Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

DiffER: Categorical Diffusion for Chemical Retrosynthesis

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Categorical diffusion can predict retrosynthetic reactants by denoising an entire SMILES string at once, and an eight-model ensemble called DiffER reaches 57.6% top-1 accuracy on USPTO-50K, the best among template-free methods.

desk verdict First categorical-diffusion retrosynthesis model with strong USPTO-50K top-1, but the ensemble may have been tuned on test and diversity reporting conflicts. read the letter →

arxiv 2505.23721 v2 pith:ETUHDAAI submitted 2025-05-29 cs.LG

Sean Current , Ziqi Chen , Daniel Adu-Ampratwum , Xia Ning , Srinivasan Parthasarathy This is my paper · ORCID
classification cs.LG
keywords retrosynthesispredictioncategoricaldiffusiontemplate-freemodelsSMILESgenerationnon-autoregressivedecodinglengthensemblevotingUSPTO-50K
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 tries to establish that categorical diffusion — generating the entire reactant SMILES string (a text encoding of a molecule) in parallel from noise rather than token by token — is a serious alternative to autoregressive transformers for template-free retrosynthesis prediction. Its model, DiffER, is an ensemble of eight multinomial diffusion models; on the USPTO-50K benchmark it reaches 57.6% top-1 accuracy, ahead of every other template-free method, with competitive top-3 (79.0%), top-5 (84.1%), and top-10 (87.4%) results. The central technical problem the paper solves is that diffusion needs the output length fixed in advance: it trains models with a uniformly random number of padding tokens so that lengths near the predicted length are treated as roughly equally probable, and this length-variance trick lifts accuracy from 40.4% with rigid length prediction to the ensemble figure. A companion experiment with oracle length prediction reaches 77.0% top-1, better than every compared method, which the paper reads as evidence that output length estimation is the main bottleneck for the diffusion approach. The case studies also show the model learns many viable reactions, including sensible alternatives to the patented ground truth.

What carries the argument

The central object is the multinomial diffusion process over categorical SMILES tokens: a forward Markov chain replaces the one-hot reactant sequence with uniform noise at rate $\beta_t$ per step, and the reverse step is a categorical posterior $p_\theta(y_{t-1} \mid y_t, x_0)$ whose estimate of the clean sequence $y_0$ comes from a transformer decoder conditioned on the product encoder's representation. The paper's own contribution is a length-prediction-with-variance scheme: a LENGTH token's representation predicts the length difference between product and reactant, and training appends a uniform random number $n \sim U(1, N)$ of padding tokens, making lengths near the prediction approximately equally probable while still allowing the model to deviate. An ensemble of eight such models with padding limits $N \in \{20, 30, 40, 50, 60, 70, 80, 90\}$, combined with sampling-frequency ranking and ranked-choice tie-breaking, carries the final accuracy figures.

What would settle it

Inspect the distribution of output lengths that DiffER actually produces across the test set: if the length-variance mechanism is doing the advertised work, lengths should spread across the window around each predicted length and length errors should track prediction errors, whereas a distribution concentrated on the maximum padded length, or one that ignores the predicted length entirely, would show the model is emitting padding tokens rather than chemistry and would refute the claim that length variance is the source of the gain.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is that a categorical diffusion model trained to denoise a fully noised reactant SMILES sequence, conditioned on an encoder representation of the product, can compete with — and at top-1 beat — autoregressive template-free retrosynthesis models. The model follows the multinomial diffusion construction: a forward chain gradually turns one-hot reactant tokens into uniform noise, and the decoder learns the posterior denoising step, with the source sequence processed by a transformer encoder and supplied as memory. The novel piece is length prediction with variance: during training a uniform random number of padding tokens is appended to the target, so that the model assigns approximately equal probability to sequences whose lengths fall within a window of the predicted length, avoiding the collapse to over-short outputs that a rigid length predictor causes. At inference, twenty root-aligned SMILES augmentations per product are sampled from each of eight models trained with different padding limits, and canonicalized outputs are ranked by sampling frequency with ranked-choice tie-breaking. The paper reports 57.6% top-1 accuracy on USPTO-50K, the best among template-free methods, and shows that a single diffusion model supplied with the true output length reaches 77.0% top-1, outperforming all compared methods and demonstrating that accurate length estimation is the key to further gains.

