RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis

We would like to thank the reviewer for the time and effort they spent on reviewing our work. We are glad that the reviewer found our work to be conceptually clean, technically sound, and our use of synthons to be well motivated. We now address the main questions raised by the reviewer.

Benchmarking on other datasets

While the gains in exact‑match accuracy are impressive, the experimental study is limited to USPTO‑50k; it remains unclear how the approach scales to larger, noisier corpora or to multi‑product reactions.

We thank the reviewer for their suggestion on testing on the USPTO Full and other datasets. We believe exploring how our methods scale to datasets such as USPTO Full is an interesting direction for future work. In particular, we follow the standard experimental practice in many prior recent works, including other diffusion-based approaches and synthon-based approaches, which solely report retrosynthesis results on the USPTO-50k dataset [2-8]. Furthermore, we have conducted comprehensive ablations on USPTO-50k, and we argue that our results are comprehensive in empirical caliber. Nevertheless, we will include a discussion on these points as part of our conclusion in our updated draft.

GFlowNet vs Flow Matching

The claim of being the first “flow‑matching framework for retrosynthesis” would benefit from a clearer differentiation from the recent GFlowNet.

We appreciate the reviewer’s observation and the opportunity to clarify how our approach differs from GFlowNets. GFlowNets and flow matching are fundamentally distinct frameworks. GFlowNets aim to sample trajectories whose probabilities are proportional to a given reward function. They model how probability mass flows through a directed acyclic graph (DAG) of partial trajectories. In contrast, flow matching models a continuous-time Markov chain that transports mass between arbitrary source and target distributions. Unlike GFlowNets, flow matching does not rely on a reward function to learn this mapping. Given these conceptual differences, we believe it is appropriate to state that our work is the first to apply flow matching to retrosynthesis.

The work inherits a strong dependency on a separately trained reaction‑centre predictor whose errors are not quantified; the end‑to‑end robustness of the pipeline is therefore difficult to assess.

Although the two-stage framework relies on identifying synthons, this is not a problem in practice since reaction-center prediction can be performed with high accuracy. An off-the-shelf reaction center prediction model achieves a top-2 and top-3 accuracy of 85.1% and 89.5% respectively. The more challenging aspect of retrosynthesis is synthon completion [9], which is the main aspect addressed by our work.

Compute time Comparison

The following table provides the average inference time to sample 100 reactants for a given product in the USPTO-50k test set. We benchmark on an RTX 3090 with 24 GB of memory.

We note that 30 seconds for sampling a set of 100 reactants for a given product is a completely feasible time for applications of models like ours. In particular, this speed does not prevent the use of our method as a component of a multistep retrosynthesis planning pipeline. Additionally, the sampling time for RetroBridge with T=500 steps and 100 reactants is ~50 seconds.

Concluding remarks

We thank the reviewer again for their valuable feedback. We hope that our rebuttal addresses their questions and concerns, and we kindly ask the reviewer to consider a fresher evaluation of our paper if the reviewer is satisfied with our responses. We are also more than happy to answer any further questions that arise.

References

[1] Dai et al. “Retrosynthesis Prediction with Conditional Graph Logic Network” Neurips 2019.

[2] Laabid et al. “Equivariant Denoisers Cannot Copy Graphs: Align Your Graph Diffusion Models”, ICLR 2025.

[3] Igashov et al. “RetroBridge: Modeling Retrosynthesis with Markov Bridges”, ICLR 2024 [4] Somnath et al. “Learning Graph Models for Retrosynthesis Prediction” NeurIPS 2021.

[5] Shi et al. “A Graph to Graphs Framework for Retrosynthesis Prediction”, ICML 2021.

[6] Zhong et al. “Retrosynthesis prediction using an end-to-end graph generative architecture for molecular graph editing” 2023.

[7] Gianski et al. “RetroGFN: Diverse and Feasible Retrosynthesis using GFlowNets”, 2024.

[8] Chen and Jung. “Deep Retrosynthetic Reaction Prediction using Local Reactivity and Global Attention.” 2021

[9] Smith. “Organic Synthesis”, 2017.

We thank the reviewer for the time and effort spent reading and engaging with our work. We are glad that the reviewer found our work to be a novel application of discrete flow matching, demonstrating strong results with comprehensive evaluations and ablations. We now address the reviewer's questions.

