Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Chemist-aligned retrosynthesis by ensembling diverse inductive bias models

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Retrosynthesis ensemble beats all single models and training data

desk verdict RetroChimera is a real advance in single-step retrosynthesis via a simple count-based ensemble with learned rank weights; the accuracy claims hold for k>1, but the expert-preference claim needs a statistical re-analysis and the abstract overstates top-1. read the letter →

arxiv 2412.05269 v2 pith:URYCINJ4 submitted 2024-12-06 cs.LG cs.AIq-bio.QM

Krzysztof Maziarz , Guoqing Liu , Hubert Misztela , Austin Tripp , Junren Li , Aleksei Kornev , Piotr Gaiński , Holger Hoefling
show 3 more authors
Mike Fortunato Rishi Gupta Marwin Segler
This is my paper · ORCID
classification cs.LGcs.AIq-bio.QM
keywords retrosynthesisensemblelearningtoranktemplate-basedmodelsde-novogenerationtransformergraphneuralnetworkchemistpreference
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 argues that the fastest route to better retrosynthesis models is not a single larger architecture but a principled fusion of models with deliberately different inductive biases. It introduces RetroChimera, an ensemble of a template-based edit model (NeuralLoc) that stays close to known chemistry and a de-novo Transformer (R-SMILES 2) that can propose less precedented transformations, merged by learned rank-dependent weights that reward reactions on which both models agree. Across USPTO-50K, USPTO-FULL, and the larger Pistachio dataset, RetroChimera outperforms all major baselines, with the margin growing as more predictions are considered, and it transfers zero-shot to an internal industrial dataset. The paper further reports that expert organic chemists prefer RetroChimera's top-1 predictions over the literature ground-truth reactions it was trained on, indicating that the ensemble has learned to denoise noisy reaction data rather than merely memorize it.

What carries the argument

The load-bearing mechanism is a count-based ensemble score computed in molecule space: each unique reactant set $r$ receives $$\operatorname{score}(r)=\sum_{i=1}^{m}\sum_{k=1}^{k_{\max}}\mathbf{1}[r=r_{i,k}]\.,\theta_{i,k},$$ where $r_{i,k}$ is the $k$-th ranked prediction of model $i$ and $\theta_{i,k}$ is a rank-dependent weight learned on the validation set by minimizing a smooth pairwise ranking loss, with the constraint that each model's weights decrease and are convex across ranks. This rewards reactant sets that appear high in several diverse prediction lists and allows the relative importance of models to vary with rank, which the ablation shows is necessary for large gains. The two fused models are NeuralLoc, a template-classification graph neural network that also localizes template atoms on the product to disambiguate symmetric template applications, and R-SMILES 2, a Transformer de-novo model using root-aligned SMILES, group-query attention, RMSNorm, SwiGLU activations, and a retrosynthesis-specific beam search termination rule.

What would settle it

A decisive check would be to have expert chemists manually correct or flag every ground-truth reaction in a random sample of the 146,393 time-split Pistachio test reactions, recompute RetroChimera's top-k accuracy and its win rate against the corrected labels, and see whether the preference advantage over ground truth persists; if the win rate falls toward 50% once labels are corrected, the reported 'chemist alignment' is mostly an artifact of denoising a noisy benchmark.

Watch

Extended reading notes

Core claim

The central claim is that ensembling two strong models with complementary inductive biases—an edit-based model that applies learned reaction templates and a de-novo model that generates reactant SMILES from scratch—produces a retrosynthesis predictor that is more accurate, more robust, and more aligned with expert chemists than any single model. Concretely, the paper shows that RetroChimera sets a new state of the art for k>1 on both USPTO-50K and USPTO-FULL (top-10 accuracy gains of 1.7% and 1.6% over the best baselines), matches the top-50 accuracy of a strong de-novo model using only ten predictions on the time-split Pistachio test set, and achieves the highest solve rates in multi-step search on both the SimpRetro benchmark and a newly constructed set of 800 hard Pistachio targets. In a double-blind evaluation with nine PhD-level organic chemists, RetroChimera's top-1 output was preferred over the dataset ground-truth reaction about 64% of the time (P<0.05), and the paper demonstrates cases where the ensemble recovers the correct stereochemistry or missing reagents that the noisy database entry got wrong. On rare reaction classes, RetroChimera reaches close-to-optimal recall on well-precedented reactions while outperforming both constituents on rare templates, effectively behaving like a soft reaction database. The paper also reports that the same ensembling recipe, with weights fixed after Pistachio training, transfers zero-shot to an internal industrial reaction dataset, supporting the claim that the gains reflect genuine generalization rather than benchmark overfitting.

Load-bearing premise

The load-bearing premise is that the time-split Pistachio test set and the internal industrial test set are genuinely clean and free of leakage, since the paper's own denoising examples show that the underlying reaction database contains stereochemical and assignment errors.