Load-bearing premise

The load-bearing premise is that padding the training targets with a uniformly random number of tokens makes the model treat a window of lengths around its predicted length as roughly equally probable, rather than collapsing toward padding tokens or short outputs; if that premise fails, the entire gap between DiffER's 57.6% and the rigid length-prediction baseline's 40.4% collapses.

Editorial extensions

If this is right

  • Non-autoregressive decoding becomes a credible design choice for SMILES-based retrosynthesis: the ensemble matches or beats sequential transformers at top-1 without beam search, supporting the paper's hypothesis that parallel decoding captures global structure such as rings and branches.
  • Diffusion gives a natural probabilistic reading of candidate reactants: the fraction of samples mapping to a given canonicalized molecule approximates its posterior probability, so ranking by sampling frequency doubles as an interpretable confidence score.
  • Output length estimation, not denoising power, is the binding constraint for categorical diffusion on discrete sequences; the gap between the ensemble's length prediction and the oracle (57.6% versus 77.0% top-1) marks the clearest route to further gains.

Reading between the lines

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

  • Because the oracle result so sharply outruns the ensemble, the fastest path to higher accuracy is probably a better length model — for instance a learned distribution over length differences or length-conditioned noise initialization — rather than a stronger denoising network; the paper lists this as future work but does not test it.
  • A sampling strategy that deliberately trades confidence for diversity, such as temperature or top-p filtering on the categorical posterior, could plausibly close DiffER's top-3/5/10 gap to Root-Aligned SMILES without sacrificing its top-1 edge; the paper's ensemble buys diversity only by averaging over padding limits.
  • Scoring DiffER's non-ground-truth top-k reactants with a forward-reaction predictor or expert labels would quantify how often its 'misses' are actually viable syntheses, directly addressing the paper's own point that matching patented reactions undercounts chemically sound alternatives.
  • The same categorical-diffusion-plus-length-variance recipe should transfer to other constrained discrete generation tasks — forward reaction prediction, de novo molecular design, or protein and DNA sequence design — where output length is also unknown in advance.
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. The paper proposes DiffER, an ensemble of categorical diffusion models for template-free single-step retrosynthesis on USPTO-50K. The method encodes the product with a transformer encoder, predicts the target SMILES length, and decodes the entire reactant sequence in parallel using multinomial diffusion. A novel length-variation scheme adds a uniform random number of padding tokens during training, and eight models with different padding limits are ensembled and ranked by sampling frequency. The paper reports 57.6% top-1 accuracy, competitive top-3/5/10 accuracy, an oracle-length diffusion upper bound of 77.0%, and qualitative case studies.

Significance. If the empirical claims are reliable, the paper makes a meaningful contribution: it establishes categorical diffusion as a viable non-autoregressive alternative to transformer-based template-free retrosynthesis, introduces a practical length-variation technique, and provides an oracle-length analysis that isolates the role of length prediction. The manuscript is well-structured, includes code and data availability, and follows standard evaluation protocols (USPTO-50K, root-aligned SMILES, top-k accuracy). The main risk is that the headline result depends on ensemble hyperparameters whose selection procedure is not described as validation-based, and there is an internal inconsistency in the reported diversity statistics. With proper documentation and correction, the contribution would be solid.