Novelty of RetroSynFlow

We acknowledge the reviewers' concern that the core novelty of RetroSynFlow may not be initially apparent. We would like to politely push back against this assertion, as discrete flow matching is not an established approach for single-step retrosynthesis. Consequently, adapting the discrete flow matching framework—whose primary success has been with masked/absorbing states—to a setting where the prior is a set of reactants/synthons is a simple yet effective novel design contribution of our work. Furthermore, to the best of our knowledge, the use of FK steering with a discrete flow matching model operating on synthons is also novel. In particular, our proposed use of the forward-synthesis model as a reward for inference-time steering is a novel insight that is specifically tailored to the single-step retrosynthesis setup. In contrast, prior work like RetroGFN [7] had only considered reward models during training. We argue that these design decisions lead to an overall approach whose combination of components is novel and drives the current empirical success over prior SOTA template free approaches.

RetroSynFlow vs. template-based approaches

The advantage of RETRO SYNFLOW over template-based approaches is questionable. Local Retro is competitive and, in some cases, better. The authors claim that template free approaches are more flexible and generalize better, but it is unclear if better generalization shows up in the evaluation.

We appreciate the reviewer’s nuanced feedback on the distinction between template-free and template-based approaches. Template-based methods rely on a fixed library of templates extracted from datasets, which encode only the transformation rules observed in those datasets. Thus, these methods introduce a strong inductive bias, highlighted by the strong performance of LocalRetro, which outperforms many template-free baselines. However, this reliance on predefined templates limits the chemical space these methods can explore, making it difficult for them to generalize to unseen reactions [14–16]. For instance, GLN achieves top-1 accuracy of 52.5% on USPTO-50k but drops to 39.5% on USPTO-Full. Also, extracting reaction templates can be expensive, either requiring expert chemists to handcraft rules or relying on computationally intensive subgraph matching algorithms [17]. Automatically extracted template sets may contain errors, as they are often less curated and validated [18]. Finally, we note that RetroSynFlow outperforms existing template-free baselines, further demonstrating its effectiveness.

Use of round-trip accuracy

The round trip metrics seem flawed because the method directly optimizes for them, whereas others do not. We can see this in the Figure 5 ablations, where round trip metrics increase with the number of particles, but accuracy does not. Round trip metrics appear to be proxies for quality, but do not have obvious direct applications.

We use round-trip accuracy to address the limitations of exact-match accuracy, which overlooks the fact that multiple valid reactant sets can produce the same product but may not appear in the dataset. Top-$k$ round-trip accuracy captures this by measuring the proportion of feasible reactants (via a forward model) among the top-$k$ predictions. Importantly, it can be optimized independently of top-$k$ accuracy. Our reward-steered methods improve round-trip accuracy by generating diverse, feasible reactant sets while maintaining competitive top-$k$ performance by including ground-truth reactants among their outputs. Even without reward steering, our model performs on par with state-of-the-art methods on round-trip metrics. Although not a perfect measure, round-trip accuracy is widely used in the literature [1–3, 7–9, 11].

Benchmarking on other datasets

We thank the reviewer for their suggestion on testing on the USPTO Full and other datasets. We believe exploring how our methods scale to datasets such as USPTO Full is an interesting direction for future work. In particular, we follow the standard experimental practice in many prior recent works, including other diffusion-based approaches and synthon-based approaches, which solely report results on the USPTO-50k dataset [1-9]. Furthermore, we have conducted comprehensive ablations on USPTO-50k, and we argue that our results are comprehensive in empirical caliber. Nevertheless, we will include a discussion on these points as part of our conclusion in our updated draft

Questions

What is # of parameters of the model? How does inference time scale with the number of particles? Please provide inference wall time versus template-based baselines, and discuss GPU memory needs for SMC.

The model has 3.38 million parameters. The following table provides the average inference time to sample 100 reactants for a given product in the USPTO-50k test set. We benchmark on an RTX 3090 with 24 GB of memory.

We note that 30 seconds for sampling a set of 100 reactants for a given product is a completely feasible time for applications of models like ours. In particular, this speed does not prevent the use of our method as a component of a multistep retrosynthesis planning pipeline. Additionally, the sampling time for RetroBridge with T=500 steps and 100 reactants is ~50 seconds.

Molecular Transformer's errors may bias results but are not quantified.

We evaluate the results of Molecular Transformer on the USPTO-50k dataset and obtain the following results.

We refer the reviewer to Appendix C, specifically Table 6 and Figures 8 and 9, for a brief discussion and visualization of the potential bias introduced by the forward-synthesis model. In particular, there are cases where the reward-steered model may generate reactants that are not feasible and do not match the ground truth, whereas the non-steered version produces a correct prediction.

To clarify, are synthons just the product graph with deleted edges? If so, why is it such a good inductive bias? How many edges are being deleted on average?

Synthons correspond to the product graph with deleted or modified edges. In the USPTO-50k dataset, the reactions have one edge deleted on average. Identifying synthons is a key step in how human chemists approach retrosynthetic analysis, as it helps them visualize how a complex molecule can be constructed from simpler, idealized fragments that point to the actual reactants required for synthesis. [12, 13] Employing synthons provides more information about the reaction center and, along with the product context, simplifies the generation task by allowing the denoiser model to potentially “localize” the changes that need to be made to transform the current noisy graph into the reactant molecule graph.

