acceptodds
Under review as a conference paper at ICLR 2027

RouteDiff: Molecular Design via Discrete Diffusion over Synthetic Routes

10 views

Abstract

Molecular generative models can propose molecules with desirable properties, but the resulting molecules are not necessarily synthesizable. Synthesis-aware molecular generation addresses this gap by choosing both a molecule and an executable route that constructs it. Existing sequence-based route generators typically commit route decisions autoregressively or refine completed pathways in separate stages. We introduce RouteDiff, a masked discrete diffusion model over complete postfix routes. A bidirectional Transformer predicts unresolved building-block and reaction decisions from the currently visible route context, while diffusion progressively determines when those decisions are revealed and committed. To scale to a 223,243-entry building-block catalog, we combine frozen molecular representations with learned low-rank identity components, and compose supported property conditions at inference time without task-specific retraining. On Enamine, RouteDiff combines a 92.8 reconstruction rate and 0.987 similarity with the highest building-block diversity (0.783) among the compared methods. Under a 10,000-oracle-call budget, it achieves the highest synthesis-constrained AUC Top-10 on eight of ten PMO tasks, with gains of up to 13.2 points. Denoising-trajectory analysis further finds pre-commit changes in unresolved building-block or reaction predictions in 36/48 analyzed successful trajectories.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
discrete diffusion, molecular generation, synthetic route generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Qz9JOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RouteDiff for synthesis-aware molecular design. By incorporating a masked diffusion model and a bidirectional transformer, the proposed method enables arbitrary generation orders, rather than the left-to-right generation scheme used in conventional methods. Additionally, the method combines ChemBERTa representations with individual embeddings to represent each building block. The method is evaluated on reconstruction and PMO benchmarks.

Critical strengths

None.

Critical weaknesses
  1. The primary motivation of this paper—the decoupling of execution and generation—is unclear. The masked diffusion modeling strategy has already been employed by GenMol, which generates permutation-invariant SAFE representations. Moreover, in reconstruction tasks, where building block composition is more important than generation order, ReaSyn outperforms RouteDiff.

  2. The proposed method does not support unseen building blocks. The Enamine stock library used in this paper is updated monthly. Various synthesis-aware methods, such as SynFormer and RxnFlow, can utilize unseen building blocks that were not observed during training. In contrast, RouteDiff relies on a library-dependent representation scheme, meaning that a user who downloads the latest Enamine library would be unable to directly use the pretrained model with newly added building blocks.

Questions for authors

Since SMARTS-based reaction templates cannot fully capture chemical reactivity and reactant compatibility, it is important to impose additional synthetic constraints during generation to ensure real-world synthesizability. While conventional left-to-right synthetic route generation schemes naturally allow such constraints to be incorporated, it is unclear how the proposed method addresses this issue. Could the authors clarify how RouteDiff can incorporate additional synthetic constraints during generation?

3 comments

We thank the reviewer for the opportunity to clarify the execution-order motivation, the catalog setting, and inference-time constraint conditioning.

W1: Motivation: decoupling execution, prediction context, and commitment.

Execution order refers to how a completed postfix route is executed—not how its tokens must be generated. For example, in [BB_A, BB_B, R_1, BB_C, R_2], the executor pushes BB_A and BB_B, applies R_1 to form an intermediate, then applies R_2 with BB_C. These stack dependencies determine the chemical meaning of the sequence. RouteDiff preserves them unchanged (Sec. 2.1).

The key observation is that execution order need not determine prediction context or commitment order. A postfix autoregressive model must commit an upstream BB before generating downstream reactions. RouteDiff instead denoises the entire masked postfix route: the bidirectional Transformer predicts still-masked positions using currently revealed tokens on either side, and the reverse diffusion process selects positions for commitment independently of their postfix indices. Thus a revealed downstream reaction may inform an unresolved upstream BB, while the completed route still executes in postfix order. Once a token is revealed it remains fixed. Our revision analysis concerns predictions before commitment (Sec. 2.2).

This is distinct from GenMol and ReaSyn. GenMol diffuses SAFE molecular-fragment representations, rather than generating complete executable synthetic routes specifying purchasable BBs, reaction-template identities, and stack dependencies. ReaSyn combines bottom-up/top-down autoregressive pathway generation with iterative subtree refinement and a separately trained Edit Bridge for whole-pathway editing. In contrast, RouteDiff performs bidirectional prediction and non-postfix commitment of masked route tokens within a single whole-route reverse-diffusion process, while preserving the same postfix execution semantics. Thus, it allows unresolved route decisions to incorporate global context before commitment within one whole-route generation.

