acceptodds
Under review as a conference paper at ICLR 2027

Navigating Route Latent Space for Synthesizable Molecular Design

10 views

Abstract

Goal-directed molecular design has advanced rapidly, yet a substantial proportion of designed molecules remain difficult to synthesize in practice, limiting their real-world utility. Prior synthesizability-aware methods either project generated molecules back to synthesizable analogs that deviate from the intended target, or optimize directly in discrete synthesis spaces that lack a continuous landscape for efficient search. We argue that this limitation mainly comes from the search space rather than the optimizer. To address this, we propose **RouteFlow**, a framework that reformulates synthesizable molecular design as a search over a continuous route latent space, where each latent maps back to a complete synthesis route and synthesizability is inherently preserved. To navigate this space, we adopt reward-guided flow matching as an efficient sampler that steers toward high-property regions. Since reward optimization may push latents off the manifold of real synthesis routes, where decoding becomes unreliable, we further introduce a cycle-consistency mechanism to stabilize fine-tuning. Across 16 optimization tasks from Therapeutic Data Commons, **RouteFlow** achieves the best sample efficiency among synthesizability-aware baselines, with the best synthetic accessibility and the highest retrosynthesis success rate. Our results also confirm that the proposed cycle-consistency reliably keeps optimization on-manifold while improving target properties, supporting effective synthesizable molecular discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Synthesizable Molecular Design, Route Latent Space, Reward-Guided Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_QUjoOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

I think this is a solid and well-motivated work. The paper identifies two key limitations of existing approaches: (1) discrete synthesis-based optimization follows realistic chemical reaction constraints but can suffer from limited search efficiency in large combinatorial spaces; and (2) continuous molecular optimization can generate promising molecules, but the resulting candidates may not be directly synthesizable and often require additional searches for synthesizable alternatives.

To address these limitations, the authors propose a continuous route latent space that enables optimization over synthesis routes. They also introduce a cycle-consistency mechanism to discourage exploration of unreliable regions in the latent space and improve decoding stability.

The proposed method achieves strong overall performance across molecular optimization benchmarks, with state-of-the-art results on several synthesizability and molecular quality metrics. The paper also provides meaningful ablation studies that generally support its main arguments.

Critical strengths

1.Well-motivated and coherent methodological design.

The paper introduces an interesting perspective on synthesizable molecular optimization by moving the search from discrete synthesis routes to a continuous route latent space. The proposed representation and chemistry-constrained decoder allow latent variables to be mapped into executable synthesis routes. The authors also recognize the potential issue of unreliable latent-space decoding and introduce cycle consistency to mitigate it. Overall, the framework is reasonably designed, and its components are well connected to the central motivation.

2.Practical relevance and strong empirical performance.

The work addresses a meaningful gap between discrete synthesis-based optimization and continuous molecular optimization. While the former naturally incorporates chemical synthesis constraints but faces challenges in search efficiency, the latter enables flexible optimization but does not necessarily ensure synthesizability. RouteFlow attempts to combine the advantages of both approaches and achieves strong results across 16 optimization tasks. I also appreciate the search-space comparisons in Table 5 and the ablation studies in Figure 4 and Table 6, which provide useful evidence for the effectiveness of the proposed design.

Critical weaknesses
  1. Limited isolation of the benefits of continuous route latent optimization.

The introduction emphasizes the limitations of existing discrete synthesis-based optimization methods. However, the experiments do not fully isolate the advantages of continuous route latent optimization over discrete route-based search. Although Table 5 provides a useful comparison between molecular and route latent spaces under different optimizers, a more controlled comparison with discrete optimization using the same structured route representation would help better establish the specific contribution of the proposed continuous formulation.

  1. Insufficient comparison of computational efficiency.

Although RouteFlow demonstrates strong performance in terms of oracle efficiency, the framework introduces additional computational costs through the encoder, decoder, and pretraining stages. The paper reports computational costs for individual components, but a direct quantitative comparison of end-to-end computational efficiency with existing synthesis-constrained methods is still limited. Such a comparison would help clarify the practical efficiency gains of the proposed approach.

Overall, despite these limitations, I believe the proposed framework is well motivated, technically sound, and supported by strong empirical results. The remaining concerns do not substantially affect my positive assessment of the paper.

Questions for authors
  1. Multiple distinct synthesis routes may lead to the same final molecule. How does the proposed route latent space represent this many-to-one relationship? Although the replay buffer includes molecular similarity-based deduplication, could route multiplicity still affect the learned latent geometry or introduce sampling biases?