major comments (3)
  1. [Section 1.4 and Table 1] The manuscript does not state whether the ensemble composition (the selected set of N values and the voting scheme) was chosen on a validation split. Section 1.4 says the authors tested N in {5,10,...,90} and various voting combinations, then fixed the final ensemble to N in {20,30,...,90}. If this selection was performed on the test set, the reported 57.6% top-1 accuracy is an optimistically biased maximum over many configurations, and the 1.3-percentage-point gap over R-SMILES (56.3%) is not a reliable estimate of generalization. Please specify the model-selection protocol, report the validation-set performance for all configurations considered, and, ideally, provide confidence intervals or significance tests for the headline comparison.
  2. [Section 3.1 vs. Figure 2] The diversity statistics are internally inconsistent. Section 3.1 states the ensemble produces a median of 5 samples, with 17.9% of reactions producing less than 5 and 57.1% producing less than 10. The Figure 2 caption states a mean of 10.0 and a median of 9. These two descriptions cannot both be correct for the same distribution. Please correct the text, figure, or caption and re-verify the evaluation pipeline, since this discrepancy undermines confidence in the reported statistics.
  3. [Section 1.3] The claim that adding n ~ U(1,N) padding tokens makes the model assign approximately equal probability to sequences whose lengths fall near the predicted length is presented without empirical support. The uniform padding alone does not guarantee that the trained model will weight lengths in that range uniformly; the model could still be biased toward predicting padding tokens or toward shorter outputs. Please provide evidence about the predicted length distribution (e.g., a histogram of predicted lengths versus target lengths, or an analysis of where the probability mass lies) or temper the claim accordingly.
minor comments (5)
  1. [Eq. (8)] The loss expression appears to contain a typo: '||y2_0 − yhat_0^2||' should likely be '||y0 − yhat0||^2'.
  2. [Abstract and Section 4.4] The phrase 'we prove that DiffER is a strong baseline' is misleading; no formal proof is provided, and 'demonstrate' or 'show' would be more appropriate.
  3. [Section 1.6] The sampling procedure is unclear: it says 20 random Root-aligned SMILES strings are generated per product and that each model outputs 20 random samples, but Section 1.4 refers to drawing multiple samples per model without specifying whether diffusion stochasticity is used in addition to input augmentation. Please clarify the exact number of diffusion samples per augmented input per model.
  4. [Section 3.3] The oracle-length experiment is run with a single diffusion model, not the ensemble; the text calls this an upper limit of 'the diffusion approach,' which is imprecise. An ensemble with oracle length could potentially perform even better. Please rephrase to avoid overclaiming.
  5. [Section 4.3] There is a typo: 'UPSTO-50K' should be 'USPTO-50K'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular dependency: DiffER's benchmark claims rest on external evaluations and controlled ablations, not on self-referential definitions or fitted predictions.

full rationale

DiffER's central claims are empirical benchmark results, not derivations from the method's own outputs. The method is specified by explicit equations (Sections 1.1-1.3): multinomial noising (Eqs. 1-2), posterior denoising (Eqs. 3-6), length prediction (Eq. 7), and losses (Eqs. 8-10). None of these equations contains the reported accuracy or the ground-truth reactant as an input; the model must be trained and evaluated on external USPTO-50K data. The headline 57.6% top-1 accuracy is compared against published template-free baselines (Table 1) using standard top-k accuracy, so it is not a fitted parameter renamed as a prediction. The oracle-length experiment is explicitly labeled an upper limit that uses ground-truth length at inference, not a claim of predictive performance. The improvement of DiffER over the baseline length-prediction model is a controlled ablation (same diffusion architecture, with versus without the length-variation padding), so the attribution to the proposed mechanism is empirical, not tautological. The only co-authored citation is G2Retro [20], used as a baseline for reaction-type comparison and as an example of similar per-class analysis; it is not load-bearing for the method's validity, is not a uniqueness theorem, and its numbers are externally established benchmark results. The paper's own limitations discussion (Section 4.3) acknowledges the method's sensitivity to length prediction and limited output diversity, which is consistent with an honest empirical report, not a circular derivation. Two non-circular caveats remain: the paper does not state whether ensemble and voting hyperparameters were selected on a validation split rather than the test set, which is a statistical unbiasedness concern; and the text (median 5) and Figure 2 caption (median 9) disagree on sample diversity. Neither caveat is a self-referential reduction, so they do not affect the circularity score.

