{"id":"6c2b7be3-86d3-47cc-9bef-b03d76e025a0","arxiv_id":"2507.08854","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"DiffNMR uses a discrete graph diffusion model conditioned on NMR spectra to predict molecular structures, achieving 68.26% top-1 accuracy with formula on molecules up to 15 heavy atoms.","lead":"A team introduces DiffNMR, a diffusion model that generates molecular structures directly from NMR spectra, and reports up to 68% top-1 accuracy on small molecules from a simulated dataset. The paper proposes a promising automation for chemistry, but its evaluation lacks baselines and has unresolved details.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No direct baseline comparison: 'competitive' is the central claim, but the paper never evaluates DiffNMR against a baseline on the same split, so the main claim is currently unfalsifiable.","rationale":"The reader's designated weakest assumption is that MestReNova-simulated spectra transfer to experimental NMR, which is a legitimate external-validity concern. However, the most load-bearing gap for the central claim as stated is that 'competitive' is a comparative judgment, and the paper provides no same-benchmark baseline. The reader's rationale does flag missing baselines as a serious concern, so there is partial agreement, but they did not make it the central weakest assumption. I set aside the retrieval-database issue because Table 1, the strongest claimed result, does not use retrieval initialization, and I set aside formula-conditioning ambiguity because the narrative comparison with published numbers would still be insufficient even if formula conditioning were perfectly clear. The proposed check is concrete and directly targets the advertised conclusion. The paper does have internal support, including ablation studies and a public code link, so a conditional acceptance with a mandatory baseline table is appropriate rather than rejection.","tokens_in":17586,"tokens_out":7927,"duration_ms":84265,"concrete_test":"Run a direct baseline comparison on the same MSD split and formula condition: (a) train a Transformer encoder-decoder (e.g., the architecture of [24] or [22]) on 1H+13C peak tokens; (b) evaluate it with the same top-k and Tanimoto metrics on the same ≤15/20/25 HAC test subsets, with and without formula. If DiffNMR's top-1 is not within roughly 2-3 points of the best baseline, the 'competitive' wording in the abstract and conclusion should be revised or qualified. Reporting this comparative table is the minimal check that would settle whether the central claim holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract and conclusion is that DiffNMR is \"competitive\" for NMR-based structure elucidation. The evidence offered is Table 1 (top-1 68.26% on ≤15 heavy atoms with formula) and Table 2, but neither table contains a single baseline run on the same train/test split. The only prior-work numbers appear as narrative in the Introduction: [22] reports 69.6% top-15 on molecules up to 19 heavy atoms and [24] reports 67% top-1 on a million-scale dataset, but these use different datasets, splits, formula conditions, and evaluation setups, so they cannot establish competitiveness. If, for example, a standard transformer encoder-decoder obtained 75% top-1 on the same MSD split, DiffNMR's number would not be competitive, and the paper would not detect this. The claim is therefore unfalsifiable as reported. This is not a minor quibble: the contribution is explicitly framed as improving on autoregressive models, yet the paper's ablations (Fig. 6) only compare DiffNMR variants against each other, not against any external method. The missing baseline is the load-bearing gap between the reported accuracy and the advertised conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"DiffNMR proposes a conditional discrete diffusion model, built on the DiGress framework, for de novo molecular structure elucidation from 1H and 13C NMR spectra. The system has three main components: a graph-transformer molecular encoder, a specialized NMR encoder that uses RBF encoding for chemical shifts and bidirectional cross-attention to fuse 1H and 13C information, and a graph diffusion decoder conditioned on the NMR representation via FiLM. Training proceeds through two pretraining stages (a diffusion autoencoder for molecular graphs, then contrastive alignment of the NMR encoder to the frozen molecular encoder) followed by joint fine-tuning. At inference, the method can use retrieval initialization from the training set and cosine-similarity-based filtering of generated candidates. Experiments on the simulated MSD dataset report Top-1 accuracies up to 68.26% for molecules with ≤15 heavy atoms when the molecular formula is provided, along with ablations showing benefits of pretraining, RBF encoding, similarity filtering, and retrieval initialization.","tokens_in":17856,"tokens_out":3528,"duration_ms":39118,"significance":"If the central claim of competitiveness with autoregressive approaches were properly supported, DiffNMR