  2. Could the authors provide a more controlled comparison between continuous route latent optimization and discrete reward-guided optimization using the same structured route representation, route data, reaction templates, building blocks, and oracle budgets? This would help clarify the independent benefits of continuous latent-space optimization.

6 comments

We thank the reviewer for the positive assessment of our work.

W1:

Thank you for suggesting this comparison. We additionally evaluated discrete Bayesian optimization directly over the 300K training-route library on the three docking tasks. RouteFlow's continuous route latent space consistently outperforms discrete optimization over this library.

MethodSearch spaceEGFRDRD3A2ARAverage
RouteFlowRoute latent space (continuous)0.950 ± 0.0100.944 ± 0.0150.960 ± 0.0070.951 ± 0.010
Discrete BO (300K)300K library (discrete)0.929 ± 0.0210.923 ± 0.0190.932 ± 0.0200.928 ± 0.019

Q1:

Thank you for the question. We use a replay buffer to manage multiple routes leading to the same molecule. At each iteration, 80% of the samples used to update the flow model are freshly drawn. This design intentionally trades off some route multiplicity to improve molecular properties during optimization.

W2:

For the concern about computational cost, routeFlow requires autoencoder pretraining, which adds an upfront cost compared with GFlowNet-based methods. Its total computational cost still remains lower than the pretraining costs of ReaSyn and SynFormer. The tables below summarize total compute cost across all 16 tasks and oracle budget across methods.

MethodTrained parametersCompute (all 16 tasks, including pretraining)
SynthesisNet492.1M~6 GPU-h
ReaSyn342.1M~1,900 A100 GPU-h
SynFormer~100M~1,200 A100 GPU-h
TRACER44.2M~32 GPU-h
SynFlowNet3.2M~2.1 GPU-h
RxnFlow3.2M~1.3 GPU-h
RouteFlow (ours)21.5M*~76 GPU-h (RTX 6000)*
MethodBatch per roundIterationsTotal oracle calls
ReaSyn10010010,000
SynFormer10010010,000
SynFlowNet64157≈10,000
RxnFlow64157≈10,000
SynthesisNet10010010,000
TRACER10010010,000

Thanks again for your time and thoughtful review.

Thank you for the detailed response and the additional experiments. I appreciate the authors' efforts, especially the new discrete BO comparison and the analysis of molecular novelty.

After reading the rebuttal, I have a better understanding of what RouteFlow actually demonstrates. Initially, I thought that learning a continuous route latent space would also help the model explore more complex, multi-step synthesis routes.

However, the new results on route lengths made me reconsider this. Around 81.4% of the training routes are one-step routes, and more than 92% of the top-100 molecules in each of the three docking tasks also come from one-step routes.

I understand that shorter routes are useful in practice, and I don't think this is necessarily a problem with the method itself. But I still feel that the experiments haven't fully shown the value of learning a continuous latent space for complete synthesis routes, especially when it comes to more complex routes.

So my current view is that RouteFlow is a solid framework for finding high-scoring molecules with computationally executable synthesis routes. The additional experiments support this, and I still think the work has good practical value. However, I am less convinced that some of the broader advantages of the proposed route latent space have been fully demonstrated.

For these reasons, I have adjusted my rating from 4 to 3, while remaining positive about the paper.

Thank you for the clarification. We agree that the current results, dominated by one-step routes, do not yet establish an advantage for multi-step synthesis. We think that increasing the sampling of multi-step synthesis routes during offline data generation could help balance the route length distribution in the training data and potentially improve the model's ability to explore more complex synthesis routes. We will elaborate on this potential direction in the discussion section of the revised manuscript.

Thank you again for your continued engagement.

Reviewer_3YcaOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses a worthwhile problem by encoding synthesis routes in a continuous latent space and using flow matching to optimize molecular properties through that space. The motivation is clear: synthesizability is built into the search procedure rather than imposed afterwards by projecting generated molecules onto a synthesizable analog. RouteFlow represents each complete route as a token sequence, learns a 256-dimensional latent route space, samples from it with a reward-weighted flow model, decodes the samples into routes, and uses a cycle-aware sample weight to reduce the influence of poorly reconstructed samples. The benchmark coverage is extensive, and the implementation details are presented with reasonable completeness. At this stage, however, the evidence does not yet establish the paper’s central claims. The reconstruction comparison, the cycle-consistency analysis, and the case studies all have important gaps in experimental design or support. Some statements in the abstract and main text also claim more than the reported results support.

Critical strengths

A route-level latent space is a sensible and worthwhile design choice. Compared with searching directly over a discrete route-action space, it provides a natural interface for continuous optimization. The authors combine stack-based route serialization with building-block feasibility and reaction-template constraints, and the overall design is reasonably complete. Table 5, which crosses search space with optimizer, is the most informative experiment in the paper. It begins to disentangle the effect of the route space from that of the optimizer.

