Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

ASKCOS: an open source software suite for synthesis planning

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read ASKCOS bundles retrosynthesis, condition recommendation, and outcome prediction into one open-source suite, and demonstrates the suite's breadth on FDA-approved drug targets.

desk verdict A solid, honest software release note for an important open-source CASP platform; the missing end-to-end evaluation matters, but the paper deserves review if positioned as a software paper. read the letter →

arxiv 2501.01835 v1 pith:EFSBJ2SK submitted 2025-01-03 cs.AI

Zhengkai Tu , Sourabh J. Choure , Mun Hong Fong , Jihye Roh , Itai Levin , Kevin Yu , Joonyoung F. Joung , Nathan Morgan
show 11 more authors
Shih-Cheng Li Xiaoqi Sun Huiqian Lin Mark Murnin Jordan P. Liles Thomas J. Struble Michael E. Fortunato Mengjie Liu William H. Green Klavs F. Jensen Connor W. Coley
This is my paper · ORCID
classification cs.AI
keywords synthesisplanningretrosynthesisreactionconditionrecommendationoutcomepredictionopensourcesoftwaremachinelearningcheminformaticsmicroservices
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

ASKCOS is an open-source software suite that aims to cover the full range of synthesis planning tasks in one freely available platform. The paper reports the latest version, which integrates four one-step retrosynthesis models, two multi-step tree-search algorithms, reaction condition recommendation, reaction outcome prediction, and auxiliary tools such as solubility and quantum-mechanical descriptor prediction. The central claim is that this breadth, combined with modular design and permissive licenses, makes ASKCOS a practical complement to expert chemists rather than a replacement. A case study on FDA-approved drugs from 2019 to 2023 illustrates how automatic route-finding, model retraining, and interactive editing can be combined to propose putative synthetic routes.

What carries the argument

The central object is the ASKCOS software suite itself: a containerized microservice architecture in which each prediction module runs as an independent service behind a centralized API gateway, with a Vue-based front end. The load-bearing mechanisms are the four one-step retrosynthesis models (template relevance, Retrosim retrieval, SMILES-to-SMILES Transformer, and graph-to-SMILES Graph2SMILES), the MCTS and Retro* tree-search loops, and the condition-recommendation model that treats agents, temperature, and equivalences as separate classification and regression subproblems. The modularity is what lets multiple one-step strategies be combined and lets users retrain or replace individual models.

What would settle it

Deploy ASKCOS, run automatic retrosynthesis on a held-out set of novel drug-like targets that are absent from its training corpora, and attempt to execute the proposed routes in the laboratory; the claim of practical breadth predicts that most targets yield buyable routes whose steps pass plausibility filtering and whose recommended conditions produce the desired product, so a large fraction of failures at any stage would undercut the claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that ASKCOS is distinct among synthesis planning tools in its breadth: it attempts to cover a wide range of tasks, including interactive and automatic retrosynthesis, condition recommendation, forward outcome prediction, pathway scoring, and auxiliary utilities, rather than focusing mostly on retrosynthetic analysis. The suite builds on four one-step strategies (template relevance, Retrosim, Transformer, and Graph2SMILES) that can be used individually or together, with a fast plausibility filter and reranking by buyability and complexity. Multi-step planning is handled by MCTS and Retro*, and a case study demonstrates that these components can propose hypothetical routes for many FDA-approved drugs, with some targets requiring interactive intervention. The paper also argues that the 2023 microservice refactor makes the suite maintainable and extensible for community-driven development.

Load-bearing premise

The suite's usefulness rests on the assumption that its integrated predictive models, trained largely on published reaction data, give chemically plausible suggestions for targets outside their training data; the paper notes that some suggested steps have precedents but may or may not be achievable experimentally.

Editorial extensions

If this is right

  • Chemists can run the entire planning workflow—retrosynthesis, conditions, outcomes, and route ranking—inside one free platform rather than stitching together separate tools.
  • Combining template-based and template-free one-step models in a single search can surface routes that neither would find alone, as the case study shows for pralsetinib.
  • Retraining modules on proprietary reaction data is supported, so companies can adapt the suite to internal chemistry behind a firewall.
  • Because each module is a separate microservice, the suite can be partially deployed to save resources or extended with new models without rewriting the core.

