acceptodds
Under review as a conference paper at ICLR 2027

RetroGEF: Dynamic Graph Edit Flow for Single-Step Retrosynthesis

12 views

Abstract

Retrosynthesis enables the discovery of viable synthetic routes to target molecules. It plays a central role in modern drug discovery and materials design. Retrosynthesis involves molecular graph transformations that can change both connectivity and graph size. These transformations may introduce reactant components absent from the target while revising the product-derived structure. To model these transformations, we propose *RetroGEF*, a flow-based generative model for single-step retrosynthesis. Starting from the target molecule, it constructs possible reactants by adding atoms and changing bonds in the molecular graph. *RetroGEF* models molecular transformations and changes in graph size within the same generative process, rather than relying on a fixed-size graph canvas. It learns this process directly from product–reactant pairs without requiring a prescribed edit order. Experiments on representative retrosynthesis benchmarks demonstrate that *RetroGEF* achieves state-of-the-art performance. Code is available at https://anonymous.4open.science/r/retrogef-E7F8/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Retrosynthesis, AI for Science

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_E8HHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes RetroGEF, a single-step retrosynthesis method based on continuous-time Markov chains. During inference, the model starts out by taking the product molecule (in graph format) as input and iteratively adding atoms and bonds, changing atom properties, and removing bonds, until it ends up in a reactant. The network takes the partially formed graph as input and outputs 1) a total rate of "how many changes will happen per unit time", and 2) probabilities for which changes will happen next. These outputs are trained by 1) taking reactant-product pairs from the data, and making edits from the product side until we end up in the reactants 2) for a given intermediate state, making the network predict the next possible additions to the graph (e.g., atoms to attach directly to the existing structure or non-connected atoms that start new molecules), and also the total rate based on how many such possible actions still exist. After reaching the product, the amount of such actions goes to zero, and the network learns this, and this ultimately ends the generation of new atoms.

Critical strengths
  • The USPTO-50k and USPTO-Full benchmark numbers seem quite good, and improve over the baselines.
  • The core idea of restricting updates to incremental edits on the current molecular graph seems sensible
Critical weaknesses
  • I think that the writing needs substantive improvement for acceptance. As it is, it was not easy to clearly understand what is the proposed method and what were the novel components. As an example, section 4.5 introduces ablations on multiple method/architecture variants that are introduced in single sentences in the same section. It's unclear what is the takeaway of Figure 5, aside from some cases existing where the proposed method matches with the ground-truth better than RetroDiT. Figures 1-3 have lots of notation that is not defined in the figures themselves, making them difficult to read. The paper should include a description of the basic generative framework (with relevant citations for the work that this is building on clearly laid out), and then clearly show how it adapts the framework for retrosynthesis, and how it differs from other related work such as RetroDiT and Retrobridge. And how is the data actually represented in the network. Right now the paper goes into the weeds too fast. Smaller writing issues: Table 4 is referenced before Table 3 in the text.
  • Right now, the visibility to why the method is better than the alternative diffusion/flow based methods is not great. The method seems to make multiple changes compared to, e.g., RetroDiT, and it is difficult to say which one mattered here. In particular, it would be important for the key claim of the paper to include an ablation where each atom/bond is generated independently vs. in the coupled fashion presented in this paper, while keeping everything else fixed.
  • There is no discussion on inference speed in the paper.
Questions for authors
  • What are all the differences between the method in the paper and RetroDiT?
  • What is the concrete benefit of variable-sized graph generation? Is it inference speed, or an inductive bias for better retrosynthesis modelling?
4 comments
AuthorsOpenReview ↗

Response to Reviewer#E8HH 1/2

Thank you for recognizing the performance and the value of incremental graph edits. We respond to your comments as below:

