acceptodds
Under review as a conference paper at ICLR 2027

NMR-CRAFT: Structure Elucidation through Bidirectional Modeling and Fragment Recombination

Abstract

Nuclear magnetic resonance (NMR) spectroscopy is fundamental to molecular structure determination, yet recovering complete structures from experimental spectra remains challenging, especially for unseen molecular scaffolds. Motivated by recent advances in large language models (LLMs), we first show that adapting pretrained LLMs to NMR data yields strong structure elucidation performance. A closer analysis of prediction errors reveals two complementary patterns: the correct structure might be generated but not ranked first, while incorrect predictions also preserve substantial parts of the correct structure. These observations suggest that better structure elucidation requires not only stronger generation, but also the ability to reuse partial hypotheses and verify candidate structures. We therefore introduce **NMR-LM**, a unified language model trained in both directions: spectrum-to-structure prediction generates structural hypotheses, while structure-to-spectrum modeling evaluates their consistency with the observed spectrum. Building on these capabilities, we develop **NMR-CRAFT**, an efficient test-time procedure that recombines useful fragments from generated and retrieved candidates into new hypotheses and ranks them by spectral likelihood. **NMR-LM** achieves state-of-the-art structure elucidation performance on NMRGym and NMRSpec, with competitive spectral prediction. **NMR-CRAFT** further improves Top-1 accuracy by 120% over direct generation on scaffold-split NMRGym, with a speedup exceeding over the reference search. These results highlight our method's potential as a practical tool for NMR structure elucidation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
NMR, Structure elucidation, Language models, Constrained search

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_sAmqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper studies structure elucidation from experimental ¹H/¹³C NMR chemical-shift lists. It fine-tunes Qwen3.5-4B jointly on spectrum→SMILES and SMILES→spectrum (NMR-LM), then applies DPO with ground-truth positives and the most similar formula-matched incorrect sample as the negative. On top of NMR-LM, NMR-CRAFT is a test-time search: candidates from beam search (32), sampling (32) and spectral retrieval from SimNMR-PubChem (50, using NMR-Solver's database and retrieval) are fragmented; fragment pairs whose heavy-atom counts sum to the target formula are reassembled; up to 100 candidates are preselected by a rank fusion of parent-source priors and recombination support; and the final ranking mixes this prior with the mean log-likelihood of the observed spectrum under NMR-LM's forward direction. On scaffold-split NMRGym with Exclude-Scaffold retrieval, Top-1 rises from 3.75% (NMR-LM, beam 10) to 8.23%; on random-split NMRSpec, from 47.80% to 58.51%. On a 1,000-molecule NMRGym subset, NMR-CRAFT reaches 8.1% Top-1 in 14.2 s per molecule versus 2.9% and 222–260 s for a reproduced NMR-Solver under database exclusion. NMR-LM is also evaluated for spectrum prediction with set- and vector-similarity metrics.

Critical strengths
  1. Careful leakage control and unusually complete reproducibility. The paper recognises that SimNMR-PubChem contains 74.9% of NMRGym and 33.2% of NMRSpec test molecules (l. 340), applies Exclude-Scaffold / Exclude-Molecule masks during retrieval (Appendix C.1), and reports how sharply NMR-Solver drops when test molecules are removed (Table 5: 12.50% → 2.90% Top-1). The authors release the checkpoint, the full NMR-CRAFT outputs and the code. This makes the numbers checkable and is a real service to a field where leakage through large spectral databases is common.

  2. The forward-likelihood verification signal is shown to matter. Removing spectral likelihood from the final ranking halves Top-1 on the 1,000-molecule subset (8.1% → 4.1%, Table 22), and both endpoints of the λ_rank sweep (4.1% and 4.3%) are far below the mixed setting (Fig. 9f). Together with the NMRSpec stratification by molecular size and training-set similarity with bootstrap CIs (Fig. 3a/b, Table 14), this gives credible evidence that a bidirectionally trained model provides a useful candidate-verification signal, including for larger and less familiar molecules.

Critical weaknesses

1. The comparison with NMR-Solver is not "under the same protocol" (as claimed in l. 087, 099 and 483), so the accuracy and 15× speed claims are not attributable to the method.

  • Formula access differs. NMR-CRAFT uses the molecular formula as a hard constraint inside the search: fragment pairs must sum to the target heavy-atom counts and products must match the formula exactly (Eq. 7; l. 218–228). For NMR-Solver, only the allowed element types are given, and the formula is applied only as an output filter that does not modify the internal FB-MO search (Appendix B.6, l. 902–911). In structure elucidation the formula is a very strong constraint, so this asymmetry alone could explain much of the gap.
  • A constraint of NMR-Solver is switched off. Hydrogen-multiplicity constraints are disabled (l. 903).
  • Budgets and hardware differ. NMR-Solver is run with 1,000 retrieved candidates, up to 200,000 fragment pairs and a 1,000-molecule pool (Table 10), whereas NMR-CRAFT uses 50 retrieved candidates, 1,000 pairs and 100 candidates. NMR-CRAFT's two model services run on separate A100s of a 4-GPU server (l. 829–831), while the NMRNet ensemble runs on a single A100 (l. 899–900). The 15× speed-up therefore largely reflects budget choices rather than an algorithmic advantage; an accuracy–time curve at matched budgets is needed.
  • The reproduction looks weak. Unrestricted full NMR-Solver obtains 12.50% Top-1 (Table 5), lower than its own search-only component with formula filtering in Table 1 (17.92%). The subsets differ, but a full pipeline performing much worse than its retrieval stage suggests a configuration problem that needs to be explained.

2. Search hyper-parameters appear to have been tuned on the same test molecules that are used to report results, and most workflow ablations are within noise.

  • Appendix D.2 (l. 1263–1273) states that the five search budgets and λ_rank were varied on the same 1,000 NMRGym test molecules used for Table 5. The defaults (λ_rank = 0.32, source weights 1.00/0.65/0.35, δ = 4, κ = 60; l. 841, 871, 884) are not said to be chosen on validation data, and the chosen λ_rank is exactly the best value in that test-set sweep. If so, Tables 1, 5 and 22 are optimistically biased.
  • On 1,000 molecules at p ≈ 0.08, the standard error of Top-1 is about 0.86 points. The ablations "w/o database" (−0.9), "w/o sampling" (−0.7) and "w/o support" (−1.0) in Table 22 are each about one standard error (7–10 molecules), and no paired test is reported. Only the likelihood ablation is clearly significant, so the claim that database retrieval, sampling and the support signal are complementary (l. 466–468) is not established.
  • The headline "120% improvement over direct generation" compares against beam-10 decoding. Table 4 shows that 32-beam candidates with formula filtering and likelihood re-ranking already give 5.36%, so retrieval plus recombination adds about 2.9 points (≈54%). With 8.23% Top-1 and 17.34% Top-10 on scaffold-split NMRGym, the description as a potentially practical tool (l. 032) overstates the current capability.
Questions for authors

Clarifications that could change my assessment

  1. Were λ_rank, the source weights, δ, κ, B_pair and B_cand chosen on the NMRGym validation split or on test molecules? Please report test results with all search hyper-parameters selected on validation data only.
  2. Can NMR-Solver's official implementation take the molecular formula as a search constraint (and hydrogen multiplicities)? If so, please re-run Table 5 with these enabled and with candidate/pair budgets matched to NMR-CRAFT on the same hardware, and report accuracy versus wall-clock time.
  3. Why does full NMR-Solver without exclusion (12.50% Top-1, Table 5) underperform its own search-only + formula component (17.92%, Table 1)?
  4. Please report paired confidence intervals (e.g., bootstrap or McNemar) for the Table 22 ablations.
  5. Spectrum prediction: for per-atom predictors (CASCADE, DetaNet, NMRNet, Mnova), were symmetry-equivalent atoms merged to match the reference format of one ¹³C shift per distinct environment? How many molecules failed for each method, and what are the results on the subset where all methods produce a prediction? Please also report per-atom MAE (ppm), which is the standard metric in this area.
  6. On NMRSpec, do the NMRTrans/NMRMind numbers (taken from the original papers) use the molecular formula and the multiplicity/J-coupling information that NMR-LM receives (Appendix B.1)? Since you reproduced NMRTrans (Appendix D.1), please also report it on NMRGym.

