Pith. sign in

REVIEW 4 major objections 6 minor 54 references

DeepRetro: Retrosynthetic Pathway Discovery using Iterative LLM Reasoning

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read DeepRetro combines a large language model with a template-based retrosynthesis engine and chemist feedback to discover viable synthetic routes to complex natural products, solving 183 of 190 benchmark targets and proposing two novel routes.

desk verdict Open, honest hybrid retrosynthesis system with real benchmarks, but the 'novel pathway discovery' claim is human-guided and unvalidated; worth a serious read, not at face value. read the letter →

arxiv 2507.07060 v2 pith:4A6OSEAZ submitted 2025-07-07 q-bio.QM cs.AIcs.CLcs.LGq-bio.BMq-bio.MN

Shreyas Vinaya Sathyanarayana , Sharanabasava D. Hiremath , Rahil Shah , Rishikesh Panda , Rahul Jana , Riya Singh , Rida Irfan , Ashwin Murali
show 1 more author
Bharath Ramsundar
This is my paper · ORCID
classification q-bio.QMcs.AIcs.CLcs.LGq-bio.BMq-bio.MN
keywords retrosynthesiscomputer-aidedsynthesisplanninglargelanguagemodelshuman-in-the-loopnaturalproductreactionvalidationLLMhallucinationsmulti-stepbenchmark
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

DeepRetro is a retrosynthesis planner that tries to show that large language models, when forced through per-step chemical validation and a recursive template-based search, can plan viable routes to molecules that automated planners previously could not handle. In fully automatic mode it solves 183 of 190 multi-step benchmark targets, matching or exceeding the best template/search baselines, and in five expert-guided case studies it produces complete routes to complex natural products. Two of those case-study pathways are new, not found in the literature. The paper argues that the hybrid loop—LLM proposes, validators filter, template engine resolves, chemist steers—is the right working model for LLM-driven scientific discovery.

What carries the argument

The central object is the recursive control loop that alternates between a template/MCTS retrosynthesis solver and an LLM. If the template engine fails, the LLM proposes possible single-step disconnections, molecule checkers filter them for chemical validity, structural stability, and hallucination-like inconsistencies, and surviving precursors are recursively fed back into the loop until they reach purchasable stock. A graphical interface lets a chemist selectively regenerate segments, edit molecular structures directly, or add protecting groups, so the loop can absorb human strategic input at the steps where automated reasoning is weakest.

What would settle it

Run DeepRetro on a fresh set of five complex natural products with no human guidance beyond target entry and feasibility checks, and require the proposed routes to be validated by independent chemists: if no complete, chemically sound route can be produced without a human supplying intermediates, fragment boundaries, or protecting-group choices, the central claim fails. A cheaper version: take the published erythromycin route and remove the human-suggested biosynthetic intermediate; if the pipeline never terminates at a viable pathway, the single indispensable human step is confirmed.

Watch

Extended reading notes

Core claim

The paper claims DeepRetro enables discovery of viable synthetic routes for complex molecules typically out of reach of existing retrosynthetic methods. Its central finding is that an iterative loop in which an LLM proposes single-step disconnections, each subject to validity, stability, and hallucination filters, then recursively handed to a template/MCTS engine, yields state-of-the-art benchmark performance while in automatic mode, reaching 183 of 190 solved on the standard multi-step benchmark with the strongest configuration. With expert human guidance, the system proposes routes to five complex natural products, and two of those routes are novel: an unprecedented early esterification strategy for Ohauamine C and a convergent three-fragment strategy for Discodermolide. Human intervention was required in the case studies, typically six to fourteen runs per target, including one indispensable suggestion of a biosynthetic intermediate for Erythromycin B; the paper treats this as a feature of human-machine collaboration rather than a weakness.

Load-bearing premise

The load-bearing premise is that the successful case-study routes reflect DeepRetro's planning ability rather than the strategic decisions supplied by the human chemists, so if the human inputs carry the essential reasoning, the claim that DeepRetro discovers novel pathways is overstated.

Editorial extensions

If this is right

  • DeepRetro in automatic mode can match or exceed leading template-based planners on standard multi-step benchmarks, suggesting that LLM-guided search is a practical alternative for moderately complex targets without human intervention.
  • The two novel case-study routes indicate that the system can generalize beyond reaction template libraries and propose disconnections not represented in historical reaction databases.
  • Human-in-the-loop guidance remains necessary for complex natural products, with six to fourteen runs per target and at least one case where a human-suggested intermediate was indispensable.
  • Performance scales with the underlying LLM, with successive model generations improving both single-step and multi-step results, implying continued gains as language models improve.
  • The open-source release with documented prompts, validation checks, and intervention types offers a template for applying validated LLM reasoning to scientific discovery problems beyond retrosynthesis.