W1: RetroGEF learns complete edits from reaction pairs. RetroGEF builds on Generator Matching [1], cited in Sections 2 and 3, and learns the rates of complete edits that transform a product graph into a reactant graph. Sections 3.1–3.4 and Appendices A–C describe the graph state, bridge, rate network and training objective. We will add a short method overview before the technical definitions. The initial state is the product graph. Nodes represent inherited or newly created atoms, and edges carry the current bond attributes. Paired reactants supply intermediate bridge states and corrective edit rates, accommodating alternative orders and equivalent missing-atom assignments. A graph Transformer encodes atom and bond features, their changes relative to the product, time and the center condition. It predicts total edit intensity and conditional distributions over edit type, location and attributes. Generation samples an event time and a complete edit, updates the graph, and recomputes rates on the new state; the terminal graph is decoded into reactants. The bridge endpoint is the reactant graph, where the target outgoing rate is zero (Equation (4)). Figures 1–3 illustrate the formulation; we will define the graph states and network quantities directly in their captions. Q1 compares the state, transitions and training target with RetroDiT and RetroBridge. Figure 5 demonstrates reconstruction of the complete reactant set: in example (b), both methods recover the ring-opened amino alcohol, while RetroGEF additionally reconstructs the complete triphosgene component (Section 4.6). We will make this takeaway explicit in the caption. Tables 3–4 quantify the corresponding demands of fragment construction and scaffold revision. Section 4.5 defines the ablated factors: constant addition fixes the total growth rate to one, while the edge-update variant removes Transformer edge updates and retains bond inputs and bond-edit actions. We will expand these variant descriptions and introduce Table 3 before Table 4.

W2: Complete edits improve prediction. Complete edits improve Top-1/3/5/10 over separate atom/bond updates under both Pred. and Oracle conditions. Under the same USPTO-50K test setting as in the paper, we compare Top-k accuracy (%) for RetroGEF and two variants: one removes the new atom's learned embedding from first-bond prediction; the other creates the atom and its attachment through separate graph updates. RetroGEF achieves Top-1 of 62.4% versus 59.6% under Pred. and 72.3% versus 72.0% under Oracle. It also exceeds the variant without learned atom-to-bond conditioning at Top-3/5/10 under both center conditions. These results support creating the atom and its first bond together and using the new atom to predict that bond.

ConditionConfigurationTop-1Top-3Top-5Top-10
Pred.Complete Edits62.482.888.691.9
Pred.Without learned atom-to-bond conditioning60.081.186.990.7
Pred.Separate atom/bond updates59.681.387.291.1
OracleComplete Edits72.392.996.097.0
OracleWithout learned atom-to-bond conditioning73.292.695.296.3
OracleSeparate atom/bond updates72.092.795.496.5

W3: Complete edits reduce graph-update cost. Profiling shows 41.7% and 40.1% fewer graph-edit events under Pred. and Oracle, respectively, with 29.6% and 31.1% less graph-update time per trajectory than separate atom/bond updates. With batched inference on one H200 GPU, the complete-edit generator produces 100 trajectories per product in an average of 0.49 seconds under Pred. and 0.48 seconds under Oracle, including graph sampling, RDKit decoding and prediction collection (Appendix D.5).

ConditionConfigurationEdits per trajectoryGraph-update time per trajectory (ms)
Pred.Complete Edits3.380.147
Pred.Separate updates5.800.209
OracleComplete Edits3.400.148
OracleSeparate updates5.680.215
AuthorsOpenReview ↗

Response to Reviewer#E8HH 2/2

Q1: RetroGEF predicts complete-edit rates on evolving graphs. The main differences are the graph state, the transition parameterization and the training target. RetroGEF's Equation (3) conditions the first bond on the new atom, applies both as one graph transition, and uses the resulting graph for the next decision. Its bridge also aggregates latent atom assignments that correspond to the same observable edit. The baseline descriptions follow Section 3.2 of RetroBridge [2] and Section 4 and Appendix C of RetroDiT [3].

MethodState and transitionTraining target
RetroBridgeProduct graph augmented with dummy nodes; categorical atom/bond transitionsEndpoint-conditioned Markov-bridge transitions
RetroDiTCenter-rooted graph ordering with dummy nodes; discrete flow over atom/bond variablesFactorized endpoint-category predictions
RetroGEFCurrent graph expanded and revised through complete editsObservable edit rates from product–reactant bridges

Q2: New structure guides subsequent edits. RetroGEF provides a construction bias for retrosynthesis and reduces graph updates. Each decision operates on the current graph, and a newly created atom and its first bond immediately become context for subsequent edits. This supports interleaving fragment construction, connectivity changes and scaffold revision within one process. The structural analysis also covers products requiring more than twenty graph changes: Oracle Top-1 is 42.5% for RetroGEF versus 33.6% for RetroDiT (Table 3). This grouping counts attribute updates and bond changes alongside atom additions, complementing the growth analysis. In W2, RetroGEF outperforms separate updates at Top-1/3/5/10 under both center conditions. W3 quantifies the reduction in graph-edit events and graph-update time.