Can you report the proportion of FK steered outputs that are actually chemically valid (validated with RDKit) but misscored by the oracle? A sensitivity analysis on oracle noise would clarify reliability.

For the non-steering outputs, 25% of the chemically valid reactants were misscored by the oracle. For the FK-steered outputs, only 14% of the chemically valid reactants were misscored by the oracle.

Concluding remarks

We hope that our responses were sufficient in clarifying all the great questions asked by the reviewer. We thank the reviewer again for their time, and we politely encourage the reviewer to consider updating their score if they deem that our responses in this rebuttal, along with the new experiments, merit it.

References

[1] Wan et al. “Retroformer: Pushing the Limits of End-to-end Retrosynthesis Transformer”, ICML 2022.

[2] Laabid et al. “Equivariant Denoisers Cannot Copy Graphs: Align Your Graph Diffusion Models”, ICLR 2025.

[3] Igashov et al. “RetroBridge”, ICLR 2024

[4] Somnath et al. “Learning Graph Models for Retrosynthesis Prediction” NeurIPS 2021.

[5] Shi et al. “A Graph to Graphs Framework for Retrosynthesis Prediction”, ICML 2021.

[6] Zhong et al. “Retrosynthesis prediction using an end-to-end graph generative architecture for molecular graph editing” 2023.

[7] Gianski et al. “RetroGFN”, 2024.

[8] Chen and Jung. “Deep Retrosynthetic Reaction Prediction using Local Reactivity and Global Attention.” 2021

[9] Qiao et al. “Advancing Retrosynthesis with Retrieval-Augmented Graph Generation”. AAAI 2025.

[10] Schwaller et al. “Molecular Transformer: A Model for Uncertainty-Calibrated Chemical Reaction Prediction”. 2019

[11] Zeng et al. “Ualign: pushing the limit of template-free retrosynthesis prediction with unsupervised SMILES alignment”. 2024

[12] Corey, “General Methods for the Construction of Complex Molecules”. 1967.

[13] Warren, “Designing Organic Syntheses: A Programmed Introduction to the Synthon Approach”. 1978.

[14] Yan et al. “RetroXpert”. 2020

[15] Segler et al. “Modelling Chemical Reasoning to Predict and Invent Reactions”. 2016.

[16] Thakkar et al. “Datasets and their influence on the development of computer assisted synthesis planning tools in the pharmaceutical domain”. 2020

[17] Liu et al. “Retrosynthetic reaction prediction using neural sequence-to-sequence models”. 2017

[18] Yao et al. “Node-Aligned Graph-to-Graph: Elevating Template-free Deep Learning Approaches in Single-Step Retrosynthesis. 2024.

We would like to thank the reviewer for the time and effort they spent engaging with our work. We appreciate that the reviewer finds our application of discrete flow matching to retrosynthesis novel and the use of FK-steering to improve round-trip accuracy a strength of our work. We now address the questions and concerns.

Template-free vs semi-template

The authors repeatedly describe their methods, including RSF, as "template-free", which is somewhat misleading. I would encourage the authors to consider repositioning the model as a semi-template method.

We appreciate the reviewer’s constructive feedback on the positioning of our work as a semi-template method. Semi-template approaches aim to combine the interpretability of template-based models with the flexibility and generalization ability of template-free methods. These methods typically operate in two stages: the first stage involves predicting the reaction center to identify synthons, and the second stage, which we focus on, completes the synthons into valid reactants. While some prior works (e.g., [2]) do not explicitly differentiate between semi-template and template-free methods, the consensus in the literature is to classify such two-stage frameworks as semi-template approaches [3]. In light of this, we will revise the introduction and related work sections to clarify the characteristics of semi-template methods and more accurately position our contribution within this framework.

Using imperfect forward synthesis model as the reward oracle

The reward oracle for FK-steering is an imperfect forward synthesis model. A discussion of this limitation, and perhaps a small-scale analysis of the forward model's accuracy on a subset of the generated reactions, would add some nuance.

This is indeed a fair point. Forward-synthesis is an easier task than retrosynthesis, and correspondingly, forward models such as Molecular Transformer [1] have strong performance. However, they are still subject to false positives and false negatives, which can potentially bias the steering process towards invalid reactants and limit performance in round-trip accuracy. We will add this to the limitation sections at the end of our paper.

We evaluate the results of Molecular Transformer on the USPTO-50k dataset and obtain the following results.