Editorial extensions

If this is right

  • RetroChimera improves multi-step retrosynthetic search: it attains close to 100% solve rate on SimpRetro targets under the longest time limit and outperforms its constituents on 800 hard Pistachio targets for medium-to-long search times.
  • The ensemble addresses the rare-reaction problem: it performs well on templates with little or no precedent in training data, where de-novo models degrade less than edit-based ones, while also reaching near-optimal recall on well-precedented reactions.
  • Because chemist preference beats ground truth, the model can act as a denoiser: it ignores likely incorrect stereochemical assignments and infers missing reagents in noisy database records, pointing toward a use in data curation.
  • The ensembling framework is not tied to autoregressive token spaces, so it can in principle combine any set of prediction sources, including non-ML sources such as reaction database lookups or chemist-drawn proposals, at negligible extra inference cost when paired with a fast edit model.
  • Zero-shot transfer to an internal industrial dataset, with ensembling weights kept from Pistachio, shows that the gains are not confined to the training distribution and that deployment without fine-tuning is viable.

Reading between the lines

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

  • If the expert-preference result replicates in larger, more diverse chemist panels, retrosynthesis evaluation should shift from recall against noisy literature labels toward preference-based and forward-validated metrics, because accuracy benchmarks may systematically underestimate models that correct database errors.
  • The learned rank-dependent weights, whose curves for the two models cross as k grows, suggest a general principle for fusing set-valued predictors: optimal weighting is a function of rank, not just of overall model quality; this could transfer to other domains where multiple generators return ranked lists, such as code synthesis or document retrieval.
  • A testable extension of the framework is to deliberately increase submodel diversity, for example by training the edit model only on templates with high precedence and the de-novo model only on rare reactions, and checking whether the ensemble improves further, since the paper's own ablation shows that combining similar models yields little benefit.
  • Because the ensemble score is computed from saved model outputs on a validation set, the fusion layer is cheap to retrain; one could use the same weight-learning procedure to dynamically add newly available reaction databases or human-in-the-loop suggestions to RetroChimera without retraining the neural submodels.
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. The paper proposes RetroChimera, a retrosynthesis model formed by ensembling a new edit-based template model (NeuralLoc) and a new de-novo Transformer model (R-SMILES 2). The ensemble uses learned rank-dependent weights theta, optimized on a validation set with a rank loss, to merge the ranked output lists of the two models (Eqs. 1-2). The authors report state-of-the-art results for k>1 on USPTO-50K and USPTO-FULL, strong performance on a time-split Pistachio test set, zero-shot transfer to an internal Novartis dataset, improved multi-step search solve rates, and an expert preference study in which chemists prefer RetroChimera's top-1 predictions over Pistachio ground-truth reactions (P<0.05, mean preference rate about 64%). They also present qualitative evidence that the ensemble denoises erroneous stereochemical assignments in the training data.

Significance. If the claims hold, the count-based learned ensembling framework is a useful and general contribution to retrosynthesis, and the two component models are competitive in their respective classes. The paper's strengths include systematic benchmarking in syntheseus, ablations across 55 model pairs, a time-split test set, a zero-shot internal transfer set, and multi-step search experiments with tuned temperatures. The expert preference study, with blinding and a dummy baseline, is a valuable step beyond top-k accuracy. However, the headline claims are broader than the evidence: the abstract's 'outperforms all major models by a large margin' is contradicted by the paper's own top-1 tables, and the statistical analysis of the expert study ignores clustering by rater and product. These issues are fixable, but they affect the central claims as currently stated.