[1] Holderrieth et al., Generator Matching: Generative modeling with arbitrary Markov processes (2025).

[2] Igashov et al., RetroBridge: Modeling Retrosynthesis with Markov Bridges (2024).

[3] Wang et al., Order Matters in Retrosynthesis: Structure-aware Generation via Reaction-Center-Guided Discrete Flow Matching (2026).

I thank the authors for the clarifications and additional experiments. They partially address my concerns, but I still see issues with the paper.

  • I appreciate the clarifications to the paper figures. Note that Figure 1 still doesn't define e or m.
  • I'm still not sure what Figure 5 demonstrates exactly: this could be a special case where RetroGEF happens to work better? Is it supposed to show something general, e.g., that RetroDiT can't handle these types of reactions?
  • The results on W2 look good, but what exactly is "the other creates the atom and its attachment through separate graph updates"? can you provide details on this?
  • W3: I think the time should be included in the paper, with comparisons to other methods.
  • Q1: Thank you for the clarification, but it would still be useful to have a clear conceptual handle on what are e.g. the core conceptual components that are different from RetroDiT spelled out. E.g., is the data representation different? How about the architecture? Generator matching methodology / loss function? The sampling process? As in, how many distinct changes to RetroGEF would we have to make to turn it into RetroDiT.
  • Q2: Thank you for clarifying. Regarding "Each decision operates on the current graph, and a newly created atom and its first bond immediately become context for subsequent edits" -> the basic fact here holds also for RetroBridge and RetroDiT in that each decision operates on the current graph, changing the graph for the next edit. It indeed seems possible that RetroGEF provides a useful inductive bias for retrosynthesis, but it still seems difficult to be confident in what factors are actually providing the improvement on benchmarks.

In general, I think the paper would still benefit from an overhaul in the writing and experiment design to clearly pinpoint the factors driving the improvements over previous methods.

AuthorsOpenReview ↗

Follow-up Response to Reviewer E8HH

Dear Reviewer E8HH,

Thank you for following up. We address your questions below.

W1. The revised Figure 1 caption defines m as the attributes of the new atom u and e as those of its first bond to the existing atom j (Eq. (3)). Figure 5 shows two examples in which matching the reference reactant set requires reconstructing a separate molecule. In panel (b), both methods recover the ring-opened amino alcohol, and RetroGEF also reconstructs triphosgene. On USPTO-Full, RetroGEF leads RetroDiT in Oracle Top-1 in every added-atom group in Table 4. For the 5,624 products requiring 11–20 added atoms, as in panel (a), Oracle Top-1 is 33.4% for RetroGEF versus 26.0% for RetroDiT.

W2. In RetroGEF’s separate-update variant (Table 8), one edit adds an isolated atom and a later bond edit connects it to an existing atom. With complete edits, RetroGEF chooses the attachment site and the new atom’s element together, predicts the first bond conditioned on the atom’s attributes, and adds the atom and bond in one edit (Figure 2B).

W3. We timed the original RetroGEF and RetroDiT implementations on the USPTO-50K test set in the same computing environment, using 100 trajectories per product. Mean generation time per worker was 0.49 s/product for RetroGEF and 2.62 s/product for RetroDiT. Section 4.5 reports these times, and Table 15 gives the breakdown. RetroDiT reserves positions for possible new atoms and predicts atom and bond labels for those positions at every sampling step. RetroGEF adds atoms as needed and can form the first bond in the same edit, saving the extra forward pass required by separate atom and bond updates.

ModelGraph updates (s/product)Forward passes (s/product)Total (s/product)
Complete Edits0.01470.47390.49
Separate updates0.02090.48780.51
RetroDiT0.631.992.62

Q1. To construct reactants of unknown size and connectivity, RetroGEF starts from the product and learns rates for complete edits that grow and revise the graph (Section 3). RetroDiT [1] appends dummy nodes before sampling and predicts final atom and bond labels on this fixed node set. Its graph Transformer encodes atom order from the reaction center with rotary position embeddings. Its sampler derives update rates from the label distributions learned with cross-entropy. Our graph Transformer encodes changes relative to the product and predicts the total edit intensity and complete-edit probabilities (Figure 3). We fit the resulting rates to bridge targets along alternative construction paths using the generalized KL loss in Eq. (4). During generation, we sample edit times and complete edits from the learned rates (Figure 1 and Algorithm 2).

