acceptodds
Under review as a conference paper at ICLR 2027

RetroVirtSyn: Adaptive Multi-Center Virtual Synthons for Single-Step Retrosynthesis

Abstract

Two-stage single-step retrosynthesis is susceptible to error propagation because reactant generation depends critically on reaction-center predictions made in the preceding stage. Conventional semi-template methods typically retain a single reaction-center candidate and construct synthons through hard bond disconnection based on that prediction. When the predicted center is incorrect, synthon construction converts prediction errors into structural deviations, amplifying their impact on downstream reactant generation. To address this limitation, we propose RetroVirtSyn, a retrosynthesis framework based on adaptive multi-center virtual synthons. RetroVirtSyn avoids hard bond disconnection; instead, it adaptively retains multiple candidate reaction centers based on prediction confidence and encodes them as confidence-weighted probabilistic conditions to construct virtual synthons directly on the intact product graph. Empirically, RetroVirtSyn achieves 63.2% Top-1 accuracy on USPTO-50K, outperforming the previous state of the art by 3.2 percentage points, while also achieving the best Top-3, Top-5, and Top-10 accuracies among the compared methods. These results demonstrate that virtual synthons provide an effective interface between reaction-center prediction and reactant generation, mitigating error propagation and thereby improving retrosynthesis accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Retrosynthesis, Virtual Synthons, Chemistry, Drug Discovery, Discrete Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GvRhOpenReview ↗
Rating
4(+3)
Confidence
3
Informativeness
2(+1)
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This proposes a two stage pipeline for the single-step retrosynthesis. At the first stage, the pipeline predicts the reaction centers to get the synthons. A dynamic selection rule is applied to preserve the RC candidate for the next stage. Unlike previous works. the pipeline do not breaks the bonds according to the RC predictions, but merge and mark the predictions on the input product mole as virtual synthons. A flow matching conditioned on the virtual synthon is applied finally for the generation of the products.

Critical strengths
  • The motivation that hard construction of the synthons might have the multi-stage error accumulation is very reasonable. The paper is well motivated.
  • The presentation is clear, the paper is easy to follow.
Critical weaknesses
  • The paper's core idea is not new, see the questions for details.
  • The evaluation in the article is chemically unreasonable, and the performance is overestimated. see the questions for details.
Questions for authors
  • According to the supplementary material, the molecular graph representation includes atom types and chemical bond types but does not explicitly encode stereochemical information. However, baselines such as Graph2SMILES and Chemformer use stereochemistry-aware canonical SMILES exact matching for evaluation. Since approximately 20% of USPTO-50K test reactions involve stereochemistry, ignoring it during evaluation may substantially inflate the reported performance and lead to unfair comparisons.

    • Could the authors clarify whether stereochemistry is considered during evaluation and,
    • if so, how it is recovered from stereochemistry-free graph predictions? The authors should also provide results under a consistent stereochemistry-aware evaluation protocol.
  • The evaluation lacks many powerful baselines like [3, 4, 5, 6]

  • The idea that mark the reaction center on the reactant the perform conditional generation for product are not new too. Most recently, the one combines the soft RC mark with the conditional flow matching for single-step retrosynthesis are provided at ICML2026[3]. For Smiles generation we also have the previous work like [2] and can even date back to 2023[1]. The author do not discuss them, and the major contribution of this paper do not holds.

[1] Thakkar, Amol, et al. "Unbiasing retrosynthesis language models with disconnection prompts." ACS Central Science 9.7 (2023): 1488.

[2] Xiong, Jiacheng, et al. "Bridging chemistry and artificial intelligence by a reaction description language." Nature Machine Intelligence 7.5 (2025): 782-793.

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

[4] Zhong, Zipeng, et al. "Root-aligned SMILES: a tight representation for chemical reaction prediction." Chemical Science 13.31 (2022): 9023-9034.

[5] Han, Yuqiang, et al. "Retrosynthesis prediction with an iterative string editing model." Nature Communications 15.1 (2024): 6404.