Against published synthesis-constrained baselines, RouteDiff leads on 8/10 original PMO objectives (Table 3). In the matched AR comparison, it retains comparable Enamine reconstruction (92.8% vs. 92.6%) with higher BB diversity (0.783 vs. 0.756; Table 2). In a success-conditioned cohort, 36/48 trajectories show pre-commit BB/reaction prediction revisions (Appendix E.2).

W2: Building-block catalogs.

A specified building-block catalog and reaction-template set are fundamental to synthesis-constrained generation. Together they define which routes can be executed. Dynamic vendor-catalog expansion is a separate design dimension from the order-decoupling question. It is notable that RouteDiff operates at substantial scale, with 223,243 purchasable Enamine building blocks and 115 reaction templates, enabling exact selection among over 223k identities at BB positions.

The semantic–identity interface is designed for this scale. ChemBERTa prototypes capture chemical relationships, while learned low-rank identity components distinguish exact catalog entries (Sec. 2.3). The controlled ablation gives 92.8% Enamine reconstruction for the full interface versus 88.1% without both components (Table 2). ChemBERTa can encode new molecular entries. Updating exact catalog identities concerns the indexed input/output interface, rather than the whole-route diffusion mechanism.

Q: Incorporating additional synthetic constraints.

Experiment 3 implements multi-constraint guidance during generation, not merely post-hoc ranking. RouteDiff accepts molecular fingerprints and exact-value, upper-bound, and lower-bound scalar conditions. At every reverse diffusion step, the shared denoiser evaluates each condition on the same partially masked route, and positive-weight sums of the resulting token-type, BB, and reaction logits define the sampling distributions before positions are committed (Sec. 2.4 and Appendix C.3). Thus multiple constraints jointly guide route-token decisions without task-specific retraining.

Synthesis compatibility can be addressed through the executable route representation. The stack executor applies SMARTS reactions to the instantiated intermediates and checks whether a sampled route can be executed successfully. For further catalog/reaction restrictions, the typed BB/reaction distributions provide a direct interface for masking disallowed choices, and additional chemistry-specific validators can be connected to route acceptance. Neither mechanism requires left-to-right commitment. RouteDiff therefore preserves postfix execution semantics while incorporating molecular constraints into the reverse generation process.

Thank you for the clarifications and additional experiments. The matched AR comparison provides useful evidence supporting the proposed formulation. I acknowledge that this strengthens the empirical case for decoupling token commitment order from postfix execution order. However, two concerns remain.

Practical usability of the pretrained checkpoint and reproducibility

My concern is about the practical usability of the pretrained checkpoint. The Enamine catalog version used in this work may no longer be available for download, and redistributing it requires permission. How can the users use the pretrained checkpoint with the currently available catalog? Since the semantic–identity parameterization is presented as a contribution of this work, its implications for pretrained checkpoint reuse need to be clarified. Given the model’s reliance on catalog-specific identity components, wouldn’t users need to retrain the model on their own catalog rather than directly use the released checkpoint? Do the authors plan to release pretrained checkpoints, and if so, how would they address this issue? Relatedly, how can other researchers reproduce the reported results if they cannot obtain the catalog used in the experiments?

Enforcing additional synthetic constraints during generation

I understand that the synthesis executor can check whether a route is executable under the specified SMARTS reaction templates. My question concerns additional constraints on the reactivity of the actual reactants beyond these template-based checks, such as computational chemistry-based reactivity filters.

Using the example in your response, suppose only BB_A and R_2 have been committed. How can the remaining choices, particularly R_1, be restricted using such filters when the other building blocks and the intermediate are still unknown? These filters may require the actual reactant structures to evaluate whether a proposed reaction is feasible. In left-to-right route generation, they can be evaluated using the instantiated intermediates and candidate reactants before committing to the next action. How is this handled when tokens are committed in an order different from the postfix execution order, particularly since revealed tokens remain fixed?

Your response mentions connecting chemistry-specific validators to route acceptance. If these additional reactivity filters are applied only after route generation, how does this differ from generation followed by filtering? Please clarify whether such constraints can be enforced before token commitment or only checked afterward.

In recognition of the additional evidence from the matched AR comparison, I will raise my rating to 2 (Weak rejection). The concerns above remain unresolved.

We sincerely thank the reviewer for the detailed follow-up and for acknowledging the additional evidence from the matched-AR comparison. We address the two remaining concerns below.