Q2. Table 8 compares construction choices within RetroGEF under both center conditions on USPTO-50K. The complete-edit generator outperforms the separate-update variant at all four cutoffs. Among the variants that create the atom and bond together, conditioning the first bond on the new atom’s learned embedding improves Top-3/5/10. In the training bridge, setting the atom-addition rate to the number of atoms still to add raises Pred. Top-10 on USPTO-Full by 7 percentage points over a constant rate (Table 5).

We have submitted a revised manuscript that includes these clarifications and the timing comparison. Thank you again for your time and consideration. We hope our responses and the additional experiments have addressed your concerns.

[1] Wang et al., Order Matters in Retrosynthesis: Structure-aware Generation via Reaction-Center-Guided Discrete Flow Matching (2026).

Reviewer_w7hgOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents RetroGEF, a continuous-time graph jump process for single-step retrosynthesis on variable-size graphs. Complete edits jointly specify a new atom's attributes, attachment site, and initial bond. Training uses stochastic product-reactant bridges that accommodate alternative construction orders and aggregate equivalent edits arising from latent atom correspondences. Generator matching learns the edit rates. Experiments on USPTO-50K and USPTO-Full report improvements over RetroDiT under both predicted and ground-truth reaction-center conditioning.

Critical strengths
  1. The formulation combines graph growth with revisions to existing structure, while the bridge supports learning from endpoint pairs without prescribing a single construction order. Aggregating equivalent edits over latent atom correspondences aligns supervision with the observable edits used at inference. These design choices provide a coherent way to model reactant reconstruction.
  2. The reported improvements over RetroDiT hold across two benchmarks and both reaction-center settings. Analyses by added atom count, transformation complexity, and reaction class provide further support for the method's usefulness across different reconstruction requirements.
Critical weaknesses
  1. The motivation for dynamic graph growth and complete edits needs clearer justification. Placeholder-based methods can also generate reactants with different atom counts, so variable graph size alone does not explain the advantage of the proposed formulation. The paper needs to identify the specific limitations it addresses and connect these to the design and results. This affects how much methodological progress can be attributed to the approach.
  2. Tables 6 and 10 evaluate the reference model on the test set but the training variants on the validation set. This prevents direct comparison and leaves the conclusions about partial hints and recovery difficult to assess.
Questions for authors
  1. Which limitations of placeholder-based approaches are addressed by dynamic graph growth and complete edits, and which results support these benefits? To clarify the contribution, please also discuss the differences from Morph [1] in the complete edits, product-reactant bridge, and treatment of latent atom correspondences. A conceptual comparison would be sufficient for this recent work.
  2. Can the reference model and training variants in Tables 6 and 10 be evaluated on a common split? Please explain whether the conclusions about partial hints and recovery still hold under that comparison. [1] Franke et al. Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport. arXiv:2606.07239, 2026.
1 comment
AuthorsOpenReview ↗

Response to Reviewer#w7hg

Thank you for recognizing the method's formulation and the consistent improvements across benchmarks. We address your concerns below:

W1: RetroGEF grows and revises the current graph. RetroGEF constructs reactants on the evolving graph: it creates an attached atom and its first bond as one transition, then uses the updated graph to guide further growth, bond changes and scaffold revision. A placeholder canvas represents atom presence and connectivity through preallocated node and edge variables. The product–reactant bridge supervises RetroGEF's edits from reaction endpoints, using product atom mappings for inherited atoms while accommodating alternative construction orders and equivalent assignments of missing atoms. Q1 gives the supporting results and comparison with Morph [1].

W2: The test results support both supervision strategies. We compare Top-k accuracy (%) of the supervision variants and full RetroGEF under the same USPTO-50K test setting as in the paper. Partial hints raise Pred. Top-1 from 59.8% to 61.4%, and full RetroGEF reaches 62.4%, leading at all four Pred. cutoffs. Recovery improves Oracle Top-3 from 93.0% to 93.5%; full RetroGEF achieves the highest Oracle Top-5/Top-10, 96.0/97.0%.