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

The central claim rests on benchmark validity, the diffusion approximation, and the ad hoc random-padding length mechanism. No new physical or conceptual entities are introduced. The main free parameters are the padding limit N and the ensemble composition, both selected to maximize benchmark accuracy.

free parameters (4)
  • per-model random padding limit N = {20, 30, 40, 50, 60, 70, 80, 90}
    Values below 20 were tested and discarded because they over-rely on the length prediction component and perform poorly (Section 1.4). The selection is not documented as validation-based, so the chosen set may reflect test-set observations.
  • ensemble composition = 8 models, one per selected N
    The number of models and their N values were chosen jointly with the padding limit; the paper does not report a search on a separate validation set.
  • diffusion steps T = 200
    Standard hyperparameter chosen for the diffusion process; not derived from data and not central to the method's novelty.
  • inference samples per model per input = 20
    Number of root-aligned SMILES augmentations per product at inference, inherited from R-SMILES [6]; affects diversity and inference cost.
assumptions (4)
  • domain assumption USPTO-50K with root-aligned SMILES splits is a valid benchmark for single-step retrosynthesis.
    All accuracy claims are measured on this dataset and these splits; if the benchmark or splits are unrepresentative, the SOTA claim is unsupported.
  • domain assumption The multinomial diffusion model trained with Equation 6 approximates the true posterior distribution over reactant SMILES.
    This is the core modeling assumption behind DiffER; the paper does not provide a formal guarantee, only empirical validation.
  • ad hoc to paper Adding a uniform random number of padding tokens (n ~ U(1,N)) during training gives approximately equal probability to sequences of lengths near the predicted target length.
    Introduced in Section 1.3 to justify the length-variation method; it is an informal claim about the model's learned behavior, not a derived theorem.
  • domain assumption Product and reactant SMILES lengths are strongly correlated, so predicting the length difference is easier than predicting absolute length.
    Section 1.3 relies on this correlation to shift from total length prediction to difference prediction; if false, the length predictor would be less reliable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffER: Categorical Diffusion for Chemical Retrosynthesis." pith.science (2026). https://pith.science/paper/ETUHDAAI

@misc{pith2026250523721,
  author       = {Pith},
  title        = {Pith review of: DiffER: Categorical Diffusion for Chemical Retrosynthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ETUHDAAI}},
  note         = {Machine review of arXiv:2505.23721}
}
read the original abstract

Methods for automatic chemical retrosynthesis have found recent success through the application of models traditionally built for natural language processing, primarily through transformer neural networks. These models have demonstrated significant ability to translate between the SMILES encodings of chemical products and reactants, but are constrained as a result of their autoregressive nature. We propose DiffER, an alternative template-free method for retrosynthesis prediction in the form of categorical diffusion, which allows the entire output SMILES sequence to be predicted in unison. We construct an ensemble of diffusion models which achieves state-of-the-art performance for top-1 accuracy and competitive performance for top-3, top-5, and top-10 accuracy among template-free methods. We prove that DiffER is a strong baseline for a new class of template-free model, capable of learning a variety of synthetic techniques used in laboratory settings and outperforming a variety of other template-free methods on top-k accuracy metrics. By constructing an ensemble of categorical diffusion models with a novel length prediction component with variance, our method is able to approximately sample from the posterior distribution of reactants, producing results with strong metrics of confidence and likelihood. Furthermore, our analyses demonstrate that accurate prediction of the SMILES sequence length is key to further boosting the performance of categorical diffusion models.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Discrete Diffusion Models: A Unified Framework from Tokenization to Generation

    cs.LG 2026-07 unverdicted novelty 4.0 of 10

    Discrete diffusion models are re-framed as instances of a tokenization-centric, four-component design space (corruption, denoiser, objective, sampler) in a broad survey with no new experimental or theoretical results.