Reading between the lines

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

  • If the human interventions, rather than the LLM, supplied the essential strategic reasoning, then the two 'novel' pathways may trace primarily to the expert chemists' design choices; a clean test would run the same targets with human help restricted to feasibility checks, not strategy, and count how many novel routes survive.
  • The cost profile, roughly one dollar per molecule for the strongest LLM plus six to fourteen runs per case study, suggests that economically sensible use today lies in high-value medicinal chemistry and natural product programs rather than high-throughput screening.
  • The claim that the validation loop, not LLM quality alone, provides reliability could be tested by ablating the validity, stability, and hallucination checks and measuring how many generated pathways fail or terminate in chemically implausible intermediates.
  • The design suggests a general pattern for LLM-driven scientific discovery: generative proposal from the model, cheap automated filtering, and rare expert intervention, a pattern that might transfer to materials synthesis or experimental design.
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. DeepRetro is presented as an open-source, hybrid retrosynthetic planning framework that combines a template/MCTS-based CASP engine (an AiZynthFinder adaptation) with iterative single-step disconnection proposals from large language models, followed by chemical validity, stability, and hallucination checks, and optional human-in-the-loop intervention. The paper reports single-step accuracy on a custom 250-molecule subset of USPTO-50k and multi-step pathway success on USPTO-190, where the best Claude 4 Opus/Pistachio configuration solves 183/190 molecules in automatic mode. The central qualitative claim is that, with expert human guidance, DeepRetro discovers novel synthetic pathways to complex natural products; two of the five case-study molecules (Ohauamine C and a tetracyclic azepine derivative) are claimed to receive pathways not previously reported. The manuscript is transparent about costs, prompts, and many limitations, and explicitly states that fully automated retrosynthesis for complex natural products remains out of reach.

Significance. If the central claim were sustained, DeepRetro would be a useful demonstration that LLM-based reasoning can extend template-based retrosynthesis into difficult natural-product territory, and the detailed release of prompts, pseudocode, and cost analysis would be valuable to the CASP community. The paper also has genuine strengths: the benchmarks are run in automatic mode with no human intervention; the authors explicitly distinguish automatic and human-in-the-loop settings; the appendices provide unusually complete algorithmic and cost transparency; and the open-source release, including the exact prompts and datasets, supports reproduction. The main significance is therefore as a human-AI co-design study rather than as a fully automated discovery system, and the manuscript's own acknowledgements repeatedly support that more modest reading.

major comments (4)
  1. [Section 2.5, Table 3] The claim that DeepRetro 'discovers' novel pathways is not supported by the evidence as presented, because the human contributions documented in Table 3 and Appendix B are the strategic disconnections that define novelty. For Ohauamine C the human 'identified the basic building blocks that constitute the core of the molecule'; for discodermolide the human 'suggested initial fragment boundaries (C1–C7, C8–C16, C17–C24)'; and for erythromycin the asterisk states the route 'could not be generated without one key human intervention' (Section 2.5.3). The abstract and Section 3.1 attribute 'novel pathway discovery' to the system, but the current evidence supports a weaker claim: an expert chemist using DeepRetro as an interactive editor can elaborate a plan the chemist has substantially set. To support the discovery claim, the authors should provide either fully automatic runs on the same targets with the same stock and LLM, or an ablation that removes each human intervention and shows the route is still found; without this, the novelty claims should be reworded as human-guided co-design.
  2. [Section 2.4, Table 2] The multi-step comparison does not match computational budgets across systems, which weakens the claim that DeepRetro 'can match or exceed the success rates of state-of-the-art tools' (Section 3). Section 2.2.2 defines pathway success within 'a given computational budget (time limits, API & Compute Cost requirements),' but Table 2 reports only solved counts with no time limits, expansion counts, API costs, or hardware budgets for Retro*, PVDN, or the DeepRetro configurations. In addition, the best DeepRetro result uses the Pistachio 2024Q4 dataset, a much larger reaction database than the USPTO data used for Retro* and PVDN, so the 183/190 vs 177/190 difference could reflect database size rather than the hybrid architecture. The authors should report matched-budget results and, ideally, an ablation of the LLM component with the same template tool and database.
  3. [Section 2.2.1, Table 1] The single-step evidence for 'outperforming strong baselines' is weaker than the abstract suggests. The 'Any Correct Accuracy' metric counts a prediction as correct if any single proposed reactant matches any ground-truth reactant, which can reward systems that propose many alternative disconnections; the authors themselves describe this metric as rewarding alternative valid routes (Section 2.2.3). The 'All Correct' results are close (44% vs 42% for Claude 4 Opus/Pistachio versus ASKCOS), and the 250-molecule subset was selected by the authors rather than being the standard full USPTO-50k test set. The manuscript explicitly calls these results 'qualitative comparisons' (Section 2.3), so the single-step benchmark claims should be presented as indicative rather than as a decisive state-of-the-art comparison, or the full test set should be evaluated.
  4. [Abstract and Section 2.5] The abstract's phrase 'viable synthetic routes' is not supported by the evidence in the case studies. The case-study pathways are assessed by manual review of chemical plausibility and by reproducibility of runs (Section 2.2.4), not by experimental synthesis; Section 2.5 itself says the case studies required 6–14 human-guided runs. Calling the routes 'viable' conflates computationally plausible with experimentally validated. The authors should either replace 'viable' with 'chemically plausible' or 'potentially viable', or provide experimental validation for at least one case study; the current wording overstates the strength of the qualitative evidence.
minor comments (6)
  1. [Section 1] The first paragraph contains a typo: 'developent' should be 'development'.
  2. [Section 3.1] The sentence 'again with Claude 4 Opus yielding the top performance of 54% (135/250)' is inconsistent: 135/250 is the single-step Any Correct count from Table 1, while the surrounding text discusses the multi-step task in Table 2, where the corresponding result is 183/190 (96.3%). The two results should be labeled and cited separately.
  3. [Table 2] The table lists the baseline as 'PDVN' while the text refers to 'PVDN'; the spelling should be consistent and verified against the cited source.
  4. [Algorithm 3] The recursive function in Algorithm 3 is named 'RecursivePrithvi,' which does not match the described DeepRetro search procedure; this is confusing and should be renamed.
  5. [Section 5] The text references 'Table 7' for hallucination details, but no Table 7 appears in the manuscript; the cross-reference should be corrected or the table included.
  6. [Section 2.5.3] The claim that the erythromycin route is 'novel' from intermediate 3b onward should be stated more cautiously, since the authors note the transformations are 'precedented' individually and the route was seeded by a literature-based intermediate chosen by a human chemist.