1. Practical usability of the pretrained checkpoint and reproducibility.

We agree that semantic representations alone do not enable exact prediction of unseen catalog identities. ChemBERTa can encode a newly introduced building block, but the pretrained checkpoint does not contain learned identity parameters for that entry. When adapting to an updated catalog, learned identity parameters for shared BBs can be retained after identity alignment, while newly introduced identities require adaptation of the catalog-specific components. This does not necessarily require training RouteDiff from scratch. The pretrained checkpoint can serve as an initialization for such adaptation.

We plan to release the pretrained checkpoint and its reproducibility package by the camera-ready stage if accepted. Importantly, the chemical-space file used in our experiments, enamine2310_rxn115.chemspace, is already publicly available. Other researchers can therefore access the original released resources without relying on the currently available catalog from Enamine. The released checkpoint will be accompanied by instructions for using the matching chemical space. As clarified above, adapting to a different catalog requires updating the catalog-specific identity components rather than directly reusing the original checkpoint without modification.

2. Enforcing additional synthetic constraints during generation.

We appreciate the reviewer's continued input on this point. In the example where only BB_A and R_2 are committed, a reactivity filter requiring the structures of the remaining reactants or intermediates cannot be evaluated before those structures become available. Our standard sampler does not guarantee enforcement of such filters before the initial commitment of those tokens. The inference-time multi-constraint guidance acts before commitment for molecular-property conditions, while the SMARTS executor checks route executability. A chemistry-specific validator applied only to a completed route is indeed post-generation filtering. We appreciate the opportunity to make this boundary explicit.

A possible inference-time extension is validator-guided re-masking and resampling. Once the required reactant and intermediate structures become available, an external reactivity validator could identify reaction steps that violate the specified constraints. Importantly, previously committed BB identities can be revisited through inference-time re-masking without retraining the model. The affected BB positions could then be re-masked, and the same pretrained bidirectional denoiser could resample them conditioned on the retained route context. Repeated validation and targeted resampling could incorporate chemical feedback into subsequent token decisions, rather than only accepting or rejecting complete routes.

To evaluate the feasibility of inference-time revision, we implemented a lightweight, confidence-guided BB revision mechanism using the original RouteDiff checkpoint. At two late steps of the 16-step reverse-diffusion process (t = 6 and 3), the sampler selectively re-masks previously committed BB positions based on prediction confidence and commitment age, and resamples their identities using the same bidirectional denoiser. A replacement is accepted only if its conditional BB log-probability exceeds that of the original identity.

In an additional Experiment 1 evaluation, the inference-time BB revision variant achieves 93.0% reconstruction (vs. 92.8% for the original RouteDiff sampler), 0.988 mean-best similarity (vs. 0.987), 0.098 pathway diversity (vs. 0.113), and 0.804 BB diversity (vs. 0.783) on Enamine, without retraining. On ChEMBL, it achieves 23.3% reconstruction (vs. 21.0%), 0.748 mean-best similarity (vs. 0.733), 0.037 pathway diversity (vs. 0.040), and 0.316 BB diversity (vs. 0.291). These results provide evidence that inference-time BB revision can be incorporated into RouteDiff while maintaining strong reconstruction and diversity performance using the original model parameters.

The evaluated revision mechanism is guided by model confidence rather than external chemical validation, and currently revises BB identities only. Nevertheless, the ability to revisit committed BB decisions using the original checkpoint provides a concrete starting point for incorporating reactivity feedback during inference. Integrating reactant-dependent validators to identify incompatible reaction steps and guide targeted re-masking of the implicated BB and reaction positions remains future work.

We thank the reviewer again for the constructive discussion and will clarify these points in the revised manuscript.

Reviewer_kEqLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies synthesis-constrained molecular design within a fixed building-block catalog and reaction-template set. Given a target molecule or property requirements, it generates executable synthetic routes and their products, while an external optimization loop handles objectives defined by molecular scoring functions.

Motivated by the fact that postfix autoregressive generation commits decisions before later route context becomes visible, RouteDiff retains the existing postfix representation and execution framework but introduces bidirectional masked diffusion over complete routes, allowing predictions at uncommitted positions to evolve as context emerges.

Semantic prototypes and low-rank identity components jointly form the denoiser’s building-block input and output interfaces, aiming to combine shared chemical information with exact catalog selection; multiple conditions influence route generation through logit composition at each denoising step.