We refer the reviewer to Appendix C, specifically Table 6 and Figures 8 and 9, for a brief discussion and visualization of the potential bias introduced by the forward-synthesis model. In particular, there are cases where the reward-steered model may generate reactants that are not feasible and do not match the ground truth, whereas the non-steered version produces a correct prediction.

Furthermore, for the non-steering outputs, 25% of the chemically valid reactants were misscored by the forward-model oracle. For the FK-steered outputs, only 14% of the chemically valid reactants were misscored by the oracle.

Miscellaneous

The finding in Table 3 that removing the product context improves synthon completion accuracy is counter-intuitive.

We apologize for the confusion regarding this typo. Indeed, using the product context increases accuracy, and we have fixed this typo so the columns are labeled correctly.

Concluding remarks

We hope that our responses were sufficient in clarifying all the great questions asked by the reviewer. We thank the reviewer again for their time, and we politely encourage the reviewer to consider updating their score if they deem that our responses in this rebuttal, along with the new experiments, merit it.

References

[1] Schwaller et al. “Molecular Transformer: A Model for Uncertainty-Calibrated Chemical Reaction Prediction” 2025.

[2] Igashov et al. “RetroBridge: Modeling Retrosynthesis with Markov Bridges”. 2024

[3] Somnath et al. “Learning Graph Models for Retrosynthesis Prediction”. 2021

We would like to thank the reviewer for the time and effort they spent on reviewing our work! We appreciate that the reviewer found our model formulation and use of the forward model for FK steering to be interesting. We now address the main points raised in the review.

Ablation without FK steering

We acknowledge the reviewer's comment that our ablation should include a result that does not make use of FK steering. We believe that there may be a slight confusion, which we now hope to clarify. In our main results, we report our initial model, RetroProdFlow, which directly transports product molecules to their corresponding reactants, and importantly, is devoid of any FK steering. As evidenced by Table 1 results, we find RetroProdFlow to be a performant model on par with previous template-free methods like RetroBridge [6], but short in performance compared to RetroSynFlow, which is our primary model. Consequently, we argue that the baseline that the reviewer requests is already present in the main paper.

In addition, we further argue that our proposed approach RetroSynFlow, is a combination of flow matching with the distribution of synthons serving as a strong inductive prior and FK steering to facilitate improvements in round-trip accuracy. Therefore, it is natural to consider these two components in one system rather than in isolation.

We hope that this clarification here fully alleviates the reviewers' valid concern, and we will update the wording in the main paper to highlight that RetroProdFlow does not employ FK steering.

Citations to related work

We thank the reviewer for pointing us to these related works. In our updated draft, we will include [1,2,3] in our related works section to discuss the use of a forward-synthesis model to filter reactants in retrosynthesis. In particular, a key difference between our proposed RetroSynFlow and these related works is the use of a forward-synthesis model to aid in multi-step retrosynthesis planning, while our work pertains to single-step retrosynthesis prediction. Consequently, these works do not report results on the USPTO-50k dataset, which we use in our evaluation.

With regards to related works [4] to the related works section and [5] to our baselines. The key differences between our work and [4] are that the denoiser model in [4] requires the atom-mapping as input, while our flow-matching denoiser does not. Furthermore, [4] employs absorbing state diffusion, which assumes that the source distribution is an absorbing state. In stark contrast, discrete flow matching allows us to learn a map between synthons/products and reactants directly.

With regards to Chimera [5], we argue that this is a framework that ensembles many different retrosynthesis models across both graph and SMILE string-based methods. As a result, our method is complementary to their framework as it can be employed in the ensemble. Additionally, since Chimera ensembles various models, it may potentially be more resource-intensive than our single method in RetroSynFlow. We update the main table with following new entries:

Questions

Which Codebase have the authors used to compute the metrics?

We used the evaluation code from RetroBridge [6] to compute metrics.

Concluding remarks

We thank the reviewer for their time and effort in reviewing our paper. We hope that our rebuttal here answers all the great points raised by the reviewer, allowing them to potentially upgrade their score as they initially suggested. We are also more than happy to answer any further questions that arise in the rebuttal period. Please let us know.

References

[1] Segler et al. “Planning chemical syntheses with deep neural networks and symbolic AI”, 2018.

[2] Coley et al. “A robotic platform for flow synthesis of organic compounds informed by AI planning”, 2019.

[3] Schwaller et al. “Predicting retrosynthetic pathways using transformer-based models and a hyper-graph exploration strategy”, 2020.

[4] Laabid et al. “Equivariant Denoisers Cannot Copy Graphs: Align Your Graph Diffusion Models”, 2025.

[5] Maziarz et al. “Chimera: Accurate retrosynthesis prediction by ensembling models with diverse inductive biases”, 2024.

[6] Igashov et al. “RetroBridge: Modeling Retrosynthesis with Markov Bridges”. 2024