Circularity Check

1 steps flagged · score 4.0 of 10

Main benchmarks are empirical and self-contained; one local case-study novelty claim reduces to the human-supplied fragment boundaries.

  1. fitted input called prediction [Section 2.5.5 (Molecule 5: Discodermolide) and Table 3]
    "Table 3: 'Suggested initial fragment boundaries (C1–C7, C8–C16, C17–C24) and guided strategic convergence based on literature precedent.' Section 2.5.5: 'The human intervention guided DeepRetro to divide the molecule systematically into three fragments' and 'The DeepRetro proposed retrosynthesis of discodermolide introduces a novel, convergent three-fragment strategy (C1–C7, C8–C16, C17–C24) not explicitly reported in existing syntheses.'"

    The claimed novel output of this case study is the three-fragment disconnection strategy, but those fragment boundaries are exactly what the human chemist supplied as input. The paper's own description states that the human 'guided DeepRetro to divide the molecule systematically into three fragments,' so the adjacent claim that 'DeepRetro proposed ... a novel ... three-fragment strategy' attributes to the system a strategic choice that was provided by the human. The novelty claim therefore reduces to the human hint by construction. This is local rather than global: Table 3 itself marks Discodermolide as 'No' for novel pathway, and the main benchmark results were run in automatic mode without human intervention.

full rationale

The paper's quantitative claims are not circular: single-step and multi-step benchmarks in Tables 1 and 2 compare DeepRetro against external baselines on standard datasets, with no parameter fitted to the test set and then reported as a prediction. The 'Any Correct Accuracy' metric does reward the LLM's multiple alternative disconnections, but the paper also reports 'All Correct Accuracy' and explicitly discusses the metric's limitations, so this is a transparent metric-design choice rather than a derivation from the conclusion. The headline 'discovery of novel pathways' rests on human-in-the-loop case studies, and the authors are unusually candid that human guidance was needed, that case studies took 6-14 runs, that Erythromycin was unsolvable without one key human intervention, and that fully automated retrosynthesis of complex natural products 'is still out of reach.' No load-bearing uniqueness theorem or self-citation chain is present; the self-citations (ChemBERTa) are ordinary background references. The only genuine circularity I can exhibit is the Discodermolide fragment-boundary claim, where the 'novel three-fragment strategy' attributed to DeepRetro is precisely the fragment split the human chemist supplied. Because that molecule is not among the two 'novel pathway: Yes' cases in Table 3, and because the other novel-pathway claims (Ohauamine C, tetracyclic azepine) involve LLM-proposed disconnections that the human did not specify, the central claim retains independent content. Overall, the paper is mostly self-contained and honest about human contributions, with a partial attribution circularity in one case study.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No fitted parameters in the sense of derivation; the central claim is empirical. The free parameters listed are engineering hyperparameters that affect results but are not fitted to benchmarks. The key domain assumptions are the reliability of the template tool and the sufficiency of validation filters, both acknowledged as imperfect in the paper.

