Pith. sign in

REVIEW 3 major objections 5 minor 39 references

NPGPT: Natural Product-Like Compound Generation with GPT-based Chemical Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read Fine-tuning a GPT model on natural product SMILES yields molecules closer to natural products than a prior LSTM, and docking finds hits beating gefitinib.

desk verdict Useful but uncontrolled GPT fine-tuning study for natural-product generation; the FCD improvement over Tay et al. is plausible but not statistically established. read the letter →

arxiv 2411.12886 v1 pith:3PL7WD5D submitted 2024-11-19 q-bio.BM cs.LGq-bio.QMstat.ML

classification q-bio.BMcs.LGq-bio.QMstat.ML
keywords naturalproductchemicallanguagemodelSMILESSELFIESfine-tuningFréchetChemNetDistancemoleculargenerationprotein-liganddocking
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

This paper attempts to show that fine-tuning a GPT-style chemical language model on a large natural product corpus generates molecules whose distribution closely matches those of natural products, more closely than a previous LSTM-based model. The authors fine-tune two pretrained models, one using SMILES strings and one using SELFIES, on the COCONUT natural product dataset, and report that the SMILES-based model reaches a Fréchet ChemNet Distance of 1.290 to the natural product distribution, beating the LSTM baseline's 1.794. They also report that docking 1,000 generated molecules against EGFR yields 1,216 conformers (9.8% of 12,930) with GlideScore better than gefitinib's -7.02 kcal/mol. If valid, the method would let researchers explore natural-product-like chemical space cheaply and generate lead candidates for drug discovery.

What carries the argument

The machinery is a chemical language model: a decoder-only Transformer (GPT architecture) that learns to autoregressively generate molecular string representations, either SMILES or SELFIES. The transfer mechanism is fine-tuning: a model pretrained on PubChem-10M molecular strings is further trained on approximately 3.6 million SMILES strings created by enumerating and augmenting the ~400,000 natural products of COCONUT. The central comparison metric is the Fréchet ChemNet Distance (FCD), which measures the distance between the distribution of generated molecules and the training distribution, and the central application is Glide docking of generated compounds against the EGFR structure with PDB ID 2ITY, where gefitinib's reported GlideScore of -7.02 kcal/mol serves as the reference.

What would settle it

Retrain the fine-tuned smiles-gpt and the LSTM baseline on the identical pretraining weights and matched fine-tuning budgets across several random seeds, and compare FCD distributions; if the gap between 1.290 and 1.794 does not reliably reproduce, the claimed superiority collapses. Additionally, redock gefitinib with the same Glide version and conformer-generation protocol; if 1,216 of 12,930 conformers no longer exceed its score, the drug-candidate claim weakens.

Watch

Extended reading notes

Core claim

The central claim is that fine-tuning a pretrained, decoder-only chemical language model on natural product SMILES yields a generator of natural product-like molecules. Specifically, the fine-tuned smiles-gpt model produces a 100-million-molecule set with FCD 1.290 to the COCONUT distribution, closer than the 1.794 reported for the LSTM baseline [11], while the SELFIES-based ChemGPT model, fine-tuned on the same data, reaches only FCD 14.28 and is judged not to have captured the natural product distribution. The paper further claims that a subset of generated molecules are viable drug candidates: of 12,930 docked conformers from 1,000 generated molecules, 1,216 have GlideScore better than gefitinib's -7.02 kcal/mol, with the best at -11.51 kcal/mol. The authors argue that the SMILES representation, combined with fine-tuning, is responsible for the improvement over existing methods.

Load-bearing premise

The load-bearing premise is that the FCD gap between the fine-tuned smiles-gpt model and the LSTM baseline reflects genuine model superiority, even though the models differ in pretraining corpus, molecular representation, batch size, learning rate, and possibly training duration, and no variance or statistical significance is reported.

