later 8.4k checkpoint. This descriptive comparison does not match training budgets or isolate the effect of model size.

The main text retains the encoding/conditioning analysis: auxiliary reconstruction has different reported effects under different interfaces, with optimisation and scoring confounds. There is no clear cross-attention gain on measured SpecTeach spectra before further training. ChefNMR already associates lower simulated-to-measured hydrogen similarity with failures. Our paired input replacements test model-recovery sensitivity, which is much greater for measured hydrogen than carbon. NMRViT also shows that fine-tuning on small measured subsets helps.

3. Experimental recovery after continued checkpoint training

We continued training of saved ChefNMR checkpoints on width-modified simulations from 6,000 SpectraBase molecules, measured spectra from 60 molecules selected from the Biological Magnetic Resonance Data Bank (BMRB) collection used in Hu et al.'s structure-elucidation study, and both sources jointly. Those 60 measured molecules are disjoint from SpecTeach. The table uses our corrected, solvent-bin-blanked 213-molecule SpecTeach evaluation copy.

ModelBefore further trainingAfter combined-data fine-tuning
ChefNMR-S59.94%73.55%
Cross-attention ChefNMR-S59.00%80.91%

The cross-attention own-start gain is 21.91 percentage points. Its final result is 20.97 points above ChefNMR-S before further training and 7.36 points above the separately fine-tuned baseline. Continued-training and shuffled-pair controls support useful measured-recovery gains.

4. Empirical contribution and assessment

The study finds useful per-element conditioning gains in two backbone families, shows that gains from simulated spectra do not directly transfer to measured inputs, and tests how further training changes experimental recovery. A separately trained query control and continued-training controls address specific alternatives. We ask that the contribution be assessed on this task-specific knowledge and its stated limits.

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

This paper explores an alternative conditioning mechanism for the NMR structure elucidation task. They use two existing models from Xiong et al, NeurIPS 2025. For NMR structure elucidation, conditional generative models are trained to generate structures conditioned on an input spectra. The published models encode the input spectra into a single vector, which is then used to modulate the generation process. In this work, they use a different conditioning mechanism that allows for cross attention between atom tokens and the spectral tokens. The authors find that this improves structure prediction performance.

Critical strengths

The proposed conditioning mechanism improves ChefNMR and NMR-Digress performance on simulated spectra.

Critical weaknesses

The work is very incremental. The authors take two existing published models and modify their conditioning mechanisms using established cross-attention components. While the empirical performance improves, the methodological novelty is limited, and the results provide relatively little new insight into NMR structure elucidation or conditional generative modeling. In my assessment, the paper does not sufficiently establish a broader contribution beyond improving these particular architectures.

Questions for authors

What new insight about spectral conditioning does this work establish beyond showing that standard cross-attention improves these two architectures?

The substantial gains on simulated spectra do not translate into consistent improvements on experimental spectra. Do you have evidence explaining this discrepancy?

1 comment
AuthorsOpenReview ↗

Reply to Reviewer CeFX: pooled conditioning limits two models on simulated spectra, and on measured spectra the 1H spectrum causes most of the top-1 drop

Thank you for your review.

"What new insight about spectral conditioning does this work establish beyond showing that standard cross-attention improves these two architectures?"

NMR-DiGress and ChefNMR-S give every atom the same pooled summary of the spectrum, and our results show that on simulated spectra this limits them. A larger model or more pooled vectors do not remove this limit, and giving each atom its own view of the spectrum does. We agree that the change itself is simple and uses a standard component. Three results support the finding, and points 1 and 2 of the summary comment have the tables.

  1. It is not specific to one architecture. On simulated spectra, the same change improves a graph diffusion model and a coordinate diffusion model.
  2. A larger pooled model does not reach this accuracy. Compared with ChefNMR-S, ChefNMR-L of Xiong et al. adds 328 M parameters for about 3 points, our model (CA-ChefNMR-S) 19 M for about 7.
  3. A new check: our trained models depend on the per-atom views. If at inference every atom gets the average of these views, top-1 of CA-ChefNMR-S falls from about 77% to about 1% on 3,000 random SpectraBase test molecules. This does not show that a pooled model must fail, because the models were never trained on such inputs.
Top-1 on simulated spectraPooled conditioningCross-attention
NMR-DiGress, our reproduction (generates the bond graph)43.3%52.1%
ChefNMR-S (generates 3D coordinates)69.2%76.6%

We think this is useful beyond these two models. DiffNMR and DiffSpectra, two other diffusion models, also condition on a pooled vector. We also find that a reconstruction loss, which makes the encoder retain spectral detail, hurts with one pooled vector (NMR-DiGress loses about 3 points) and helps with cross-attention. With the loss, our ChefNMR model (CA-ChefNMR-S) gains 1.65 points over the same model trained at least as long without it.

"The substantial gains on simulated spectra do not translate into consistent improvements on experimental spectra. Do you have evidence explaining this discrepancy?"

Yes, two results, with the tables in point 3 of the summary comment.

  • 165 of the 213 SpecTeach molecules are also in the training set. ChefNMR-S and our model solve about 98% of them from simulated spectra, and 94% to 96% of the 44 others with a simulated spectrum, so little improvement could carry over.
  • Most of the drop on measured SpecTeach data comes from the measured 1H spectrum: replacing only the simulated 1H spectrum of a molecule by the measured one costs about 32 to 34 of the roughly 41 to 42 points in total, replacing only the 13C input about 1 point. One difference is the line width: broadening the simulated lines to the measured width, with the simulated 13C input also given, costs 14 to 15 points.