free parameters (3)
  • LLM suggestion count k = 3-5 (prompt instructed)
    LLM pipeline requests 3-5 disconnections per step; a hyperparameter not swept.
  • stability and hallucination check cutoffs = not specified
    Algorithm 5 applies thresholds for rejection; cutoff values are not given in the paper, so exact filtering behavior is unstated.
  • max_run in LLM pipeline = 1.5 if stability/hallucination checks on, else 0.6
    Algorithm 5 sets run limit; affects how many LLM attempts are made before giving up.
assumptions (3)
  • domain assumption Template-based tool T (AiZynthFinder adaptation) reliably solves subtargets within its template coverage
    Algorithm 3 treats T's success as ground truth for termination of recursive search; relied on throughout. No error rates for T on the benchmark are reported separately.
  • domain assumption The LLM's single-step disconnections, after validation, are chemically plausible enough for recursive search to succeed
    Core premise of the hybrid loop (Section 5); paper acknowledges hallucinations persist and many edge cases are missed (Section 3.2).
  • domain assumption The stock set and termination criteria define feasible synthesis
    Pathway success depends on reaching stock molecules; the exact stock set used for USPTO-190 is not specified in main text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeepRetro: Retrosynthetic Pathway Discovery using Iterative LLM Reasoning." pith.science (2026). https://pith.science/paper/4A6OSEAZ

@misc{pith2026250707060,
  author       = {Pith},
  title        = {Pith review of: DeepRetro: Retrosynthetic Pathway Discovery using Iterative LLM Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4A6OSEAZ}},
  note         = {Machine review of arXiv:2507.07060}
}
read the original abstract

The synthesis of complex natural products remains one of the grand challenges of organic chemistry. We present DeepRetro, a major advancement in computational retrosynthesis that enables the discovery of viable synthetic routes for complex molecules typically considered beyond the reach of existing retrosynthetic methods. DeepRetro is a novel, open-source framework that tightly integrates large language models (LLMs), traditional retrosynthetic engines, and expert human feedback in an iterative design loop. Prior approaches rely solely on template-based methods or unconstrained LLM outputs. In contrast, DeepRetro combines the precision of template-based methods with the generative flexibility of LLMs, controlled by rigorous chemical validity checks and enhanced by recursive refinement. This hybrid system dynamically explores and revises synthetic pathways, guided by both algorithmic checks and expert chemist feedback through an interactive user interface. While DeepRetro achieves strong performance on standard retrosynthesis benchmarks, its true strength lies in its ability to propose novel, viable pathways to highly complex natural products-targets that have historically eluded automated planning. Through detailed case studies, we illustrate how this approach enables new routes for total synthesis and facilitates human-machine collaboration in organic chemistry. Beyond retrosynthesis, DeepRetro represents a working model for how to leverage LLMs in scientific discovery. We provide a transparent account of the system's design, algorithms, and human-feedback loop, enabling broad adaptation across scientific domains. By releasing DeepRetro as an open-source tool, we aim to empower chemists to tackle increasingly ambitious synthetic targets, accelerating progress in drug discovery, materials design, and beyond.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 41 canonical work pages

  1. [1]

    Nature Chemistry10(4), 383–394 (2018) https://doi.org/10.1038/ s41557-018-0021-z

    Blakemore, D.C., Castro, L., Churcher, I., Rees, D.C., Thomas, A.W., Wil- son, D.M., Wood, A.: Organic synthesis provides opportunities to transform drug discovery. Nature Chemistry10(4), 383–394 (2018) https://doi.org/10.1038/ s41557-018-0021-z . Accessed 2024-10-21

    work page 2018

  2. [2]

    Nature Reviews Drug Discovery17(2), 97–113 (2018) https://doi.org/10.1038/nrd.2017.232

    Schneider, G.: Automating drug discovery. Nature Reviews Drug Discovery17(2), 97–113 (2018) https://doi.org/10.1038/nrd.2017.232 . Accessed 2024-10-21

    work page doi:10.1038/nrd.2017.232 2018

  3. [3]

    Science166(3902), 178–192 (1969) https://doi.org/10.1126/science.166.3902.178

    Corey, E.J., Wipke, W.T.: Computer-Assisted Design of Complex Organic Syn- theses: Pathways for molecular synthesis can be devised with a computer and equipment for graphical communication. Science166(3902), 178–192 (1969) https://doi.org/10.1126/science.166.3902.178 . Accessed 2025-06-12 57 Fig. G31: DeepRetro Human-in-the-loop editor. The user can kick ...

    doi:10.1126/science.166.3902.178 1969

  4. [4]

    Science228(4698), 408–418 (1985) https://doi.org/10.1126/ science.3838594

    Corey, E.J., Long, A.K., Rubenstein, S.D.: Computer-Assisted Analysis in Organic Synthesis. Science228(4698), 408–418 (1985) https://doi.org/10.1126/ science.3838594 . Accessed 2025-06-12

    work page 1985

  5. [5]

