Pith. sign in

REVIEW 4 major objections 6 minor 3 cited by

NMR-Solver: Automated Structure Elucidation via Large-Scale Spectral Matching and Physics-Guided Fragment Optimization

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that automated structure elucidation from routine 1H and 13C NMR is achievable as a physics-guided fragment search, reporting 52.89% top-1 recall on experimental spectra.

desk verdict Useful integrated system and a genuinely new experimental benchmark, but the headline accuracy result is not yet proven to be elucidation beyond fragment-level retrieval; the ablation meant to address this doesn't purge target fragments and contradicts its own table. read the letter →

arxiv 2509.00640 v1 pith:PXQ7XEWT submitted 2025-08-30 physics.chem-ph cs.AI

Yongqi Jin , Jun-Jie Wang , Fanjie Xu , Xiaohong Ji , Zhifeng Gao , Linfeng Zhang , Guolin Ke , Rong Zhu
show 1 more author
Weinan E
This is my paper · ORCID
classification physics.chem-phcs.AI PACS 82.56.-b
keywords NMRspectroscopyautomatedstructureelucidationchemicalshiftpredictionfragment-basedmolecularoptimizationspectralmatchinginverseproblemsmall-moleculeidentification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

NMR-Solver tries to establish that the inverse NMR problem—recovering a molecular structure from routine 1D 1H and 13C spectra—can be solved in practice by turning it into a guided search rather than a black-box translation. It starts from a 106-million-molecule library with predicted spectra, retrieves spectrally similar candidates, fragments and recombines them under constraints that preserve local chemical environments, and scores each product by how well its predicted spectrum matches the target. On roughly 450 experimental product spectra curated from recent literature, with molecular formula provided, it achieves 52.89% top-1 and 67.33% top-10 recall with 1H and 13C spectra—about 3.5 times the top-1 recall of an end-to-end transformer baseline. If the result holds, routine 1D NMR plus a formula can produce a short list that usually contains the true structure, and the method's interpretable, evidence-traceable steps make it usable in real synthetic workflows.

What carries the argument

The central mechanism is FB-MO, fragment-NMR-based molecular optimization: a candidate pool is fragmented, fragment pairs are recombined under cleavage-bond pairing rules that preserve each carbon's radius-1 environment and most hydrogens' radius-2 environment, and the combined spectrum is estimated by inheriting parent-fragment shifts (Eq. 8). Scoring uses a fast vector similarity (Gaussian-smoothed 128-bin histograms per nucleus, concatenated to 256 dimensions) for screening and a set similarity based on optimal bipartite matching with a Gaussian kernel, solved by the Hungarian assignment algorithm, for accurate ranking. The forward model provides precise shifts for final scoring and for b

What would settle it

Measure NMR spectra of a cohort of recombination products whose fragment spectra are known, then test whether Eq. 8's inherited shifts predict the observed shifts within the ranking threshold. A complementary discriminator: run the optimizer on strained or sterically crowded targets such as ortho-disubstituted biaryls and small rings; if top-1 recall collapses selectively on these while remaining high on flexible analogues, the locality premise—not the forward model—is the weak link.

Watch

Extended reading notes

Core claim

The paper's central claim is that structure elucidation from routine 1D NMR can be reduced to a directed, evidence-driven search: retrieve known molecules whose predicted spectra resemble the target, fragment them at permissible bonds, recombine fragments into new candidates whose spectra are estimated by inheriting each atom's shift from its parent (Eq. 8, XM = XF1 ∪ XF2), then rank by precise forward prediction and spectral similarity. The claim is supported by results on simulated benchmarks, on roughly 450 experimental product spectra from recent literature (top-1 recall 52.89% with 1H and 13C plus formula, versus 14.44% for the transformer baseline), by cases where NMR-Solver solved str

Load-bearing premise

The load-bearing premise is that when two fragments are joined, the NMR shift of every atom remains what it was in its parent fragment because shifts are local; if the junction causes steric, conformational, or long-range electronic changes, the inherited-shift ranking (Eq. 8) can select the wrong structure.