Reference graph

Works this paper leans on

50 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [1]

    Chemistry–A European Journal 23(25), 5966–5971 (2017)

    Segler, M.H., Waller, M.P.: Neural-symbolic machine learning for retrosynthe- sis and reaction prediction. Chemistry–A European Journal 23(25), 5966–5971 (2017)

    work page 2017

  2. [2]

    Nature 555(7698), 604–610 (2018)

    Segler, M.H., Preuss, M., Waller, M.P.: Planning chemical syntheses with deep neural networks and symbolic ai. Nature 555(7698), 604–610 (2018)

    work page 2018

  3. [3]

    Nature Reviews Methods Primers 1(1), 1–23 (2021) 21

    Shen, Y., Borowski, J.E., Hardy, M.A., Sarpong, R., Doyle, A.G., Cernak, T.: Automation and computer-assisted planning for chemical synthesis. Nature Reviews Methods Primers 1(1), 1–23 (2021) 21

    work page 2021

  4. [4]

    Weininger, D.: Smiles, a chemical language and information system. 1. introduc- tion to methodology and encoding rules. Journal of chemical information and computer sciences 28(1), 31–36 (1988)

    work page 1988

  5. [5]

    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)

    work page 2017

  6. [6]

    Chemical Science 13(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 prediction. Chemical Science 13(31), 9023–9034 (2022)

    work page 2022

  7. [7]

    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)

    work page 2022

  8. [8]

    Advances in Neural Information Processing Systems 34, 12454–12465 (2021)

    Hoogeboom, E., Nielsen, D., Jaini, P., Forr´ e, P., Welling, M.: Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems 34, 12454–12465 (2021)

    work page 2021

Show all 50 references
  1. [9]

    arXiv preprint arXiv:2210.08933 (2022)

    Gong, S., Li, M., Feng, J., Wu, Z., Kong, L.: Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933 (2022)

    2022 arXiv

  2. [10]

    arXiv preprint arXiv:2310.05793 (2023)

    Gong, S., Li, M., Feng, J., Wu, Z., Kong, L.: Diffuseq-v2: Bridging discrete and continuous text spaces for accelerated seq2seq diffusion models. arXiv preprint arXiv:2310.05793 (2023)

    2023 arXiv

  3. [11]

    arXiv preprint arXiv:1904.09324 (2019)

    Ghazvininejad, M., Levy, O., Liu, Y., Zettlemoyer, L.: Mask-predict: Parallel decoding of conditional masked language models. arXiv preprint arXiv:1904.09324 (2019)

    2019 arXiv

  4. [12]

    arXiv preprint arXiv:1910.13461 (2019)

    Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461 (2019)

    2019 arXiv

  5. [13]

    Lowe, D.: Chemical reactions from US patents (1976-Sep2016) (2017) https:// doi.org/10.6084/m9.figshare.5104873.v1

    2017 doi

  6. [14]

    arXiv preprint arXiv:1606.08415 (2016)

    Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)

    2016 arXiv

  7. [15]

    arXiv preprint arXiv:2301.10972 (2023)

    Chen, T.: On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972 (2023)

    2023 arXiv

  8. [16]

    Wiley Interdisciplinary Reviews: 22 Computational Molecular Science 2(1), 79–107 (2012)

    Cook, A., Johnson, A.P., Law, J., Mirzazadeh, M., Ravitz, O., Simon, A.: Computer-aided synthesis design: 40 years on. Wiley Interdisciplinary Reviews: 22 Computational Molecular Science 2(1), 79–107 (2012)

    2012

  9. [17]

    Science 228(4698), 408–418 (1985)

    Corey, E.J., Long, A.K., Rubenstein, S.D.: Computer-assisted analysis in organic synthesis. Science 228(4698), 408–418 (1985)

    1985

  10. [18]

    Current Opinion in Chemical Engineering 35, 100721 (2022)