    Pure and Applied chemistry14(1), 19–38 (1967)

    Corey, E.J.: General methods for the construction of complex molecules. Pure and Applied chemistry14(1), 19–38 (1967). Publisher: De Gruyter 58 Fig. G32: DeepRetro Retrosynthesis pathway viewer with metadata. The user is able to see the entire pathway generated starting from Step 0 being the target molecule. Hovering on any molecule would show the user th...

    work page 1967

  6. [6]

    (eds.): Computer-Assisted Organic Synthe- sis

    Wipke, W.T., Howe, W.J. (eds.): Computer-Assisted Organic Synthe- sis. ACS Symposium Series, vol. 61. AMERICAN CHEMICAL SOCI- ETY, W ASHINGTON, D. C. (1977). https://doi.org/10.1021/bk-1977-0061 . https://pubs.acs.org/doi/book/10.1021/bk-1977-0061Accessed 2025-04-29

    work page doi:10.1021/bk-1977-0061 1977

  7. [7]

    Zhong, Z., Song, J., Feng, Z., Liu, T., Jia, L., Yao, S., Hou, T., Song, M.: Recent advances in artificial intelligence for retrosynthesis. arXiv. arXiv:2301.05864 [cs] (2023). https://doi.org/10.48550/arXiv.2301.05864 . http://arxiv.org/abs/2301. 05864 Accessed 2025-06-12

    work page Pith review arXiv doi:10.48550/arxiv.2301.05864 2023

  8. [8]

    Science365(6453), 1566 (2019) https://doi.org/10.1126/science

    Coley, C.W., Thomas, D.A., Lummiss, J.A.M., Jaworski, J.N., Breen, C.P., Schultz, V., Hart, T., Fishman, J.S., Rogers, L., Gao, H., Hicklin, R.W., Ple- hiers, P.P., Byington, J., Piotti, J.S., Green, W.H., Hart, A.J., Jamison, T.F., Jensen, K.F.: A robotic platform for flow synthesis of organic compounds informed by AI planning. Science365(6453), 1566 (20...

    doi:10.1126/science 2019

Show all 54 references
  1. [10]

    Angewandte Chemie International Edition55(20), 5904–5937 (2016)

    Szymku´ c, S., Gajewska, E.P., Klucznik, T., Molga, K., Dittwald, P., Startek, M., Bajczyk, M., Grzybowski, B.A.: Computer-assisted synthetic planning: the end of the beginning. Angewandte Chemie International Edition55(20), 5904–5937 (2016). Publisher: Wiley Online Library

    2016

  2. [11]

    Chemistry–A European Journal23(25), 5966–5971 (2017)

    Segler, M.H., Waller, M.P.: Neural-symbolic machine learning for retrosynthe- sis and reaction prediction. Chemistry–A European Journal23(25), 5966–5971 (2017). Publisher: Wiley Online Library

    2017

  3. [12]

    Journal of chemical information and modeling60(7), 3398– 3407 (2020)

    Fortunato, M.E., Coley, C.W., Barnes, B.C., Jensen, K.F.: Data augmentation and pretraining for template-based retrosynthetic prediction in computer-aided synthesis planning. Journal of chemical information and modeling60(7), 3398– 3407 (2020). Publisher: ACS Publications

    2020

  4. [13]

    ACS central science3(12), 1237–1245 (2017)

    Coley, C.W., Rogers, L., Green, W.H., Jensen, K.F.: Computer-assisted retrosyn- thesis based on molecular similarity. ACS central science3(12), 1237–1245 (2017). Publisher: ACS Publications

    2017

  5. [14]

    Journal of chemical information and modeling (2022)

    Seidl, P., Renz, P., Dyubankova, N., Neves, P., Verhoeven, J., Wegner, J.K., Segler, M., Hochreiter, S., Klambauer, G.: Improving Few-and Zero-Shot Reac- tion Template Prediction Using Modern Hopfield Networks. Journal of chemical information and modeling (2022). Publisher: AC...

    2022

  6. [15]

    Journal of chemical information and modeling59(12), 5026–5033 (2019)

    Ishida, S., Terayama, K., Kojima, R., Takasu, K., Okuno, Y.: Prediction and interpretable visualization of retrosynthetic reactions using graph convolutional networks. Journal of chemical information and modeling59(12), 5026–5033 (2019). Publisher: ACS Publications

    2019

  7. [16]

    Advances in Neural Information Processing Systems 32, 8872–8882 (2019)

    Dai, H., Li, C., Coley, C., Dai, B., Song, L.: Retrosynthesis Prediction with Condi- tional Graph Logic Network. Advances in Neural Information Processing Systems 32, 8872–8882 (2019)

    2019

  8. [17]

    JACS Au1(10), 1612–1620 (2021)