would be a useful contribution: it brings discrete graph diffusion to NMR structure elucidation, a domain currently dominated by autoregressive transformers, and it includes several well-motivated components (RBF shift encoding, two-stage pretraining, retrieval-based sampling) that are clearly ablated. The paper also releases code on GitHub, which aids reproducibility. However, the main empirical claim is currently unfalsifiable as stated because no same-split baseline is included, and the retrieval-initialization experiments use the training set as the retrieval database, which risks leaking structural information. These issues, together with the absence of error bars and the reliance on simulated spectra only, substantially weaken the significance of the reported accuracies.","major_comments":[{"comment":"The central claim that DiffNMR achieves \"competitive performance relative to autoregressive approaches\" is not supported by any baseline evaluated on the same data and split. The only prior-work numbers cited in the Introduction (refs [22] and [24]) come from different datasets, different evaluation protocols, and different formula conditions, so they cannot establish competitiveness. Without a direct comparison—e.g., a transformer encoder-decoder or DeepSPInN variant trained and tested on the same MSD split—the headline accuracy of 68.26% (Table 1) and the retrieval-initialization gains (Table 2) have no reference point. This is a load-bearing gap: the paper's motivation and conclusions are framed specifically in opposition to autoregressive models.","section":"§2.3, Tables 1-2; abstract and §3"},{"comment":"Retrieval initialization draws the starting molecule from the training set, which can leak information about test molecules because training and test molecules in MSD are drawn from the same USPTO-derived distribution. The reported improvements from retrieval initialization (e.g., accuracy rising from 35.76% to 49.70% for ≤25 HAC without filtering) may partly reflect nearest-neighbor memorization rather than the diffusion model's generative capability. To quantify the true contribution, the retrieval database must be disjoint from the training set (e.g., a held-out validation set), and the retrieval-only baseline (nearest neighbor without any diffusion steps) should be reported.","section":"§2.4, Table 2, text \"Here, we use the training set as the retrieval database.\""},{"comment":"The mechanism by which the molecular formula is provided as a condition is never described. Table 1 compares results with and without the formula, but Section 4, which defines the condition vector y and the NMR encoder, does not state how the formula (e.g., as a count vector, SMILES fragment, or embedding) enters the model. This omission makes the main table irreproducible, and it is particularly important because the formula condition is a major source of the reported accuracy gain (e.g., 58.83% to 68.26% for ≤15 HAC). Please specify the exact conditioning implementation and, ideally, include an ablation on how the formula is encoded.","section":"§4.4 and Table 1"},{"comment":"No error bars, standard deviations, or significance tests are reported for any accuracy or Tanimoto value. Diffusion sampling is stochastic—the method generates multiple candidates and, with random initialization, the denoising process is not deterministic—so the difference between 60.13% and 61.18% in Table 2, or between 20% improvements in Figure 6, could easily be within sampling noise. Please report mean ± std over at least three independent runs (or fix and report all random seeds) for all quantitative claims.","section":"§2.3, Tables 1-2 and Figure 6"},{"comment":"All reported results use simulated spectra from MSD (generated with MestReNova default parameters), and the paper never evaluates on experimental NMR data. The conclusion that DiffNMR \"offers a promising framework for automating NMR-based structure elucidation\" is therefore a claim about practical utility that rests on the unvalidated premise that simulation-to-experiment transfer is trivial. At minimum, the paper should acknowledge this domain-shift risk explicitly, and ideally it should include a small evaluation on experimental spectra or a discussion of how the model is expected to handle noise, peak-picking errors, and solvent differences.","section":"§2.2, §3 and the abstract"}],"minor_comments":[{"comment":"The FiLM formulation appears mistyped: \"FiLM(X1, X2) = X1W1 + (X1W2) ⊙ X2 + W2\" suggests that the bias term is a parameter matrix W2 rather than a bias vector, and the standard FiLM uses a modulation vector computed from the condition, not from X2. Please clarify the dimensions and correct the equation.","section":"§4.5, Eq. (15)"},{"comment":"The cosine schedule for α_t is written with unbalanced parentheses: αt = cos(0.5π(t/T + s)/(1 + s))^2. Please check the formula against the original DiGress paper and correct the typesetting, as the current expression is ambiguous.","section":"§4.2, Eq. (2) and the cosine schedule"},{"comment":"The