Reading between the lines

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

  • If the microservice architecture proves maintainable, ASKCOS could become a shared infrastructure where new one-step models are contributed as drop-in services, lowering the barrier for CASP research.
  • The case study's manual corrections suggest that template coverage and buyable-database size, not model accuracy alone, are the main bottlenecks for automatic route-finding; a testable extension would be measuring how route success scales with template-set size.
  • The paper does not experimentally validate any proposed route, so the practical claim about assisting chemists depends on human judgment in filtering suggestions; a prospective study tracking how many ASKCOS-proposed steps survive wet-lab testing would sharpen that claim.
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

3 major / 5 minor

Summary. This paper describes the latest version of ASKCOS, an open-source computer-aided synthesis planning suite. It details four one-step retrosynthesis models, two multi-step tree search algorithms, reaction condition recommendation, reaction outcome and selectivity predictors, pathway scoring and visualization utilities, and a 2023 microservice-based refactor. The paper claims that ASKCOS is distinct among open-source CASP tools in its breadth of coverage and that it has assisted hundreds of chemists, and it illustrates usage with a case study applying the Tree Builder to 75 FDA-approved small-molecule drugs from 2019 to 2023, with selected route examples and recommended conditions.

Significance. If the software is as functional and broadly adopted as claimed, this release is valuable for the community: it provides a permissively licensed, modular, extensible platform integrating several independently published models, with deployment tooling, model retraining pipelines, API documentation, and an archived code snapshot. The strengths of the paper are the software artifacts rather than a new scientific method; the modules are cited from prior publications and Table 1 compiles previously reported one-step benchmark accuracies. The central utility claim—that the integrated pipeline gives useful suggestions on novel targets—is not established by quantitative end-to-end evidence: the case study is qualitative, does not report aggregate success rates, and explicitly documents chemically implausible or uncertain suggestions. The manuscript would be substantially strengthened by repositioning it as a software and community-resource description and by reporting quantitative performance of the integrated system on the case-study targets.

major comments (3)
  1. [Discussion; SI §S5] The claim that ASKCOS is a broadly useful synthesis planning platform rests on the integrated pipeline's performance on novel targets, yet no aggregate evaluation is reported. Table 1 summarizes one-step retrosynthesis accuracies from previous papers, and the 75-target FDA case study (SI §S5.1) does not state how many targets yielded any route, how many routes were judged chemically plausible, or how many steps required manual correction. The displayed routes include steps described as 'counter-intuitive or seemingly implausible' (SI §S5.2) and 'may or may not be achievable experimentally' (SI §S5.3), and Figure S10 shows a route assembled with manual expansion and filtering. Please add quantitative route-finding success rates, a step-level plausibility assessment (ideally with expert adjudication or literature support), and a statement of how much human intervention was required.
  2. [SI §S5.3] The case study's re-runs are presented as demonstrating flexibility, but the manuscript does not report the number of targets solved by each configuration or a combined success count. The text says 'Hypothetical routes could then be found for more targets' without numbers, and Figure S11 lists five targets that still failed. Reporting these counts would let readers judge how much of the claimed utility comes from automated search versus human editing; without them, the statement that ASKCOS 'has assisted hundreds of chemists' is supported only by anecdotal citations.
  3. [§4.2, Eq. (1)-(2)] The UCT formula as written appears inconsistent with the stated definitions. With s_r a model score, v_r an average buyability (both presumably in [0,1]), and n_r a visit count, the exploitation term Q_r = (s_r * v_r) / n_r decays as n_r grows, which is not the usual UCT average-reward behavior; if Q_r is instead meant to be cumulative, the text should say so and give the update rule. Since the Tree Builder is one of the two central search modes, please correct the equation or define the quantities so that the algorithm is reproducible.
minor comments (5)
  1. [§4.10] In the Code availability subsection, 'has be archived' should be 'has been archived'.
  2. [Figure 1 caption] The figure caption contains the typo 'QM decriptors'; it should read 'QM descriptors'.
  3. [SI §S5.3] The text 'newly solved prasetinib' should read 'newly solved pralsetinib'.
  4. [SI §S4.4] The phrase 'SectionAdvanced Usage' is missing a space and should read 'Section Advanced Usage'.
  5. [References [8], [20], [26]] Several reference entries contain corrupted author-name fragments (e.g., 'Wo/suppress los', 'M/suppress la˙ z'); these should be repaired in the final typeset version.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an integration report whose component models rest on external benchmarks and prior independent publications.