Critical weaknesses
  1. The reconstruction experiment does not isolate the source of the reported difference. A route latent is given the building blocks, reaction templates, and their order, whereas the molecule autoencoder receives only the final product. The two validity measures also assess different things. The SELFIES comparison is useful, but it rules out only invalid syntax; it does not match information content, model capacity, or training effort. I therefore cannot reconcile the claims that every latent decodes to a complete route and that structural validity is guaranteed with the route validity of 0.811 to 0.867 in Table 2 and the 5.2% decoding failure rate reported in Section 4.1.
  2. The cycle term is described as a penalty, but the equation applies it as an exponential sample weight. This distinction matters: the mechanism does not use gradients to push a latent back toward the route manifold; it only assigns lower weight to samples with high error. Figure 4 also appears inconsistent with the intended interpretation. A stronger beta=10 setting should increase cycle consistency, yet the plotted re-decode similarity is lower than at beta=0 or beta=1. The paper attributes this to duplicate molecules in the top-100 pool, but reports no unique-molecule counts, clustering results, or diversity values to support that explanation. Seed-level uncertainty and significance tests are also absent.
Questions for authors
  1. What counts as a valid route, and how are reconstruction and similarity computed? The response should specify the denominator, how decoding failures are handled, and the results when the 5.2% failed decodes are not resampled.

  2. Were the SMILES and SELFIES autoencoders matched to the route autoencoder in parameter count, training data, training time, early stopping, and hyperparameter tuning? An information-matched comparison would help separate the effect of route-level supervision from that of the constraints.

  3. Is the cycle term intended as an importance weight, a loss, or a projection? I would like a gradient-level explanation and data showing why beta=10 lowers re-decode similarity.

  4. Figure 4 should include per-seed values, a definition of the shaded intervals, and counts of unique molecules and clusters. It should also be clear whether the reported 12% is based on final scores or trajectory AUC. A relative change in negative kcal/mol values is ambiguous.

  5. The route grammar needs a more precise specification, including reaction arity, legal POP transitions, handling of START, canonicalization of equivalent trees, duplicate routes, padding, and near-duplicate splits. Comparisons with flat reaction sequences, DAG encoders, or graph encoders at matched capacity would help justify the chosen representation.

  6. How were the case studies selected? The paper should provide complete molecular structures and route tokens, building-block identifiers, reaction conditions, and a manual chemistry check. Does AiZynthFinder validate the proposed forward route, or only the final product? At a minimum, I would expect a median case, a failure case, and a three-to-five-step route.

  7. The abstract currently reads as though SA and retrosynthesis success were best on all 16 tasks, while Table 4 evaluates only the docking tasks. Is that the intended scope? If so, the abstract should be narrowed; otherwise, the corresponding results for the other tasks should be added.

  8. How were routes split across training, validation, and test at the level of products, scaffolds, reaction sets, and complete routes?

5 comments

We appreciate the thoughtful evaluation of our manuscript. We address each point below.

W1:

We acknowledge that route and molecule autoencoders receive different forms of information. Our main point here is to assess the reliability of the proposed space for downstream optimization in spite of different input information. We also appreciate raising concern about the statement that “every latent decodes to a complete route.” We will qualify this claim in the revised manuscript to reflect the limitations more precisely.

Q1:

An invalid route can occur when the model selects the stop action prematurely, leaving at least two molecules on the stack. Reconstruction and similarity are evaluated by comparing the molecule reconstructed by AE with the product of the route. When decoding fails during sampling, the flow model will draw replacement samples again. These previous failed routes do not require oracle evaluations and therefore do not consume the oracle budget. This keeps the total oracle budget consistent with that of the baselines.

W2 & Q3:

The cycle term serves as a sample reweighting mechanism to guide navigation in the route latent space. The coefficient β controls the contribution of the cycle penalty to sample weighting within each batch, and cycle_error is standardized within the batch before weighting. Setting β = 10 scales its contribution to the exponent by a factor of 10, which can cause the resulting weights to concentrate heavily on a single sample. Each update is then effectively driven by a single sample, leading to high variance in the update direction and potentially pushing the flow model into poorly behaved regions of the latent space. It turns out that some latent points in these regions cannot be decoded into valid routes during optimization. Such failures are assigned a Tanimoto similarity of 0.0 when computing the average. The shaded bands in Figure 4 show the range across random seeds, and this instability explains the wide bands observed for β = 10.

Q2:

For the reconstruction comparison, we used the same model architecture, training, hyperparameter-tuning procedures, and molecules for the different AE. RouteFlow exclusively uses the building-block and reaction-template information associated with each molecule.