Suggestions (not required for my assessment)

  1. How are acyclic molecules (empty Bemis–Murcko scaffold) handled by Exclude-Scaffold, and is Exclude-Molecule based on stereo-aware SMILES or InChIKey connectivity? Does exact match require correct stereochemistry?
  2. Please qualify "state-of-the-art" in the abstract as in the introduction ("among the evaluated generative models"), and state the absolute Top-1/Top-10 next to the relative gain.
8 comments
AuthorsOpenReview ↗

Response to Reviewer sAmq (Part 1/3)

We thank the reviewer for the detailed and thoughtful review. We have addressed the key concerns and provide our responses below.

W1-1: Formula access differs. NMR-CRAFT uses the molecular formula as a hard constraint inside the search: fragment pairs must sum to the target heavy-atom counts and products must match the formula exactly (Eq. 7; l. 218–228). For NMR-Solver, only the allowed element types are given, and the formula is applied only as an output filter that does not modify the internal FB-MO search (Appendix B.6, l. 902–911). In structure elucidation the formula is a very strong constraint, so this asymmetry alone could explain much of the gap.

Thank you for pointing out this asymmetry. Our original NMR-Solver reproduction followed its original FB-MO search procedure, applying molecular formula filtering only to the final outputs. To assess the impact of this difference, we further incorporated exact molecular formula constraints into NMR-Solver's fragment pairing and product validation. We evaluated both variants on the same 1,000 NMRGym test molecules under the Exclude-Scaffold protocol.

SettingTop-1Top-5Top-10Time (s/sample)
NMR-Solver (Original)2.905.205.90260.34
NMR-Solver + Formula Constraints6.008.008.00195.96
NMR-CRAFT8.1014.8017.5014.11

Internal formula constraints substantially improve NMR-Solver, confirming that the original asymmetry accounts for part of the performance gap. Nevertheless, NMR-CRAFT remains stronger across all Top-k metrics and substantially faster. We have added the formula-aware comparison and search-budget analysis to Appendix D.6.

W1-2: A constraint of NMR-Solver is switched off. Hydrogen-multiplicity constraints are disabled (l. 903).

Thank you for raising this concern. We followed the official NMR-Solver configuration (use_H_split=False), consistent with its original evaluation protocol, which excludes multiplicity and J-coupling information. NMRGym provides chemical-shift lists without explicit splitting-pattern or J-coupling annotations. Thus, enabling multiplicity-aware scoring would require information unavailable in this benchmark.

W1-3: Budgets and hardware differ. NMR-Solver is run with 1,000 retrieved candidates, up to 200,000 fragment pairs and a 1,000-molecule pool (Table 10), whereas NMR-CRAFT uses 50 retrieved candidates, 1,000 pairs and 100 candidates. NMR-CRAFT's two model services run on separate A100s of a 4-GPU server (l. 829–831), while the NMRNet ensemble runs on a single A100 (l. 899–900). The 15× speed-up therefore largely reflects budget choices rather than an algorithmic advantage; an accuracy–time curve at matched budgets is needed.

Q2: Can NMR-Solver's official implementation take the molecular formula as a search constraint (and hydrogen multiplicities)? If so, please re-run Table 5 with these enabled and with candidate/pair budgets matched to NMR-CRAFT on the same hardware, and report accuracy versus wall-clock time.

Thank you for these questions. Our original NMR-Solver reproduction follows its official protocol, with formula filtering applied after search. We additionally incorporated the exact formula into its fragment-based molecular optimization (FB-MO) and evaluated different search budgets on the same 1,000 NMRGym test molecules under Exclude-Scaffold.

MethodBudget (R/P/C)Top-1Top-5Top-10Time (s)
NMR-Solver (original)1000/200k/10002.95.25.9260.34
NMR-Solver (formula-aware)50/1k/1000.00.00.010.10
NMR-Solver (formula-aware)500/10k/5002.52.72.744.81
NMR-Solver (formula-aware)1000/200k/10006.08.08.0195.96
NMR-CRAFT50/1k/1008.114.817.514.11

R/P/C denotes the limits on retrieved candidates, fragment pairs, and candidate pool size, respectively. For NMR-Solver, R applies only to the initial retrieval, while the same P and C limits apply to each FB-MO iteration. Accuracy is reported in percent and time in seconds per molecule. Formula-aware search improves NMR-Solver, but NMR-CRAFT achieves a favorable accuracy–runtime trade-off. A complete Pareto analysis would require more extensive budget sweeps.

Both methods ran on the same four-A100 server. NMR-Solver used one GPU, while NMR-CRAFT originally used separate GPUs for sequential inverse and forward inference. Running both services on one GPU yielded 14.11 s/sample, comparable to the original 14.24 s/sample and approximately 14× faster than formula-aware NMR-Solver at its original budget.

As noted in W1-2, we retain the official use_H_split=False setting because NMRGym lacks explicit splitting-pattern and J-coupling annotations. We have clarified the evaluation configurations and added the budget comparisons to Appendix D.6.

AuthorsOpenReview ↗

Response to Reviewer sAmq (Part 2/3)

W1-4: The reproduction looks weak. Unrestricted full NMR-Solver obtains 12.50% Top-1 (Table 5), lower than its own search-only component with formula filtering in Table 1 (17.92%). The subsets differ, but a full pipeline performing much worse than its retrieval stage suggests a configuration problem that needs to be explained.

Q3: Why does full NMR-Solver without exclusion (12.50% Top-1, Table 5) underperform its own search-only + formula component (17.92%, Table 1)?

Thank you for raising this concern. Table 1 reports results on the full NMRGym test set (27,054 molecules), whereas Table 5 uses a 1,000-molecule subset. For a direct comparison, we evaluated the formula-filtered search-only baseline on the same subset.

SettingTop-1Top-5Top-10
Search-only + Formula13.932.735.7
Full NMR-Solver12.531.536.5

The Top-1 difference is 1.4 percentage points. Among the 31 cases correctly ranked first by search-only but not by the full pipeline, all ground-truth structures remain in the final candidate lists but are ranked lower. Conversely, the full pipeline correctly ranks 17 additional molecules first and achieves slightly higher Top-10 accuracy. Thus, the difference primarily reflects changes in candidate ranking rather than the loss of correctly retrieved structures.

W2-1: Appendix D.2 (l. 1263–1273) states that the five search budgets and λ_rank were varied on the same 1,000 NMRGym test molecules used for Table 5. The defaults (λ_rank = 0.32, source weights 1.00/0.65/0.35, δ = 4, κ = 60; l. 841, 871, 884) are not said to be chosen on validation data, and the chosen λ_rank is exactly the best value in that test-set sweep. If so, Tables 1, 5 and 22 are optimistically biased.