Experiments cover target reconstruction, fixed-budget optimization, direct compositional generation, and denoising trajectories. The matched autoregressive comparison shows greater building-block diversity with comparable reconstruction performance; RouteDiff achieves the highest reported mean on eight of ten PMO tasks among the synthesis-constrained systems compared, while compositional queries and trajectory analyses demonstrate multi-condition generation and pre-commit revision behavior. (§§2.1–2.4; Tables 1–4; Appendix E)

Critical strengths

1. The central modeling idea is clear and directly addresses the stated motivation. Separating postfix execution semantics from prediction context and commitment timing constitutes a concrete change to the generation process. Beyond individual examples, the trajectory analysis observes pre-commit semantic revisions in 36 of 48 routes from a preselected, success-conditioned cohort, providing direct evidence that this behavior occurs in practice. (§2.2; Appendices B.2 and E.2)

2. The paper includes targeted controlled comparisons. The matched autoregressive counterpart retains the model size, conditioning, and complete building-block interface, among other settings. RouteDiff achieves reconstruction rates of 92.8% versus 92.6% and building-block diversity of 0.783 versus 0.756, supporting broader exploration of building-block choices at comparable target fidelity. The interface ablations also support the contribution of the chosen semantic–identity configuration to target recovery. (Table 2; Appendix B.6)

3. The application evaluation and diversity analyses are substantial. Alongside fixed-budget optimization, the five physicochemical composite queries achieve feasible-route rates of 37.62%–99.54% while maintaining high fingerprint diversity among feasible products. The authors further distinguish product-, scaffold-, and reaction-program-level diversity and report multi-product rates and single-product analyses, enabling a more specific interpretation of the generated outputs. (Tables 3–7; Appendix E.1)

Critical weaknesses

W1. Key details of the external PMO optimization loop remain unclear. Although the paper specifies the oracle budget, population sizes, and supported conditioning inputs, it does not clearly explain the source of the initial molecules or how selected parents are converted into conditions for generating offspring. The product-selection and scoring rules for multi-product routes, including how the corresponding evaluations count toward the oracle budget, also remain unclear. Multi-product handling is described for reconstruction and compositional generation, but these rules do not specify the PMO procedure. These details are needed to understand and reproduce the optimization process underlying Table 3. (Sections 2.4 and 3.2; Appendices B.1, C, and F) (Q1 addresses this concern.)

W2. The set used to compute the best task score in Table 4 is unclear. Section 3.3 describes the best task score as the highest score among sampled outputs, whereas Appendix F uses the feasible-product set for several product-level statistics. It remains unclear whether the maximum score is computed over all sampled products or only those satisfying all hard constraints. For the four drug-anchored queries, similarity is not part of hard feasibility, so a high maximum score and a nonzero feasible-route rate need not reflect the same product. The Cobimetinib examples already demonstrate joint success; clarification is needed to interpret the aggregate results across the queries. (Section 3.3; Tables 4–5; Figure 6; Appendix F) (Q2 addresses this concern.)

Questions for authors

Please prioritize Questions 1&2. The answer of them can serve as the response to the corresponding weaknesses; there is no need to answer both Weaknesses sections and Questions section separately.

Q1. Please clarify the external PMO optimization loop.Where do the initial 500 molecules come from, and how are selected parents converted into RouteDiff conditioning inputs? When a route produces multiple products, what selection and objective-scoring rules are used, and how are the corresponding evaluations counted toward the budget? Brief pseudocode or a description of the relevant implementation would suffice.

Q2. Over which set is the best task score computed for the four drug-anchored queries in Table 4? If it is already restricted to products satisfying all hard constraints, please make this definition explicit. If it uses all sampled products, please consider reporting the highest score within the feasible set using the existing generation records, to show how much target similarity is retained after satisfying the constraints. This would not require additional generation or training.


Questions 3 is just a minor related-work clarification. A brief response is sufficient.

Q3. Please cite and briefly discuss PrexSyn (arXiv:2512.00384), which also uses exact building-block catalog classification and composition of condition-specific logits (Sections 2.1 and 2.4). A brief comparison would help distinguish these shared design choices from RouteDiff's whole-route diffusion and semantic–identity parameterization. This is a minor related-work clarification; no additional experiment is requested.

3 comments

We thank the reviewer for the thoughtful assessment and recognition of our whole-route formulation, controlled AR comparison, and diversity analyses. We address Q1 and Q2 directly, followed by the related-work clarification.

W1/Q1: PMO optimization procedure.