[6] Zeng, Kaipeng, et al. "Ualign: pushing the limit of template-free retrosynthesis prediction with unsupervised SMILES alignment: K. Zeng et al." Journal of Cheminformatics 16.1 (2024): 80.

15 comments
AuthorsOpenReview ↗

Response to Reviewer GvRh — Critical Weakness 1: Novelty

We thank the reviewer for carefully reviewing our work and for the valuable suggestions on the experimental analysis and methodological presentation. We have revised the manuscript to address all the issues you raised and marked the corresponding changes in blue for ease of reference. Our point-by-point responses follow.

Critical Weakness 1

The central idea of the paper is not novel. Please see the Questions section for details.

Thank you for raising the concern about novelty. Our central contribution addresses the practical problem of prediction errors in the first stage by converting uncertainty in reaction-center predictions into a structural probability input for the second stage. This representation is used both as a generation condition and to construct the starting distribution of the probability bridge. Describing our method as simply marking reaction centers and then performing conditional generation does not capture this distinction.

By a “soft interface,” we mean that the input preserves predictive uncertainty as a probability distribution over possible structural states, rather than directly converting a selected center into a deterministic structural modification. Softness does not depend on whether centers are marked through atom labels, sequence ordering, or positional encodings. For a candidate bond to be broken, ee, the virtual synthon converts the center weight qeq_e intoGS(e)=(1−qe)δbond+qeδnone.\displaystyle G_S(e)=(1-q_e)\delta_{\mathrm{bond}}+q_e\delta_{\mathrm{none}}.

When 0<qe<10<q_e<1, this representation retains probability mass for both the original bond and the no-bond state, rather than deciding in advance that the bond must be broken. These probabilities describe uncertainty over discrete structural states; the sampled graphs and final reactants remain discrete.

The three related studies organize reaction-center information through disconnection prompts, explicit editing instructions, and center-first ordering, respectively. Their contributions have different emphases, but they differ from our mechanism, which converts first-stage center confidence into categorical probabilities over the original-bond and no-bond states and uses these probabilities to construct the generative probability bridge. In particular, an upstream predictor producing center scores does not imply that those scores are retained as structural probability inputs to the generator after candidate selection.

The virtual synthon therefore specifies not only where a reaction may occur, but also the probability with which a predicted structural change enters the starting distribution. It serves as an explicit condition and defines the starting distribution of the linear categorical probability bridge, thereby changing the distribution of intermediate training states. Combined with adaptive retention of multiple centers, this interface addresses two sources of error. It avoids committing an incorrect center prediction to a deterministically broken bond, and it preserves opportunities to recover a true center omitted by the top-ranked candidate. This is a probabilistic structural interface designed to address error propagation between stages, rather than a replacement for a center-annotation scheme.

This design is supported by experiments. On the full test set, the single-center virtual synthon achieves a Top-1 accuracy of 63.50%, compared with 60.30% for the hard bond-breaking interface and 60.48% for the full product plus center prompts. On the subset with incorrect center predictions, its Top-10 accuracy is 53.08%, compared with 33.97% and 52.72%, respectively. We also added transfer experiments with G²Retro and RetroSynFlow to examine the adaptability of the interface across generative frameworks. The distinctions from the three studies, the three-way interface comparison, recovery from incorrect centers, the complementarity of multiple centers, and the transfer experiments are detailed below.

To address stereochemistry, we incorporated tetrahedral chirality and double-bond stereochemistry into the discrete flow-matching model in this revision, including its state representation, training supervision, and output decoding. We implemented and retrained the resulting stereochemistry-aware version, RetroVirtSyn-Stereo. Under strict evaluation that preserves stereochemistry, its Top-1 accuracies on the full test set and the stereochemical subset are 70.96% and 69.49%, respectively. The implementation and evaluation of this new extension are detailed below.