Q1: Were λ_rank, the source weights, δ, κ, B_pair and B_cand chosen on the NMRGym validation split or on test molecules? Please report test results with all search hyper-parameters selected on validation data only.

Thank you for raising this concern. The search hyperparameters were selected using validation data, not test data. Specifically, we selected Beam/Sampling/Database budgets (32/32/50), B_pair (1,000), B_cand (100), and λ_rank (0.32) based on Top-10 sequence accuracy on a subset of 2,000 NMRGym validation molecules. The source weights (1.00/0.65/0.35) and smoothing constants (δ = 4, κ = 60) were fixed heuristically rather than separately tuned.

The test-set sensitivity analysis in Appendix Fig. 9 was conducted only to assess robustness after parameter selection. We have replaced it with the validation-set results and clarified that all reported test results use hyperparameters fixed independently of test-set performance.

W2-2: On 1,000 molecules at p ≈ 0.08, the standard error of Top-1 is about 0.86 points. The ablations "w/o database" (−0.9), "w/o sampling" (−0.7) and "w/o support" (−1.0) in Table 22 are each about one standard error (7–10 molecules), and no paired test is reported. Only the likelihood ablation is clearly significant, so the claim that database retrieval, sampling and the support signal are complementary (l. 466–468) is not established.

Q4: Please report paired confidence intervals (e.g., bootstrap or McNemar) for the Table 22 ablations.

Thank you for raising this concern. We computed 95% paired bootstrap confidence intervals using 10,000 resamples of the same 1,000 NMRGym test molecules. The sequence accuracy differences (ablation minus full NMR-CRAFT, in percentage points) are reported below.

AblationTop-1 Δ [95% CI]Top-10 Δ [95% CI]
w/o Database−0.9 [−2.0, 0.2]−2.1 [−3.4, −0.8]
w/o Sampling−0.7 [−1.5, 0.1]−1.2 [−2.3, −0.1]
w/o Support−1.0 [−1.8, −0.2]−0.8 [−1.7, 0.1]
w/o Likelihood−4.0 [−5.6, −2.5]−3.5 [−4.8, −2.3]

Removing spectral likelihood produces the largest and most consistent degradation. The other components show smaller, metric-dependent effects, with some confidence intervals including zero. Together with the fragment-source analysis in Fig. 5(a,b), these results provide further evidence for the complementary roles of candidate sources and ranking signals. We have added the complete paired confidence intervals to Appendix D.3.

AuthorsOpenReview ↗

Response to Reviewer sAmq (Part 3/3)

W2-3: The headline "120% improvement over direct generation" compares against beam-10 decoding. Table 4 shows that 32-beam candidates with formula filtering and likelihood re-ranking already give 5.36%, so retrieval plus recombination adds about 2.9 points (≈54%). With 8.23% Top-1 and 17.34% Top-10 on scaffold-split NMRGym, the description as a potentially practical tool (l. 032) overstates the current capability.

Thank you for raising this concern. As the reviewer noted, the reported 120% improvement is calculated relative to the NMR-LM Beam-10 baseline in Table 1. We agree that the stronger Beam-32 + Formula + Likelihood baseline provides additional context, against which NMR-CRAFT improves Top-1 accuracy from 5.36% to 8.23% (+2.87 percentage points).

While absolute accuracy on scaffold-split NMRGym remains limited, NMR-CRAFT improves Top-10 accuracy from 9.51% to 17.34% on NMRGym and from 66.48% to 79.67% on NMRSpec, demonstrating its value for candidate refinement. We have clarified the comparison in the abstract and added the Beam-32 + Formula-filter baseline to Table 1.

Q5: Spectrum prediction: for per-atom predictors (CASCADE, DetaNet, NMRNet, Mnova), were symmetry-equivalent atoms merged to match the reference format of one ¹³C shift per distinct environment? How many molecules failed for each method, and what are the results on the subset where all methods produce a prediction? Please also report per-atom MAE (ppm), which is the standard metric in this area.

Regarding symmetry-equivalent atoms, we adopt the aggregate results reported by NMRGym. Its paper describes Hungarian matching for peak-set evaluation but does not explicitly specify whether symmetry-equivalent atoms are merged before matching. For the MAE comparison, we use Hungarian matching with an L1 chemical-shift distance.

According to the coverage reported by NMRGym, the failure rates are 2.04% for all three CASCADE variants (both nuclei), 14.41% for DetaNet (both nuclei), 0% for NMRNet (both nuclei), and 0.09%/0.59% for Mnova (¹H/¹³C, respectively). A common-subset comparison is not available because the published NMRGym results do not provide per-molecule predictions or validity masks for all baselines.

We additionally reproduced NMRNet using its five released checkpoints and compared it with NMR-LM on NMRGym. The MAE results are:

ModelNucleusMean MAE
NMR-LM¹H0.2237 ppm
NMRNet¹H0.1724 ppm
NMR-LM¹³C2.1925 ppm
NMRNet¹³C2.4306 ppm

NMRNet has no failed predictions for either nucleus, while NMR-LM has a failure rate of 0.314% for ¹H and 0% for ¹³C. NMR-LM has higher ¹H MAE but lower ¹³C MAE than NMRNet. We have added the additional spectral prediction comparison to Appendix D.1.

Q6: On NMRSpec, do the NMRTrans/NMRMind numbers (taken from the original papers) use the molecular formula and the multiplicity/J-coupling information that NMR-LM receives (Appendix B.1)? Since you reproduced NMRTrans (Appendix D.1), please also report it on NMRGym.

Table 2 reports NMRTrans and NMRMind results from the original NMRTrans paper. The generative baselines were trained on the NMRSpec training split using their official implementations with molecular-formula conditioning. NMRTrans explicitly uses ¹H chemical shifts, integration, multiplicity, and J-couplings, together with ¹³C shifts. However, the paper does not explicitly specify whether NMRMind uses the same multiplicity and J-coupling information.

For our NMRTrans reproduction in Appendix D.1, we used the released checkpoint and official implementation for inference, without additional training. We did not report NMRTrans on NMRGym because NMRGym provides chemical-shift lists and molecular formulas but lacks the explicit multiplicity and J-coupling information required by the original NMRTrans input representation. A comparable evaluation on NMRGym would therefore require input adaptation and retraining.

Q7: How are acyclic molecules (empty Bemis–Murcko scaffold) handled by Exclude-Scaffold, and is Exclude-Molecule based on stereo-aware SMILES or InChIKey connectivity? Does exact match require correct stereochemistry?

For Exclude-Scaffold, none of the 27,054 NMRGym test molecules has an empty Bemis–Murcko scaffold, so this case does not arise in our evaluation. Exclude-Molecule uses RDKit canonical isomeric SMILES rather than InChIKey connectivity, preserving stereochemical distinctions in both NMRGym and NMRSpec. Exact-match accuracy also accounts for stereochemistry, so predictions with incorrect stereochemistry are counted as mismatches.

Q8: Please qualify "state-of-the-art" in the abstract as in the introduction ("among the evaluated generative models"), and state the absolute Top-1/Top-10 next to the relative gain.

Thank you for the suggestion. We have qualified the state-of-the-art claim in the abstract and added the absolute Top-1/Top-10 accuracies in the Introduction (line 88).

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

In this work, Qwen3.5-4B is fine-tuned on the tasks of predicting SMILES from NMR spectra and the other way around using direct preference optimization (DPO) with max-Tanimoto negatives. They pool candidates from their beam search, sampling, database retrieval and fragment retrieval. Those candidates are then reranked by Qwen3.5-4B by the likelihood of the observed spectrum. They show an improvement in NMRGym for NMR-CRAFT over NMR-LM and most other baselines and in NMRSpec over all baselines.