We cannot show an improvement from cross-attention on experimental spectra, and in the revised paper, which we will upload before the discussion period ends, we will state this more clearly. We also found a one-bin offset in the experimental 13C inputs, which changes the SpecTeach numbers but not this conclusion.

"Presentation: 1: Difficult to follow despite careful reading."

We agree that the paper was hard to follow. In the revised paper we will state the main idea and the main result first, introduce the model names one at a time, and move details that the argument does not need to the appendix. We will also split Table 1 by backbone and give one definition of top-1 next to every table. If this answers your concerns, we would be grateful if you reconsidered your score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of our response: giving each atom its own view of the spectrum improves NMR-DiGress and ChefNMR-S by 7 to 9 points of top-1 on simulated spectra, and on measured SpecTeach spectra ChefNMR-S and our model are not distinguishable

We thank the reviewers for their careful reading. We will upload the revised paper before the discussion period ends on 18 November. Its changes are listed at the end of this comment.

1. What we find

NMR-DiGress and ChefNMR-S compress the spectrum into one pooled vector, and every atom receives the same vector. In cross-attention every atom instead reads the spectrum with its own query. On simulated spectra this helps more than twice as much as a larger pooled model (ChefNMR-L) or more pooled vectors, although cross-attention itself is a standard component.

Change to the pooled modelModelChange in top-1 on simulated spectra
ChefNMR-L instead of ChefNMR-S (Xiong et al.)ChefNMR+2.89 points
Three pooled vectors instead of oneNMR-DiGress+1.47 points
Cross-attentionChefNMR+7.24 points
Cross-attentionNMR-DiGress+8.84 points

Top-1 is the share of test molecules whose first generated structure (for NMR-DiGress the first valid one) is correct, and points are percentage points. The +7.24 compares our model after the same training time with the released ChefNMR-S, both evaluated by us. All models are trained and tested on SpectraBase, the simulated data set of Hu et al. (2024). In the paper our models are CA-NMR-DiGress and CA-ChefNMR-S, and ChefNMR-S is adaLN-ChefNMR-S. CA-ChefNMR-S is trained from scratch, without the reconstruction loss.

2. For our ChefNMR model (CA-ChefNMR-S), the improvement holds after the same training time, against a pooled model three times our size, and for unseen scaffolds

The table gives top-1 on simulated spectra, for the same test molecules as Xiong et al.

ChefNMR modelsParametersAll 14,137 test moleculesUnseen-scaffold list of Xiong et al. (3,770)
ChefNMR-S, published134 M69.15%64.09%
ChefNMR-L, published462 M72.04%66.40%
CA-ChefNMR-S (ours), about the same training FLOPs as ChefNMR-S153 M75.77%
CA-ChefNMR-S, same wall-clock training time as ChefNMR-S153 M76.48%
CA-ChefNMR-S, last checkpoint153 M76.64%71.22%
  • After the same wall-clock training time our model is about 7 points above ChefNMR-S. Table 11 of the paper already reports these three results.
  • ChefNMR-L is about 5 points below our model on all test molecules and on the unseen-scaffold list.
  • 1,788 molecules of the list have rings, and none of their ring scaffolds occurs in the training data. Of these our model solves about 54% and the released ChefNMR-S about 45% in our evaluation.
  • We do not have a pooled model of exactly our size or a comparison at matched inference cost.

3. On measured spectra we show no improvement, and we correct our numbers

SpecTeach is the experimental benchmark of Xiong et al. (2025, data of Van Bramer and Bastin, 2023). We found that its 13C inputs in the ChefNMR data release were binned with another rule than the training inputs, which shifts every signal by one bin. With the offset corrected, the two models are not distinguishable.

Top-1 on the 213 SpecTeach moleculesChefNMR-SCA-ChefNMR-S
As in the submission41.2%42.9%
With the offset corrected55.2%54.9%

Two results explain why the improvement is not seen here. Both use the corrected inputs.

First, both models already solve most SpecTeach molecules from simulated spectra, so little improvement could carry over. Ranges span ChefNMR-S and CA-ChefNMR-S.

SpecTeach moleculesTop-1 from simulated spectraTop-1 from measured spectra
165 in the SpectraBase training setabout 98%about 55%
44 not in the training set, with a simulated spectrum (the other 4 have none)94% to 96%58% to 62%

Second, most of the drop comes from the measured 1H spectrum. The last table starts from the simulated inputs of these molecules and changes one thing at a time.

Change to the simulated inputsDrop in top-1
Only the 1H spectrum replaced by the measured one32 to 34 points
Only the 13C input replacedabout 1 point
Both replaced41 to 42 points
1H lines only broadened to the measured median width14 to 15 points

One difference is the line width: Hu et al. simulated the training spectra with one line width of 0.75 Hz.

In the revised paper we will:

  1. state the comparison at matched training compute in the main text.
  2. correct the SpecTeach numbers and the label "different cohort", state that experimental spectra give no clear gain and report the corrected NMRShiftDB2 runs.
  3. state the theory and masking limits next to the lemmas and present the reconstruction-loss results as a trend.
  4. put the contribution and main result first, introduce the model names one at a time and split Table 1 by backbone.
  5. add the new results of this response.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.