AuthorsOpenReview ↗

Response to Reviewer GvRh — Critical Weakness 2: Evaluation fairness

Critical Weakness 2

The evaluation in the paper is chemically unreasonable, and the performance is overestimated. Please see the Questions section for details.

Thank you for identifying this issue. We acknowledge that the original manuscript did not clearly distinguish evaluations with and without stereochemistry, which could lead to inappropriate comparisons across different evaluation criteria. In the revised manuscript, we report and discuss these results separately and have updated the experimental descriptions, result tables, and performance conclusions accordingly.

Specifically, the original model results are now explicitly categorized as graph-structure evaluation without stereochemistry. We no longer present them as strict stereochemical matching results for direct comparison with the corresponding baselines. For RetroVirtSyn-Stereo, we explicitly model and predict atom chirality and double-bond stereochemistry, and evaluate exact matches of canonical SMILES that preserve stereochemistry. A prediction is counted as incorrect if stereochemistry is missing or incorrect, even when connectivity is correct.

Under this strict criterion, the extended model achieves Top-1/Top-10 accuracies of 70.96%/83.64% on all 5,007 test reactions and 69.49%/81.21% on the 990 reactions with explicit stereochemical annotations. We also evaluate the same predictions with their ranking fixed, both with and without stereochemistry, to isolate the effect of the evaluation criterion.

The revised manuscript explicitly distinguishes correct connectivity from correct connectivity together with correct stereochemistry, rather than conflating them into a single metric.

AuthorsOpenReview ↗

Response to Reviewer GvRh — Question 1: Stereochemical modeling and evaluation (Part 1/2)

Question 1

According to the supplementary material, the molecular graph representation includes atom and bond types but does not explicitly encode stereochemical information. However, baselines such as Graph2SMILES and Chemformer are evaluated using exact matching of canonical SMILES that account for stereochemistry. Since approximately 20% of the USPTO-50K test reactions involve stereochemistry, ignoring it during evaluation could substantially inflate the reported performance and lead to unfair comparisons. Could the authors clarify whether stereochemistry is considered during evaluation? If so, how is stereochemical information recovered from graph predictions that do not contain it? The authors should also provide results under a unified evaluation protocol that accounts for stereochemistry.

Thank you for pointing out the inconsistency between the stereochemical representation and the evaluation criterion. We acknowledge this issue and agree that reliable comparisons require consistent treatment of stereochemistry. The original manuscript did not explain this sufficiently. Below, we clarify the limitations of the original implementation and report new stereochemistry-aware experiments.

(1) Did the original evaluation account for stereochemistry?

Our inspection confirmed that the original molecular graphs did not explicitly encode atom chirality or double-bond stereochemical states, and molecules reconstructed from these graphs did not fully preserve the original stereochemical information. The original evaluation therefore cannot be regarded as strict stereochemistry-aware exact matching against the original reactants using canonical SMILES. Enabling a stereochemistry-preserving option during graph-to-SMILES conversion cannot automatically restore stereochemical labels already lost from the graph representation.

We agree that this evaluation may count predictions with identical connectivity but inconsistent stereochemistry as correct, thereby overestimating strict exact-match accuracy. To address this issue, we have revised the comparison tables to present results with and without stereochemistry separately, avoiding direct comparisons across different evaluation criteria.

(2) How can stereochemistry be recovered from graph predictions without stereochemical information?

Once stereochemical labels have been lost, atom types and connectivity alone generally cannot uniquely recover the originally specified configuration. We therefore did not attempt to repair the original model outputs through SMILES canonicalization or by copying ground-truth reactant labels. Instead, we extended the model state representation and retrained a stereochemistry-aware version, referred to below as RetroVirtSyn-Stereo.

Specifically, we added tetrahedral chirality labels to atom states and double-bond stereochemical configurations to bond states. Atom chirality uses a local representation with a consistent neighbor ordering, and double-bond stereochemistry uses a relative representation with consistent reference neighbors, to prevent node reordering from changing the meaning of the labels. This information is incorporated into the product conditions, reactant supervision labels, and model output space.