ConfigurationPred. Top-1Pred. Top-10Oracle Top-1Oracle Top-10
Complete hints59.891.472.996.7
Partial hints61.491.773.296.8
Recovery60.291.573.096.6
Partial hints + Recovery62.491.972.397.0

Q1: RetroGEF jointly models growth and scaffold revision. RetroGEF's gains extend to both large-fragment construction and scaffold revision. In Table 4, Oracle Top-1 for reactions requiring 11–20 added atoms is 33.4% versus 26.0% for RetroDiT; 823 products have a correct first prediction only from RetroGEF, versus 406 only from RetroDiT (Section 4.3). For reactions with no added atoms, Oracle Top-1 is 84.8% versus 71.9%. Under the same USPTO-50K test setting, RetroGEF achieves Top-1 of 62.4% versus 59.6% under Pred. and 72.3% versus 72.0% under Oracle, with higher Top-3/5/10 than separate updates in both conditions. Complete edits also reduce graph-edit events from 5.80 to 3.38 per trajectory under Pred. and from 5.68 to 3.40 under Oracle. RetroGEF's contribution is the product-relative edit process and its supervision from reaction pairs, including ambiguity in the identity of missing atoms. Morph studies flexible-size 3D molecular design through unbalanced optimal transport, combining coordinate evolution with structural changes and conditioning inserted-node edge predictions on the sampled node.

AspectMorphRetroGEF
Task and stateMolecular design on geometric graphs with coordinatesProduct-conditioned reconstruction of reactant graphs
Endpoint alignmentUnbalanced optimal transport between geometric graphsProduct atom mapping for inherited atoms and latent correspondences for generated atoms
Growth transitionNode insertion with conditional edge predictionsAttached atom and first bond specified as one edit, followed by further graph edits
SupervisionInterpolation through geometric alignmentObservable edit rates from product–reactant bridges, aggregating equivalent latent proposals

Q2: Partial hints and recovery improve prediction. Yes. The test results in W2 support both choices: partial hints improve Pred. Top-1, and recovery improves Oracle Top-3. Full RetroGEF leads at all four Pred. cutoffs. Recovery is enabled in the benchmark training; Appendix C.1 specifies its correction targets, including generated-atom attributes.

[1] Franke et al., Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport (2026).

Reviewer_X9ngOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RETROGEF predicts reactants through a continuous-time sequence of edits to the product molecular graph. It can create atoms, attach them to the existing structure, modify bonds and attributes, and continue editing newly generated fragments. This avoids preallocating a fixed reactant canvas. Training uses stochastic bridges between paired product and reactant graphs. The bridges sample different construction orders, while equivalent hidden atom correspondences are aggregated into observable edits. A rate-matching objective separately learns the total edit intensity and the distribution over complete actions. At inference, 100 trajectories are sampled per product, and distinct reactant sets are ranked by frequency. On USPTO-50K and USPTO-Full, RETROGEF outperforms RetroDiT under predicted and oracle reaction-center conditioning. Shared-product analyses further examine performance by added-atom count, transformation complexity, and reaction class.

Critical strengths

The representation addresses a real mismatch in existing models. Retrosynthesis can revise the product-derived scaffold while also introducing entirely new fragments. RETROGEF handles both in one evolving graph. Coupling atom creation with its first bond is chemically and computationally sensible, and subsequent edits can immediately use the new structure as context.

The supervision strategy is carefully formulated. The method does not assume that reaction records contain a preferred edit order. It also avoids treating equivalent hidden atom assignments as distinct observable targets. The factorization into intensity and action choice is intuitive, and the appendices provide a useful account of the population optimum and marginal consistency for unperturbed bridges.

The paired comparisons against RetroDiT are informative. Sections 4.3 and 4.4 use 90,598 shared USPTO-Full products and 4,930 shared USPTO-50K products, respectively. These analyses partially address the concern about comparisons based on published numbers. They also show that the gains extend across added-atom counts and graph-change complexities, including reactions requiring no added atoms.

The paper communicates the mechanism well. Figures 1–3 give a clear account of the state representation and rate prediction. The construction trajectories are helpful concrete examples, and the authors appropriately distinguish graph-construction order from chemical reaction mechanism.