major comments (4)
  1. [Abstract; Extended Data Tables 1 and 2] The abstract's claim that RetroChimera 'outperforms all major models by a large margin' is contradicted by the paper's own results. On USPTO-50K, RetroChimera has top-1 accuracy 56.7%, which is lower than EditRetro's 60.8% and also lower than its own de-novo component R-SMILES 2 (56.9%); on USPTO-FULL, RetroChimera's top-1 is 51.4%, below EditRetro (52.2%) and RetroWISE (52.3%). The defensible statement, which appears in the main text, is that RetroChimera sets a new state of the art for k>1. Please revise the abstract and any similar summary sentences to state the top-k qualification explicitly.
  2. [Methods, 'Assessment by domain experts'; Figure 3e] The expert preference analysis treats 599 comparisons as independent, but the comparisons share both raters and products. With 9 experts and multiple comparisons per product, the effective sample size is smaller than 599, and the reported P<0.05 may not survive a clustering-aware analysis. In addition, the Methods say 'we sample several test products' without specifying that products were selected uniformly at random from the test set, so selection bias cannot be excluded. Please report per-rater and per-product results, use cluster-robust or mixed-effects/permutation inference, and state the exact sampling protocol. This is load-bearing because the 'chemist-aligned' claim and the abstract's statement that chemists prefer RetroChimera's predictions over the reactions it was trained on depend on this analysis.
  3. [Methods, 'Pistachio test set'; Figure 6] The time-split Pistachio test set is used to support 'robust generalization under distribution shift', but the paper itself demonstrates in Figure 6 that Pistachio ground-truth reactions contain stereochemical, assignment, and missing-reactant errors. The test set is also filtered with a single Morgan similarity threshold of 0.95. Please quantify how many test labels are affected by the errors demonstrated in Figure 6, and report a sensitivity analysis for the similarity threshold and quality-tier inclusion rules. Without this, the reported accuracy margins on Pistachio are difficult to interpret: near-duplicate scaffolds below the threshold could inflate the margins, while erroneous labels could penalize the model for generating chemically correct alternatives.
  4. [Abstract; Figure 3b; Methods, 'Bucketing test data'] The abstract's claim of 'for the first time the ability to learn from even a very small number of examples per reaction class' is not directly evidenced. Figure 3b shows non-zero top-50 accuracy for queries whose canonical ground-truth template occurs fewer than two times in training, but the Methods note that this can arise because several distinct templates may yield the same reactants for a given product. That is not the same as learning a reaction class from one or two examples. Please either provide a direct few-shot experiment that isolates template learning or qualify the claim to what the data actually show.
minor comments (6)
  1. [Extended Data Table 3] Several entries contain formatting artifacts, e.g. '0 .0' for dropout values; please clean these up.
  2. [Ensembling section] The statement 'we find that our strategy consistently outperforms other approaches' is not supported in the main text until the reader consults Extended Data Figure 7; consider summarizing the ablation comparison in the main text.
  3. [Figure 3e and main text] The main text reports P<0.05 for the expert preference result without stating the statistical test; please specify the test, the exact p-value, and the confidence interval for the 64% preference rate.
  4. [Code and Model Availability] The paper says code is 'currently being prepared for release on GitHub'; for a paper claiming state-of-the-art results, releasing evaluation scripts and trained checkpoints at publication time is important for reproducibility.
  5. [Qualitative analysis] The sentence 'This is the first time a model is able to provide predictions that are more aligned to chemists' expectations than the actual reference reactions the model has been trained on' is a strong historical claim; please support it with a citation or soften it.
  6. [Equation 1] The indicator notation '1[r = r_{i,k}]' is standard but should be defined explicitly at first use for readers outside the immediate literature.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central ensembling result is supported by held-out test evaluations and external benchmarks.

full rationale

RetroChimera's central derivation is not circular. The only fitted parameters in the ensemble are the rank weights theta, optimized on a validation set via the rank loss in Equation 2 and evaluated on held-out USPTO and time-split Pistachio test sets, with theta frozen in the zero-shot Novartis transfer; the reported accuracy gains are therefore not forced by construction. The two submodels are trained with standard supervised losses on reaction data and benchmarked against external baselines through syntheseus, so the state-of-the-art claims rest on independent held-out labels. The Pistachio test set is a prospective 2024 time split with a fingerprint-similarity filter, not a re-use of training labels. The expert preference result is an external human judgment rather than a fitted quantity, and the quality/feasibility models used for auxiliary analysis are calibration-controlled evaluation tools, not inputs to the ensemble's training objective. Self-citations such as [45] for the syntheseus benchmark and [26] for the feasibility-model approach are implementation-level references and are not load-bearing: they do not assert or assume the paper's conclusion. No uniqueness theorem or ansatz is smuggled in via self-citation. Statistical concerns about the expert study, such as potential clustering by rater or product and non-random product selection, are validity risks rather than circularity, and would not change the conclusion that the derivation chain is self-contained.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities; its novelty is algorithmic. The main fitted quantities are the ensemble weights and a few calibration coefficients. The central assumptions are about label quality, leakage prevention, expert judgment as a gold standard, and generalization of fitted weights.