During generation, the model jointly predicts reactant structures and stereochemical states. The decoder then assigns the predicted chirality and double-bond configurations to the molecules and outputs canonical SMILES that preserve stereochemistry. Ground-truth reactant stereochemical labels are not used during inference. This is therefore explicit prediction of stereochemical states, rather than post hoc recovery of the answer from the original graphs lacking stereochemical information.

(3) Strict stereochemistry-aware evaluation and comparison with existing methods

To address the concern about evaluation fairness, we evaluate RetroVirtSyn-Stereo using exact matching of canonical SMILES that preserve stereochemistry. Ground truth is obtained from the original reactant SMILES. Predictions and ground-truth reactants undergo the same canonicalization procedure, with atom-mapping numbers removed but atom chirality and double-bond stereochemistry retained. A prediction is correct only if the complete predicted reactants exactly match the ground truth. Missing or incorrect stereochemistry is counted as a mismatch even when connectivity is correct.

Reviewer_A7XLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses error propagation in two-stage single-step retrosynthesis: incorrect reaction-center predictions lead to erroneous synthon structures through hard bond disconnection, which subsequently impairs reactant generation. The authors propose RetroVirtSyn, which encodes candidate reaction centers and their confidence scores on the intact product graph to construct virtual synthons that preserve product topology. Based on the highest reaction-center confidence, the method adaptively switches between candidate rank cutoffs of (m=1) and (m=3), and generates reactants through conditional discrete graph flow matching. On USPTO-50K under the reaction-class-unknown setting, the adaptive model achieves 63.2% Top-1 and 86.5% Top-10 exact-match accuracy. The authors further provide round-trip evaluation, ablations of synthon representations, reaction-center input interventions, and confidence-stratified analyses to support the claim that virtual synthons mitigate the effects of reaction-center prediction errors.

Critical strengths
  1. The paper proposes a reasonable improvement to information transfer in two-stage retrosynthesis, reducing the impact of reaction-center prediction errors on downstream predictions.
  2. The experiments include analyses of the error propagation mechanism.
Critical weaknesses
  1. The second stage of RetroVirtSyn employs conditional flow matching. Existing implementations of this approach discard atom- and bond-level stereochemical information and therefore cannot generate candidate reactants with specified stereochemistry, such as chirality.
Questions for authors
  1. When computing the reported accuracy, are predictions with missing stereochemical information counted as incorrect?
  2. How are the candidate reactant sets ranked to obtain the Top-k predictions?
  3. Could the authors report the performance of the proposed virtual-synthon representation when incorporated into existing two-stage retrosynthesis methods?
5 comments
AuthorsOpenReview ↗

Response to Reviewer A7XL — Critical Weakness: Stereochemical modeling (Part 1/2)

We thank the reviewer for the careful review and constructive comments. We have revised the manuscript to address all the issues you raised and marked the corresponding changes in blue for ease of reference. Our point-by-point responses follow.

Critical weakness (summarized): Lack of atom- and bond-level stereochemical modeling.

We acknowledge that the original implementation did not explicitly preserve or predict atom- and bond-level stereochemical states and therefore could not fully support reactant generation with specified stereochemistry. To address this limitation, we modified the molecular representation, training objective, and output decoding, and retrained a stereochemistry-aware extension, RetroVirtSyn-Stereo, while retaining the virtual-synthon mechanism and discrete conditional flow-matching framework.

1. Explicit stereochemical states for atoms and bonds

Each atom state jointly represents its element, formal charge, isotope, explicit hydrogen count, implicit hydrogen setting, aromaticity, and tetrahedral chirality. These attribute combinations form 67 discrete atom states, including unknown and placeholder states, allowing the model to distinguish atoms with otherwise identical attributes but different chiral configurations.

Because tetrahedral chirality tags depend on neighbor ordering, we convert them into a local representation relative to a consistent neighbor order rather than directly copying SMILES chirality symbols. An odd permutation of the neighbors swaps the two chirality tags, preventing node reordering from changing the represented stereochemistry. Input node-order features are included in the conditioning to encode this local ordering convention.