full rationale

This paper is a software-suite description rather than a derivation of a scientific result, so the circularity patterns enumerated (self-definitional predictions, fitted inputs called predictions, imported uniqueness theorems, ansatz smuggling, renaming) do not occur. The four one-step retrosynthesis models, the condition recommender, forward predictors, and utility models are all cited from prior publications, and the only performance evidence reproduced here (Table 1) is a compilation of previously reported top-k accuracies on the external USPTO-50k/USPTO-full benchmarks; these benchmarks do not depend on ASKCOS or on any parameter fitted in this paper. The MCTS/UCT update (Eqs. 1-2) is the standard UCT formula and is not used to define the models' accuracy. Self-citations are present (e.g., SCScore, RetroSim, Graph2SMILES, the robotic flow-chemistry paper), but they are descriptive attributions of software components, not load-bearing justifications that reduce a claimed prediction to its own input. The FDA case study (SI S5) is explicitly qualitative and, by the authors' own admission, contains steps that are 'counter-intuitive or seemingly implausible' (SI S5.2) and 'may or may not be achievable experimentally' (SI S5.3); this is a limitation of evidence for the utility claim, not a circularity. The paper therefore contains no step in which the output is equivalent by construction to a fitted or self-cited input.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The paper contributes software integration rather than a mathematical derivation. It inherits assumptions and hyperparameters from prior models and manual choices. The central claim does not rest on newly fitted parameters or invented theoretical entities.

free parameters (7)
  • UCT exploration constant c = not specified
    Hyperparameter in Eq. (2) that balances exploration and exploitation in MCTS. Chosen by user, not fitted to data in this paper.
  • Maximum templates per expansion = 1000 (case study)
    Limits the number of one-step suggestions per expansion. Chosen by user in the case study, affects search behavior.
  • Minimum plausibility threshold = 0.001 (case study)
    Threshold for the binary plausibility filter that removes unlikely precursors. Chosen by hand.
  • Maximum branching factor = 25 (case study)
    Limits the number of child nodes in tree search. Chosen by user.
  • Maximum search depth = 6 (case study)
    Limits the depth of retrosynthetic trees in the case study.
  • Maximum chemical nodes = 5000 (case study)
    Caps the number of explored chemical nodes in tree search.
  • Maximum buyable price = $100/g (case study)
    Defines buyability cutoff for termination of routes.
assumptions (3)
  • domain assumption The machine learning models trained on historical reaction data generalize to novel target molecules.
    The utility of ASKCOS depends on this assumption, inherited from the cited model papers but not re-validated in this work. The case study shows that some suggestions are implausible, indicating imperfect generalization.
  • domain assumption The buyable building block database is representative of commercially available chemicals.
    Tree search terminates when intermediates are found in the buyable database. An incomplete database could prevent finding valid routes or falsely stop search.
  • domain assumption Algorithmically extracted reaction templates and their associated precedents provide a reliable signal for chemical plausibility.
    The paper claims template-based suggestions are explainable via precedents, but the olutasidenib example shows templates can be too general and suggest invalid steps.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ASKCOS: an open source software suite for synthesis planning." pith.science (2026). https://pith.science/paper/EFSBJ2SK

@misc{pith2026250101835,
  author       = {Pith},
  title        = {Pith review of: ASKCOS: an open source software suite for synthesis planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EFSBJ2SK}},
  note         = {Machine review of arXiv:2501.01835}
}
read the original abstract

The advancement of machine learning and the availability of large-scale reaction datasets have accelerated the development of data-driven models for computer-aided synthesis planning (CASP) in the past decade. Here, we detail the newest version of ASKCOS, an open source software suite for synthesis planning that makes available several research advances in a freely available, practical tool. Four one-step retrosynthesis models form the basis of both interactive planning and automatic planning modes. Retrosynthetic planning is complemented by other modules for feasibility assessment and pathway evaluation, including reaction condition recommendation, reaction outcome prediction, and auxiliary capabilities such as solubility prediction and quantum mechanical descriptor prediction. ASKCOS has assisted hundreds of medicinal, synthetic, and process chemists in their day-to-day tasks, complementing expert decision making. It is our belief that CASP tools like ASKCOS are an important part of modern chemistry research, and that they offer ever-increasing utility and accessibility.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Fast and scalable retrosynthetic planning with a transformer neural network and speculative beam search

    cs.LG 2025-08 conditional novelty 5.0 of 10

    Speculative beam search with Medusa heads accelerates SMILES-to-SMILES transformers for multi-step retrosynthesis, solving 26-86% more molecules within time limits.