Critical strengths
  1. Their results consistently show the benefits of NMR-CRAFT over other baselines and provide many ablations. We see a consistent gain over NMRGym and NMRSpec (Table 1/2) for Top-1, Top-5, and Top-10 (Table 1/2), and for molecule size (Figure 3). An interesting phenomenon shown by the ablation results is that using only preliminary ranking or only the NMR-spectrum likelihood under the proposed SMILES performs only about half as well as using a linear combination of both (Figure 9). A mechanism that would be worth studying.

  2. The authors use the same model (Qwen3.5-4B) to generate SMILES from NMR spectra and the other way around, for which they show that there is no measurable cost (Figure 4 and Table 16/17).

Critical weaknesses
  1. Most in the paper is already known. While the authors are honest about it, this still limits the relevance of the paper. A bidirectional spectrum-structure LM has already been proposed (Li et al. 2026, "SpecMol: A unified framework for spectroscopy-grounded molecular modeling and evaluation"). The training objectives are known, e.g., DPO (Rafailov et al. 2023, "Direct Preference Optimization: Your Language Model is Secretly a Reward Model") or L^+_SFT (seems to be the NLL loss from Pang et al. 2024, "Iterative Reasoning Preference Optimization"). The fragmentation with recombination and spectral matching/verification was already proposed by NMR-Solver

  2. The one true novel behavior mentioned in "Critical Strengths" in point 1, is not studied. Also, the improvement from 3.75 to 8.23 (Table 1) is never cleanly attributed. Table 4 shows that a beam of 32 with a formula filter already gives 5.13, but the recombination stage is never ablated on its own.

Questions for authors
  1. Why do preliminary ranking and NMR spectra likelihood rank poorly separately but improve when combined linearly?
  2. What are the confidence intervals for Tables 4/5 and Figures 4/5(c)?
  3. Were the baselines from Table 3 retrained on NMRGym or evaluated off-the-shelf?
  4. Why did you decide to use the max-Tanimoto negatives for equation (3)? Did you try other criteria? Also, what fraction of the negatives are Enantiomers? Enantiomers give identical spectra in achiral solvents (Dale and Mosher 1973, "Nuclear magnetic resonance enantiomer Reagents..."). Also, are enantiomers considered misses for the ground truth?
  5. How was the pair/candidate scoring decided on (Eq. 10-13)? There would be other possibilities, like the sum or mean over the parents instead of the max in Eq. 11).
6 comments
AuthorsOpenReview ↗

Response to Reviewer jHrT (Part 1/2)

We appreciate the reviewer's insightful observations and constructive suggestions. We have carefully considered these points and respond below.

W1: Most in the paper is already known. While the authors are honest about it, this still limits the relevance of the paper. A bidirectional spectrum-structure LM has already been proposed (Li et al. 2026, "SpecMol: A unified framework for spectroscopy-grounded molecular modeling and evaluation"). The training objectives are known, e.g., DPO (Rafailov et al. 2023, "Direct Preference Optimization: Your Language Model is Secretly a Reward Model") or L^+_SFT (seems to be the NLL loss from Pang et al. 2024, "Iterative Reasoning Preference Optimization"). The fragmentation with recombination and spectral matching/verification was already proposed by NMR-Solver.

Thank you for this thoughtful comment. We acknowledge that bidirectional modeling, preference optimization, and fragment recombination build on prior work. Our contribution is not a fundamentally new formulation of these individual techniques, but a budgeted, LLM-guided framework for reusing and verifying imperfect molecular hypotheses.

NMR-CRAFT develops a test-time structure refinement procedure that reuses and verifies imperfect molecular hypotheses, which SpecMol does not explore. Compared with NMR-Solver, our framework introduces (1) LLM-generated hypotheses as additional fragment sources alongside database retrieval, (2) source-aware, formula-constrained recombination with support-guided prioritization for efficient, budgeted search, and (3) direct forward-likelihood verification using the same bidirectional LLM, without relying on inherited chemical shifts.

Together, these design choices enable effective integration of molecular hypothesis generation, fragment recombination, and spectral verification, yielding consistent improvements on NMRGym and NMRSpec (Tables 1–2). Removing likelihood-based reranking reduces Top-1 accuracy from 8.1% to 4.1% (Fig. 5(c)), highlighting the value of spectral verification. Our methodological contribution lies in this efficient test-time refinement framework rather than in the novelty of its individual components.

W2: The one true novel behavior mentioned in "Critical Strengths" in point 1, is not studied. Also, the improvement from 3.75 to 8.23 (Table 1) is never cleanly attributed. Table 4 shows that a beam of 32 with a formula filter already gives 5.13, but the recombination stage is never ablated on its own.

Thank you for raising this concern. We further analyzed the complementarity of preliminary ranking and spectral likelihood in our response to Q1. We agree that the overall improvement from 3.75% to 8.23% involves multiple components and does not isolate the contribution of fragment recombination.

To address this, we conducted a controlled ablation on all 27,054 NMRGym test molecules. The w/o Recombination variant retains the same initial proposals (Beam + Sampling + Database), formula filtering, candidate budget (up to 100), and likelihood-based reranking.

SettingTop-1Top-5Top-10
NMR-LM (Beam-10)3.757.759.51
NMR-LM (Beam-32) + Formula5.139.9511.37
NMR-CRAFT (w/o Recombination)6.5811.4712.60
NMR-CRAFT8.2315.1417.34
Full − w/o, Δ [95% paired CI]+1.64 [1.40, 1.87]+3.67 [3.41, 3.94]+4.73 [4.45, 5.01]

Recombination also recovered 2,190 ground-truth structures absent from the initial candidate pools, of which 1,345 reached the final Top-10. These results show that fragment recombination expands candidate coverage and provides additional accuracy gains beyond the other workflow components. We have added the controlled ablation and stage-wise coverage analysis to Appendix D.5.

Q1: Why do preliminary ranking and NMR spectra likelihood rank poorly separately but improve when combined linearly?

Thank you for the question. Preliminary ranking and spectral likelihood capture complementary aspects of candidate quality: the former reflects candidate provenance and fragment-level support, while the latter evaluates consistency with the observed spectrum.

We further analyzed their complementarity on the same 1,000 NMRGym test molecules.

RankingTop-1Top-10
Preliminary only4.114.0
Likelihood only4.314.5
Combined8.117.5

The combined ranking improves Top-1 accuracy by 4.0 percentage points over preliminary ranking (95% paired CI: [2.5, 5.6]) and by 3.8 points over likelihood-only ranking (95% paired CI: [2.4, 5.3]). Among the 199 molecules whose ground-truth structures appear in the candidate pool, conditional Top-1 accuracy increases from 20.6%/21.6% to 40.7%. Notably, 22 molecules are ranked correctly only by the combined score, supporting the complementary roles of the two signals. We have added this analysis to Appendix D.2.

AuthorsOpenReview ↗

Response to Reviewer jHrT (Part 2/2)

Q2: What are the confidence intervals for Tables 4/5 and Figures 4/5(c)?

Thank you for the suggestion. We computed molecule-level 95% confidence intervals for Tables 4–5 and Figures 4(a–c)/5(c), including paired confidence intervals for key method comparisons. The main results are summarized below (differences in percentage points).