free parameters (4)
  • Ensemble rank weights theta_i,k = Learned; not tabulated as scalar values (see Extended Data Figure 7b for example curves)
    The central ensembling method in Eq. 1 uses per-model, per-rank weights optimized by minimizing Lrank on Dval. Without these weights the ensemble would be an unweighted union; they are the main learned component of the method.
  • NeuralLoc localization weight w_loc = 2.25
    Inference reranks outputs by score sclf + wloc * sloc, with wloc = 2.25 chosen empirically (Methods, 'Editing submodel (NeuralLoc)'). This coefficient affects the edit model's ranking and therefore the ensemble input.
  • Search temperature T = Tuned per model from {0.25,0.35,0.5,0.71,1.0,1.41,2.0,2.83,4.0} on 151 validation targets
    Multi-step search results in Figure 3d depend on the temperature selected for Retro*; the authors tune this per model on validation, which is fair but makes the search comparison partly dependent on a fitted hyperparameter.
  • Forward and feasibility model acceptance thresholds = Calibrated to accept ~95% of ground-truth reactions (forward: product in top-2; feasibility: score > 0.1)
    Quality assessment in Figure 3c uses thresholds chosen on the ground-truth reactions of the filtered test set; this is a calibration step, not a prediction.
assumptions (5)
  • domain assumption Reaction databases (USPTO, Pistachio, Novartis) provide correct ground-truth reactions for training and evaluation.
    The paper measures accuracy as exact match to these labels, yet the qualitative 'denoising' section shows several labeled Pistachio reactions contain stereochemical or assignment errors. The assumption of label correctness is thus partially violated, and accuracy numbers may penalize models for predicting chemically better alternatives.
  • domain assumption Morgan fingerprint Tanimoto similarity below 0.95 between test and training products removes leakage in the time-split Pistachio test set.
    Used to construct the 146,393-reaction test set; the threshold is an unvalidated cutoff, and the paper does not report a sensitivity analysis for it.
  • domain assumption Expert chemists' preferences are a valid gold standard for reaction quality.
    The 'chemist-aligned' claim is built on 599 pairwise preferences from 9 chemists; the paper assumes these preferences measure quality rather than stylistic or route-choice conventions.
  • standard math Standard molecular graph featurization and template extraction tools (RDKit, rdchiral, dgllife) correctly encode chemical semantics.
    NeuralLoc depends on rdchiral template extraction and dgllife/WeaveAtom featurizers; the authors note template extraction errors can occur (Extended Data Figure 13).
  • domain assumption The learned ensembling weights theta fitted on validation generalize to the test distribution.
    If validation and test distributions differ, the learned weights could underperform hand-tuned schemes; the paper tests one zero-shot transfer to Novartis, which supports the assumption but does not prove it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Chemist-aligned retrosynthesis by ensembling diverse inductive bias models." pith.science (2026). https://pith.science/paper/URYCINJ4

@misc{pith2026241205269,
  author       = {Pith},
  title        = {Pith review of: Chemist-aligned retrosynthesis by ensembling diverse inductive bias models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/URYCINJ4}},
  note         = {Machine review of arXiv:2412.05269}
}
read the original abstract

Chemical synthesis remains a critical bottleneck in the discovery and manufacture of functional small molecules. AI-based synthesis planning models could be a potential remedy to find effective syntheses, and have made progress in recent years. However, they still struggle with less frequent, yet critical reactions for synthetic strategy, as well as hallucinated, incorrect predictions. This hampers multi-step search algorithms that rely on models, and leads to misalignment with chemists' expectations. Here we propose RetroChimera: a frontier retrosynthesis model, built upon two newly developed components with complementary inductive biases, which we fuse together using a new framework for integrating predictions from multiple sources via a learning-based ensembling strategy. Through experiments across several orders of magnitude in data scale and splitting strategy, we show RetroChimera outperforms all major models by a large margin, demonstrating robustness outside the training data, as well as for the first time the ability to learn from even a very small number of examples per reaction class. Moreover, industrial organic chemists prefer predictions from RetroChimera over the reactions it was trained on in terms of quality, revealing high levels of alignment. Finally, we demonstrate zero-shot transfer to an internal dataset from a major pharmaceutical company, showing robust generalization under distribution shift. With the new dimension that our ensembling framework unlocks, we anticipate further acceleration in the development of even more accurate models.

Figures

Figures reproduced from arXiv: 2412.05269 by the authors.

Figure 1
Figure 1. Our framework for ensemble-based retrosynthesis with learned reranking which underpins RetroChimera. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. a, Architecture of the editing model (NeuralLoc). Product and templates are encoded through Graph Neural Network encoders to produce contextualized atom representations. Template scores are computed by multiplying product representation with template representations. Localization scores are computed as products of product atom representations and template left-hand side atom representations. All templates in the bat… view at source ↗
Figure 3
Figure 3. Benchmarking Pistachio-trained models (ours shown as solid lines, baselines as dashed). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Example routes identified by RetroChimera. Targets were selected from the Pistachio test set, and represent [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visualization of how predictions from R-SMILES 2 and NeuralLoc are combined by RetroChimera. Molecule [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Examples of the denoising behaviour of RetroChimera. Left column ( [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

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. URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment

    cs.LG 2026-07 accept novelty 6.0 of 10

    Specialized retrosynthesis models outperform LLMs on chemically plausible multi-step routes when scored by the new URSA Solv-2 protocol using ChemCensor.

  2. Nature Language Model: Deciphering the Language of Nature for Scientific Discovery

    cs.AI 2025-02 conditional novelty 6.0 of 10

    A single sequence-based model, pretrained across molecules, proteins, materials, nucleotides and text, outperforms specialist models on several generation tasks and enables cross-domain design.