  2. A Multi-Layered Framework for Modeling Human Biology: From Basic AI Agents to a Full-Body AI Agent

    q-bio.TO 2025-08 reject novelty 4.0 of 10

    The paper proposes, but does not implement or validate, a multi-agent AI framework for cross-scale modeling of human biology from molecules to whole body, with sketches of metastasis scoring and drug development.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages · cited by 2 Pith papers

  1. [1]

    Vermeire, F., Wu, H., J

    Chung, Y., H. Vermeire, F., Wu, H., J. Walker, P., Abraham, M.H., Green, W.H.: Group Contribution and Machine Learning Approaches to Predict Abra- ham Solute Parameters, Solvation Free Energy, and Solvation Enthalpy. Journal of Chemical Information and Modeling 62(3), 433–446 (2022) https://doi.org/ 10.1021/acs.jcim.1c01103 . Publisher: American Chemical Society

    work page doi:10.1021/acs.jcim.1c01103 2022

  2. [2]

    Chemical Science 12(6), 2198–2208 (2021) https://doi.org/10.1039/ D0SC04823B

    Guan, Y., Coley, C.W., Wu, H., Ranasinghe, D., Heid, E., Struble, T.J., Pattanaik, L., Green, W.H., Jensen, K.F.: Regio-selectivity prediction with a machine-learned reaction representation and on-the-fly quantum mechanical descriptors. Chemical Science 12(6), 2198–2208 (2021) https://doi.org/10.1039/ D0SC04823B . Publisher: The Royal Society of Chemistry

    work page 2021

  3. [3]

    Publisher: American Chemical Society

    Li, S.-C., Wu, H., Menon, A., Spiekermann, K.A., Li, Y.-P., Green, W.H.: When Do Quantum Mechanical Descriptors Help Graph Neural Networks to Pre- dict Chemical Properties? Journal of the American Chemical Society 146(33), 23103–23120 (2024) https://doi.org/10.1021/jacs.4c04670 . Publisher: American Chemical Society

    work page doi:10.1021/jacs.4c04670 2024

  4. [4]

    The Jour- nal of Chemical Physics 156(8) (2022)

    Stuyver, T., Coley, C.W.: Quantum chemistry-augmented neural networks for reactivity prediction: Performance, generalizability, and explainability. The Jour- nal of Chemical Physics 156(8) (2022)

    work page 2022

  5. [5]

    https://www.rdkit.org/ Accessed 2024-07-25

    RDKit: Open-Source Cheminformatics Software. https://www.rdkit.org/ Accessed 2024-07-25

    work page 2024

  6. [6]

    https://fastapi.tiangolo.com/ Accessed 2024-07-25

    FastAPI. https://fastapi.tiangolo.com/ Accessed 2024-07-25

    work page 2024

  7. [7]

    https://pytorch.org/serve/ Accessed 2024-07-25

    TorchServe. https://pytorch.org/serve/ Accessed 2024-07-25

    work page 2024

  8. [8]

    https://gitlab.com/mlpds mit/askcosv2/askcos-docs/-/ wikis/home Accessed 2024-07-09

    The ASKCOS wiki. https://gitlab.com/mlpds mit/askcosv2/askcos-docs/-/ wikis/home Accessed 2024-07-09

    work page 2024

Show all 26 references
  1. [9]

    https://vuejs.org/ Accessed 2024-09-22

    Vue.js. https://vuejs.org/ Accessed 2024-09-22

    2024

  2. [10]

    https://vuetifyjs.com/ Accessed 2024-09-22

    Vuetify. https://vuetifyjs.com/ Accessed 2024-09-22

    2024

  3. [11]

    https://www.cypress.io/ Accessed 2024-09-22

    Cypress. https://www.cypress.io/ Accessed 2024-09-22

    2024

  4. [12]

    https://pinia.vuejs.org/ Accessed 2024-09-22

    Pinia: The intuitive store for Vue.js. https://pinia.vuejs.org/ Accessed 2024-09-22

    2024

  5. [13]

    https://router.vuejs.org/ Accessed 2024-09-22

    VueRouter. https://router.vuejs.org/ Accessed 2024-09-22

    2024

  6. [14]

    https://cactus.nci.nih.gov/chemical/ structure Accessed 2024-11-04 27

    Chemical Identifier Resolver from NIH. https://cactus.nci.nih.gov/chemical/ structure Accessed 2024-11-04 27

    2024

  7. [15]

    https://www.fda.gov/drugs/drug-approvals-and-databases/ compilation-cder-new-molecular-entity-nme-drug-and-new-biologic-approvals/ Accessed 2024-10-14

    Compilation of CDER New Molecular Entity (NME) Drug and New Bio- logic Approvals. https://www.fda.gov/drugs/drug-approvals-and-databases/ compilation-cder-new-molecular-entity-nme-drug-and-new-biologic-approvals/ Accessed 2024-10-14

    2024

  8. [16]

    Anderson, H.J., Loader, C.E., Xu, R.X., Lˆ e, N., Gogan, N.J., McDonald, R., Edwards, L.G.: Pyrrole chemistry. XXVIII. Substitution reactions of 1- (phenylsulfonyl)pyrrole and some derivatives. Canadian Journal of Chem- istry 63(4), 896–902 (1985) https://doi.org/10.1139/v85-1...