We expand the five original bond categories (no bond, single, double, triple, and aromatic) to eight, adding same-side, opposite-side, and unspecified double-bond stereochemical states relative to consistently selected reference neighbors. Encoding aligns the reference neighbors at both ends and transforms the stereo labels accordingly; decoding restores the reference atoms and double-bond configuration from the predicted states. Thus, stereochemical information is retained through both molecule-to-graph and graph-to-molecule conversion.

2. Learning and generating stereochemical states

These attributes are included in the product conditioning, intermediate graph states, ground-truth reactant labels, and output categories. The model jointly predicts reactant atom and bond states, including tetrahedral chirality and double-bond stereochemistry.

For atoms already present in the product, we retain the element-type constraint while allowing chirality, charge, and hydrogen states to change during generation. This avoids fixing the entire product atom state and thereby restricting the chemical states needed for the reactants. Predicted states are decoded into molecules and canonical isomeric SMILES.

All input attributes come from the given product. Ground-truth reactant stereochemistry is used only for training supervision and evaluation, never as an inference condition or to repair generated outputs. The stereochemical predictions therefore come from the model rather than post-processing based on test answers.

During data preparation, we also checked encoding-decoding round trips for the original products and reactants. The reconstructed canonical isomeric SMILES matched the canonical forms of the original molecules.

Strict evaluation results are provided in Part 2/2.

AuthorsOpenReview ↗

Response to Reviewer A7XL — Critical Weakness: Stereochemical modeling (Part 2/2)

This comment continues our response to the stereochemical modeling concern in Part 1/2.

3. Strict evaluation on the full test set and stereochemical subset

We evaluated RetroVirtSyn-Stereo on all 5,007 USPTO-50K test reactions and separately on the 990 reactions whose ground-truth reactants contain explicit stereochemical annotations (19.77% of the test set).

Predicted and ground-truth reactants are converted to canonical isomeric SMILES with atom-map numbers removed and stereochemistry retained. A prediction is correct only if the complete reactant set matches the ground truth. Missing or incorrect stereochemistry counts as an error even when atom connectivity is correct. Reactions with generation failures remain in the evaluation denominator.

We generated 100 samples per reaction. Valid candidates were ranked by generation frequency before computing Top-kk accuracy.

Evaluation subsetNTop-1 (%)Top-3 (%)Top-5 (%)Top-10 (%)
Full test set5,00770.9682.4883.4883.64
Stereochemical subset99069.4979.7081.0181.21

On the stereochemical subset, the model correctly predicted 688 reactions at Top-1 and 804 at Top-10. These matches require both correct reactant structure and correct stereochemistry, providing direct evidence that the extension can generate reactants with the correct stereochemical configurations.

As an evaluation control, we rescored the same predictions with their ranking unchanged, removing stereochemistry only during matching.

Evaluation subsetMatching criterionTop-1 (%)Top-3 (%)Top-5 (%)Top-10 (%)
Full test setStereo-aware70.9682.4883.4883.64
Full test setStereo-agnostic72.6283.8084.7884.94
Stereochemical subsetStereo-aware69.4979.7081.0181.21
Stereochemical subsetStereo-agnostic74.6583.9485.1585.35

This control shows that requiring correct stereochemistry affects the measured accuracy. We therefore use the stereo-aware metrics as the primary results for this extension.

Together, these changes and results show that the virtual-synthon mechanism can be combined with explicit stereochemical states to generate reactants containing tetrahedral chirality and double-bond stereochemistry. The original limitation arose from the chosen state representation and decoding scheme; extending these components enables stereochemical prediction without replacing the core conditional flow-matching framework.

AuthorsOpenReview ↗

Response to Reviewer A7XL — Question 1: Accuracy and stereochemistry

Question 1

When computing the reported accuracy, are predictions with missing stereochemical information counted as incorrect?