Reference graph

Works this paper leans on

83 extracted references · 71 canonical work pages · cited by 2 Pith papers

  1. [1]

    & Segler, M

    Stanley, M. & Segler, M. Fake it until you make it? generative de novo design and virtual screen- ing of synthesizable molecules.Current Opinion in Structural Biology82, 102658 (2023)

    2023

  2. [2]

    D.et al.Aizynth impact on medicinal chemistry practice at astrazeneca.RSC Medici- nal Chemistry15, 1085–1095 (2024)

    Shields, J. D.et al.Aizynth impact on medicinal chemistry practice at astrazeneca.RSC Medici- nal Chemistry15, 1085–1095 (2024)

    2024

  3. [3]

    & Coley, C

    Tu, Z., Stuyver, T. & Coley, C. W. Predictive chemistry: machine learning for reaction deploy- ment, reactiondevelopment, andreactiondiscov- ery. Chemical science14, 226–244 (2023)

    2023

  4. [4]

    Concerning one system of classifica- tion and codification of organic reactions.Infor- mation Storage and Retrieval1, 117–146 (1963)

    Vleduts, G. Concerning one system of classifica- tion and codification of organic reactions.Infor- mation Storage and Retrieval1, 117–146 (1963)

    work page 1963

  5. [5]

    Corey, E. J. & Wipke, W. T. Computer-assisted design of complex organic syntheses: Pathways for molecular synthesis can be devised with a computer and equipment for graphical commu- nication. Science 166, 178–192 (1969)

    work page 1969

  6. [6]

    & Gasteiger, J

    Ihlenfeldt, W.-D. & Gasteiger, J. Computer- assisted planning of organic syntheses: the sec- ondgenerationofprograms. Angewandte Chemie International Edition in English34, 2613–2633 (1996)

    work page 1996

  7. [7]

    Hastedt, F.et al.Investigatingthereliabilityand interpretability of machine learning frameworks for chemical retrosynthesis.Digital Discovery3, 1194–1212 (2024)

    work page 2024

  8. [8]

    Schwaller, P. et al. Machine intelligence for chemical reaction space.Wiley Interdisciplinary Reviews: Computational Molecular Science 12, e1604 (2022)

    work page 2022

Show all 83 references
  1. [9]

    Klucznik, T. et al. Efficient syntheses of di- verse, medicinally relevant targets planned by computer and executed in the laboratory.Chem 4, 522–532 (2018)

    2018

  2. [10]

    Nature 588, 83–88 (2020)

    Mikulak-Klucznik, B.et al.Computational plan- ning of the synthesis of complex natural prod- ucts. Nature 588, 83–88 (2020)

    2020

  3. [11]

    Segler, M. H. & Waller, M. P. Neural-symbolic machine learning for retrosynthesis and reaction prediction. Chemistry–A European Journal23, 5966–5971 (2017)

    2017

  4. [12]

    H., Preuss, M

    Segler, M. H., Preuss, M. & Waller, M. P. Plan- ning chemical syntheses with deep neural net- works and symbolic ai. Nature 555, 604–610 (2018). 10

    2018

  5. [13]

    Coley, C. W.et al. A robotic platform for flow synthesis of organic compounds informed by ai planning. Science 365, eaax1566 (2019)

    2019

  6. [14]

    Retrosynthetic reaction prediction using neural sequence-to-sequence models.ACS central science3, 1103–1113 (2017)

    Liu, B.et al. Retrosynthetic reaction prediction using neural sequence-to-sequence models.ACS central science3, 1103–1113 (2017)

    2017

  7. [15]

    Schwaller, P. et al. Molecular transformer: a model for uncertainty-calibrated chemical reac- tion prediction. ACS central science 5, 1572– 1583 (2019)

    2019

  8. [16]

    & Song, L

    Dai, H., Li, C., Coley, C., Dai, B. & Song, L. Retrosynthesis prediction with conditional graph logic network. Advances in Neural Information Processing Systems32 (2019)

    2019

  9. [17]

    V., Karpov, P., Van Deursen, R

    Tetko, I. V., Karpov, P., Van Deursen, R. & Godin, G. State-of-the-art augmented nlp trans- former models for direct and single-step ret- rosynthesis. Nature communications 11, 5575 (2020)

    2020

  10. [18]

    Sacha, M. et al. Molecule edit graph atten- tion network: modeling chemical reactions as se- quences of graph edits.Journal of Chemical In- formation and Modeling61, 3273–3284 (2021)

    2021

  11. [19]

    & Jung, Y

    Chen, S. & Jung, Y. Deep retrosynthetic reac- tion prediction using local reactivity and global attention. JACS Au1, 1612–1620 (2021)

    2021

  12. [20]

    & Coley, C

    Tu, Z. & Coley, C. W. Permutation invariant graph-to-sequence model for template-free ret- rosynthesis and reaction prediction. Journal of chemical information and modeling 62, 3503– 3513 (2022)

    2022

  13. [21]

    Zhong, Z. et al. Root-aligned smiles: a tight representation for chemical reaction prediction. Chemical Science13, 9023–9034 (2022)

    2022

  14. [22]

    & Correia, B

    Igashov, I., Schneuing, A., Segler, M., Bronstein, M. & Correia, B. Retrobridge: Modeling ret- rosynthesis with markov bridges.arXiv preprint arXiv:2308.16212 (2023)