Our PMO uses ECFP4-guided genetic search: crossover of scored parent-product fingerprints proposes offspring conditions, and RouteDiff samples synthetic-route candidates conditioned on each target.

Initialization: We sample 1,000 random ECFP4 fingerprints, generate eight routes per fingerprint, and randomly retain up to 500 deduplicated valid products before oracle scoring.

Offspring generation: From the scored initialization pool, we initially select up to 50 parent products by score-weighted sampling without replacement (softmax temperature 0.5). Bitwise Bernoulli crossover of their actual product ECFP4 fingerprints generates 50 offspring fingerprint targets, each conditioning eight RouteDiff route samples using 16 diffusion steps. Following product selection and oracle evaluation, parents and offspring are merged, deduplicated, and the top 50 retained by oracle score to form the parent pool for subsequent iterations.

Multi-product handling: For each fingerprint target, the representative product is selected across sampled routes by maximum ECFP4 Tanimoto similarity to that target, before calling the PMO oracle. Thus, product selection does not use additional objective evaluations.

Oracle accounting: Each previously unseen canonical SMILES is scored once. Repeated molecules reuse cached scores, and initialization is included in the 10,000-evaluation budget. The same accounting applies to all 13 reported PMO objectives (Appendix B.7).

W2/Q2: Best task score and hard-feasible products.

For the four drug-anchored queries, the best task score is computed over unique products from valid sampled routes, instead of only those satisfying all hard constraints. We additionally report the highest task score restricted to products satisfying every specified hard constraint. The results across five seeds are shown below.

Drug queryBest score (all)Best score (hard-feasible)Feasible seeds
Cobimetinib0.913 ± 0.0120.902 ± 0.0165/5
Osimertinib0.762 ± 0.0310.711 ± 0.0655/5
Amlodipine0.827 ± 0.0480.827 ± 0.0485/5
Perindopril0.523 ± 0.2170.679 ± 0.0523/5

Best score (all) is averaged over all five seeds. Best score (hard-feasible) is averaged only over seeds containing at least one hard-feasible product. The final column gives this count. Values are mean ± sample SD over the corresponding seeds.

The feasible-best results provide direct evidence that favorable task scores can be retained among products satisfying all hard constraints. For Cobimetinib, the feasible-best score reaches 0.902, compared with 0.913 over all valid products. Amlodipine maintains 0.827 under both definitions, while Osimertinib attains 0.711 within the feasible set, compared with 0.762 over all valid products. For Perindopril, the feasible-best score reaches 0.679 across the three seeds containing feasible products. Overall, at least one product satisfying all hard constraints is recovered in 18/20 query–seed runs, providing direct evidence of feasible-set task performance across the four drug-anchored queries.

Q3: Relation to PrexSyn.

We thank the reviewer for pointing out this relevant work. PrexSyn (Luo and Coley, 2026) likewise uses postfix synthesis routes, exact catalog selection, and inference-time composition of condition-specific logits. Unlike PrexSyn's autoregressive decoding, RouteDiff performs bidirectional masked diffusion over complete routes with non-postfix token commitment, together with a semantic-prototype/low-rank identity parameterization for exact building-block prediction. We have added the PrexSyn citation and comparison to the revised manuscript (Section 1, Section 2.4, and Appendix A).

Reference: Shitong Luo and Connor W. Coley. Efficient and programmable exploration of synthesizable chemical space. Transactions on Machine Learning Research, 2026. URL: https://openreview.net/forum?id=xDlIer2UnI.

Thanks for the clarifications and additional results. These address my questions, and I’m happy to keep my weak accept recommendation.

We are pleased that our clarifications and additional results have addressed your questions, and we appreciate your continued support. Your thoughtful feedback and constructive engagement throughout the review process have helped us improve the clarity and presentation of our work.

Reviewer_pT2BOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RouteFidd is a masked discrete diffusion model over complete postfix synthetic routs. Model generates molecules with desired properties within a predefined synthesis space. The paper evaluates target reconstruction, molecular property optimization, and multiple design constraints conditioning, assessing both the quality and diversity of the generated solutions.