Critical weaknesses
  1. The ablations do not isolate the main contributions. The paper does not directly test complete edits against separate atom/bond edits, quotient aggregation against fixed atom correspondences, multiple construction orders against a fixed order, or rate matching against next-edit classification. The current experiments show that the full system works, but not which design choices explain its gains. A matched fixed-canvas baseline would also strengthen the comparison.

  2. The main benchmark comparisons are not fully controlled. Tables 1–2 combine new results with published scores without establishing identical test products, preprocessing, model capacity, and candidate budgets. The shared-product RetroDiT analyses partly address this concern, but not for the complete tables. Moreover, RSGPT outperforms RETROGEF under predicted-center evaluation, so the abstract’s unqualified “state-of-the-art” claim should be narrowed to the reaction-center-conditioned comparison.

  3. The supervision ablation compares different evaluation splits. Tables 6 and 10 report the reference model on the test set but the training variants on validation, without an unchanged validation baseline. These numbers cannot establish the benefits of partial hints or recovery augmentation. All variants should use the same split and protocol, and the manuscript should clarify whether recovery augmentation is enabled in the main reference model.

  4. The variability of the results is unknown. The paper reports no training-seed variation, confidence intervals, or significance tests. This is important for modest gains such as the 1.2-point USPTO-50K Top-1 improvement and for small reaction classes. Paired per-product intervals and repeated decoding would help distinguish genuine improvements from training and sampling variation.

  5. The reaction-class discussion overlooks substantial regressions. Table 11 shows predicted-center Top-1 decreases for oxidation, from 69.5% to 56.1%, and deprotection, from 56.4% to 51.2%; oracle deprotection also declines. RetroDiT’s 6.7% predicted-center versus 69.7% oracle accuracy on heterocycles is another unexplained anomaly. These results deserve analysis alongside the improvements, with uncertainty estimates for small classes.

Questions for authors

1 Which proposed component accounts for most of the improvement? Please provide controlled results for complete versus separated atom/bond edits, quotient versus fixed correspondences, multiple versus fixed edit orders, and rate matching versus next-edit classification.

2 Can you provide a fully matched RETROGEF–RetroDiT comparison on the main evaluation sets, including the same products, center predictions, candidate budget, topology matcher, and comparable model capacity?

3 What explains the oxidation and deprotection regressions and RetroDiT’s unusually low predicted-center accuracy on heterocycles?

4 What fraction of trajectories yields RDKit-valid reactant sets? Can you additionally report full isomeric-SMILES accuracy and round-trip accuracy or another measure of chemical plausibility?

2 comments
AuthorsOpenReview ↗

Response to Reviewer#X9ng 1/2

We thank the reviewer for the thorough evaluation and constructive suggestions, and for recognizing the complete-edit representation and bridge supervision. We address each concern below:

W1: Complete edits improve prediction. RetroGEF outperforms separate atom/bond updates at Top-1/3/5/10 under both Pred. and Oracle conditions. Under the same USPTO-50K test setting as in the paper, we compare Top-k accuracy (%) for complete edits and two variants: one removes the new atom's learned embedding from first-bond prediction; the other separates atom creation and bond attachment into successive graph updates. RetroGEF achieves Top-1 of 62.4% versus 59.6% under Pred. and 72.3% versus 72.0% under Oracle. It also exceeds the variant without learned atom-to-bond conditioning at Top-3/5/10 under both conditions. Complete edits reduce graph-edit events by 41.7% under Pred. and 40.1% under Oracle. The RetroDiT comparisons in Sections 4.2–4.4 provide the fixed-canvas reference.