sentence \"We employ a graph transformer network[] as denoising network\" contains an empty citation placeholder. Please fill in the appropriate reference or remove the brackets.","section":"§4.5, first sentence"},{"comment":"The text in §2.5 refers to \"Figure 6(a)\" for the RBF-versus-discrete comparison, but Figure 6(a) is described in the caption as the pretraining comparison. The order of panels in the figure appears to be inconsistent with the in-text references; reorder the panels or adjust the text so that the descriptions match.","section":"Figure 6 and its caption"},{"comment":"The value \"61.18 %\" contains a space before the percent sign, which is inconsistent with other entries. Please make the formatting uniform.","section":"Table 2, ≤20 row, \"√ √\" column"},{"comment":"The phrase \"It’s obvious\" is informal for a research paper; consider replacing it with a more neutral phrasing such as \"The results show\" or \"As expected\".","section":"§2.3, \"It’s obvious\""},{"comment":"The strings beginning with \"/uni000000...\" in the captions appear to be encoding artifacts from the PDF generation. They should be removed or replaced with the intended text, as they are not readable chemical information.","section":"Figures 2 and 3 captions"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be of interest to the NMR/ML community, but the missing baselines and the training-set retrieval database are substantive issues that may reflect a need for a more careful evaluation rather than a fundamental flaw in the method. I would encourage the editor to require the authors to run proper baselines on the same split and to redo the retrieval experiments with a disjoint database before reconsidering the paper. The code release is a positive sign, but the current manuscript does not yet substantiate the advertised claim of competitiveness."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read DiffNMR. The core idea is a discrete graph diffusion model conditioned on NMR spectra, with contrastive pretraining aligning spectra and molecules, plus retrieval initialization and similarity filtering at inference. The specific combination is new for NMR elucidation, and the authors build it carefully on DiGress. The internal ablations are the most convincing part: pretraining both encoder and decoder gives roughly 10% over no pretraining, and RBF encoding of chemical shifts beats discrete encoding by over 20%. Those are real, controlled comparisons, and they suggest the components do what the paper says.\n\nThe big problem is that the central claim—DiffNMR is 'competitive' with autoregressive transformers—is never actually tested. Table 1 reports top-1 accuracy up to 68.26%, but there is no baseline run on the same train/test split. The only comparisons are narrative references to prior work on different datasets, different size limits, and different formula conditions. A standard transformer might get 75% on this split; the paper would not detect that. The ablations in Figure 6 only compare DiffNMR variants against each other. So the advertised conclusion is unfalsifiable as reported. This is the load-bearing gap.\n\nTwo more soft spots, in proportion. First, the retrieval database is the training set. For each test molecule, they retrieve the most similar training structure and start diffusion from it. If that structure is close to the target, it trivially helps; the paper doesn't quantify how often retrieval returns the target's near-neighbor, so the gains in Table 2 are hard to interpret. Second, all evaluations are on simulated MestReNova spectra; there is no experimental NMR test, so practical transfer remains an assumption. I'd also want the data split described (random vs scaffold), error bars or seeds, and a precise explanation of how molecular formula is injected as a condition. These are fixable, but they need to be in the paper.\n\nWho is this for? Groups working on ML-based structure elucidation will want to know this architecture exists. The method itself deserves referee time; the evaluation currently does not. I'd send it to review with major-revision expectations: add direct baselines, clarify retrieval leakage, report variance, and ideally test on a small real-NMR set. If those come back, the paper could be solid.","headline":"Useful new architecture for NMR elucidation, but the missing head-to-head baselines make the central 'competitive' claim unsupported as written.","tokens_in":18332,"tokens_out":2050,"would_cite":false,"duration_ms":48754,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DiffNMR claims that NMR spectra can be decoded into molecular structures by a conditional discrete diffusion process over molecular graphs, reaching 68.26% top-1 accuracy on molecules up to 15 heavy atoms when the molecular formula is…","keywords":["NMR structure elucidation","conditional discrete diffusion","molecular graph generation","radial basis function encoding","contrastive learning","retrieval-augmented inference","Tanimoto similarity","simulated NMR