Critical strengths
  1. The paper addresses a relevant problem and presents a concrete generative modelling approach over a large building block catalog.
  2. The reported results are promising, with the highest reported performance among the compared synthesis constrained methods on eight of ten PMO objectives.
  3. The representation ablations in Table 2 provide evidence that the proposed components improve reconstruction and the matched autoregressive counterpart in Table 2 is a useful control.
Critical weaknesses
  1. Limited methodological novelty. The main contribution is adapting existing masked discrete diffusion to synthetic route generation. The bidirectional transformer is the denoiser architecture within the model, and progressive token revelation is a part of its sampling procedure. The paper does not clearly establish a methodological advance beyond this adaptation and the ability to generate routes independently of postfix execution order.
  2. Comparisons are difficult to interpret. RouteDiff and the baselines use different sampling procedures and building block catalogs, without clear comparison in inference budgets. RoteDiff samples 256 routes per target, whereas ReaSyn uses beam search and iterative refinement. The optimization experiments match oracle calls, but use different search procedures and synthesis spaces. The optimization procedure also needs clearer specification, particularly how parents are used to generate offsprings and how oracle calls are counted.
  3. Evidence for the benefit of diffusion is limited. RouteDiff has lower reconstruction than ReaSyn on both datasets and lower values on all reported ChEMBL metrics. The matched autoregressive counterpart achieves similar Enamine reconstruction, with modest gains from diffusion in diversity.
  4. Important capabilities are insufficiently tested. The evaluation omits DRD2, GSK3β and JNK3 PMO activity objectives reported by ReaSyn without explaining the selection. For compositional generation, feasible route rates for the queries anchored to reference drugs are only 0.06% to 1.78%. A comparison baseline would be useful to assess the performance of RouteDiff.
Questions for authors

Q1: Could you clarify the differences in catalogs and inference budgets across methods, and specify the optimization procedure, particularly how parents determine offspring generation and how oracle calls are counted?

Q2: Why were the DRD2, GSK3β and JNK3 activity objectives reported by ReaSyn excluded from the optimization evaluation? Can you provide results on these tasks to asses whether the gains extend on activity optimization?

Q3: How does the matched autoregressive counterpart perform on the optimization under the same optimization procedure, oracle budget, and synthesis space?

3 comments

Thank you for your careful assessment and valuable comments. We address your concerns and questions point-by-point below.

W1/W3: Methodological novelty and benefits of diffusion.

Our key contribution is decoupling execution order, prediction context, and commitment order in complete synthetic-route generation. ReaSyn's reconstruction results involve iterative pathway refinement, whereas our model uses direct whole-route diffusion sampling. The matched AR control demonstrates higher building-block diversity (0.783 vs. 0.756) at comparable reconstruction (92.8% vs. 92.6%).

W2/Q1: Synthesis spaces, inference, and optimization.

  • Catalogs (BBs/templates): RouteDiff 223,243/115; SynFormer/Graph GA-SF 223,244/115; ReaSyn/Graph GA-ReaSyn 211,220/115; SynNet/SynthesisNet 147,505/91. DoG uses USPTO synthesis DAGs; Graph GA uses unconstrained molecular search. Notably, RouteDiff and SynFormer share 115 templates and similar catalog sizes, with 92.8% vs. 66.3% Enamine reconstruction.
  • Reconstruction inference: RouteDiff draws 256 target-conditioned complete routes (16 denoising steps each), without separate refinement. SynNet uses width 12; SynFormer width 64/exhaustiveness 24 (ReaSyn, App. E.1). ReaSyn uses a full search-and-refinement pipeline: beam width 8/exhaustiveness 4, 12 Enamine (24 ChEMBL) refinement cycles, and 100 Edit Bridge samples per cycle. Edit Bridge further improves reconstruction (ReaSyn, Fig. 4). Both report similarity 0.987, while RouteDiff has higher BB diversity (0.783 vs. 0.753).
  • PMO budget and baseline: All methods use up to 10,000 oracle calls. Graph GA-ReaSyn uses 100/100 population/offspring, projection width 1 or 2, exhaustiveness 4, one refinement cycle, and four Edit Bridge samples/cycle (ReaSyn, App. E.2).
  • Our search: Initialize with 1,000 random ECFP4 targets; retain up to 500 distinct projected products. In the first iteration, select up to 50 parents by score-weighted sampling (temperature 0.5); form 50 offspring fingerprints via bitwise crossover of two parent-product fingerprints, sampling eight routes each. The representative product is selected by maximum ECFP4 Tanimoto similarity before oracle evaluation. Parents and offspring are then merged, deduplicated, and pruned to the top 50. Oracle calls count each previously unseen canonical SMILES once, and initialization counts toward the 10,000-call budget.

W4/Q2: Additional activity PMO.

Under the same 10,000-call protocol, we evaluate DRD2, GSK3β and JNK3.