    Chen, S., Jung, Y.: Deep retrosynthetic reaction prediction using local reactiv- ity and global attention. JACS Au1(10), 1612–1620 (2021). Publisher: ACS Publications

    2021

  9. [18]

    Journal of Chemical Information and Modeling60(1), 47–55 (2019)

    Zheng, S., Rao, J., Zhang, Z., Xu, J., Yang, Y.: Predicting retrosynthetic reactions using self-corrected transformer neural networks. Journal of Chemical Information and Modeling60(1), 47–55 (2019). Publisher: ACS Publications

    2019

  10. [19]

    Chen, B., Shen, T., Jaakkola, T.S., Barzilay, R.: Learning to make generalizable and diverse predictions for retrosynthesis (2019)

    2019

  11. [20]

    Chemical communications55(81), 12152–12155 (2019)

    Yang, Q., Sresht, V., Bolgar, P., Hou, X., Klug-McLeod, J.L., Butler, C.R.,et al.: Molecular transformer unifies reaction prediction and retrosynthesis across pharma chemical space. Chemical communications55(81), 12152–12155 (2019). Publisher: Royal Society of Chemistry 60

    2019

  12. [21]

    Chemical science11(12), 3355–3364 (2020)

    Lin, K., Xu, Y., Pei, J., Lai, L.: Automatic retrosynthetic route planning using template-free models. Chemical science11(12), 3355–3364 (2020). Publisher: Royal Society of Chemistry

    2020

  13. [22]

    Nature communications11(1), 1–11 (2020)

    Tetko, I.V., Karpov, P., Van Deursen, R., Godin, G.: State-of-the-art aug- mented NLP transformer models for direct and single-step retrosynthesis. Nature communications11(1), 1–11 (2020). Publisher: Nature Publishing Group

    2020

  14. [23]

    Proceedings of the AAAI Conference on Artificial Intelligence35(1), 531–539 (2021)

    Seo, S.-W., Song, Y.Y., Yang, J.Y., Bae, S., Lee, H., Shin, J., Hwang, S.J., Yang, E.: GTA: Graph Truncated Attention for Retrosynthesis. Proceedings of the AAAI Conference on Artificial Intelligence35(1), 531–539 (2021)

    2021

  15. [24]

    Journal of Chemical Information and Modeling61(1), 123–133 (2021)

    Kim, E., Lee, D., Kwon, Y., Park, M.S., Choi, Y.-S.: Valid, Plausible, and Diverse Retrosynthesis Using Tied Two-Way Transformers with Latent Variables. Journal of Chemical Information and Modeling61(1), 123–133 (2021). Publisher: ACS Publications

    2021

  16. [25]

    Machine Learning: Science and Technology 3(1), 015022 (2022)

    Irwin, R., Dimitriadis, S., He, J., Bjerrum, E.J.: Chemformer: a pre-trained trans- former for computational chemistry. Machine Learning: Science and Technology 3(1), 015022 (2022). Publisher: IOP Publishing

    2022

  17. [26]

    Chemical Science13(31), 9023–9034 (2022)

    Zhong, Z., Song, J., Feng, Z., Liu, T., Jia, L., Yao, S., Wu, M., Hou, T., Song, M.: Root-aligned SMILES: a tight representation for chemical reaction predic- tion. Chemical Science13(31), 9023–9034 (2022). Publisher: Royal Society of Chemistry

    2022

  18. [27]

    Journal of Chemical Information and Modeling61(7), 3273–3284 (2021)

    Sacha, M., B laz, M., Byrski, P., Dabrowski-Tumanski, P., Chrominski, M., Loska, R., W lodarczyk-Pruszynski, P., Jastrzebski, S.: Molecule edit graph attention network: modeling chemical reactions as sequences of graph edits. Journal of Chemical Information and Modeling61(7), ...

    2021

  19. [28]

    Neurocomputing457, 193– 202 (2021)

    Mao, K., Xiao, X., Xu, T., Rong, Y., Huang, J., Zhao, P.: Molecular graph enhanced transformer for retrosynthesis prediction. Neurocomputing457, 193– 202 (2021). Publisher: Elsevier

    2021

  20. [29]

    Computers & Chemical Engineering155, 107533 (2021)

    Mann, V., Venkatasubramanian, V.: Retrosynthesis prediction using grammar- based neural machine translation: An information-theoretic approach. Computers & Chemical Engineering155, 107533 (2021). Publisher: Elsevier

    2021

  21. [30]

    Journal of cheminformatics13(1), 1–15 (2021)

    Ucak, U.V., Kang, T., Ko, J., Lee, J.: Substructure-based neural machine transla- tion for retrosynthetic prediction. Journal of cheminformatics13(1), 1–15 (2021). Publisher: BioMed Central

    2021

  22. [31]

    Nature communications13(1), 1–10 (2022)