spectra"],"falsifier":"Run DiffNMR on a held-out set of experimental 1H and 13C NMR spectra with known structures, recorded under a different solvent or reference than the simulated CDCl3 defaults, and compare top-1 accuracy and Tanimoto similarity with the reported benchmark numbers; a large drop in accuracy would contradict the transferability premise.","tokens_in":17434,"feed_emoji":"🧪","tokens_out":8182,"duration_ms":85225,"temperature":0.7,"pith_summary":"The paper tries to establish that NMR-based structure elucidation can be posed as conditional discrete diffusion over molecular graphs, rather than as sequential SMILES generation. DiffNMR encodes 1H and 13C peaks, fuses them with cross-attention, and denoises a molecular graph under that conditioning, reaching top-1 accuracy of 68.26% on molecules with at most 15 heavy atoms when the molecular formula is supplied (58.83% without it). The authors argue this is competitive with transformer-based autoregressive methods while avoiding the cascading errors of left-to-right generation. A reader should care because automated, end-to-end structure elucidation would shorten the synthesis-analysis loop in chemistry.","feed_headline":"Diffusion model turns NMR spectra into molecules: 68% top-1","feed_subtitle":"A graph-diffusion model reconstructs molecular structures from 1H and 13C spectra without autoregressive error buildup.","key_machinery":"The central object is a conditional discrete graph diffusion model in the style of DiGress: a Markov noise process independently corrupts node and edge types, and a graph transformer denoiser $\\phi_\\theta(G_t, t, y)$ predicts the clean graph $G_0$ with the conditioning vector $y$ injected through FiLM layers. The conditioning vector is produced by an NMR encoder that embeds chemical shifts with radial basis functions (Gaussian kernels centered across the shift range), encodes 1H and 13C peaks in separate transformers, and fuses them with bidirectional cross-attention. Two pretraining stages support the pipeline: a diffusion autoencoder establishes the molecular latent space, and contrastive learning (InfoNCE) aligns NMR embeddings to that space; at inference, cosine-similarity filtering and retrieval initialization from the training database steer the denoising process.","core_discovery":"The paper's central claim is that a conditional discrete diffusion model can perform de novo structure elucidation from 1H and 13C NMR spectra with accuracy comparable to autoregressive transformer models. The model treats a molecule as a graph and learns a reverse diffusion process that iteratively repairs noise-corrupted atom and bond types, conditioned on an NMR-derived vector through FiLM modulation. The NMR encoder uses radial basis functions for chemical shifts, separate transformers for each nucleus, and bidirectional cross-attention to fuse them; two pretraining stages first learn molecular representations with a diffusion autoencoder and then align the NMR space to the molecular space with an InfoNCE contrastive loss. On the MSD dataset of simulated spectra, the authors report top-1 accuracy of 68.26% with molecular formula and 58.83% without it for molecules up to 15 heavy atoms, and they report that retrieval initialization plus similarity filtering raises top-1 accuracy on the no-formula, up-to-25-heavy-atom set from 35.76% to 58.47% while raising average Tanimoto similarity from 0.71 to 0.82.","pith_inferences":["A direct test of the framework's practical value would be fine-tuning or evaluating on experimental spectra recorded under conditions outside the simulated CDCl3 default; the paper's RBF encoding is designed to be continuous, but cross-domain transfer is not shown.","The contrastive alignment that powers similarity filtering is a reusable spectrum-to-structure retrieval mechanism; it could support database search independently of the diffusion decoder.","The two-stage pretraining suggests a modular route to multi-modal elucidation: additional spectroscopies such as IR, MS, or 2D NMR could be attached as aligned encoders without retraining the molecular decoder.","If simulated-to-experimental transfer fails, the headline accuracies would overstate real-world performance; a held-out experimental benchmark with known structures would settle the question."],"forward_implications":["NMR-based elucidation no longer needs predefined fragment libraries or hand-coded assembly rules; the same conditional diffusion decoder generates the whole graph.","Because the molecular and spectral spaces are aligned by contrastive learning, the decoder can in principle be reconditioned by any encoder that maps into the same latent space.","Retrieval initialization both improves accuracy and shortens the denoising chain, which points to database-augmented inference as a scalable route for larger molecules.","The accuracy drop between 20 and 25 heavy atoms is attributed to data density rather than model capacity, so larger and more diverse training sets are a direct lever for improvement.","Even incorrect top-1 