ActivityRouteDiffMatched ARGraph GA-ReaSyn
DRD20.9290.8980.977
GSK3β0.8960.9010.889
JNK30.8310.7350.695

RouteDiff exceeds the strongest published synthesis-constrained baseline on 10/13 PMO objectives (8/10 original plus GSK3β and JNK3), and matched AR on 2/3 activity tasks (DRD2, JNK3).

W3/Q3: Matched AR optimization.

We evaluate the matched autoregressive counterpart under the same synthesis space, PMO optimization procedure, and 10,000-oracle-call budget.

ObjectiveRouteDiffAR
Amlodipine MPO0.7860.782
Celecoxib rediscovery0.7800.752
Fexofenadine MPO0.8120.875
Median10.3650.352
Median20.2950.271
Osimertinib MPO0.8270.852
Perindopril MPO0.6930.680
Ranolazine MPO0.7640.803
Sitagliptin MPO0.4230.475
Zaleplon MPO0.5240.454

RouteDiff scores higher on 8/13 tasks overall (6/10 original, 2/3 activity). Gains include JNK3 (+0.096) and Zaleplon MPO (+0.070).

W4: Compositional generation.

We evaluate the matched autoregressive counterpart on compositional generation. For the drug-anchored queries, we additionally report the highest score among products satisfying all hard constraints.

Experiment 3: RouteDiff vs. matched AR (RouteDiff / AR)

QueryValid %Feasible %Best scoreFeasible div.
Lipinski+Veber99.94/100.0098.82/99.321.000/1.0000.936/0.931
Lipinski Ro5100.00/100.0099.54/99.381.000/1.0000.938/0.932
Veber95.30/99.9292.94/99.481.000/1.0000.909/0.914
Congreve Ro3100.00/100.0076.60/73.621.000/1.0000.937/0.934
Ghose96.98/98.5637.62/14.281.000/1.0000.914/0.855
Cobimetinib91.56/87.041.78/0.500.913/0.8940.800/0.747
Osimertinib95.74/91.460.42/1.360.762/0.8130.834/0.817
Amlodipine98.42/98.520.16/5.520.827/0.8690.463/0.704
Perindopril99.80/99.960.06/0.000.523/0.1040.230/-

Ghose (37.62% vs. 14.28%) and Cobimetinib (1.78% vs. 0.50%) show higher feasibility and diversity than AR.

Drug-anchored queries: best score among hard-feasible products

DrugRouteDiff feasible bestAR feasible bestSeeds (RD;AR)
Cobimetinib0.9020.8755/5;5/5
Osimertinib0.7110.7145/5;5/5
Amlodipine0.8270.8695/5;5/5
Perindopril0.679-3/5;0/5

Hard-feasible products retain favorable scores. RouteDiff recovers hard-feasible products in 18/20 query–seed runs, compared with 15/20 for matched AR.

Dear Reviewer,

We hope this message finds you well. As the discussion period is nearing its end, we wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues so we can improve our work.

Thank you for your time and effort in reviewing our paper.

Thank you for the clarifications and additional experiments. These substantially address my concerns about experimental details and matched comparisons.

My concern about methodological novelty remains. The bidirectional denoiser and progressive token commitment are components of the established masked diffusion framework, so the claimed decoupling follows from applying this framework to postfix routes.

The new matched AR results show benefits on some tasks, but also advantages for AR on others. RouteDiff also retains lower reconstruction than ReaSyn on both datasets and lower values on all reported ChEMBL metrics. Although the rebuttal strengthens the paper, I remain unconvinced that the empirical benefits sufficiently compensate for the limited methodological advance. I therefore maintain my weak rejection recommendation.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear Reviewers and Chairs,

We sincerely thank the reviewers for their careful assessments and constructive suggestions, and the chairs for their time and effort in coordinating the review process.

We are encouraged by the positive comments on the paper's clear presentation, the relevance of synthesis-constrained molecular design, the targeted ablations and matched-AR comparison, and the PMO and diversity results. We address the main review concerns and summarize the changes in our revised manuscript below.

Summary of key clarifications and additional evidence

  • Contribution and evidence for diffusion (pT2B, Qz9J). The reviewers ask how RouteDiff differs from existing route-generation and masked-diffusion methods, and whether the formulation provides measurable benefits. RouteDiff preserves executable postfix routes while using bidirectional whole-route context and non-postfix token commitment for building-block and reaction predictions. Compared with a matched autoregressive (AR) counterpart, RouteDiff achieves 92.8% vs. 92.6% Enamine reconstruction and 0.783 vs. 0.756 building-block diversity (Table 2). Under the same synthesis space, optimization procedure, and 10,000-oracle-call budget, RouteDiff achieves higher AUC Top-10 on 8/13 objectives (6/10 original and 2/3 activity objectives).