Editorial extensions

If this is right

  • If correct, a chemist with only 1D 1H and 13C spectra, a molecular formula, and optionally reactant structures can obtain a ranked shortlist that usually contains the true structure, reducing reliance on 2D NMR and manual expertise in many cases.
  • The ablation removing the target from retrieval candidates shows less than a 2% recall drop, so the optimization stage—not database coincidence—does most of the work; the method should find structures not present in the simulated library.
  • Providing reactant structures lifts top-1 recall to 60.22% and top-10 to 76.22%, making reaction-aware deployment in synthesis validation a direct use case.
  • Prediction success rises monotonically with spectral match score, so the score can serve as a practical confidence indicator for accepting or re-examining a prediction.
  • Because every recombination step is traceable to spectral evidence, the framework supports human-in-the-loop validation rather than black-box output, and can flag suspicious literature assignments.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the authors leave implicit: benchmark accuracy on a cohort of strained or sterically crowded products, such as ortho-disubstituted biaryls and small fused rings. If failures concentrate where recombination changes local geometry, Eq. 8's locality premise—not the forward model—is the binding constraint, and a junction-induced shift correction would be needed.
  • The same retrieval–fragment–match loop generalizes to any modality with a forward predictor, such as IR, MS/MS, or joint multi-spectral queries; only the forward model and similarity kernel would change. This is my inference, not the paper's claim.
  • Since inherited shifts make candidate scoring cheap, the fragment library could be reused as a fast scoring oracle for library enumeration—precompute fragment spectra once and score millions of recombination products without invoking the full forward model.
  • The monotone score-accuracy relationship suggests an active-curation use: compounds whose top-1 similarity is borderline are precisely the ones worth measuring by 2D NMR or HRMS, directing experimental effort where the confidence signal is weakest.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. NMR-Solver combines large-scale retrieval over a ~106-million-molecule simulated NMR database (SimNMR-PubChem) with fragment-based molecular optimization (FB-MO) to derive small-molecule structures from 1H and 13C NMR spectra, optionally with molecular formula or reactant context. The framework scores candidates by comparing NMRNet-predicted shifts with experimental spectra via vector and set similarity, then iteratively fragments, recombines, and filters candidate structures. The paper reports competitive results on the Alberts et al. simulated benchmark, 52.89% top-1 and 67.33% top-10 recall on a manually curated JACS 2024 experimental dataset under 1H+13C+formula conditions (stereochemistry ignored), improvements when reactants are provided, and several real-world case studies, including correction of two literature misassignments.

Significance. If the central claims hold, NMR-Solver would be a practically useful, interpretable tool for routine structure elucidation from common 1D NMR data, with substantial advantages over end-to-end generative baselines in realistic settings. The manuscript ships open-source code, model weights, and a large public database, which are concrete contributions. The method is not circular in the derivation sense: the spectral scoring functions are not constructed from the target structure. However, the experimental evaluation does not currently establish de novo elucidation as opposed to database-aided retrieval, because the retrieval database includes essentially all published small molecules and the ablation does not purge target-specific fragments. The fragment-inheritance assumption underlying the search is also only weakly validated. These are fixable evaluation gaps rather than fundamental flaws.