    Ucak, U.V., Ashyrmamatov, I., Ko, J., Lee, J.: Retrosynthetic reaction pathway prediction through neural machine translation of atomic environments. Nature communications13(1), 1–10 (2022). Publisher: Nature Publishing Group 61

    2022

  23. [32]

    ACS central science3(10), 1103–1113 (2017)

    Liu, B., Ramsundar, B., Kawthekar, P., Shi, J., Gomes, J., Luu Nguyen, Q., Ho, S., Sloane, J., Wender, P., Pande, V.: Retrosynthetic reaction prediction using neural sequence-to-sequence models. ACS central science3(10), 1103–1113 (2017). Publisher: ACS Publications

    2017

  24. [33]

    Journal of Cheminformatics12(1), 70 (2020) https: //doi.org/10.1186/s13321-020-00472-1

    Genheden, S., Thakkar, A., Chadimov´ a, V., Reymond, J.-L., Engkvist, O., Bjer- rum, E.: AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning. Journal of Cheminformatics12(1), 70 (2020) https: //doi.org/10.1186/s13321-020-00472-1 . Access...

    2020 doi

  25. [34]

    Advances in neural informa- tion processing systems30(2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, k., Polosukhin, I.: Attention is all you need. Advances in neural informa- tion processing systems30(2017)

    2017

  26. [35]

    Chithrananda, S., Grand, G., Ramsundar, B.: ChemBERTa: Large-Scale Self-Supervised Pretraining for Molecular Property Prediction. arXiv. arXiv:2010.09885 [cs] (2020). https://doi.org/10.48550/arXiv.2010.09885 . http://arxiv.org/abs/2010.09885 Accessed 2025-06-12

    2020 doi arXiv

  27. [36]

    Ahmad, W., Simon, E., Chithrananda, S., Grand, G., Ramsundar, B.: ChemBERTa-2: Towards Chemical Foundation Models. arXiv. arXiv:2209.01712 [cs] (2022). https://doi.org/10.48550/arXiv.2209.01712 . http://arxiv.org/abs/ 2209.01712 Accessed 2025-06-12

    2022 doi arXiv

  28. [37]

    https://arxiv.org/abs/2505.02639

    Lin, X., Liu, Q., Xiang, H., Zeng, D., Zeng, X.: Enhancing Chemical Reaction and Retrosynthesis Prediction with Large Language Model and Dual-task Learning (2025). https://arxiv.org/abs/2505.02639

    2025 arXiv

  29. [38]

    https://arxiv

    Liu, X., Guo, Y., Li, H., Liu, J., Huang, S., Ke, B., Lv, J.: DrugLLM: Open Large Language Model for Few-shot Molecule Generation (2024). https://arxiv. org/abs/2405.06690

    2024 arXiv

  30. [39]

    https://arxiv.org/abs/2204.11817

    Edwards, C., Lai, T., Ros, K., Honke, G., Cho, K., Ji, H.: Translation between Molecules and Natural Language (2022). https://arxiv.org/abs/2204.11817

    2022 arXiv

  31. [40]

    Chem- ical Science15(27), 10600–10611 (2024) https://doi.org/10.1039/D4SC00924J

    Zhang, W., Wang, Q., Kong, X., Xiong, J., Ni, S., Cao, D., Niu, B., Chen, M., Li, Y., Zhang, R., Wang, Y., Zhang, L., Li, X., Xiong, Z., Shi, Q., Huang, Z., Fu, Z., Zheng, M.: Fine-tuning large language models for chemical text mining. Chem- ical Science15(27), 10600–10611 (20...

    2024 doi

  32. [41]

    Bran, A.M., Cox, S., Schilter, O., Baldassari, C., White, A.D., Schwaller, P.: ChemCrow: Augmenting large-language models with chemistry tools. arXiv. arXiv:2304.05376 [physics] (2023). https://doi.org/10.48550/arXiv.2304.05376 . http://arxiv.org/abs/2304.05376 Accessed 2025-06-12

    2023 doi arXiv

  33. [42]

    Wang, H., Guo, J., Kong, L., Ramprasad, R., Schwaller, P., Du, Y., Zhang, 62 C.: LLM-Augmented Chemical Synthesis and Design Decision Programs. arXiv. arXiv:2505.07027 [cs] (2025). https://doi.org/10.48550/arXiv.2505.07027 . http: //arxiv.org/abs/2505.07027 Accessed 2025-06-12

    2025 doi arXiv

  34. [43]

    Journal of Natural Products88(3), 871–876 (2025) https://doi.org/10.1021/acs

    Chen, D., Rouvier, F., Keyzers, R., Bourguet-Kondracki, M.-L., Brunel, J.M., Cadelis, M.M., Copp, B.R.: Ohauamines A–D, Structurally Unprecedented Tri- cyclic Depsi-tripeptides from the New Zealand AscidianPycnoclavella kottae. Journal of Natural Products88(3), 871–876 (2025) ...