    2023 arXiv

  15. [23]

    Retrosynthesis prediction with an interpretable deep-learning framework based on molecular assembly tasks.Nature Communi- cations 14, 6155 (2023)

    Wang, Y.et al. Retrosynthesis prediction with an interpretable deep-learning framework based on molecular assembly tasks.Nature Communi- cations 14, 6155 (2023)

    2023

  16. [24]

    In Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Xie, S.et al.Retrosynthesis prediction with local template retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, 5330–5338 (2023)

    2023

  17. [25]

    & Garg, V

    Laabid, N., Rissanen, S., Heinonen, M., Solin, A. & Garg, V. Alignment is key for applying dif- fusion models to retrosynthesis. arXiv preprint arXiv:2405.17656 (2024)

    2024 arXiv

  18. [26]

    Gaiński, P. et al. Retrogfn: Diverse and feasi- ble retrosynthesis using gflownets. InWorkshop on Generative and Experimental Perspectives for Biomolecular Design (ICLR-W 2024)(2024)

    2024

  19. [27]

    M.et al.Do chemformers dream of organic matter? evaluating a transformer model for multistep retrosynthesis

    Westerlund, A. M.et al.Do chemformers dream of organic matter? evaluating a transformer model for multistep retrosynthesis. Journal of Chemical Information and Modeling 64, 3021– 3033 (2024)

    2024

  20. [28]

    & Yang, Y

    Zhang, X., Mo, Y., Wang, W. & Yang, Y. Retrosynthesis prediction enhanced by in-silico reaction data augmentation. arXiv preprint arXiv:2402.00086 (2024)

    2024 arXiv

  21. [29]

    Nature Commu- nications 15, 6404 (2024)

    Han, Y.et al.Retrosynthesis prediction with an iterative string editing model. Nature Commu- nications 15, 6404 (2024)

    2024

  22. [30]

    Rsgpt: a generative transformer model for retrosynthesis planning pre-trained on ten billion datapoints

    Deng, Y.et al. Rsgpt: a generative transformer model for retrosynthesis planning pre-trained on ten billion datapoints. Nature Communications 16, 7012 (2025)

    2025

  23. [31]

    & Song, L

    Chen, B., Li, C., Dai, H. & Song, L. Retro*: learning retrosynthetic planning with neural guided a* search. InInternational conference on machine learning, 1608–1616 (PMLR, 2020)

    2020

  24. [32]

    Xie, S. et al. Retrograph: Retrosynthetic plan- ning with graph search. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2120–2129 (2022)

    2022

  25. [33]

    InInternational Conference on Machine Learning, 22266–22276 (PMLR, 2023)

    Liu, G.et al.Retrosynthetic planning with dual value networks. InInternational Conference on Machine Learning, 22266–22276 (PMLR, 2023)

    2023

  26. [34]

    & Hernández-Lobato, J

    Tripp, A., Maziarz, K., Lewis, S., Segler, M. & Hernández-Lobato, J. M. Retro-fallback: retrosynthetic planning in an uncertain world. arXiv preprint arXiv:2310.09270(2023)

    2023 arXiv

  27. [35]

    Genheden, S.et al.Aizynthfinder: a fast, robust and flexible open-source software for retrosyn- thetic planning.Journal of cheminformatics12, 70 (2020)

    2020

  28. [36]

    Askcos: Open-source, data-driven synthesis planning

    Tu, Z.et al. Askcos: Open-source, data-driven synthesis planning. Accounts of Chemical Re- search (2025). 11

    2025

  29. [37]

    Strieth-Kalthoff, F., Sandfort, F., Segler, M. H. & Glorius, F. Machine learning the ropes: prin- ciples, applications and directions in synthetic chemistry. Chemical Society Reviews49, 6154– 6168 (2020)

    2020

  30. [38]

    E., Coley, C

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

    2020

  31. [39]

    Journal of chemical information and modeling 62, 2111–2120 (2022)

    Seidl, P.et al.Improving few-and zero-shot reac- tion template prediction using modern hopfield networks. Journal of chemical information and modeling 62, 2111–2120 (2022)

    2022

  32. [40]

    Hassen, A. K. et al. Mind the retrosynthe- sis gap: bridging the divide between single-step and multi-step retrosynthesis prediction. arXiv preprint arXiv:2212.11809 (2022)

    2022 arXiv

  33. [41]

    & Duarte, F

    Tanovic, S., Wieczorek, E. & Duarte, F. An ex- ploration of dataset bias in single-step retrosyn- thesis prediction. chemrxiv (2025)

    2025

  34. [42]

    Ensemble learning

    Sewell, M. Ensemble learning. RN 11, 1–34 (2008)

    2008

  35. [43]

    & Mahoney, M

    Theisen, R., Kim, H., Yang, Y., Hodgkinson, L. & Mahoney, M. W. When are ensembles really effective? Advances in Neural Information Pro- cessing Systems36 (2024)