Q4:

We apologize for the lack of clarity. The shaded bands in Figure 4 represent variation across random seeds. The molecules considered in this experiment are already the top 100 unique molecules for each task. The final docking scores reported below are raw affinity scores, where lower values indicate stronger binding affinity. We will make these details explicit in the revised manuscript.

TaskβSeed 1Seed 2Seed 3Mean ± SD
DRD30-10.54-10.84-10.62-10.667 ± 0.155
DRD31-11.94-11.32-12.49-11.917 ± 0.585
DRD310-11.01-10.29-11.04-10.780 ± 0.425
EGFR0-10.73-10.90-10.78-10.803 ± 0.087
EGFR1-11.34-12.25-11.34-11.643 ± 0.525
EGFR10-11.54-10.34-10.46-10.780 ± 0.661
A2AR0-11.08-11.14-10.89-11.037 ± 0.131
A2AR1-14.18-13.51-11.85-13.180 ± 1.200
A2AR10-11.92-11.06-11.60-11.527 ± 0.435

Q5:

Regarding the route grammar. The following table details the grammar rules used in our implementation.

ItemActual implementation
Token vocabularySTART, END, BRANCH, REACT, POP, RXN (115 templates), BB
Reaction arity115 templates: 13 one-reactant, 90 two-reactant, and 12 three-reactant
BRANCH semanticsEvery reactant slot must be filled from BB
REACT semanticsImplicitly takes the top molecule from the stack; the remaining slots may be filled by BB or POP
POP legalityaction == REACT, the stack is non-empty, and the stack-top molecule matches the template requirement for that reactant slot
STARTFixed at position 0
END legalityThe preceding token is in {BB, POP}
PaddingPadding positions are masked out of the loss

Regarding alternative representations. We additionally trained a DAG encoder based on a three-layer GNN with a matched parameter count and compared its reconstruction performance with our representation. The two approaches perform similarly on the high-synthesizability set. On the low-synthesizability set, however RouteFlow's representation achieves substantially better reconstruction.

EncoderHigh synth. ↑Low synth. ↑
RouteFlow (token sequence)0.8550.671
DAG encoder (GNN)0.8500.579

Q6:

The case studies were manually selected from the top 100 molecules obtained through optimization. We provide the complete generated token sequences for the three cases below. As discussed in the Discussion section, RouteFlow does not model reaction conditions or replace manual chemical validation. We also provide three additional cases as requested. AiZynthFinder evaluates the retrosynthetic accessibility of the final product and does not directly validate the specific forward route proposed by RouteFlow.

FigureTarget (PDB)Route token sequence
Fig. 5EGFR (2RGP)[BR] r92 b145035 b22521 [RE] r19 b255764 [END]
Fig. 6aDRD3 (3PBL)[BR] r114 b170429 [RE] r29 b81966 [END]
Fig. 6bA2AR (3EML)[BR] r114 b97459 [RE] r46 b119241 [END]
CaseTypeTarget (PDB)StepsDocking (kcal/mol)AiZynthRoute token sequenceWhat the tokens mean
1Median caseEGFR (2RGP)1−10.40Solved[BR] r20 b162278 b5940 [END]r20: benzimidazole formation, joining BB b162278 (chloro-isoquinoline carbaldehyde) with BB b5940 (3,4-diaminobenzenesulfonamide bearing a 2-methoxyphenyl group)
2Failure caseDRD3 (3PBL)2−10.50Unsolved[BR] r109 b183945 [RE] r33 b178222 [END]r109: alcohol-to-chloride conversion on BB b183945 (7-hydroxyquinoline-2-carboxylic acid)<br>→ r33: sulfone formation, merging that intermediate with BB b178222 (2-chloroquinoline bearing a difluoromethoxy group)
3Three-step routeA2AR (3EML)3−11.70Solved[BR] r60 b165722 b34578 [RE] r85 [RE] r19 b89094 [END]r60: CuAAC azide–alkyne cycloaddition, joining BB b165722 (methyl 6-(azidomethyl)pyridine-2-carboxylate) with BB b34578 (3-ethynylbiphenyl)<br>→ r85: ester hydrolysis of that intermediate (no building block)<br>→ r19: benzimidazole formation with BB b89094 (N-(3-methylphenyl)benzene-1,2-diamine)

Q7:

Thank you for pointing this out. SA and retrosynthesis success were evaluated on the three docking tasks, while AUC10 was evaluated across all 16 tasks. We will qualify the corresponding claim to make the evaluation scope more clear in the revised manuscript.