major comments (4)
  1. [Supp. Note 3 and Supp. Table 7, Section 2.3] The ablation does not establish that optimization can solve molecules absent from the database. The text claims a drop of less than 2% when target molecules are removed from initial retrieval, but Supp. Table 7 shows larger drops: 1H+formula top-1 falls 20.00 to 16.67 (-3.33), top-3 -4.00, top-10 -3.78; 13C+formula top-10 drops -4.22; the headline 1H+13C+formula row drops -1.56/-1.33/-1.77. More importantly, the protocol removes only intact target molecules; target-specific fragments remain in the fragment vector database built from 106M PubChem molecules, so the optimizer can reassemble the target even when the intact molecule is excluded. A clean temporal holdout or a fully purged fragment database is required to support the claim that NMR-Solver performs structure elucidation beyond retrieval.
  2. [Section 2.2, Table 1] The simulated-benchmark comparison is weakened by excluding fluorine-containing compounds from the evaluation because of C-F coupling artifacts, while the prior results quoted for GraphGA and NMR-to-Structure may have been computed on the full benchmark. This changes the test set and prevents a direct 'same evaluation conditions' comparison. The authors should either evaluate on the identical full benchmark or report results on the common subset for all methods.
  3. [Section 4.5, Eq. (8)] The core search heuristic assumes that chemical shifts are inherited from parent fragments via multiset union (XM = XF1 ∪ XF2) and that the local-environment argument justifies this. No direct validation is provided for recombination products. The final scoring step uses full forward prediction, which mitigates the risk, but the fragment screening and selection that guide the search depend on inherited shifts. The strained-system failure in Fig. 3f illustrates a case where this assumption breaks. The authors should quantify the error between inherited and fully predicted shifts on a set of recombination products, ideally stratified by fragment types and ring strain.
  4. [Section 2.3, Supp. Tables 4-5] The main experimental result ignores stereochemistry, and the stereochemistry-preserving evaluation shows a large drop: for 1H+13C+formula, top-1 recall is 31.56% versus 52.89% without stereochemistry. Since NMR structure elucidation in practice often must assign relative or absolute configuration, the stereo-aware numbers should be reported prominently in the main text, and the claims about practical utility should be qualified accordingly.
minor comments (6)
  1. [Section 2.1] Typo: 'an SE(3)-equivariant Transformer architecture that that predicts' — remove duplicate 'that'.
  2. [Supp. Note 3] The statement that performance drops by 'less than 2% across all settings' is inconsistent with Supp. Table 7 entries showing drops above 2% (e.g., 1H+formula top-3 -4.00). Please correct the text or the table.
  3. [Eq. (8), Section 4.1] Eq. (8) calls XM a union, but the spectra are represented as multisets (1H shifts repeated by proton count). Clarify how duplicates and multiplicities are handled in the union and in the additive vector combination of Eq. (9).
  4. [Section 4.1] The sentence 'For literature data, features are extracted from textual reports using regular expressions from textual NMR reports' contains a repetitive phrase; consider rewording.
  5. [Figure 2] The caption duplicates 'b, c' labels ('b, c. Results ...'), and 'Tanimoto' should be capitalized consistently.
  6. [Section 2.6] The real-world validation would benefit from reporting the size of the candidate pool and wall-clock runtime per case, since the claimed practical utility depends on computational feasibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: NMR-Solver's scoring and optimization are not equivalent to its inputs by construction; the main caveat is an imperfect ablation, not a derivation loop.

full rationale

The paper's claimed derivation chain is self-contained against external benchmarks. The forward model NMRNet [16] is used both to build SimNMR-PubChem and to score candidates, but it is an independently published, code-released model with benchmarked MAEs, so relying on it is a dependency, not a circularity. Eq. 8 (XM = XF1 ∪ XF2) is explicitly a fast screening approximation; the final candidate scoring is done by the full forward model (Section 4.5, 'Newly generated molecules undergo precise NMR prediction using the forward model'), so no predicted structure is defined by construction as a union of input fragments. The experimental JACS evaluation uses external spectra and ground-truth structures, not quantities fitted from the target set. The ablations in Supp. Note 3/Table 7 do raise a validation concern: only intact target molecules are removed from the initial pool, not target fragments from the fragment-vector database, and the text's 'less than 2%' claim is contradicted by Table 7 (e.g., 1H+formula top-1 drops 3.33 points). This weakens the attribution of performance to optimization rather than retrieval, but it is a control flaw, not a circular reduction of the output to the input. No equation or fitted parameter makes the target structure equivalent to the input spectra by construction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central result rests on the fidelity of NMRNet's shift predictions and on the locality of NMR chemical shifts. The paper does not fit these to the target spectra, so they are inputs rather than circular outputs, but they are unverified for novel scaffolds. The ledger contains no invented physical entities and several domain assumptions that are plausible but only partially tested.