ComparisonTop-1 Δ [95% CI]Top-10 Δ [95% CI]
Table 4: NMR-CRAFT − NMR-LM+4.48 [4.17, 4.78]+7.83 [7.49, 8.16]
Fig. 4(a): Joint SFT+DPO − Joint SFT+1.23 [0.99, 1.46]+2.51 [2.19, 2.82]
Fig. 4(c): DPO (Ours) − DPO (Formula)+0.10 [0.01, 0.19]+0.41 [0.26, 0.55]
Fig. 5(c): Full − w/o Likelihood+4.00 [2.50, 5.60]+3.50 [2.30, 4.80]

For Table 5, under Exclude-Scaffold, NMR-CRAFT achieves 8.10% [6.50, 9.90] Top-1 and 17.50% [15.10, 20.00] Top-10 accuracy, compared with 2.90% [1.90, 4.00] and 5.90% [4.50, 7.40] for NMR-Solver.

For Fig. 4(b), DPO produces only small changes relative to Joint SFT in spectral set similarity (¹H: +0.14 [−0.01, 0.29]; ¹³C: +0.02 [−0.04, 0.10]). Other workflow ablations in Fig. 5(c) show smaller effects, with some paired confidence intervals including zero.

We have added detailed confidence intervals for these results in Appendix D.3.

Q3: Were the baselines from Table 3 retrained on NMRGym or evaluated off-the-shelf?

Thank you for the question. The baseline results in Table 3 are taken directly from NMRGym [1] (Table 6), without retraining these models ourselves.

Q4: Why did you decide to use the max-Tanimoto negatives for equation (3)? Did you try other criteria? Also, what fraction of the negatives are Enantiomers? Enantiomers give identical spectra in achiral solvents (Dale and Mosher 1973, "Nuclear magnetic resonance enantiomer Reagents..."). Also, are enantiomers considered misses for the ground truth?

Thank you for these questions. We chose max-Tanimoto selection to construct challenging, formula-matched negatives that are structurally similar to the ground truth. Fig. 4(c) and Table 21 already compare DPO (Random), DPO (Formula), and DPO (Ours). We further evaluated min-Tanimoto and max-token-accuracy selection, both using formula-matched negatives. Their Top-1/5/10 sequence accuracies are summarized below.

MethodTop-1Top-5Top-10
DPO (Random)3.106.798.18
DPO (Formula)3.657.649.10
DPO (Min-Tanimoto)3.477.018.46
DPO (Max-Token)3.727.699.35
DPO (Ours)3.757.759.51

DPO (Ours), using max-Tanimoto negatives, achieves the best results among the evaluated criteria, while performing comparably to max-token-accuracy selection.

Among the 3,970 DPO negatives, 79 form strict enantiomeric pairs with their corresponding ground truths (1.99%). We acknowledge that such pairs are indistinguishable by conventional NMR in achiral solvents. Our exact-match evaluation uses stereochemistry-aware canonical SMILES; therefore, predictions with incorrect stereochemistry are counted as mismatches when the ground truth specifies the configuration. We have added these analyses to Appendix D.2.

Q5: How was the pair/candidate scoring decided on (Eq. 10-13)? There would be other possibilities, like the sum or mean over the parents instead of the max in Eq. 11).

Thank you for the question. Our scoring design considers both candidate provenance and recombination support.

  • Eq. 10 assigns each parent its best source-weighted rank score.
  • Eq. 11 selects the highest-scoring valid parent assignment based on the sum of the two parent scores.
  • Eq. 12 retains the strongest recombination path as the candidate prior, while counting additional supporting paths separately.
  • Eq. 13 converts prior and support into ranks for subsequent fusion.

To assess the max-based scoring, we further evaluated mean aggregation in Eqs. 11 and 12 on the same 1,000 NMRGym test molecules.

SettingTop-1Top-5Top-10
NMR-CRAFT8.114.817.5
Eq. 11: max → mean6.914.016.4
Eq. 12: max → mean7.013.516.0

Both alternatives reduce sequence accuracy across all three Top-k metrics, supporting our max-based scoring design. We have added these scoring ablations to Appendix D.2.

[1] Fang et al. "NMRGym: A Comprehensive Benchmark for Nuclear Magnetic Resonance Based Molecular Structure Elucidation." KDD, 2026.

Thank you for the detailed response and the additional experiments, which address most of my points.

Resolved. W2 (attribution): the controlled ablation on the full test set, including paired CIs, resolves my concern. Q2 and Q3 have also been addressed.

Remaining points

(W1) I appreciate the honest answer that the contribution lies in the framework rather than its individual components. However, the limited novelty remains. I would also appreciate it if you cited Pang et al. (2024) for the auxiliary L⁺_SFT term in Eq. 5.

(Q1) I agree that the two signals are complementary, but it remains open how exactly they complement each other. At λ_rank = 0.32, this could mean “trust the original candidates unless the likelihood strongly disagrees”, or that the likelihood changes the ranking only when the preliminary ranking is uncertain. A more detailed analysis would therefore help. This could include rank correlations between the preliminary ranking, the spectral likelihood and their combination. It could also include a closer look at the cases where the two interact, for example the 22 molecules that are ranked correctly only by the combined score.

(Q4) I appreciate the comparison with alternatives to max-Tanimoto selection and the reported fraction of enantiomers. However, the differences between the variants are small (e.g. 0.03 points Top-1 over Max-Token), and no CIs are reported for the new variants. I would therefore either add CIs or slightly soften the claim that max-Tanimoto “achieves the best results among the evaluated criteria”.

(Q5) I also appreciate these results. Again, however, CIs are missing, and the difference amounts to only about 11 molecules out of 1,000. Reporting CIs here would be appreciated as well.

Given the controlled ablation and the confidence intervals, I will raise my rating to weak acceptance. My concern about limited novelty remains, which is why I do not go higher. I will reconsider my soundness score once the remaining points above are addressed.

Reviewer_6qX9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents NMR-CRAFT, a framework for NMR-based molecular structure elucidation. It introduces NMR-LM, a 4B language model jointly trained for structure elucidation and spectral prediction, with DPO used to improve structure generation. At inference time, NMR-CRAFT combines model-generated and retrieved candidates, performs constrained fragment recombination, and reranks candidates using spectral likelihood. Experiments on NMRGym and NMRSpec show improvements over direct generation, with a reported speedup of over 15× compared to NMR-Solver.

Critical strengths
  1. The paper is well motivated by the observation that incorrect molecular predictions can still contain useful structural fragments.

  2. The experiments cover two benchmarks and include several ablation studies to evaluate the proposed framework.

Critical weaknesses
  1. NMR-Solver [1] has already explored database retrieval, fragment recombination, and spectral verification, while SpecMol [2] has investigated bidirectional spectrum-to-structure and structure-to-spectrum modeling. The proposed framework appears to primarily integrate these existing techniques with pretrained language models and DPO. It remains unclear what fundamentally new algorithmic principles are introduced beyond this combination. [1] NMR-Solver: automated structure elucidation via large-scale spectral matching and physics-guided fragment optimization [2] SpecMol: A Spectroscopy-Grounded Foundation Model for Multi-Task Molecular Learning

  2. The paper does not provide a comprehensive analysis of how candidate coverage evolves throughout the proposed pipeline. Specifically, it is unclear what fraction of ground-truth structures appears in the Beam, Sampling, Database, Recombined, and final Preselected candidate pools, respectively. More importantly, how many ground-truth structures are recovered for the first time through fragment recombination, rather than already being present among the initial candidates? Such an analysis is essential to determine whether the proposed recombination mechanism genuinely expands the set of recoverable structures or primarily reorganizes existing candidates.