Editorial extensions

If this is right

  • If the central claim holds, fine-tuning a SMILES-based GPT model on natural products yields libraries that sit closer to natural product space than the best prior generative model, as measured by FCD.
  • The docking results imply that a fraction of generated molecules can match or exceed an approved EGFR inhibitor's predicted binding affinity, making them concrete leads for synthesis and assay.
  • The sharp contrast with the SELFIES-based ChemGPT model implies that molecular representation choice materially affects whether fine-tuning captures a chemical distribution, not just validity.
  • The NP Score and SA Score distributions shifting toward COCONUT after fine-tuning suggest the generator produces scaffolds typical of natural products rather than memorized database entries.

Reading between the lines

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

  • A matched-budget retraining study, holding pretraining corpus, tokenization, batch size, and learning rate constant across architectures, would be needed to attribute the FCD difference to the model family rather than training details; the paper does not provide such a comparison.
  • The docking enrichment over gefitinib may partly reflect the already-noted correlation between molecular weight and GlideScore; testing with ligand efficiency or molecular-weight-matched controls would clarify whether natural product-likeness itself drives the better scores.
  • If validated experimentally, the same fine-tuning recipe could be applied to narrower biosynthetic classes (e.g., marine natural products or specific scaffolds), but the paper only demonstrates the broad COCONUT distribution.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper fine-tunes two GPT-based chemical language models, smiles-gpt (SMILES) and ChemGPT (SELFIES), on the COCONUT natural product database, generates 100 million molecules from each, and evaluates them with validity, uniqueness, novelty, internal diversity, and Fréchet ChemNet Distance (FCD). The authors report that fine-tuned smiles-gpt reaches FCD 1.290, which they interpret as being closer to the COCONUT distribution than the FCD 1.794 reported by Tay et al. for an LSTM model. They also visualize the distributions with t-SNE and NP/SA score kernel density estimates, and dock 1,000 smiles-gpt-generated molecules against EGFR, finding 1,216 conformers with GlideScore better than gefitinib's -7.02 kcal/mol. The paper concludes that the fine-tuned smiles-gpt generates natural product-like compounds and can serve as a source of drug candidates.

Significance. If the central comparison is reliable, the paper demonstrates a useful practical result: a pretrained GPT-2 chemistry model, after fine-tuning on a natural product corpus, can generate molecules whose aggregate distribution is close to natural products, and some of these molecules show promising docking scores. The authors have released code and generated molecule sets, which supports reproducibility. The docking section is commendable for acknowledging the known correlation between molecular weight and GlideScore. However, the main quantitative comparison against Tay et al. is confounded by multiple uncontrolled differences between models and training protocols, and no uncertainty quantification is reported, so the paper's headline claim of outperforming the previous LSTM approach is not yet established.

major comments (3)
  1. [Section 3.1, Table 2] The central claim that fine-tuned smiles-gpt generates molecules closer to natural products than Tay et al.'s LSTM (FCD 1.290 versus 1.794) is not supported by the evidence as presented. The two models differ in architecture (GPT-2 vs. LSTM), pretraining corpus (PubChem vs. none), molecular representation, batch size, learning rate schedule, and data augmentation, and Table 2 reports only a single FCD value per model with no confidence intervals, bootstrap replicates, or repeated sampling. The paper itself states in Section 3.3 that definitive statements are difficult 'due to differences in training conditions and model specifics,' yet the Conclusion restates the comparison without this caveat. To make the claim load-bearing, the authors should either run a controlled comparison (e.g., same architecture trained from scratch on the same augmented COCONUT data, multiple random seeds, and bootstrap CIs) or explicitly reframe the FCD improvement as descriptive rather than causal.
  2. [Section 3.4, Figures 8 and 9] The docking result that 1,216 conformers (9.8% of 12,930) have GlideScore better than gefitinib is presented as evidence of viable drug candidates, but the comparison is not adjusted for molecular size. The paper itself reports a weak negative correlation between molecular weight and GlideScore (r = -0.358, Figure 9), and the generated compounds are relatively large and complex, as acknowledged in the discussion of SA Scores in Table 3. A fairer evaluation would use ligand efficiency or a size-matched control set, and would report the distribution of GlideScores for a chemically matched reference library rather than a raw count of conformers.
  3. [Section 3.3, Figures 4 and 5] The claim that fine-tuned smiles-gpt produces NP Score and SA Score distributions 'closer' to COCONUT than Tay et al. is based on visual inspection of kernel density estimates, with no quantitative distance measure or significance test. Because this is one of the main supports for the conclusion, I ask the authors to add a numerical comparison, such as the Kolmogorov-Smirnov statistic, Wasserstein distance, or bootstrap confidence intervals on the score distributions.
minor comments (5)
  1. [Abstract and Section 3.1] The statement that 'the distribution of the compounds generated was similar to that of natural products' is too strong without error bars or statistical tests; consider softening to 'was closer to that of natural products than the previous model' or adding the missing uncertainty.
  2. [Section 2.4] The training description does not report the number of epochs or total optimization steps for either model, and the data augmentation seed or enumeration procedure is not specified; these details should be added for reproducibility.
  3. [Figures 2 and 3] The t-SNE visualizations do not report the perplexity, learning rate, or random seed used; since t-SNE is stochastic and sensitive to these settings, please provide them and ideally show multiple runs.
  4. [Table 3] The 'most similar natural product' column contains only images, without identifiers or SMILES strings, making it hard to verify the similarity claims; providing COCONUT IDs or SMILES would help.
  5. [Section 3.4] The docking setup omits details about receptor preparation, grid center, and docking mode (HTVS/SP/XP), which are needed to reproduce the GlideScore values; please add these or cite a protocol.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical fine-tuning study; FCD, novelty, and docking metrics are not equivalent to the model's inputs by construction.