    2024

  36. [44]

    InProceedings of the 22nd international conference on Machine learning, 89–96 (2005)

    Burges, C.et al.Learning to rank using gradient descent. InProceedings of the 22nd international conference on Machine learning, 89–96 (2005)

    2005

  37. [45]

    Maziarz, K. et al. Re-evaluating retrosynthesis algorithmswithsyntheseus. Faraday Discussions (2024)

    2024

  38. [46]

    & Bjerrum, E

    Irwin, R., Dimitriadis, S., He, J. & Bjerrum, E. J. Chemformer: a pre-trained transformer for computational chemistry.Machine Learning: Science and Technology3, 015022 (2022)

    2022

  39. [47]

    & Chen, C

    Zhong, W., Yang, Z. & Chen, C. Y.-C. Retrosyn- thesis prediction using an end-to-end graph gen- erative architecture for molecular graph editing. Nature Communications14, 3009 (2023)

    2023

  40. [48]

    H., Tu, Z

    Lin, M. H., Tu, Z. & Coley, C. W. Improving the performance of models for one-step retrosynthe- sis through re-ranking. Journal of cheminfor- matics 14, 15 (2022)

    2022

  41. [49]

    & Jastrzębski, S

    Sacha, M., Sadowski, M., Kozakowski, P., van Workum, R. & Jastrzębski, S. Molecule-edit templates for efficient and accurate retrosynthe- sis prediction. arXiv preprint arXiv:2310.07313 (2023)

    2023 arXiv

  42. [50]

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

    2017

  43. [51]

    Touvron, H. et al. Llama 2: Open foundation and fine-tuned chat models (2023). URLhttps: //arxiv.org/abs/2307.09288. 2307.09288

    2023 arXiv

  44. [52]

    Jiang, A. Q. et al. Mistral 7b (2023). URL https://arxiv.org/abs/2310.06825. 2310. 06825

    2023 arXiv

  45. [53]

    & Rush, A

    Klein, G., Kim, Y., Deng, Y., Senellart, J. & Rush, A. OpenNMT: Open-source toolkit for neural machine translation. In Proceedings of ACL 2017, System Demonstrations, 67–72 (As- sociation for Computational Linguistics, Van- couver, Canada, 2017). URL https://www. aclweb.org/an...

    2017

  46. [54]

    Ainslie, J. et al. GQA: Training generalized multi-query transformer models from multi-head checkpoints. In Bouamor, H., Pino, J. & Bali, K. (eds.)Proceedings of the 2023 Conference on Empirical Methods in Natural Language Process- ing, 4895–4901 (Association for Computational...

    2023

  47. [55]

    & Sennrich, R

    Zhang, B. & Sennrich, R. Root mean square layer normalization. In Wallach, H. et al. (eds.) Advances in Neural Information Pro- cessing Systems, vol. 32 (Curran Associates, Inc., 2019). URL https://proceedings. neurips.cc/paper_files/paper/2019/file/ 1e8a19426224ca89e83cef47f1...

    2019

  48. [56]

    GLU variants improve transformer

    Shazeer, N. GLU variants improve transformer. CoRR abs/2002.05202 (2020). URL https: //arxiv.org/abs/2002.05202. 2002.05202

    2020 arXiv

  49. [57]

    & Thost, V

    Tu, H., Shorewala, S., Ma, T. & Thost, V. Retrosynthesis prediction revisited. InNeurIPS 2022 AI for Science: Progress and Promises (2022)

    2022

  50. [58]

    H., Petraglia, R

    Schwaller, P., Nair, V. H., Petraglia, R. & Laino, T. Evaluation metrics for single-step retrosyn- thetic models. InSecond Workshop on Machine Learning and the Physical Sciences (NeurIPS Vancouver, Canada, 2019)

    2019

  51. [59]

    & Lai, L

    Li, J., Lin, K., Pei, J. & Lai, L. Challenging complexity with simplicity: Rethinking the role 12 of single-step models in computer-aided synthe- sis planning. Journal of Chemical Information and Modeling64, 5470–5479 (2024)

    2024

  52. [60]

    & Schuffenhauer, A

    Ertl, P. & Schuffenhauer, A. Estimation of syn- thetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. Journal of cheminformatics1, 1– 11 (2009)

    2009

  53. [61]

    ring breaker

    Thakkar, A., Selmi, N., Reymond, J.-L., En- gkvist, O. & Bjerrum, E. J. “ring breaker”: neu- ral network driven synthesis prediction of the ring system chemical space. Journal of medic- inal chemistry63, 8791–8808 (2020)

    2020

  54. [62]

    Levin, I., Liu, M., Voigt, C. A. & Coley, C. W. Merging enzymatic and synthetic chemistry with computational synthesis planning.Nature Com- munications 13, 7747 (2022)

    2022

  55. [63]

    Kim, T. et al. Readretro: natural product biosynthesis predicting with retrieval-augmented dual-view retrosynthesis. New Phytologist (2024)