    Sun, Y., Sahinidis, N.V.: Computer-aided retrosynthetic design: fundamentals, tools, and outlook. Current Opinion in Chemical Engineering 35, 100721 (2022)

    2022

  11. [19]

    ACS central science 3(12), 1237–1245 (2017)

    Coley, C.W., Rogers, L., Green, W.H., Jensen, K.F.: Computer-assisted ret- rosynthesis based on molecular similarity. ACS central science 3(12), 1237–1245 (2017)

    2017

  12. [20]

    Communications Chemistry 6(1), 102 (2023)

    Chen, Z., Ayinde, O.R., Fuchs, J.R., Sun, H., Ning, X.: G 2 retro as a two- step graph generative models for retrosynthesis prediction. Communications Chemistry 6(1), 102 (2023)

    2023

  13. [21]

    Advances in Neural Information Processing Systems 34, 9405–9415 (2021)

    Somnath, V.R., Bunne, C., Coley, C., Krause, A., Barzilay, R.: Learning graph models for retrosynthesis prediction. Advances in Neural Information Processing Systems 34, 9405–9415 (2021)

    2021

  14. [22]

    Journal of chemical information and modeling 62(15), 3503–3513 (2022)

    Tu, Z., Coley, C.W.: Permutation invariant graph-to-sequence model for template- free retrosynthesis and reaction prediction. Journal of chemical information and modeling 62(15), 3503–3513 (2022)

    2022

  15. [23]

    Journal of Chemical Information and Modeling 61(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 Modeling 61(7),...

    2021

  16. [24]

    Greg Landrum 8(31.10), 5281 (2013)

    Landrum, G., et al.: Rdkit: A software suite for cheminformatics, computational chemistry, and predictive modeling. Greg Landrum 8(31.10), 5281 (2013)

    2013

  17. [25]

    Advances in neural information processing systems 30 (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)

    2017

  18. [26]

    Towards Data Science (2020)

    Joshi, C.K.: Transformers are Graph Neural Networks. Towards Data Science (2020). https://towardsdatascience.com/ transformers-are-graph-neural-networks-bca9f75412aa

    2020

  19. [27]

    In: International Confer- ence on Machine Learning, pp

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: International Confer- ence on Machine Learning, pp. 2256–2265 (2015). PMLR

    2015

  20. [28]

    Advances in neural information processing systems 33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840–6851 (2020)

    2020

  21. [29]

    In: 23 International Conference on Machine Learning, pp

    Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: 23 International Conference on Machine Learning, pp. 8162–8171 (2021). PMLR

    2021

  22. [30]

    Advances in neural information processing systems 35, 36479–36494 (2022)

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text- to-image diffusion models with deep language understanding. Advances in neural information processing systems 35...

    2022

  23. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., Guo, B.: Vector quantized diffusion model for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10696–10706 (2022)

    2022

  24. [32]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 3836–3847 (2023)

    2023

  25. [33]

    arXiv preprint arXiv:2212.10325 (2022)

    Yuan, H., Yuan, Z., Tan, C., Huang, F., Huang, S.: Seqdiffuseq: Text diffusion with encoder-decoder transformers. arXiv preprint arXiv:2212.10325 (2022)

    2022 arXiv

  26. [34]

    arXiv preprint arXiv:2211.15089 (2022)

    Dieleman, S., Sartran, L., Roshannai, A., Savinov, N., Ganin, Y., Richemond, P.H., Doucet, A., Strudel, R., Dyer, C., Durkan, C., et al.: Continuous diffusion for categorical data. arXiv preprint arXiv:2211.15089 (2022)

    2022 arXiv

  27. [35]

    Advances in Neural Information Processing Systems 34, 17981–17993 (2021)

    Austin, J., Johnson, D.D., Ho, J., Tarlow, D., Van Den Berg, R.: Struc- tured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems 34, 17981–17993 (2021)

    2021

  28. [36]

    arXiv preprint arXiv:2211.15029 (2022)