Q8:

We ensured that the molecules in each evaluation set do not overlap with the training molecules. We then used molecular SA scores to divide the test data by synthesizability, covering both easy- and hard-to-synthesize molecules. We also evaluated patent-derived routes from PaRoutes as an external test set. These evaluations assess the reliability of the latent space across different levels of synthesizability.

We appreciate the reviewer's detailed feedback and will use it to improve the clarity and precision of the revised manuscript.

Dear Reviewer 3Yca,

Thank you again for your detailed comments. We have provided additional analyses and clarifications regarding the reconstruction evaluation, cycle-consistency mechanism, and case studies as suggested. We have also uploaded a revised manuscript and summarized the changes in “Summary of Manuscript Revisions.” As the discussion period draws to a close, we would appreciate hearing whether any concerns remain that we could further clarify. Thank you for your time and consideration.

Reviewer_PQNYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RouteFlow, a generative architecture combining autoencoders and flow matching to perform molecule optimization in the latent space of synthesis routes. An autoencoder is trained to encode and decode synthesis routes constructed from Enamine US catalog building blocks and 115 reaction templates used in previous work. After the autoencoder is trained, a flow-matching model is trained in the latent space of the autoencoder, and is fine-tuned to perform goal-directed molecular optimization. The authors introduce several architectural modifications to overcome problems introduced by their method: a classifier on the decoder to select tokens in the sequence to support sampling from large token spaces (i.e., many building blocks), with masking to restrict sampling to chemically feasible reaction steps; a cyclic-penalty term and W2-distance term in the fine-tuning objective encouraging the fine-tuned flow-matching model to stay on the learned latent space data manifold during fine-tuning, as well as encouraging the updated flow-matching model to stay relatively close to its pre-trained version before fine-tuning; a replay buffer to focus training on previously discovered high-reward molecules.

The authors include several experiments to evaluate their model: (1) an evaluation of the reconstruction accuracy of synthesis routes vs. autoencoder baselines trained on other representations such as SMILES and SELFIES strings; (2) an evaluation of goal-directed molecular optimization against other synthesizability-aware baselines, focusing on AUC-top 10, which characterizes sample efficiency and reward accumulated in the top-ranking molecules observed, as well as molecular properties such as SAScore, AiZynth success rate, diversity, and MW. They also perform further isolated experiments to evaluate the individual components of their model.

Critical strengths

The architecture appears to be methodologically novel and achieves strong performance on evaluated baselines, although there are significant concerns on the fairness of evaluation (see below). The authors also provide interesting experiments examining different components of the generative architecture, such as reconstruction rate and latent space geometry, as well as comparing against other latent-space optimization methods. They also report molecular properties such as SAScore, AiZynthFinder success rate, diversity, and molecular weight, which help to verify that the generated molecules have reasonably high quality.

Critical weaknesses

I see this work as a promising direction and methodologically interesting, but I have several concerns to be addressed before recommending this work for acceptance:

(W1) The authors attribute the reconstruction fidelity in Table 2 to the choice of representation (synthesis routes vs. SMILES vs. SELFIES). However, it appears that the route AE differs from the SMILES/SELFIES AEs beyond simply choice of token representation. Namely, the route decoder executes each reaction with RDKit to update the state of the decoded sequence, and uses the current state to mask out chemically infeasible building blocks. Thus, it's hard to separate the performance in Table 2 into contributions from the choice of representation as the authors claim vs. other architectural design choices.

(W2) The comparison to other baselines seems to be insufficient. Importantly, the authors do not give details on what building blocks and reaction templates are used for each baseline, which can have a dramatic impact on optimization performance as well as AiZynthFinder success rate. This could also explain the poor AiZynthFinder success rate of certain methods such as SynFlowNet and RxnFlow. The authors claim that "search space determines synthesizability" (line 367), but models trained on the same search space appear to suffer from poor synthesizability. Additionally, the authors do not report specific details on how baselines are trained and evaluated. It would also be helpful to include [1], a strong baseline on the PMO benchmark.

(W3) The autoencoder is trained on a library of 300k enumerated routes, and the flow is pretrained on their latent-space representations. It is therefore unclear whether fine-tuning discovers molecules meaningfully beyond this library, or mainly recovers high-reward members within it.

(W4) It is not clear which architectural components have been proposed in previous work and thus hard to evaluate novelty. For example, the fine-tuning objective appears to be taken directly from ORW-CFM-W2. Several components in the decoder seem to resemble SynNet [2] (not cited), such as masked softmax classification over invalid actions. DoG-AE (Bradshaw et al., 2020) also constructs a latent space over synthesis routes, though it is not used for optimization; the authors cite DoG-Gen as "generator fine-tuning", but do not highlight DoG-AE presented in the same work. It would be helpful if the authors explicitly stated their contributions relative to this prior work.