  3. Although the paper provides several workflow ablations, it lacks a direct comparison that disables fragment recombination while keeping the initial candidate pool, candidate budget, and reranking procedure unchanged. This is a particularly important omission given that fragment recombination is presented as a central contribution of NMR-CRAFT.

  4. The comparison with NMR-Solver raises concerns regarding fairness in both information availability and computational resources. NMR-CRAFT explicitly incorporates the target molecular formula during fragment pairing and assembly, whereas the reproduced NMR-Solver applies exact molecular-formula filtering only after its internal search. Furthermore, the two methods employ substantially different search budgets and GPU deployment configurations. Consequently, the reported accuracy improvements and over 15× speedup may be partially attributable to differences in search constraints, computational budgets, and system configurations rather than the proposed algorithm itself. The authors should provide more carefully controlled comparisons and report accuracy–computational-cost Pareto curves under comparable information constraints and hardware resources.

  5. The paper does not sufficiently demonstrate why structure elucidation and spectral prediction must be jointly modeled within a single shared-parameter language model. In fact, the reported ablations indicate that joint SFT does not improve structure elucidation accuracy over single-task SFT. Although the forward modeling capability is useful for candidate reranking, this does not establish that parameter sharing is necessary or advantageous. A critical missing baseline is a framework combining independently trained spectrum-to-structure and structure-to-spectrum models under the same candidate generation and reranking protocol. Such a comparison would clarify whether the proposed bidirectional modeling provides meaningful benefits in prediction accuracy, computational efficiency, or both, beyond simply combining two separately trained models.

Questions for authors
  1. How much of the reported improvement can be attributed specifically to fragment recombination? Could the authors provide a controlled ablation without recombination, keeping the initial candidate pool, computational budget, and reranking procedure unchanged, along with the number of ground-truth structures uniquely recovered through recombination?

  2. How would the accuracy and efficiency comparisons with NMR-Solver change under comparable molecular-formula constraints, search budgets, and GPU resources? Given the differences in the current experimental settings, it is unclear whether the reported improvements and 15× speedup primarily reflect the proposed method or the evaluation configuration.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 6qX9 (Part 1/3)

We sincerely thank the reviewer for the valuable feedback. We provide point-by-point responses to the comments below.

W1: NMR-Solver [1] has already explored database retrieval, fragment recombination, and spectral verification, while SpecMol [2] has investigated bidirectional spectrum-to-structure and structure-to-spectrum modeling. The proposed framework appears to primarily integrate these existing techniques with pretrained language models and DPO. It remains unclear what fundamentally new algorithmic principles are introduced beyond this combination.

[1] NMR-Solver: automated structure elucidation via large-scale spectral matching and physics-guided fragment optimization

[2] SpecMol: A Spectroscopy-Grounded Foundation Model for Multi-Task Molecular Learning

Thank you for this important comment. We acknowledge that bidirectional modeling, preference optimization, and fragment recombination have precedents. Our contribution is not a fundamentally new formulation of these individual techniques, but a budgeted, LLM-guided framework for reusing and verifying imperfect molecular hypotheses.

NMR-CRAFT develops a test-time structure refinement procedure that reuses and verifies imperfect molecular hypotheses, which SpecMol does not explore. Compared with NMR-Solver, our framework introduces (1) LLM-generated hypotheses as additional fragment sources alongside database retrieval, (2) source-aware, formula-constrained recombination with support-guided prioritization for efficient, budgeted search, and (3) direct forward-likelihood verification using the same bidirectional LLM, without relying on inherited chemical shifts.

Together, these design choices enable effective integration of molecular hypothesis generation, fragment recombination, and spectral verification, yielding consistent improvements on NMRGym and NMRSpec (Tables 1–2). Removing likelihood-based reranking reduces Top-1 accuracy from 8.1% to 4.1% (Fig. 5(c)), highlighting the value of spectral verification. Our methodological contribution lies in this efficient test-time refinement framework rather than in the novelty of its individual components.

W2: The paper does not provide a comprehensive analysis of how candidate coverage evolves throughout the proposed pipeline. Specifically, it is unclear what fraction of ground-truth structures appears in the Beam, Sampling, Database, Recombined, and final Preselected candidate pools, respectively. More importantly, how many ground-truth structures are recovered for the first time through fragment recombination, rather than already being present among the initial candidates? Such an analysis is essential to determine whether the proposed recombination mechanism genuinely expands the set of recoverable structures or primarily reorganizes existing candidates.

Thank you for raising this concern. While Fig. 5(a) analyzes the fragment sources of recovered structures, we further conducted a stage-wise coverage analysis on all 27,054 NMRGym test molecules.

The initial pool combines Beam, Sampling, and Database candidates. The recombined pool contains all valid recombination products, while the post-recombination pool combines them with the initial candidates. Final preselection retains up to 100 candidates for spectral-likelihood reranking.

Candidate PoolGround Truth CoveredCoverage (%)
Beam (Top-32)3,28112.13
Sampling (32 candidates)2,5829.54
Database (Top-50)00.00
Initial Pool3,56913.19
Recombined Pool5,65420.90
Post-Recombination Pool5,75921.29
Final Preselected (≤100)5,47320.23

Database coverage is zero under the Exclude-Scaffold retrieval setting, which excludes database molecules sharing the same Bemis–Murcko scaffold as the test molecules. Importantly, recombination newly recovered 2,190 ground-truth structures (8.09%) absent from the initial pool. Among these, 1,904 survived preselection and 1,345 reached the final Top-10. These results demonstrate that recombination expands candidate coverage rather than merely reorganizing existing candidates. We have added this stage-wise coverage analysis to Appendix D.5.

AuthorsOpenReview ↗

Response to Reviewer 6qX9 (Part 2/3)

W3: Although the paper provides several workflow ablations, it lacks a direct comparison that disables fragment recombination while keeping the initial candidate pool, candidate budget, and reranking procedure unchanged. This is a particularly important omission given that fragment recombination is presented as a central contribution of NMR-CRAFT.

Q1: How much of the reported improvement can be attributed specifically to fragment recombination? Could the authors provide a controlled ablation without recombination, keeping the initial candidate pool, computational budget, and reranking procedure unchanged, along with the number of ground-truth structures uniquely recovered through recombination?

Thank you for raising this concern. We conducted a controlled w/o Recombination ablation on all 27,054 NMRGym test molecules, retaining the same initial candidate pool (Beam + Sampling + Database), formula filtering, candidate budget (up to 100), and likelihood-based reranking (λ = 0.32), while disabling fragment recombination.

SettingTop-1Top-5Top-10
NMR-CRAFT (w/o Recombination)6.5811.4712.60
NMR-CRAFT8.2315.1417.34
Δ [95% paired CI]+1.64 [1.40, 1.87]+3.67 [3.41, 3.94]+4.73 [4.45, 5.01]

Compared with the ablated variant, NMR-CRAFT yields a net increase of 1,281 correct Top-10 predictions, demonstrating the contribution of fragment recombination. We have added this controlled ablation and its paired confidence intervals to Appendix D.5.

W4: The comparison with NMR-Solver raises concerns regarding fairness in both information availability and computational resources. NMR-CRAFT explicitly incorporates the target molecular formula during fragment pairing and assembly, whereas the reproduced NMR-Solver applies exact molecular-formula filtering only after its internal search. Furthermore, the two methods employ substantially different search budgets and GPU deployment configurations. Consequently, the reported accuracy improvements and over 15× speedup may be partially attributable to differences in search constraints, computational budgets, and system configurations rather than the proposed algorithm itself. The authors should provide more carefully controlled comparisons and report accuracy–computational-cost Pareto curves under comparable information constraints and hardware resources.