    2025 doi

  35. [44]

    Organic Process Research & Development8(5), 754–768 (2004) https: //doi.org/10.1021/op0402026

    Gala, D., Dahanukar, V.H., Eckert, J.M., Lucas, B.S., Schumacher, D.P., Zavialov, I.A., Buholzer, P., Kubisch, P., Mergelsberg, I., Scherer, D.: Development of an Efficient Process for the Preparation of Sch 39166:Aziridinium Chemistry on Scale. Organic Process Research & Deve...

    2004 doi

  36. [45]

    Chemical Reviews97(7), 2611–2630 (1997) https://doi.org/10.1021/cr9600316

    Staunton, J., Wilkinson, B.: Biosynthesis of Erythromycin and Rapamycin. Chemical Reviews97(7), 2611–2630 (1997) https://doi.org/10.1021/cr9600316 . Publisher: American Chemical Society

    1997 doi

  37. [46]

    Chemical Reviews105(12), 4671–4706 (2005) https://doi.org/10.1021/ cr050521a

    Chen, F.-E., Huang, J.: Reserpine: A Challenge for Total Synthesis of Natural Products. Chemical Reviews105(12), 4671–4706 (2005) https://doi.org/10.1021/ cr050521a . Accessed 2025-08-07

    2005

  38. [47]

    Journal of the American Chemical Society78(9), 2023–2025 (1956) https://doi.org/10.1021/ja01590a079

    Woodward, R.B., Bader, F.E., Bickel, H., Frey, A.J., Kierstead, R.W.: THE TOTAL SYNTHESIS OF RESERPINE. Journal of the American Chemical Society78(9), 2023–2025 (1956) https://doi.org/10.1021/ja01590a079 . Accessed 2025-08-07

    1956 doi

  39. [48]

    The Jour- nal of Organic Chemistry63(22), 7885–7892 (1998) https://doi.org/10.1021/ jo9811423

    Marshall, J.A., Johns, B.A.: Total Synthesis of (+)-Discodermolide. The Jour- nal of Organic Chemistry63(22), 7885–7892 (1998) https://doi.org/10.1021/ jo9811423 . Accessed 2025-08-07

    1998

  40. [49]

    figshare

    Lowe, D.: Chemical reactions from US patents (1976-Sep2016). figshare. Artwork Size: 1494665893 Bytes Pages: 1494665893 Bytes (2017). https://doi.org/10.6084/ M9.FIGSHARE.5104873.V1 . https://figshare.com/articles/dataset/Chemical reactions from US patents 1976-Sep2016 /510487...

    2017

  41. [50]

    In: Cappellato, L., Eickhoff, C., Ferro, N., N´ ev´ eol, A

    Lowe, D.M., Mayfield, J.: Extraction of reactions from patents using grammars. In: Cappellato, L., Eickhoff, C., Ferro, N., N´ ev´ eol, A. (eds.) Working Notes of CLEF 2020 - Conference and Labs of the Evaluation Forum, Thessaloniki, Greece, September 22-25, 2020. CEUR Worksho...

    2020

  42. [51]

    PhD thesis, Apollo - University of Cambridge Repository (2012)

    Lowe, D.M.: Extraction of chemical structures and reactions from the literature. PhD thesis, Apollo - University of Cambridge Repository (2012). https://doi.org/ 10.17863/CAM.16293 . https://www.repository.cam.ac.uk/handle/1810/244727 63

    2012 doi

  43. [52]

    NextMove Software

    NextMove Software: Pistachio: Reaction Data, Querying and Analytics. NextMove Software. https://www.nextmovesoftware.com/pistachio.html

  44. [53]

    Accessed: 2025-06-30 (2024)

    Anthropic: Introducing the next generation of Claude. Accessed: 2025-06-30 (2024). https://www.anthropic.com/news/claude-3-family

    2024

  45. [54]

    https://arxiv.org/abs/2501.12948

    DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., Zhang, X., Yu, X., Wu, Y., Wu, Z.F., Gou, Z., Shao, Z., Li, Z., Gao, Z., Liu, A., Xue, B., Wang, B., Wu, B., Feng, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., Da...

    2025 arXiv

  46. [55]

    Tetrahedron63(26), 5709–5729 (2007) https://doi.org/10.1016/ j.tet.2007.02.044

    Breton, P., Hergenrother, P.J., Hida, T., Hodgson, A., Judd, A.S., Kraynack, E., Kym, P.R., Lee, W.-C., Loft, M.S., Yamashita, M., Martin, S.F.: Total synthesis of erythromycin b. Tetrahedron63(26), 5709–5729 (2007) https://doi.org/10.1016/ j.tet.2007.02.044 . Tetrahedron 50th...

    2007

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.