free parameters (5)
  • Gaussian kernel width for vector similarity = 1H: 0.3 ppm; 13C: 2 ppm
    Controls spectral smoothing and discretization in Supplementary Table 1; retrieval and ranking depend on it.
  • Gaussian kernel width for set similarity = 1H: 1 ppm; 13C: 10 ppm
    Controls peak matching tolerance in Eq. 7; no sensitivity analysis is reported.
  • Multiplicity mismatch weights = w1=1.0, w2=0.8
    Supplementary Note 4; weights modify set similarity when multiplicity patterns match or mismatch; defaults chosen by hand.
  • FB-MO search and pool parameters = num_search=1000, num_pool=1000, num_filter_pair=200000, num_filter_mol=1000
    Supplementary Table 3; algorithm capacity parameters that affect exploration versus computational cost.
  • Spectral sampling grid = 128 points; 1H [-1,15] ppm; 13C [-10,230] ppm
    Supplementary Table 1; discretization for the vector similarity representation.
assumptions (6)
  • domain assumption Chemical shifts are primarily determined by local chemical structure; long-range effects decay rapidly and are negligible.
    Invoked in Section 4.5 to justify fragment shift inheritance (Eq. 8) and cleavage-bond constraints. This is the load-bearing locality assumption.
  • domain assumption NMRNet predicts 1H and 13C chemical shifts with enough accuracy (MAE 0.181 ppm and 1.098 ppm) to rank candidate structures.
    NMRNet is used for database construction and final scoring; the authors admit in Section 2.6 and Fig. 3f that a strained-system prediction error broke the method.
  • domain assumption The SimNMR-PubChem database of 106M molecules with NMRNet-predicted spectra provides sufficiently close candidates to seed optimization.
    Section 4.4; if the database lacks relevant scaffolds or predicted shifts are too inaccurate, retrieval cannot initialize useful pools.
  • domain assumption Unordered multiset peak representation, with integration encoded as repeated shifts and symmetry-equivalent carbons averaged, captures the experimentally relevant information in 1H and 13C spectra.
    Sections 4.1-4.3; this discards coupling fine structure and peak shapes, which is deliberate but limits information use.
  • domain assumption The manually curated JACS reactant-product pairs are correctly extracted and assigned.
    Section 4.1 and Section 2.3; ground truth comes from literature reports and manual peak curation, so errors would propagate to recall estimates.
  • standard math The Kuhn-Munkres algorithm gives an optimal one-to-one matching for the set similarity metric.
    Section 4.3, Eq. 6; standard combinatorial optimization, not contested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NMR-Solver: Automated Structure Elucidation via Large-Scale Spectral Matching and Physics-Guided Fragment Optimization." pith.science (2026). https://pith.science/paper/PXQ7XEWT

@misc{pith2026250900640,
  author       = {Pith},
  title        = {Pith review of: NMR-Solver: Automated Structure Elucidation via Large-Scale Spectral Matching and Physics-Guided Fragment Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PXQ7XEWT}},
  note         = {Machine review of arXiv:2509.00640}
}
abstract

Nuclear Magnetic Resonance (NMR) spectroscopy is one of the most powerful and widely used tools for molecular structure elucidation in organic chemistry. However, the interpretation of NMR spectra to determine unknown molecular structures remains a labor-intensive and expertise-dependent process, particularly for complex or novel compounds. Although recent methods have been proposed for molecular structure elucidation, they often underperform in real-world applications due to inherent algorithmic limitations and limited high-quality data. Here, we present NMR-Solver, a practical and interpretable framework for the automated determination of small organic molecule structures from $^1$H and $^{13}$C NMR spectra. Our method introduces an automated framework for molecular structure elucidation, integrating large-scale spectral matching with physics-guided fragment-based optimization that exploits atomic-level structure-spectrum relationships in NMR. We evaluate NMR-Solver on simulated benchmarks, curated experimental data from the literature, and real-world experiments, demonstrating its strong generalization, robustness, and practical utility in challenging, real-life scenarios. NMR-Solver unifies computational NMR analysis, deep learning, and interpretable chemical reasoning into a coherent system. By incorporating the physical principles of NMR into molecular optimization, it enables scalable, automated, and chemically meaningful molecular identification, establishing a generalizable paradigm for solving inverse problems in molecular science.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Data Fusion and Contrastive Alignment for Unconstrained IR Molecular Structure Elucidation

    cs.LG 2026-07 conditional novelty 6.0 of 10

    A formula-free transformer with fuzzy MoE aggregation and contrastive alignment improves IR-to-SMILES elucidation, reaching 31.8% Top-1 on experimental NIST spectra, but the >10-point gains occur only on the QM9S subset.