Q2: How would the accuracy and efficiency comparisons with NMR-Solver change under comparable molecular-formula constraints, search budgets, and GPU resources? Given the differences in the current experimental settings, it is unclear whether the reported improvements and 15× speedup primarily reflect the proposed method or the evaluation configuration.

Thank you for raising these concerns. Our original NMR-Solver reproduction follows its default protocol, applying molecular-formula filtering only after search. To examine the effect of formula access, we additionally incorporated the exact molecular formula into NMR-Solver's fragment-based molecular optimization (FB-MO). We also evaluated different search budgets on the same 1,000 NMRGym test molecules under Exclude-Scaffold.

MethodBudget (R/P/C)Top-1Top-5Top-10Time (s)
NMR-Solver (original)1000/200k/10002.95.25.9260.34
NMR-Solver (formula-aware)50/1k/1000.00.00.010.10
NMR-Solver (formula-aware)500/10k/5002.52.72.744.81
NMR-Solver (formula-aware)1000/200k/10006.08.08.0195.96
NMR-CRAFT50/1k/1008.114.817.514.11

R/P/C denotes the limits on retrieved candidates, fragment pairs, and candidate pool size, respectively. For NMR-Solver, R applies only to the initial retrieval, while the same P and C limits apply to each FB-MO iteration. Accuracy is reported in percent, and time is the mean wall-clock time per molecule. Incorporating the formula improves NMR-Solver's original configuration, while NMR-CRAFT achieves higher accuracy with substantially lower runtime at that configuration. While a full Pareto analysis would require more extensive budget sweeps, these comparisons demonstrate a favorable accuracy–runtime trade-off for NMR-CRAFT under the evaluated settings.

Regarding hardware, both methods were evaluated on the same four-A100 server. NMR-Solver's NMRNet ensemble runs on one GPU, while our original NMR-CRAFT deployment used separate GPUs for inverse generation and forward verification. Since these stages execute sequentially, we additionally deployed both services on a single GPU, obtaining 14.11 s/sample versus the originally reported 14.24 s/sample. Compared with formula-aware NMR-Solver at its original search budget, NMR-CRAFT remains approximately 14× faster (14.11 vs. 195.96 s/sample).

We have updated the manuscript to report the single-GPU results (14.11 s/sample) and added the budget comparisons to Appendix D.6.

AuthorsOpenReview ↗

Response to Reviewer 6qX9 (Part 3/3)

W5: The paper does not sufficiently demonstrate why structure elucidation and spectral prediction must be jointly modeled within a single shared-parameter language model. In fact, the reported ablations indicate that joint SFT does not improve structure elucidation accuracy over single-task SFT. Although the forward modeling capability is useful for candidate reranking, this does not establish that parameter sharing is necessary or advantageous. A critical missing baseline is a framework combining independently trained spectrum-to-structure and structure-to-spectrum models under the same candidate generation and reranking protocol. Such a comparison would clarify whether the proposed bidirectional modeling provides meaningful benefits in prediction accuracy, computational efficiency, or both, beyond simply combining two separately trained models.

Thank you for raising this important point. To further investigate the benefit of parameter sharing, we compared the shared bidirectional model with two independently trained models: an inverse-only model for candidate generation and a forward-only model for likelihood scoring. Both configurations use identical candidate generation and reranking protocols on the same 1,000 NMRGym test molecules.

SettingTop-1Top-5Top-10
Independent models7.014.416.8
Shared bidirectional model8.114.817.5

A paired analysis shows that both settings correctly recover 151 molecules within the Top-10, while the shared and independent models additionally recover 24 and 17 distinct molecules, respectively.

The shared model achieves slightly higher sequence accuracy while requiring only one 4B model instead of two independent 4B models (8B in total). This reduces model storage and memory requirements and allows both directions to share a single model instance, improving deployment efficiency. We have added this comparison to Appendix D.2.

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

This paper introduces NMR-LM and NMR-CRAFT, a unified framework for molecular structure elucidation from NMR spectra. The authors first demonstrate that fine-tuning a general-purpose 4B-parameter LLM (Qwen3.5) bidirectionally on both spectrum-to-structure and structure-to-spectrum tasks yields state-of-the-art performance. Building on the observation that incorrect predictions often contain valid substructures, they propose NMR-CRAFT, a test-time refinement procedure. This method recombines fragments from model-generated candidates and database retrievals under chemical constraints, then reranks the resulting hypotheses using the model’s own forward spectral likelihood. Validation on NMRGym (scaffold-split) and NMRSpec benchmarks shows NMR-CRAFT achieves ~2.2× the Top-1 accuracy of direct generation and over 15× speedup compared to the NMR-Solver baseline, while NMR-LM sets new SOTAs among generative models for structure elucidation and remains competitive in spectral prediction.

Critical strengths
  1. The integration of forward spectral prediction as a verification signal within the same model used for generation is a consequential architectural insight. Unlike prior methods requiring separate predictors or expensive DFT calculations, this allows efficient, self-consistent reranking of recombined candidates. The ablation in Figure 5(c) confirms that removing spectral likelihood causes the largest performance drop (Top-1 acc: 8.1% → 4.1%), validating that the bidirectional capability is essential to the framework's success rather than incidental.
  2. The evaluation protocol strictly enforces scaffold-based splitting on NMRGym and applies Exclude-Scaffold filtering to the retrieval database, preventing data leakage common in spectral matching tasks. Despite this stricter setting, NMR-CRAFT significantly outperforms specialized baselines like NMR-Solver and NMRTrans. The comprehensive source analysis (Figure 5a/b) further strengthens validity by demonstrating that gains arise specifically from recombining complementary fragments across model and database sources, rather than simple retrieval memorization.
Critical weaknesses
  1. Dependence on Molecular Formula: The NMR-CRAFT workflow currently requires the molecular formula as input for fragment recombination and validation (Eq. 7). While formula-conditioned results are strong, real-world NMR elucidation often lacks exact formulas initially. The limitation section acknowledges this but does not provide empirical evidence of how gracefully the method degrades without formulas or whether the bidirectional model could be adapted to jointly infer them, which limits immediate practical applicability for fully de novo unknown identification.
  2. Limited Modality Scope: The method relies exclusively on 1D ^1$H and $^{13} NMR. In practice, structure elucidation for novel scaffolds heavily depends on 2D NMR (HSQC, HMBC) to resolve connectivity ambiguities that 1D spectra cannot. While the LLM-based approach is extensible, the current validation does not demonstrate efficacy on multimodal inputs, leaving open the question of whether the fragment recombination strategy remains effective when spectral ambiguity is higher or when richer structural constraints are available.
Questions for authors
  1. How does NMR-CRAFT perform in a formula-free setting, or when provided with only an approximate mass/range? Could the bidirectional model be prompted to generate candidate formulas as part of the "Propose" stage to enable formula-free search?
  2. Have you evaluated the sensitivity of the spectral likelihood reranker to systematic errors in predicted chemical shifts (e.g., solvent effects or concentration-dependent shifts) that might penalize correct structures? Is there any calibration or normalization applied to the log-likelihood scores beyond the z-score standardization in Eq. 14?
  3. Given that NMR-LM is trained on serialized peak lists, how robust is the tokenization and likelihood scoring to variations in peak reporting precision or missing low-intensity signals in experimental data compared to the training distribution?
4 comments
AuthorsOpenReview ↗

Response to Reviewer tKsN (Part 1/2)

We thank the reviewer for the careful evaluation and helpful suggestions. We have addressed the raised concerns as detailed below.