[1] A Genetic Algorithm for Navigating Synthesizable Molecular Spaces, Lo et al. (2025) [2] Amortized Tree Generation for Bottom-Up Synthesis Planning and Synthesizable Molecular Design, Gao et al. (2022)

Questions for authors

Relating to W1: Can the authors perform an ablation on the route AE in which the source softmax is unmasked (scoring all building blocks), holding everything else fixed? If the reconstruction rate remains high, this would further support the authors' claim that the choice of representation is the main component contributing to reconstruction.

Figure 3 evaluates continuity on only the high-synthesizability hold-out set. Can the authors report the results on the low-synthesizability and/or the PaRoutes sets? Additionally, the off-manifold control uses random directions, which provide a weak baseline to compare against. Could the authors repeat the analysis along directions taken from a fine-tuning run with β=0\beta=0 (i.e., no cycle penalty)? This would show how the cycle penalty term behaves in regions of the latent space that are actually reached during optimization, rather than those chosen randomly.

Relating to W2: Can the authors provide further clarification on how baselines are evaluated? Importantly, are all template-based methods using the same building blocks and reaction sets? How many iterations of training are performed for each baseline? The authors train RouteFlow using 200 iterations of 50 observations on the tasks with 10k oracle calls, and 20 iterations of 50 observations on the docking tasks with 1k oracle calls. It would be helpful to know the corresponding batch sizes and training iterations for each baseline. Can the authors clarify why they chose to initialize projection-based methods such as SynthesisNet, SynFormer, and ReaSyn on ZINC250k as opposed to the same molecules corresponding to the 300k routes used to train RouteFlow?

Relating to W3: (1) How does the RouteFlow compare to simply virtual screening the 300K enumerated library? It would be helpful to run Bayesian optimization over the discrete enumerated set, as this would establish whether the reported performance comes from the continuous latent-space optimization as the authors suggest. (2) Could the authors report nearest-neighbor Tanimoto similarity between top-100 discovered molecules and the training set as well as scaffold count to quantify whether the framework is able to generate novel molecules beyond the training set? (3) All three case studies are performed with 2-step synthesis routes, while the training routes allow up to 5 steps. Could the authors report the distribution of route lengths among the top-100 discovered molecules? If the optimization concentrates on short routes, this would reflect how much of the enumerated space is reached, and give context on the AiZynthFinder comparison, as shorter routes would be easier to solve.

7 comments

We thank the reviewer for the constructive feedback and address each point below.

W1 & Q1:

We further evaluated the effect of removing reaction-template constraints when generating building-block tokens at inference time. Masking provides a modest improvement on the high-synthesizability set, but substantially improves reconstruction on the low-synthesizability set. These results support the role of masking in improving the fidelity of the route latent space.

SplitSource softmaxExact routeProduct matchValidity
High synth.Masked0.85030.85520.8668
High synth.Unmasked (softmax over 300K)0.83930.84400.8513
Low synth.Masked0.66470.67050.8111
Low synth.Unmasked (softmax over 300K)0.55100.55470.5689

W2 & Q2:

Regarding reaction templates and building blocks. GFlowNet-based methods such as RxnFlow and SynFlowNet do not natively support three-reactant reactions. We therefore implemented baselines using the reaction templates and building blocks specified in their original papers. Following ReaSyn and SynFormer, RouteFlow uses 115 reaction templates and building-block library. The batch sizes, iteration counts, and total oracle calls for the baselines are listed below.

MethodBatch per roundIterationsTotal oracle calls
ReaSyn10010010,000
SynFormer10010010,000
SynFlowNet64157≈10,000
RxnFlow64157≈10,000
SynthesisNet10010010,000
TRACER10010010,000

Regarding the initial molecules. We evaluated the effect of changing the initial molecules for ReaSyn, the strongest projection-based baseline. We found the algorithm to be relatively stable, no significant effect from the choice of initial molecules.

TaskInitial moleculesAUC10
DRD2ZINC250K0.978 ± 0.005
DRD2300K route products0.942 ± 0.021
JNK3ZINC250K0.701 ± 0.008
JNK3300K route products0.712 ± 0.053
GSK3BZINC250K0.831 ± 0.133
GSK3B300K route products0.837 ± 0.170

Regarding SynGA. Thank you for suggesting this additional baseline. We evaluated SynGA on all 16 tasks and compared the quality of the resulting molecules as shown below. Overall, RouteFlow still achieves better optimization performance, synthesizability and molecular diversity than SynGA.