full rationale

The paper does not present a derivation chain in which an output is defined in terms of an input. The central claim is that smiles-gpt fine-tuned on COCONUT produces molecules with FCD 1.290 to the COCONUT distribution, closer than Tay et al.'s 1.794. FCD is a distributional distance evaluated after training, not a fitted parameter or a quantity defined by the training objective; the model could have failed to match COCONUT, as ChemGPT did with FCD 14.28. Novelty is a post-hoc count of generated molecules absent from COCONUT, and although it references the training set, the metric does not force any particular value. The docking experiment uses an external GlideScore baseline (-7.02 kcal/mol for gefitinib, cited to [33]) and an independent docking pipeline; even if the baseline is a self-citation, it is an externally falsifiable empirical value rather than an assumption that entails the generated compounds' scores. The paper also acknowledges that training-condition differences limit definitive comparisons to Tay et al. No equation or fitted parameter reduces to its own input, so no significant circularity is found.

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

The paper introduces no new entities or mathematical derivations. Its central claims rest on standard cheminformatics metrics and the implicit assumption that fine-tuning a pretrained language model on COCONUT produces a distribution close to the training data. The free parameters listed are training and evaluation choices that differ between models and are not quantified with uncertainty.

free parameters (6)
  • Data augmentation factor = approximately 9x
    SMILES enumeration randomly rewrites each molecule about nine times; the exact factor changes the training set size and distribution, affecting all downstream metrics.
  • Size filtering thresholds = atom count > 150, ring count > 10 removed
    Large compounds were removed from COCONUT before training, altering the target distribution and therefore the FCD and NP Score comparisons.
  • Learning rate schedule = 5e-4 to 5e-8 cosine (smiles-gpt); 5e-5 fixed (ChemGPT)
    The models were trained with different learning rates and schedules, confounding the FCD comparison between them.
  • Batch size = 256 (smiles-gpt), 32 (ChemGPT)
    Batch size differed due to GPU memory constraints, creating another uncontrolled difference between models.
  • Number of generated molecules = 100 million
    The evaluation metrics are computed on 100M generated molecules per model; this large but finite sample affects uniqueness, novelty, and FCD stability.
  • Docking sample size = 1,000 molecules
    Only 1,000 of the generated molecules were randomly selected for docking, leading to 12,930 conformers; the small sample limits the statistical power of the docking analysis.
assumptions (6)
  • domain assumption Pretrained GPT models on PubChem-10M have learned valid chemical syntax and structural patterns.
    The authors hypothesize in Section 2.1 that pretrained models 'have already learned chemical structures,' so fine-tuning can efficiently adapt them to natural products.
  • domain assumption Fréchet ChemNet Distance (FCD) is a valid and sufficient measure of similarity between generated and real molecule distributions.
    FCD is used as the primary evidence for natural product likeness in Section 3.1; no alternative distributional metrics or statistical tests are reported.
  • domain assumption NP Score and SA Score are reliable indices for natural product likeness and synthetic accessibility.
    Section 3.3 uses these scores to claim that smiles-gpt approaches COCONUT distributions; the validity of these scores is taken from the literature.
  • domain assumption GlideScore predicts binding affinity and drug candidacy.
    Section 3.4 interprets docking scores as indicating 'viable drug candidates' without experimental validation, and the MW correlation is acknowledged as a confound.
  • domain assumption Mean Tanimoto similarity of Morgan fingerprints to COCONUT compounds measures natural product similarity.
    Used in Section 3.4 to correlate NP similarity with GlideScore; the choice of fingerprint radius and bit length affects the result.
  • standard math SMILES enumeration preserves molecular identity while varying string order.
    Data augmentation relies on this property from Bjerrum (2017), cited in Section 2.2; it is a standard but unverified-for-COCONUT assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NPGPT: Natural Product-Like Compound Generation with GPT-based Chemical Language Models." pith.science (2026). https://pith.science/paper/3PL7WD5D