predictions retain high Tanimoto similarity to the target, so the sampled candidates can serve as chemically relevant hypotheses for manual refinement."],"supporting_citations":[{"why":"Supplies the Multimodal Spectroscopic Dataset (794,403 simulated molecules with 1H and 13C NMR) used for all training and evaluation.","marker":"[23]"},{"why":"Provides the DiGress discrete denoising graph diffusion method from which the graph decoder is adapted.","marker":"[28]"},{"why":"Defines the autoregressive transformer baseline that translates NMR peak tokens to SMILES with 67% top-1 accuracy, the main comparison point.","marker":"[24]"},{"why":"Defines the multitask transformer baseline that predicts substructures and complete structures from raw 1H and 13C spectra, a competing autoregressive approach.","marker":"[22]"},{"why":"Defines the reinforcement-learning/MCTS baseline DeepSPInN whose small-molecule accuracy and scaling limits motivate the diffusion design.","marker":"[19]"},{"why":"Provides the diffusion autoencoder concept used in the first pretraining stage to learn molecular representations.","marker":"[37]"},{"why":"Provides the contrastive-learning formulation for cross-modal NMR-structure retrieval used in the second pretraining stage.","marker":"[43]"},{"why":"Defines the circular fingerprint used to compute Tanimoto similarity between predicted and target molecules.","marker":"[44]"}],"fun_headline_variants":["Diffusion model decodes NMR spectra into molecular graphs","Graph diffusion reconstructs molecules from NMR data","NMR-to-molecule: diffusion beats autoregressive error","DiffNMR: reverse diffusion for NMR structure elucidation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that NMR spectra simulated computationally with default parameters are faithful enough to real experimental spectra that the accuracy measured on the simulated MSD benchmark will transfer to practical structure elucidation; all reported results come from simulated data, and the paper does not test on experimental spectra.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model decodes NMR spectra into molecular graphs","Graph diffusion reconstructs molecules from NMR data","NMR-to-molecule: diffusion beats autoregressive error","DiffNMR: reverse diffusion for NMR structure elucidation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1246,"prompt_tokens":955,"completion_tokens":291,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":227}},"tokens_in":571,"tokens_out":291,"duration_ms":3308,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:59:17.173153+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DiffNMR on a held-out set of experimental 1H and 13C NMR spectra with known structures, recorded under a different solvent or reference than the simulated CDCl3 defaults, and compare top-1 accuracy and Tanimoto similarity with the reported benchmark numbers; a large drop in accuracy would contradict the transferability premise.","supporting_citations":[{"cited_title":"Unraveling molecular structure: A multimodal spectroscopic dataset for chemistry","cited_arxiv_id":null,"evidence_quote":"Supplies the Multimodal Spectroscopic Dataset (794,403 simulated molecules with 1H and 13C NMR) used for all training and evaluation."},{"cited_title":"Learning the language of nmr: Structure elucidation from nmr spectra using transformer models","cited_arxiv_id":null,"evidence_quote":"Defines the autoregressive transformer baseline that translates NMR peak tokens to SMILES with 67% top-1 accuracy, the main comparison point."},{"cited_title":"Accurate and efficient structure elucidation from routine one-dimensional nmr spectra using multitask machine learning","cited_arxiv_id":null,"evidence_quote":"Defines the multitask transformer baseline that predicts substructures and complete structures from raw 1H and 13C spectra, a competing autoregressive approach."},{"cited_title":"Deepspinn-multimodal deep learning for molecular structure prediction from infrared and nmr spectra","cited_arxiv_id":null,"evidence_quote":"Defines the reinforcement-learning/MCTS baseline DeepSPInN whose small-molecule accuracy and scaling limits motivate the diffusion design."},{"cited_title":"Diffusion autoen- coders: Toward a meaningful and decodable representation","cited_arxiv_id":null,"evidence_quote":"Provides the diffusion autoencoder concept used in the first pretraining stage to learn molecular representations."},{"cited_title":"Cross-modal retrieval between 13c nmr spectra and structures for compound identification using deep contrastive learning","cited_arxiv_id":null,"evidence_quote":"Provides the contrastive-learning formulation for cross-modal NMR-structure retrieval used in the second pretraining stage."},{"cited_title":"The generation of a unique machine description for chemical structures-a technique developed at chemical abstracts service","cited_arxiv_id":null,"evidence_quote":"Defines the circular fingerprint used to compute Tanimoto similarity between predicted and target molecules."}],"review_version":1}