    2024

  56. [64]

    Torren-Peraire, P. et al. Models matter: The impact of single-step retrosynthesis on synthesis planning. Digital Discovery3, 558–572 (2024)

    2024

  57. [65]

    Journal of Cheminformat- ics 16, 57 (2024)

    Saigiridharan, L.et al.Aizynthfinder 4.0: devel- opments based on learnings from 3 years of in- dustrial application. Journal of Cheminformat- ics 16, 57 (2024)

    2024

  58. [66]

    Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  59. [67]

    Pytorch: An imperative style, high-performance deep learning library

    Paszke, A.et al. Pytorch: An imperative style, high-performance deep learning library. Ad- vances in neural information processing systems 32 (2019)

    2019

  60. [68]

    Willett, P., Barnard, J. M. & Downs, G. M. Chemical similarity searching.Journal of chem- ical information and computer sciences38, 983– 996 (1998)

    1998

  61. [69]

    & Engkvist, O

    Steffen, A., Kogej, T., Tyrchan, C. & Engkvist, O. Comparison of molecular fingerprint methods on the basis of biological profile data.Journal of chemical information and modeling49, 338–347 (2009)

    2009

  62. [70]

    Dgl-lifesci: An open-source toolkit for deep learning on graphs in life science.ACS Omega (2021)

    Li, M.et al. Dgl-lifesci: An open-source toolkit for deep learning on graphs in life science.ACS Omega (2021)

    2021

  63. [71]

    W., Green, W

    Coley, C. W., Green, W. H. & Jensen, K. F. Rdchiral: An rdkit wrapper for handling stereo- chemistry in retrosynthetic template extraction andapplication. Journal of chemical information and modeling59, 2529–2537 (2019)

    2019

  64. [72]

    D., Zentgraf, M

    Christ, C. D., Zentgraf, M. & Kriegl, J. M. Min- ing electronic laboratory notebooks: analysis, retrosynthesis, and reaction based enumeration. Journal of chemical information and modeling 52, 1745–1756 (2012)

    2012

  65. [73]

    & Green, W

    Heid, E., Liu, J., Aude, A. & Green, W. H. Influ- ence of template size, canonicalization, and ex- clusivity for retrosynthesis and reaction predic- tion applications. Journal of Chemical Informa- tion and Modeling62, 16–26 (2021)

    2021

  66. [74]

    & Veličković, P

    Corso, G., Cavalleri, L., Beaini, D., Liò, P. & Veličković, P. Principal neighbourhood aggre- gation for graph nets. Advances in Neural In- formation Processing Systems33, 13260–13271 (2020)

    2020

  67. [75]

    & Lenssen, J

    Fey, M. & Lenssen, J. E. Fast Graph Rep- resentation Learning with PyTorch Geometric (2019). URL https://github.com/pyg-team/ pytorch_geometric

    2019

  68. [76]

    Rampášek, L. et al. Recipe for a general, pow- erful, scalable graph transformer. Advances in Neural Information Processing Systems 35, 14501–14515 (2022)

    2022

  69. [77]

    Beaini, D. et al. Towards foundational mod- els for molecular learning on large-scale multi- task datasets. arXiv preprint arXiv:2310.04292 (2023)

    2023 arXiv

  70. [78]

    Learning to extend molecular scaffolds with structural motifs

    Maziarz, K.et al. Learning to extend molecular scaffolds with structural motifs. InInternational Conference on Learning Representations (ICLR

  71. [79]

    & Beyer, L

    Zhai, X., Mustafa, B., Kolesnikov, A. & Beyer, L. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 11975–11986 (2023)

    2023

  72. [80]

    Retroprime: A diverse, plausi- bleandtransformer-basedmethodforsingle-step retrosynthesispredictions

    Wang, X.et al. Retroprime: A diverse, plausi- bleandtransformer-basedmethodforsingle-step retrosynthesispredictions. Chemical Engineering Journal 420, 129845 (2021)

    2021

  73. [81]

    Yan, C. et al. Retroxpert: Decompose ret- rosynthesis prediction like a chemist.Advances in Neural Information Processing Systems 33, 11248–11258 (2020). 13

    2020

  74. [82]

    Sheridan, R. P. Time-split cross-validation as a method for estimating the goodness of prospec- tive prediction.Journal of chemical information and modeling53, 783–790 (2013)

    2013

  75. [83]

    (\[[^\]]+]|Br?|Cl?|N|O|S|P|F| I|b|c|n|o|s|p|\(|\)|\.|=|#|-|\+|\\\\|\/|:|~| @|\?|>|\*|\$|\%[0-9]{2}|[0-9])

    Campello, R. J., Moulavi, D. & Sander, J. Density-based clustering based on hierarchical density estimates. InPacific-Asia conference on knowledge discovery and data mining, 160–172 (Springer, 2013). 14 Methods Prior Work on Ensembling While ensembling for reaction prediction ...

    2013

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.