  2. SpecXMaster Technical Report

    cs.LG 2026-03 conditional novelty 6.0 of 10

    An agentic reinforcement-learning system using raw FID NMR data, generate/search/repair/rerank tools, and a GRPO-trained policy reports improved hit@1 structure elucidation versus its own generation baseline and untra...

  3. NMIRacle: Multi-modal Generative Molecular Elucidation from IR and NMR Spectra

    physics.chem-ph 2025-12 conditional novelty 5.0 of 10

    NMIRacle generates molecular structures from combined raw IR and NMR spectra, improving Top-1 elucidation accuracy from 0.41 to 0.48 over the NMR2Struct baseline on the Alberts benchmark.

Reference graph

Works this paper leans on

50 extracted references · 46 canonical work pages · cited by 3 Pith papers

  1. [1]

    & Warren, S

    Clayden, J., Greeves, N. & Warren, S. Organic chemistry (Oxford university press, 2012)

    work page 2012

  2. [2]

    principles of instrumental analysis

    Skoog, D. A., Holler, F. J. & Crouch, S. R. Textbook “principles of instrumental analysis”. Cengage learning 6 (2019)

    work page 2019

  3. [3]

    & Martin, G

    Elyashberg, M., Williams, A. & Martin, G. Computer-assisted structure veri- fication and elucidation tools in NMR-based structure elucidation.Progress in Nuclear Magnetic Resonance Spectroscopy 53, 1–104 (2008)

    work page 2008

  4. [4]

    E., Agback, T

    Ermanis, K., Parkes, K. E., Agback, T. & Goodman, J. M. Doubling the power of DP4 for computational structure elucidation.Organic & Biomolecular Chemistry 15, 8998–9007 (2017)

    work page 2017

  5. [5]

    & Goodman, J

    Howarth, A., Ermanis, K. & Goodman, J. M. DP4-AI automated NMR data analysis:straightfromspectrometertostructure. Chemical science11,4351–4359 (2020)

    work page 2020

  6. [6]

    O., Zanardi, M

    Marcarino, M. O., Zanardi, M. M., Cicetti, S. & Sarotti, A. M. NMR calculations with quantum methods: development of new tools for structural elucidation and beyond. Accounts of Chemical Research 53, 1922–1932 (2020)

    work page 1922

  7. [7]

    ACD Labs

    NMR workbook suite. ACD Labs. https://www.acdlabs.com/products/ spectrus-platform/nmr-workbook-suite. Accessed: 2025-07-25

    work page 2025

  8. [8]

    MestreLab Research

    MNova. MestreLab Research. https://mestrelab.com/software/mestrenova. Accessed: 2025-07-25

    work page 2025

Show all 50 references
  1. [9]

    Ruddigkeit, L., Van Deursen, R., Blum, L. C. & Reymond, J.-L. Enumeration of 166 billion organic small molecules in the chemical universe database GDB-17. Journal of chemical information and modeling 52, 2864–2875 (2012)

    2012

  2. [10]

    & Burke, M

    Trobe, M. & Burke, M. D. The molecular industrial revolution: automated synthesis of small molecules. Angewandte Chemie International Edition 57, 4192–4214 (2018)

    2018

  3. [11]

    & Kuhn, S

    Jonas, E. & Kuhn, S. Rapid prediction of NMR spectral properties with quantified uncertainty. Journal of cheminformatics 11, 50 (2019)

    2019

  4. [12]

    Han, J. et al. Scalable graph neural network for NMR chemical shift prediction. Physical Chemistry Chemical Physics 24, 26870–26878 (2022)

    2022

  5. [13]

    Zou, Z. et al. A deep learning model for predicting selected organic molecular spectra. Nature Computational Science 3, 957–964 (2023). 18

    2023

  6. [14]

    Wolinski, K., Hinton, J. F. & Pulay, P. Efficient implementation of the gauge- independent atomic orbital method for NMR chemical shift calculations.Journal of the American Chemical Society 112, 8251–8260 (1990)

    1990

  7. [15]