TaskSynGARouteFlow
JNK30.757 ± 0.0070.726 ± 0.007
DRD20.874 ± 0.0380.987 ± 0.009
GSK3 beta0.840 ± 0.0110.976 ± 0.015
amlodipine mpo0.549 ± 0.0440.831 ± 0.009
fexofenadine mpo0.763 ± 0.0200.758 ± 0.058
osimertinib mpo0.788 ± 0.0070.837 ± 0.009
perindopril mpo0.582 ± 0.0030.711 ± 0.001
ranolazine mpo0.795 ± 0.0240.801 ± 0.011
sitagliptin mpo0.386 ± 0.0340.468 ± 0.062
zaleplon mpo0.474 ± 0.0090.544 ± 0.029
median 10.335 ± 0.0010.310 ± 0.021
median 20.338 ± 0.0040.301 ± 0.055
celecoxib redisc0.753 ± 0.0260.806 ± 0.024
DRD30.810 ± 0.0130.956 ± 0.004
EGFR0.815 ± 0.0010.961 ± 0.006
A2AR0.817 ± 0.0090.968 ± 0.009
Avg (16)0.6670.746
MethodSA ↓AiZynth ↑Diversity ↑MW ↓
SynGA3.017 ± 0.0450.348 ± 0.0490.700 ± 0.092489.4 ± 6.3
RouteFlow (paper)2.340 ± 0.0630.908 ± 0.0490.792 ± 0.021395.3 ± 7.6

Regarding the claim that “search space determines synthesizability.” Our claim focus on optimization within a continuous route latent space. Even when methods such as SynFlowNet or ReaSyn use the same building blocks and reaction templates, they optimize in discrete spaces using GFlowNets or genetic algorithms. These methods therefore do not directly address the specific claim we make.

Regarding reconstruction fidelity. Following your suggestion, we extended the analysis to the low-synthesizability set and PaRoutes to examine whether the same manifold behavior holds. We evaluated cycle-consistency error and similarity to the starting molecule at five interpolation points(on/off manifold). The overall trends are consistent with those observed on the high-synthesizability set in Sec 4.2.

MetricλLow synth. (on)Low synth. (off)PaRoutes (on)PaRoutes (off)
ecyce_{\mathrm{cyc}}0.022.6 ± 3.922.6 ± 3.922.1 ± 3.022.1 ± 3.0
ecyce_{\mathrm{cyc}}0.233.2 ± 4.035.8 ± 6.334.4 ± 4.329.6 ± 3.3
ecyce_{\mathrm{cyc}}0.575.2 ± 4.097.1 ± 7.952.6 ± 3.975.2 ± 6.2
ecyce_{\mathrm{cyc}}0.841.0 ± 7.1215.4 ± 11.935.0 ± 4.2157.7 ± 12.3
ecyce_{\mathrm{cyc}}1.027.9 ± 6.3327.1 ± 18.722.5 ± 3.5221.1 ± 17.2
Sim. to start0.00.919 ± 0.0700.919 ± 0.0700.779 ± 0.0890.779 ± 0.089
Sim. to start0.20.855 ± 0.0860.907 ± 0.0760.555 ± 0.1020.726 ± 0.095
Sim. to start0.50.221 ± 0.0500.847 ± 0.0880.161 ± 0.0310.483 ± 0.098
Sim. to start0.80.108 ± 0.0100.627 ± 0.1040.140 ± 0.0250.237 ± 0.059
Sim. to start1.00.103 ± 0.0110.416 ± 0.0960.136 ± 0.0240.177 ± 0.025

Regarding the evolution of cycle error without the cycle penalty. We tracked this error during optimization with the JNK3 oracle comparing β = 0 and β = 1. The cycle penalty substantially reduces the cycle error during optimization.

MetricOracle callsβ = 0β = 1
ecyce_{\mathrm{cyc}}1000147.2 ± 26.2118.8 ± 81.9
ecyce_{\mathrm{cyc}}2500495.8 ± 220.7353.0 ± 5.1
ecyce_{\mathrm{cyc}}50001878.0 ± 1426.5679.1 ± 419.4
ecyce_{\mathrm{cyc}}75002220.3 ± 1816.0540.0 ± 215.4
ecyce_{\mathrm{cyc}}100001741.9 ± 1267.5522.4 ± 179.9

W3 & Q3:

Regarding random search and discrete Bayesian optimization. We additionally ran random search and discrete Bayesian optimization over the 300K library on the three docking tasks using the same oracle budget. The results further support the benefit of optimization in the continuous route latent space below.