W1: Dependence on Molecular Formula: The NMR-CRAFT workflow currently requires the molecular formula as input for fragment recombination and validation (Eq. 7). While formula-conditioned results are strong, real-world NMR elucidation often lacks exact formulas initially. The limitation section acknowledges this but does not provide empirical evidence of how gracefully the method degrades without formulas or whether the bidirectional model could be adapted to jointly infer them, which limits immediate practical applicability for fully de novo unknown identification.

Q1: How does NMR-CRAFT perform in a formula-free setting, or when provided with only an approximate mass/range? Could the bidirectional model be prompted to generate candidate formulas as part of the "Propose" stage to enable formula-free search?

Thank you for raising this important point. Formula-conditioned structure elucidation is an established setting in prior NMR studies and benchmarks [1–3], although we agree that requiring an exact molecular formula limits applicability. Since molecular weight can often be readily determined using mass spectrometry [4], we further explored a less restrictive mass-constrained setting.

We evaluated a mass-constrained NMR-CRAFT variant on 1,000 NMRGym test molecules, using our bidirectionally trained formula-free NMR-LM checkpoint. Instead of an exact molecular formula, the workflow uses nominal molecular mass with a ±1 Da tolerance. Since NMRGym does not provide experimental MS data, nominal masses were calculated from the reference molecular structures. Search budgets and likelihood reranking remained unchanged.

SettingNMR-LM ConfigurationSearch ConstraintTop-1Top-5Top-10
NMR-LMFormula-free—2.46.28.3
NMR-CRAFTFormula-freeNominal mass (±1 Da)7.913.916.0
NMR-CRAFTFormula-conditionedExact formula8.114.817.5

Under mass constraints, recombination recovered 70 ground-truth structures absent from the initial candidates, of which 49 reached the final Top-10. These results suggest that NMR-CRAFT can operate under weaker chemical constraints. Extending toward fully formula-free search through model-predicted formulas is a promising direction. We have added these results to Section 4.5.

W2: Limited Modality Scope: The method relies exclusively on 1D ^1^{13} NMR. In practice, structure elucidation for novel scaffolds heavily depends on 2D NMR (HSQC, HMBC) to resolve connectivity ambiguities that 1D spectra cannot. While the LLM-based approach is extensible, the current validation does not demonstrate efficacy on multimodal inputs, leaving open the question of whether the fragment recombination strategy remains effective when spectral ambiguity is higher or when richer structural constraints are available.

Thank you for raising this important point. We agree that 2D NMR techniques such as HSQC and HMBC provide valuable connectivity information that can resolve ambiguities in 1D spectra. This work focuses on experimental 1D ¹H/¹³C NMR, investigating how imperfect molecular hypotheses can be exploited through constrained fragment recombination and spectral-likelihood verification. We acknowledge that the effectiveness of fragment recombination with 2D or multimodal inputs has not yet been validated. As noted in our limitations, incorporating 2D NMR into bidirectional modeling and structural search is an important direction for future work.

[1] Huang et al. "A framework for automated structure elucidation from routine NMR spectra." Chemical Science, 2021.

[2] Yang et al. "NMRTrans: Structure Elucidation from Experimental NMR Spectra via Set Transformers." KDD, 2026.

[3] Fang et al. "NMRGym: A Comprehensive Benchmark for Nuclear Magnetic Resonance Based Molecular Structure Elucidation." KDD, 2026.

[4] Holčapek et al. "Basic rules for the interpretation of atmospheric pressure ionization mass spectra of small molecules." Journal of Chromatography A, 2010.

AuthorsOpenReview ↗

Response to Reviewer tKsN (Part 2/2)

Q2: Have you evaluated the sensitivity of the spectral likelihood reranker to systematic errors in predicted chemical shifts (e.g., solvent effects or concentration-dependent shifts) that might penalize correct structures? Is there any calibration or normalization applied to the log-likelihood scores beyond the z-score standardization in Eq. 14?

Thank you for raising this concern. We evaluated end-to-end sensitivity to chemical-shift perturbations on 1,000 NMRGym test molecules. We considered random perturbations applied to 20% of signals, with noise sampled uniformly from U(−0.05, 0.05)/U(−0.5, 0.5) ppm or U(−0.1, 0.1)/U(−1, 1) ppm for ¹H/¹³C, respectively, as well as systematic offsets. Repeated shifts representing the same proton signal were perturbed together. Random-perturbation results are averaged over three seeds, with Top-1/Top-10 sequence accuracy (%) reported below.

Setting (¹H / ¹³C, ppm)NMR-LM (Top-1 / Top-10)NMR-CRAFT (Top-1 / Top-10)
Clean2.8 / 8.68.1 / 17.5
Random 20%, U(−0.05, 0.05) / U(−0.5, 0.5)2.5 / 8.07.4 / 16.4
Random 20%, U(−0.1, 0.1) / U(−1, 1)2.2 / 6.56.1 / 13.8
Offset +0.05 / +0.51.9 / 6.96.3 / 14.1
Offset +0.1 / +11.1 / 5.84.0 / 12.1

NMR-CRAFT consistently outperforms direct NMR-LM generation, with modest degradation under small random perturbations and greater sensitivity to larger systematic shifts. These tests assess end-to-end robustness, although they do not fully isolate the reranker’s sensitivity. Beyond the z-score standardization in Eq. 14, no additional calibration is applied to spectral likelihoods. Training-time shift augmentation is a potential direction for improving robustness. We have added the complete robustness results to Appendix D.4.

Q3: Given that NMR-LM is trained on serialized peak lists, how robust is the tokenization and likelihood scoring to variations in peak reporting precision or missing low-intensity signals in experimental data compared to the training distribution?

Thank you for this question. We evaluated end-to-end robustness to variations in peak precision and missing signals on 1,000 NMRGym test molecules. The training spectra were serialized with up to three decimal places (trailing zeros removed). We tested rounding ¹H/¹³C shifts to two/one decimal places and randomly masking 5% or 10% of signals, restricting masking to single-entry signals while preserving repeated-shift groups. Random-masking results are averaged over three seeds. Top-1/Top-10 sequence accuracy (%) is reported below.

SettingNMR-LM (Top-1 / Top-10)NMR-CRAFT (Top-1 / Top-10)
Clean2.8 / 8.68.1 / 17.5
Round (¹H: 2, ¹³C: 1 decimal)2.4 / 7.77.1 / 16.3
Random Mask 5%2.4 / 8.06.9 / 16.2
Random Mask 10%2.0 / 6.06.0 / 12.8

NMR-CRAFT shows modest degradation under reduced precision and 5% masking, with larger declines at 10% masking. Since explicit peak intensities are unavailable, this masking strategy approximates missing signals rather than intensity-specific dropout. We have added the complete results to Appendix D.4.

Reviewer_tKsNOpenReview ↗

Response

As a researcher with an extensive publication record in the NMR spectroscopy field, I have carefully evaluated the authors' rebuttal. I believe they have thoroughly and satisfactorily addressed all my concerns, particularly with the new empirical results on mass-constrained search and robustness. Therefore, I have raised my score to 3 (Weak Accept) and recommend this paper for acceptance. Good luck.

We sincerely thank the reviewer for the encouraging feedback and thoughtful suggestions. In particular, your suggestion on mass-constrained search motivated us to explore a more practical setting and further strengthened our experimental evaluation. We greatly appreciate your constructive input and support for our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.