  • Reconstruction protocols and PMO reproducibility (pT2B, kEqL). The reviewers request clarification of cross-method inference settings and the external optimization loop. RouteDiff uses 256 direct target-conditioned route samples, whereas ReaSyn's 95.0% Enamine reconstruction is obtained with beam search, 12 pathway-refinement cycles, and a separate Edit Bridge (ReaSyn, Appendix E.1). We also clarify PMO initialization, crossover of scored parent-product fingerprints, multi-product selection, and oracle accounting with canonical-SMILES score caching, including initialization within the 10,000-call budget.

  • Activity optimization and compositional feasibility (pT2B, kEqL). The reviewers request additional activity objectives, a compositional baseline, and task scores computed within the hard-feasible product set. In response, we additionally evaluate DRD2, GSK3β, and JNK3, expanding the PMO benchmark to 13 objectives. RouteDiff achieves higher AUC Top-10 than the strongest published synthesis-constrained baseline on 10/13 tasks. In addition, we report matched-AR results for all nine composite queries and distinguish the best score over all valid products from the best score restricted to hard-feasible products. For the four drug-anchored queries, RouteDiff finds hard-feasible products in 18/20 drug-query–seed runs, compared with 15/20 for matched AR. Cobimetinib achieves 0.902 mean best score among hard-feasible products, versus 0.913 over all valid products, while Amlodipine reports 0.827 under both definitions, demonstrating favorable task scores within the hard-feasible product set.

  • Additional synthetic constraints and related work (Qz9J, kEqL). Reviewer Qz9J asks how additional synthetic constraints can be incorporated during generation, while Reviewer kEqL requests a brief comparison with PrexSyn. Condition-specific logits are composed at each reverse-diffusion step to guide building-block and reaction predictions, and the generated postfix routes are evaluated by the synthesis executor. We also distinguish RouteDiff's whole-route diffusion and semantic–identity parameterization from PrexSyn's autoregressive formulation, with the full reference provided in our response to Reviewer kEqL.

Changes in the revised manuscript

  • Overview and related work (Abstract; Sections 1 and 2.4; Appendix A): Updated the expanded PMO summary, added PrexSyn, and clarified shared catalog-classification/logit-composition choices versus RouteDiff's whole-route diffusion and semantic–identity parameterization.
  • PMO implementation (Section 3.2; Appendices B.7 and F): Detailed initialization, parent selection, fingerprint crossover, representative-product selection for multi-product routes, oracle accounting, and evaluation-metric cross-references.
  • Expanded PMO evidence (Section 3.2; Appendix G): Added the three activity objectives, the matched-AR optimization control, and complete task-level tables.
  • Compositional evaluation (Section 3.3; Appendix G): Defined the hard-feasible best score and added matched-AR results, feasible-set scores, and query–seed coverage for drug-anchored queries.

Full tables and implementation details are provided in the reviewer-specific responses and corresponding appendices. We thank the reviewers and chairs for their consideration.

Best regards,

The Authors

AuthorsOpenReview ↗

Summary of Rebuttal

Dear Chairs,

Thank you for your time and consideration. To complement our earlier General Response, we would like to provide a concise summary of the reviewers' feedback and the additional evidence presented during the discussion period.

  • Reviewer kEqL has explicitly confirmed that our clarifications and additional results addressed their questions and has maintained their weak accept recommendation.

  • Reviewer Qz9J has acknowledged that the matched-AR comparison strengthens the empirical evidence supporting our proposed formulation and has increased their rating from 1 to 2. We subsequently provided a detailed follow-up addressing the two remaining concerns regarding pretrained checkpoint reuse and reproducibility, as well as reactant-dependent synthetic constraints.

  • Reviewer pT2B has acknowledged that our clarifications and additional experiments substantially addressed their concerns regarding experimental details and matched comparisons. The reviewer nevertheless maintains concerns about methodological advance and has maintained their weak reject recommendation.

We sincerely appreciate the reviewers' constructive feedback and the Chairs' time and consideration. We hope that these developments, together with the additional evidence provided during the rebuttal, will support a favorable final assessment of our work.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.