ConditionConfigurationTop-1Top-3Top-5Top-10
Pred.Complete Edits62.482.888.691.9
Pred.Without learned atom-to-bond conditioning60.081.186.990.7
Pred.Separate atom/bond updates59.681.387.291.1
OracleComplete Edits72.392.996.097.0
OracleWithout learned atom-to-bond conditioning73.292.695.296.3
OracleSeparate atom/bond updates72.092.795.496.5

W2: RSGPT's headline results use additional data. RSGPT predicts reactants directly from the product, without a predicted-center input. Its headline result uses 10B additional synthetic reactions for pre-training, while its public release does not include the full pre-training corpus. Trained only on USPTO-50K without pre-training or RLAIF, RSGPT [1] reports 26.4% Top-1 and 46.4% Top-10, compared with 62.4% and 91.9% for RetroGEF (Pred.). These RSGPT results are included as RSGPT* in our Table 2. The comparison involves different training data and conditioning settings. RetroGEF achieves state-of-the-art performance, improving over RetroDiT at all four cutoffs on both datasets under both center conditions. The Top-1 gains are 3.0/5.2 points on USPTO-Full and 1.2/1.2 points on USPTO-50K, under Pred./Oracle respectively. Sections 4.3–4.4 also compare the methods on 90,598 shared USPTO-Full products and 4,930 shared USPTO-50K products.

W3: The test results support both supervision strategies. We compare Top-k accuracy (%) of the supervision variants and full RetroGEF under the same USPTO-50K test setting as in the paper. Recovery is enabled in the benchmark training; Appendix C.1 specifies its correction targets, including generated-atom attributes. Partial hints raise Pred. Top-1 from 59.8% to 61.4%, and full RetroGEF reaches 62.4%, leading at all four Pred. cutoffs. Recovery improves Oracle Top-3 from 93.0% to 93.5%; full RetroGEF achieves the highest Oracle Top-5/Top-10, 96.0/97.0%.

ConditionConfigurationTop-1Top-3Top-5Top-10
Pred.Complete hints59.881.087.391.4
Pred.Partial hints61.482.288.091.7
Pred.Recovery60.281.987.391.5
Pred.Partial hints + Recovery62.482.888.691.9
OracleComplete hints72.993.095.996.7
OraclePartial hints73.293.195.996.8
OracleRecovery73.093.595.796.6
OraclePartial hints + Recovery72.392.996.097.0

W4: Paired comparisons support the Top-10 gains. The paired confidence intervals support the USPTO-Full gains, the Top-10 improvements in all four settings, and the USPTO-50K Oracle Top-1 improvement. We estimate 95% intervals from 10,000 product-bootstrap replicates of the shared-product predictions. Values are percentage-point improvements over RetroDiT, with pointwise 95% confidence intervals in brackets. Table 15 in Appendix D.6 reports the classwise intervals and sample counts.

Dataset / conditionTop-1 differenceTop-10 difference
USPTO-Full Pred.+3.19 [2.91, 3.48]+1.27 [1.03, 1.50]
USPTO-Full Oracle+5.21 [4.92, 5.50]+0.53 [0.31, 0.75]
USPTO-50K Pred.+1.10 [−0.18, 2.39]+2.66 [1.83, 3.51]
USPTO-50K Oracle+1.22 [0.06, 2.41]+0.53 [0.04, 1.05]
AuthorsOpenReview ↗

Response to Reviewer#X9ng 2/2

W5: Candidate coverage improves for oxidation and deprotection. RetroGEF improves oxidation Pred. Top-10 and deprotection Oracle Top-3/5/10. For heterocycles, RetroDiT's center distribution assigns zero probability to the reference-center support for most products. Q3 gives the supporting results and diagnostic.

Q1: The bridge supervises edit choices and rates. The comparison in W1 supports joint atom/bond construction: RetroGEF achieves higher Top-1/3/5/10 than separate updates under both center conditions. Quotient aggregation sums the rates of latent proposals that produce the same observable edit. If proposals with weights wzw_z produce actions aza_z, then