@misc{pith2026241112886,
  author       = {Pith},
  title        = {Pith review of: NPGPT: Natural Product-Like Compound Generation with GPT-based Chemical Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3PL7WD5D}},
  note         = {Machine review of arXiv:2411.12886}
}
read the original abstract

Natural products are substances produced by organisms in nature and often possess biological activity and structural diversity. Drug development based on natural products has been common for many years. However, the intricate structures of these compounds present challenges in terms of structure determination and synthesis, particularly compared to the efficiency of high-throughput screening of synthetic compounds. In recent years, deep learning-based methods have been applied to the generation of molecules. In this study, we trained chemical language models on a natural product dataset and generated natural product-like compounds. The results showed that the distribution of the compounds generated was similar to that of natural products. We also evaluated the effectiveness of the generated compounds as drug candidates. Our method can be used to explore the vast chemical space and reduce the time and cost of drug discovery of natural products.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 33 canonical work pages

  1. [1]

    Metabolites 2(2), 303–336 (2012)

    Dias, D.A., Urban, S., Roessner, U.: A historical overview of natural products in drug discovery. Metabolites 2(2), 303–336 (2012)

  2. [2]

    Cragg, G.M., Newman, D.J.: Natural products: a continuing source of novel drug leads. Biochim. Biophys. Acta 1830(6), 3670–3695 (2013)

  3. [3]

    Newman, D.J., Cragg, G.M.: Natural products as sources of new drugs from 1981 to 2014. J. Nat. Prod. 79(3), 629–661 (2016)

  4. [4]

    Demain, A.L.: Importance of microbial natural products and the need to revitalize their discovery. J. Ind. Microbiol. Biotechnol. 41(2), 185–201 (2014)

  5. [5]

    Pelaez, F.: The historical delivery of antibiotics from microbial natural prod- ucts—can history repeat? Biochemical pharmacology 71(7), 981–990 (2006)

  6. [6]

    Cell 163(6), 1297–1300 (2015)

    Shen, B.: A new golden age of natural products drug discovery. Cell 163(6), 1297–1300 (2015)

  7. [7]

    Li, J.W.-H., Vederas, J.C.: Drug discovery and natural products: end of an era or an endless frontier? Science 325(5937), 161–165 (2009)

  8. [8]

    Wiley Interdiscip

    Bilodeau, C., Jin, W., Jaakkola, T., Barzilay, R., Jensen, K.F.: Generative models for molecular discovery: Recent advances and challenges. Wiley Interdiscip. Rev. Comput. Mol. Sci. 12(5) (2022) 14

Show all 39 references
  1. [9]

    Nucleic Acids Res

    Kim, S., Chen, J., Cheng, T., Gindulyte, A., He, J., He, S., Li, Q., Shoemaker, B.A., Thiessen, P.A., Yu, B., Zaslavsky, L., Zhang, J., Bolton, E.E.: PubChem 2023 update. Nucleic Acids Res. 51(D1), 1373–1380 (2023)