    Chen, H., Liang, T., Tan, K., Wu, A. & Lu, X. GT-NMR: a novel graph transformer-based approach for accurate prediction of NMR chemical shifts. Journal of Cheminformatics 16, 132 (2024)

    2024

  8. [16]

    Toward a unified benchmark and framework for deep learning-based prediction of nuclear magnetic resonance chemical shifts.Nature Computational Science 1–9 (2025)

    Xu, F.et al. Toward a unified benchmark and framework for deep learning-based prediction of nuclear magnetic resonance chemical shifts.Nature Computational Science 1–9 (2025)

    2025

  9. [17]

    Conditional molecular generation net enables automated structure elucidationbasedon13CNMRspectraandpriorknowledge

    Yao, L.et al. Conditional molecular generation net enables automated structure elucidationbasedon13CNMRspectraandpriorknowledge. Analytical chemistry 95, 5393–5401 (2023)

    2023

  10. [18]

    S., Rotskoff, G

    Hu, F., Chen, M. S., Rotskoff, G. M., Kanan, M. W. & Markland, T. E. Accurate and efficient structure elucidation from routine one-dimensional NMR spectra using multitask machine learning.ACS Central Science 10, 2162–2170 (2024)

    2024

  11. [20]

    SMILES, a chemical language and information system

    Weininger, D. SMILES, a chemical language and information system. 1. intro- duction to methodology and encoding rules.Journal of chemical information and computer sciences 28, 31–36 (1988)

    1988

  12. [21]

    Brown, N., Fiscato, M., Segler, M. H. & Vaucher, A. C. GuacaMol: benchmark- ing models for de novo molecular design.Journal of chemical information and modeling 59, 1096–1108 (2019)

    2019

  13. [22]

    & Hernández-Lobato, J

    Tripp, A. & Hernández-Lobato, J. M. Genetic algorithms are strong baselines for molecule generation. arXiv preprint arXiv:2310.09267 (2023)

    2023 arXiv

  14. [23]

    Jensen, J. H. A graph-based genetic algorithm and generative model/monte carlo tree search for the exploration of chemical space.Chemical science 10, 3567–3572 (2019)

    2019

  15. [24]

    & Jablonka, K

    Mirza, A. & Jablonka, K. M. Elucidating structures from spectra using mul- timodal embeddings and discrete optimization. ChemRxiv 10.26434/chemrxiv– 2024–f3b18–v2 (2024)

    2024 doi

  16. [25]

    C., Mazzola, E

    Burns, D. C., Mazzola, E. P. & Reynolds, W. F. The role of computer-assisted structure elucidation (CASE) programs in the structure elucidation of complex natural products. Natural product reports 36, 919–933 (2019). 19

    2019

  17. [26]

    Yang, Z. et al. Cross-modal retrieval between 13C NMR spectra and structures forcompoundidentificationusingdeepcontrastivelearning. Analytical Chemistry 93, 16947–16955 (2021)

    2021

  18. [27]

    Sun, H. et al. Cross-modal retrieval between 13C NMR spectra and structures based on focused libraries.Analytical Chemistry 96, 5763–5770 (2024)

    2024

  19. [28]

    Vaswani, A. et al. Attention is all you need. Advances in neural information processing systems 30 (2017)

    2017

  20. [29]

    Kim, S. et al. PubChem 2025 update. Nucleic acids research 53, D1516–D1525 (2025)

    2025

  21. [30]

    & Schlörer, N

    Kuhn, S. & Schlörer, N. E. Facilitating quality control for spectra assignments of small organic molecules: nmrshiftdb2–a free in-house NMR database with inte- grated LIMS for academic service laboratories.Magnetic Resonance in Chemistry 53, 582–589 (2015)

    2015

  22. [31]

    Wishart, D. S. et al. NP-MRD: the natural products magnetic resonance database. Nucleic Acids Research 50, D665–D677 (2022)

    2022

  23. [32]

    & Ramakrishnan, R

    Gupta, A., Chakraborty, S. & Ramakrishnan, R. Revving up 13C NMR shield- ing predictions across chemical space: benchmarks for atoms-in-molecules kernel machine learning with new data for 134 kilo molecules. Machine Learning: Science and Technology 2, 035010 (2021)

    2021

  24. [33]