=-\sum_a\left(\sum_{z:a_z=a}w_z\right)\log\pi_\theta(a).$$ For a fixed proposal set, aggregation preserves the objective obtained by enumerating all latent proposals, while supervising observable edits. Fixing the correspondence of generated atoms or the construction order restricts supervision to a particular assignment or sequence. Our bridge samples corrective construction orders and sums over equivalent missing-atom assignments, so reaction endpoints supervise observable edits without imposing an arbitrary target sequence. Rate matching learns **when edits occur**: Equation (4) trains total edit intensity alongside action choice, with zero target intensity at the reactant endpoint. A next-edit classifier models action choice, with timing and termination specified separately. Table 5 tests growth-rate learning: fixing the addition rate lowers Top-10 by **7.0/4.8 points on USPTO-Full** and **3.8/4.0 points on USPTO-50K**, under Pred./Oracle respectively. Appendix F.3 derives the intensity and action-choice terms, including the zero-rate endpoint case. **Q2: We compare methods on shared test products.** Sections 4.3–4.4 compare the methods on common products with the same candidate budget and topology-matching criterion, using each method's center inputs. All 4,930 paired USPTO-50K records have **identical topology-level reference reactants**. RetroGEF uses 100 trajectories per product and ranks distinct reactant sets by frequency. **Q3: The gains extend beyond Top-1.** RetroGEF achieves better candidate coverage for oxidation and deprotection. Oxidation Pred. Top-10 is **95.1% versus 91.5%** for RetroDiT; deprotection Oracle Top-3/5/10 are 89.6/94.9/96.8% versus 88.8/93.8/96.5%. In a Pred. deprotection example, RetroGEF reconstructs the trimethylsilyl protecting group and ranks the reference precursor first, while RetroDiT does not recover it within its top ten (Appendix D.3.2). For heterocycles, RetroDiT's center distribution assigns zero probability to every root in the reference-center support for **83 of 89 products** (Appendix D.3.2). RetroGEF also retains a clear advantage with Oracle centers, reaching 77.5/95.5% Top-1/Top-10 versus 69.7/84.3% for RetroDiT (Table 14). **Q4: RetroGEF achieves high molecular validity.** RetroGEF achieves **97.7–97.9% RDKit validity** under full molecular reconstruction. We assess validity with RDKit sanitization. Top-k denotes full isomeric-SMILES accuracy (%), assessing atom and bond attributes as well as connectivity. Validity counts all 100 trajectories per product: 9,328,900 per USPTO-Full condition and 494,400 for USPTO-50K Pred. Section 4.7 provides these results while Appendix D.7 gives the evaluation details. | Dataset / condition | RDKit validity | Top-1 | Top-3 | Top-5 | Top-10 | |---|---:|---:|---:|---:|---:| | USPTO-Full Pred. | 97.7 | 52.9 | 68.7 | 72.8 | 75.9 | | USPTO-Full Oracle | 97.8 | 67.2 | 79.7 | 81.9 | 83.3 | | USPTO-50K Pred. | 97.9 | 60.9 | 81.2 | 86.9 | 90.6 | [1] Deng et al., RSGPT: a generative transformer model for retrosynthesis planning pre-trained on ten billion datapoints (2025).

Comments on the paper

2 comments

We sincerely thank the reviewers for their time and thoughtful feedback. We are also encouraged that all three reviewers recognize RetroGEF's performance gains. X9ng and w7hg highlight the bridge supervision and the analyses across different reconstruction requirements, while E8HH recognizes the rationale for incremental graph edits.

Our additional experiments support the complete-edit formulation and supervision strategies. On USPTO-50K, complete edits improve Top-1/3/5/10 over separate atom/bond updates under both Pred. and Oracle conditions, with approximately 40% fewer graph-edit events. Evaluating the supervision variants on the test set supports the benefits of partial hints and recovery. We also provide paired confidence intervals, reaction-class analyses and inference profiling, together with full molecular reconstruction results showing 97.7–97.9% RDKit validity. The responses clarify the differences from RetroDiT [1], RetroBridge [2] and Morph [3], and distinguish RSGPT's results with 10B-reaction pre-training [4] from its USPTO-50K-only results in Table 2.

We will submit a revised manuscript incorporating these results and clarifications soon, with a concise method overview, expanded explanations of the ablations, and clearer figure captions and table references.

[1] Wang et al., Order Matters in Retrosynthesis: Structure-aware Generation via Reaction-Center-Guided Discrete Flow Matching (2026).

[2] Igashov et al., RetroBridge: Modeling Retrosynthesis with Markov Bridges (2024).

[3] Franke et al., Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport (2026).

[4] Deng et al., RSGPT: a generative transformer model for retrosynthesis planning pre-trained on ten billion datapoints (2025).

AuthorsOpenReview ↗

Revision update

Dear reviewers, we have uploaded the revised manuscript, incorporating the additional experimental results and clarifications discussed in our responses. We have also refined the method presentation, ablation explanations, and figure captions based on the reviewers’ feedback. We sincerely thank the reviewers again for their time and thoughtful suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.