  2. [10]

    In: Proceedings of the 37th International Conference on Machine Learning, vol

    Jin, W., Barzilay, D.R., Jaakkola, T.: Hierarchical generation of molecular graphs using structural motifs. In: Proceedings of the 37th International Conference on Machine Learning, vol. 119, pp. 4839–4848 (2020)

  3. [11]

    Tay, D.W.P., Yeo, N.Z.X., Adaikkappan, K., Lim, Y.H., Ang, S.J.: 67 million natu- ral product-like compound database generated via molecular language processing. Sci. Data 10(1), 296 (2023)

  4. [12]

    Sorokina, M., Merseburger, P., Rajan, K., Yirik, M.A., Steinbeck, C.: COCONUT online: Collection of open natural products database. J. Cheminform. 13(1), 2 (2021)

  5. [13]

    Weininger, D.: SMILES, a chemical language and information system. 1. intro- duction to methodology and encoding rules. J. Chem. Inf. Comput. Sci. 28(1), 31–36 (1988)

  6. [14]

    Krenn, M., H¨ ase, F., Nigam, A., Friederich, P., Aspuru-Guzik, A.: Self-referencing embedded strings (SELFIES): A 100% robust molecular string representation. Mach. Learn. Sci. Technol. 1(4), 045024 (2020)

  7. [15]

    arXiv preprint (arXiv:1703.07076) (2017)

    Bjerrum, E.J.: SMILES enumeration as data augmentation for neural network modeling of molecules. arXiv preprint (arXiv:1703.07076) (2017)

  8. [16]

    ChemRxiv (10.26434/chemrxiv-2021-5fwjd) (2021)

    Adilov, S.: Generative pre-training from molecules. ChemRxiv (10.26434/chemrxiv-2021-5fwjd) (2021)

  9. [17]

    Frey, N.C., Soklaski, R., Axelrod, S., Samsi, S., G´ omez-Bombarelli, R., Coley, C.W., Gadepally, V.: Neural scaling of deep chemical models. Nat. Mach. Intellig. 5, 1297–1305 (2023)

  10. [18]

    arXiv preprint (arXiv:2010.09885) (2020)

    Chithrananda, S., Grand, G., Ramsundar, B.: ChemBERTa: large-scale self- supervised pretraining for molecular property prediction. arXiv preprint (arXiv:2010.09885) (2020)

  11. [19]

    https://openai.com/ research/better-language-models

    Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Lan- guage models are unsupervised multitask learners (2019). https://openai.com/ research/better-language-models

  12. [20]

    https://doi.org/10

    Black, S., Gao, L., Wang, P., Leahy, C., Biderman, S.: GPT-Neo: Large Scale Autoregressive Language Modeling with mesh-Tensorflow. https://doi.org/10. 5281/zenodo.5297715 . https://doi.org/10.5281/zenodo.5297715

  13. [21]

    arXiv preprint 15 (arXiv:1711.05101) (2017)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint 15 (arXiv:1711.05101) (2017)

  14. [22]

    Polykovskiy, D., Zhebrak, A., Sanchez-Lengeling, B., Golovanov, S., Tatanov, O., Belyaev, S., Kurbanov, R., Artamonov, A., Aladinskiy, V., Veselov, M., Kadurin, A., Johansson, S., Chen, H., Nikolenko, S., Aspuru-Guzik, A., Zhavoronkov, A.: Molecular sets (MOSES): A benchmarkin...