    & Laino, T

    Alberts, M., Schilter, O., Zipoli, F., Hartrampf, N. & Laino, T. Unraveling molecular structure: A multimodal spectroscopic dataset for chemistry.Advances in Neural Information Processing Systems 37, 125780–125808 (2024)

    2024

  25. [34]

    https://www.bohrium.com/apps/nmr-toolbox (2025)

    NMR Toolbox. https://www.bohrium.com/apps/nmr-toolbox (2025)

    2025

  26. [35]

    & Héberger, K

    Bajusz, D., Rácz, A. & Héberger, K. Why is tanimoto index an appropriate choice for fingerprint-based similarity calculations?Journal of cheminformatics 7, 20 (2015)

    2015

  27. [36]

    Morgan, H. L. The generation of a unique machine description for chemi- cal structures-a technique developed at chemical abstracts service. Journal of chemical documentation 5, 107–113 (1965)

    1965

  28. [37]

    Fu, Y. et al. Photocatalyzed dehydroxylative amination of phenols: A ring- expansion approach for medium-sized benzolactams. Organic Letters 23, 8317–8321 (2021)

    2021

  29. [38]

    Cheng, D., Yu, C., Pu, Y. & Xu, X. DDQ-mediated oxidative coupling reaction of N, N-dimethyl enaminones with cycloheptatriene. Tetrahedron Letters 90, 153609 (2022). 20

    2022

  30. [39]

    Novitskiy, I. M. & Kutateladze, A. G. Peculiar reaction products and mechanisms revisited with machine learning-augmented computational NMR.The Journal of Organic Chemistry 87, 8589–8598 (2022)

    2022

  31. [40]

    Landrum, G. et al. RDKit: Open-source cheminformatics software (2016)

    2016

  32. [41]

    Merckmolecularforcefield.i.basis,form,scope,parameterization, and performance of MMFF94.Journal of computational chemistry 17, 490–519 (1996)

    Halgren,T.A. Merckmolecularforcefield.i.basis,form,scope,parameterization, and performance of MMFF94.Journal of computational chemistry 17, 490–519 (1996)

    1996

  33. [42]

    Kuhn, H. W. The Hungarian method for the assignment problem.Naval research logistics quarterly 2, 83–97 (1955)

    1955

  34. [43]

    Algorithms for the assignment and transportation problems.Journal of the Society for Industrial and Applied Mathematics 5, 32–38 (1957)

    Munkres, J. Algorithms for the assignment and transportation problems.Journal of the Society for Industrial and Applied Mathematics 5, 32–38 (1957). URL https://doi.org/10.1137/0105003

    1957 doi

  35. [44]

    Crouse, D. F. On implementing 2D rectangular assignment algorithms. IEEE Transactions on Aerospace and Electronic Systems 52, 1679–1696 (2016)

    2016

  36. [45]

    & Jégou, H

    Johnson, J., Douze, M. & Jégou, H. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7, 535–547 (2019)

    2019

  37. [46]

    Malkov, Y. A. & Yashunin, D. A. Efficient and robust approximate nearest neigh- bor search using hierarchical navigable small world graphs.IEEE transactions on pattern analysis and machine intelligence 42, 824–836 (2018)

    2018

  38. [47]

    Hose—a novel substructure code

    Bremser, W. Hose—a novel substructure code. Analytica Chimica Acta 103, 355–365 (1978)

    1978

  39. [48]

    Understanding NMR spectroscopy (John Wiley & Sons, 2011)

    Keeler, J. Understanding NMR spectroscopy (John Wiley & Sons, 2011)

    2011

  40. [49]

    Tanimoto, T. T. An elementary mathematical theory of classification and prediction (1958)

    1958

  41. [50]

    & Hahn, M

    Rogers, D. & Hahn, M. Extended-connectivity fingerprints.Journal of chemical information and modeling 50, 742–754 (2010)

    2010

  42. [51]

    & Guibas, L

    Rubner, Y., Tomasi, C. & Guibas, L. J. The earth mover’s distance as a metric for image retrieval.International journal of computer vision 40, 99–121 (2000). 21 Supplementary Information Supplementary Notes Supplementary Note 1. Web App The web-based application, accessible at...

    2000

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.