    He, Z., Sun, T., Wang, K., Huang, X., Qiu, X.: Diffusionbert: Improving generative masked language models with diffusion models. arXiv preprint arXiv:2211.15029 (2022)

    2022 arXiv

  29. [37]

    Advances in Neural Information Processing Systems 32 (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 (2019)

    2019

  30. [38]

    JACS Au 1(10), 1612–1620 (2021)

    Chen, S., Jung, Y.: Deep retrosynthetic reaction prediction using local reactivity and global attention. JACS Au 1(10), 1612–1620 (2021)

    2021

  31. [39]

    In: International Conference on Machine Learning, pp

    Shi, C., Xu, M., Guo, H., Zhang, M., Tang, J.: A graph to graphs framework for retrosynthesis prediction. In: International Conference on Machine Learning, pp. 8818–8827 (2020). PMLR

    2020

  32. [40]

    Advances in Neural Information Processing Systems 33, 11248–11258 (2020)

    Yan, C., Ding, Q., Zhao, P., Zheng, S., Yang, J., Yu, Y., Huang, J.: Retrox- pert: Decompose retrosynthesis prediction like a chemist. Advances in Neural Information Processing Systems 33, 11248–11258 (2020)

    2020

  33. [41]

    Chemical Engineering Journal 420, 129845 (2021)

    Wang, X., Li, Y., Qiu, J., Chen, G., Liu, H., Liao, B., Hsieh, C.-Y., Yao, X.: 24 Retroprime: A diverse, plausible and transformer-based method for single-step retrosynthesis predictions. Chemical Engineering Journal 420, 129845 (2021)

    2021

  34. [42]

    arXiv preprint arXiv:2501.08001 (2025)

    Sun, S., Yu, W., Ren, Y., Du, W., Liu, L., Zhang, X., Hu, Y., Ma, C.: Gdiffretro: Retrosynthesis prediction with dual graph enhanced molecular representation and diffusion generation. arXiv preprint arXiv:2501.08001 (2025)

    2025 arXiv

  35. [43]

    Sumner, D., He, J., Thakkar, A., Engkvist, O., Bjerrum, E.J.: Levenshtein augmentation improves performance of smiles based deep-learning synthesis prediction (2020)

    2020

  36. [44]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    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. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, pp. 531–539 (2021)

    2021

  37. [45]

    Advances in Neural Information Processing Systems 34, 10186–10194 (2021)

    Sun, R., Dai, H., Li, L., Kearnes, S., Dai, B.: Towards understanding retrosynthe- sis by energy-based models. Advances in Neural Information Processing Systems 34, 10186–10194 (2021)

    2021

  38. [46]

    In: International Conference on Machine Learning, pp

    Wan, Y., Hsieh, C.-Y., Liao, B., Zhang, S.: Retroformer: Pushing the limits of end-to-end retrosynthesis transformer. In: International Conference on Machine Learning, pp. 22475–22490 (2022). PMLR

    2022

  39. [47]

    Journal of Chemical Information and Modeling 61(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 Modeling 61(1), 123–133 (2021)

    2021

  40. [48]

    Berichte der deutschen chemischen Gesellschaft17(2), 2756–2767 (1884)

    Paal, C.: Ueber die derivate des acetophenonacetessigesters und des acetony- lacetessigesters. Berichte der deutschen chemischen Gesellschaft17(2), 2756–2767 (1884)

  41. [49]

    Berichte der deutschen chemischen Gesellschaft 17(2), 2863–2870 (1884)

    Knorr, L.: Synthese von furfuranderivaten aus dem diacetbernsteins¨ aureester. Berichte der deutschen chemischen Gesellschaft 17(2), 2863–2870 (1884)

  42. [50]

    Organic Syntheses 52, 124–124 (2003) 25

    Borch, R.F.: Reductive amination with sodium cyanoborohydride: N, n- dimethylcyclohexylamine: Cyclohexanamine, 4, 4-dimethyl-. Organic Syntheses 52, 124–124 (2003) 25

    2003

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.