    1985 doi

  9. [17]

    Journal of the American Chemical Society 129(11), 3078–3079 (2007) https://doi.org/10.1021/ja070254l

    Elliott, L.D., Berry, M., Orr-Ewing, A.J., Booker-Milburn, K.I.: The Intramolec- ular Photometathesis of Pyrroles. Journal of the American Chemical Society 129(11), 3078–3079 (2007) https://doi.org/10.1021/ja070254l . Publisher: Amer- ican Chemical Society

    2007 doi

  10. [18]

    Organic Letters 18(21), 5608–5611 (2016) https://doi.org/10

    Koovits, P.J., Knowles, J.P., Booker-Milburn, K.I.: Conformationally Driven Two- and Three-Photon Cascade Processes in the Stereoselective Photorearrange- ment of Pyrroles. Organic Letters 18(21), 5608–5611 (2016) https://doi.org/10. 1021/acs.orglett.6b02829 . Publisher: Ameri...

    2016

  11. [19]

    Boogaard, A.T., Pandit, U.K., Koome, G.-J.: Ring D modifications of ellipticine. Part 1. New ellipticine derivatives from 1-cyano-6-methylellipticine. Tetrahedron 50(8), 2551–2560 (1994) https://doi.org/10.1016/S0040-4020(01)86971-4

    1994 doi

  12. [20]

    Potent and selective N-methyl-D-aspartic acid receptor antagonists with a short duration of action

    Ornstein, P.L., Schoepp, D.D., Arnold, M.B., Leander, J.D., Lodge, D., Paschal, J.W., Elzey, T.: 4-(Tetrazolylalkyl)piperidine-2-carboxylic acids. Potent and selective N-methyl-D-aspartic acid receptor antagonists with a short duration of action. Journal of Medicinal Chemistry...

    1991

  13. [21]

    Bioorganic & Medicinal Chemistry Letters 28(10), 1958–1963 (2018) https://doi.org/10.1016/ j.bmcl.2018.03.034

    Price, D.J., Drewry, D.H., Schaller, L.T., Thompson, B.D., Reid, P.R., Maloney, P.R., Liang, X., Banker, P., Buckholz, R.G., Selley, P.K., McDonald, O.B., Smith, J.L., Shearer, T.W., Cox, R.F., Williams, S.P., Reid, R.A., Tacconi, S., Faggioni, F., Piubelli, C., Sartori, I., T...

    2018

  14. [22]

    Tetrahedron letters 30(42), 5709–5712 (1989)

    Henry, J.R., Marcin, L.R., McIntosh, M.C., Scola, P.M., Harris Jr, G.D., Weinreb, S.M.: Mitsunobu reactions of n-alkyl and n-acyl sulfonamides-an efficient route to protected amines. Tetrahedron letters 30(42), 5709–5712 (1989)

    1989

  15. [23]

    Ring Breaker

    Thakkar, A., Selmi, N., Reymond, J.-L., Engkvist, O., Bjerrum, E.J.: “Ring Breaker”: Neural Network Driven Synthesis Prediction of the Ring System Chem- ical Space. Journal of Medicinal Chemistry 63(16), 8791–8808 (2020) https:// doi.org/10.1021/acs.jmedchem.9b01919 . Publishe...

    2020 doi

  16. [24]

    https://www.nextmovesoftware.com/pistachio.html Accessed 2024-07-08

    The Pistachio dataset. https://www.nextmovesoftware.com/pistachio.html Accessed 2024-07-08

    2024

  17. [25]

    https://www.cas.org/cas-data/cas-reactions Accessed 2024-07-08

    The CAS reactions collection. https://www.cas.org/cas-data/cas-reactions Accessed 2024-07-08

    2024

  18. [26]

    Angewandte Chemie International Edi- tion 52(11), 3267–3271 (2013) https://doi.org/10.1002/anie.201209584

    Antonchick, A.P., Burgmann, L.: Direct Selective Oxidative Cross-Coupling of Simple Alkanes with Heteroarenes. Angewandte Chemie International Edi- tion 52(11), 3267–3271 (2013) https://doi.org/10.1002/anie.201209584 . eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/an...

    2013 doi

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.