  15. [23]

    Preuer, K., Renz, P., Unterthiner, T., Hochreiter, S., Klambauer, G.: Fr´ echet ChemNet distance: A metric for generative models for molecules in drug discovery. J. Chem. Inf. Model. 58(9), 1736–1741 (2018)

  16. [24]

    Landrum, G., et al.: RDKit: Open-source Cheminformatics

  17. [25]

    Ertl, P., Roggo, S., Schuffenhauer, A.: Natural product-likeness score and its application for prioritization of compound libraries. J. Chem. Inf. Model. 48(1), 68–74 (2008)

  18. [26]

    Ertl, P., Schuffenhauer, A.: Estimation of synthetic accessibility score of drug- like molecules based on molecular complexity and fragment contributions. J. Cheminform. 1(8) (2009)

  19. [27]

    Advances in neural information processing systems 35, 21342–21357 (2022)

    Gao, W., Fu, T., Sun, J., Coley, C.: Sample efficiency matters: a benchmark for practical molecular optimization. Advances in neural information processing systems 35, 21342–21357 (2022)

  20. [28]

    In: AI4Mat - NeurIPS 2023 Workshop (2023)

    Ghugare, R., Miret, S., Hugessen, A., Phielipp, M., Berseth, G.: Searching for high-value molecules using reinforcement learning and transformers. In: AI4Mat - NeurIPS 2023 Workshop (2023). https://openreview.net/forum?id= O8mZO2ri33

  21. [29]

    Normanno, N., Bianco, C., De Luca, A., Maiello, M.R., Salomon, D.S.: Target- based agents against ErbB receptors and their ligands: a novel approach to cancer treatment. Endocr. Relat. Cancer 10(1), 1–21 (2003)

  22. [30]

    Cancer Cell 11(3), 217–227 (2007)

    Yun, C.-H., Boggon, T.J., Li, Y., Woo, M.S., Greulich, H., Meyerson, M., Eck, M.J.: Structures of lung cancer-derived EGFR mutants and inhibitor complexes: mechanism of activation and insights into differential inhibitor sensitivity. Cancer Cell 11(3), 217–227 (2007)

  23. [31]

    Schr¨ odinger, LLC: LigPrep (2023)

  24. [32]

    Friesner, R.A., Banks, J.L., Murphy, R.B., Halgren, T.A., Klicic, J.J., Mainz, D.T., Repasky, M.P., Knoll, E.H., Shelley, M., Perry, J.K., Shaw, D.E., Francis, P., Shenkin, P.S.: Glide: a new approach for rapid, accurate docking and scoring

  25. [33]

    method and assessment of docking accuracy. J. Med. Chem. 47(7), 1739–1749 (2004) 16

  26. [34]

    Commun Chem 6(1), 249 (2023)

    Ochiai, T., Inukai, T., Akiyama, M., Furui, K., Ohue, M., Matsumori, N., Inuki, S., Uesugi, M., Sunazuka, T., Kikuchi, K., Kakeya, H., Sakakibara, Y.: Variational autoencoder-based chemical latent space for large molecular structures with 3D complexity. Commun Chem 6(1), 249 (2023)

  27. [35]

    Nucleic acids research47(D1), 930–940 (2019) https://doi.org/10.1093/nar/gky1075

    Mendez, D., Gaulton, A., Bento, A.P., Chambers, J., De Veij, M., F´ elix, E., Magari˜ nos, M.P., Mosquera, J.F., Mutowo, P., Nowotka, M., Gordillo-Mara˜ n´ on, M., Hunter, F., Junco, L., Mugumbate, G., Rodriguez-Lopez, M., Atkinson, F., Bosc, N., Radoux, C.J., Segura-Cabrera, ...

  28. [36]

    Reynolds, C.H., Tounge, B.A., Bembenek, S.D.: Ligand binding efficiency: trends, physical basis, and implications. J. Med. Chem. 51(8), 2432–2438 (2008)

  29. [37]

    Zhu, H., Yang, J., Huang, N.: Assessment of the generalization abilities of machine-learning scoring functions for structure-based virtual screening. J. Chem. Inf. Model. 62(22), 5485–5502 (2022)

  30. [38]

    Kengkanna, A., Ohue, M.: Enhancing property and activity prediction and inter- pretation using multiple molecular graph representations with MMGX. Commun. Chem. 7(1), 74 (2024)

  31. [39]

    Wu, Z., Wang, J., Du, H., Jiang, D., Kang, Y., Li, D., Pan, P., Deng, Y., Cao, D., Hsieh, C.Y., Hou, T.: Chemistry-intuitive explanation of graph neural networks for molecular property prediction with substructure masking. Nat. Commun. 14(1), 2585 (2023) 17

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.