MethodSearch spaceEGFRDRD3A2AR
RouteFlowRoute latent space0.950 ± 0.0100.944 ± 0.0150.960 ± 0.007
Discrete BO (300K)300K library0.929 ± 0.0210.923 ± 0.0190.932 ± 0.020
Discrete random (300K)300K library0.888 ± 0.0280.874 ± 0.0370.897 ± 0.022

Regarding similarity to the 300K training molecules. For each of the three docking tasks, we compared the top 100 optimized molecules with the training molecules using Tanimoto similarity and counted new Murcko scaffolds. As shown below, only a small fraction of the optimized molecules are close analogs of training molecules, and many contain new scaffolds. These results support the ability of optimization to generalize beyond the training molecules.

Target% ≥ 0.7 (analog)New Murcko scaffolds / 100
EGFR8.0%70.0 ± 6.6
DRD34.7%66.0 ± 16.5
A2AR7.3%56.7 ± 20.0
All6.7%64.2

Regarding the distribution of generated route lengths. Thank you for raising this point. We found that most optimized molecules are associated with one-step routes. Examining the training data revealed a similar concentration of one-step routes. A likely explanation is that one-step reaction sequences have a high success rate during offline sampling of training routes, leading to their overrepresentation in the training data. Shorter routes can be attractive for experimental synthesis since they involve fewer operations and may reduce opportunities for side reactions. We think controlling the proportions of different route lengths during offline sampling is a promising way to improve the diversity of routes and molecules generated by the model. We appreciate for pointing this out.

Sourcen1-step2-step3-step≥4-step
300K enumerated training routes300,00081.4%14.9%3.6%0.0%
RouteFlow top-100 — EGFR30092.3%7.3%0.3%0.0%
RouteFlow top-100 — DRD330098.7%1.3%0.0%0.0%
RouteFlow top-100 — A2AR30097.3%2.7%0.0%0.0%

W4:

Thank you for bringing SynNet to our attention. We will cite and discuss it in the revised manuscript. The main contribution of RouteFlow is a route latent space that provides a synthesizability-aware foundation for optimization without restricting the choice of optimizer. ORW-CFM-W2 is the best-performing optimization strategy we identified.

Regarding SynNet. SynNet trains four separate MLPs to make decisions in different states. In contrast, our special-token design handles different action types through a unified softmax. SynNet also uses a genetic algorithm to optimize through discrete mutations, whereas RouteFlow performs optimization in a continuous latent space.

Regarding DoG-AE. Although DoG-AE also provides a latent space, it directly optimizes the latent-space generator itself by fine-tuning. Our approach instead provides a general route latent space into which different optimizers can be incorporated, providing a more flexible framework for different optimizers.

We appreciate the reviewer's constructive suggestions and will incorporate these additional analyses and clarifications into the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Manuscript Revisions

Dear Area Chair and Reviewers,

Thank you for your constructive feedback. We have revised the manuscript to incorporate the additional experiments and clarifications discussed in our responses. The main revisions addressing the reviewers' concerns are summarized below:

  1. Baseline coverage and evaluation transparency. To strengthen the comparison as suggested, we added SynGA results across all 16 optimization tasks, molecular-quality comparisons on the docking tasks(Tables 3–4). To clarify differences in experimental settings, we also document baseline reaction spaces and PMO optimization schedules, and report initialization sensitivity (Appendices A.3 and D.1).

  2. Optimization beyond the training library. To examine whether RouteFlow discovers molecules beyond the enumerated training products, the new Section 4.5 compares it with discrete Bayesian optimization and random search over the 300K library under the same 1K oracle-call budget. RouteFlow achieves higher AUC₁₀ on all three docking tasks. Nearest-neighbor similarity and new Murcko scaffold counts provide additional evidence of structural novelty relative to the training products (Table 6).

  3. Latent-space behavior and cycle consistency. We added evaluations on low-synthesizability routes and PaRoutes, together with cycle-error trajectories during optimization with and without the cycle penalty (Appendices D.2–D.3). We also revised the explanation of excessive cycle weighting and clarified how decoding failures contribute to the reported re-decode similarity.

  4. Route lengths and interpretation of the retrosynthesis results. To address questions about route lengths and their implications for the AiZynthFinder comparison, Appendix D.4 compares retrosynthetic route-length distributions across methods under a common planning configuration. Among successfully solved molecules, RouteFlow's median route length is two steps, matching five of the seven baselines, and its mean is comparable to those of SynFormer and SynthesisNet. This analysis provides additional context for interpreting the retrosynthesis results.

We also expanded the related-work discussion and narrowed the described synthetic-accessibility and retrosynthesis-success claims to the docking tasks as suggested.

We would appreciate any further feedback on our revisions and responses. Thank you again for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.