Thank you for this question. For the Top-kk accuracies reported in the original manuscript, missing stereochemical information was not separately counted as an error. The original implementation reconstructed both predicted and reference reactants from graph representations without explicit stereochemical labels. Consequently, predictions with matching connectivity but missing stereochemistry could be counted as correct. We acknowledge that the original manuscript did not clearly explain this evaluation protocol and have added clarification in the revised manuscript.

To address this issue, the new RetroVirtSyn-Stereo experiments use exact matching of canonical SMILES that retain stereochemical information. When the ground-truth reactants have specified stereochemistry, missing or incorrect stereochemical information in a prediction counts as an error, even if its atom connectivity is correct. This criterion applies consistently to Top-1, Top-3, Top-5, and Top-10. A hit is counted only if at least one of the first kk candidates exactly matches the ground truth in both structure and stereochemistry. The table below compares RetroVirtSyn-Stereo with baseline methods that also use stereochemistry-aware evaluation. Under this strict criterion, the model achieves Top-1/3/5/10 accuracies of 71.0%, 82.5%, 83.5%, and 83.6%, respectively, across the full test set.

CategoryModelTop-1 (%)Top-3 (%)Top-5 (%)Top-10 (%)
Semi-template-basedGraph2Edits (Zhong et al., 2023)55.177.383.489.4
Template-freeGraph2SMILES (Tu & Coley, 2022)52.966.570.072.9
Template-freeChemformer-Large (Irwin et al., 2022)54.3—62.363.0
Template-freeR-SMILES (Zhong et al., 2022)56.379.286.291.0
Template-freeEditRetro (Han et al., 2024)60.880.686.090.3
Template-freeUAlign (Zeng et al., 2024)53.577.384.690.5
OursRetroVirtSyn-Stereo71.082.583.583.6

These results show that, under strict stereochemical matching, RetroVirtSyn-Stereo achieves higher Top-1 and Top-3 accuracies than all comparison methods in the table, demonstrating its ability to accurately predict reactant structures and stereochemistry among the leading candidates.

Reviewer_m6VQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses error propagation in two-stage retrosynthesis by retaining candidate reaction centers according to prediction confidence and encoding them as virtual synthons on the intact product graph. These representations condition a discrete graph flow-matching model for reactant generation. Experiments on USPTO-50K support the effectiveness of the approach.

Critical strengths

The paper is generally clear and easy to follow. The comparisons and ablations support the usefulness of virtual-synthon conditioning.

Critical weaknesses
  1. The confidence-based selection rule is introduced from ML implementation perspective, and experiments in Appendix B support the cutoff. However, it does not directly assess whether retained hypotheses provide complementary chemical alternatives or redundant guidance. Evidence on this distinction would strengthen the selection rationale.
  2. Appendix F addresses ground-truth-versus-predicted conditioning by training with predicted centers. Please clarify whether these training predictions come from examples used to fit the center predictor and how their error distributions compare with those at inference. Separate training does not itself demonstrate or resolve the concern on training-testing distribution-shift problem.
  3. GsG_s is a constructed representation, and Equation (7) provides the reactant-generator loss. The remaining ambiguity is how the center predictor is supervised, whether it is frozen during generator training, and which parameters receive gradients from the generation objective. This part looks unclear to me.
  4. Existing ablations test virtual conditioning, center count, and confidence-based expansion. They do not isolate the impacts of confidence socres. Comparing original scores with random or fixed weights, consistently across GsG_s and CcenterC_{\mathrm{center}}, would clarify their contribution.
Questions for authors

Major questions:

See the above weakness session.

Minor presentation comment:

Figure 1 is dense. Splitting the overview, predictor architecture, and flow-matching details into smaller figures near the corresponding text would improve readability.

7 comments
AuthorsOpenReview ↗

Response to Reviewer m6VQ — Critical Weakness 1: Candidate complementarity (Part 1/2)

We thank the reviewer for the careful review, recognition of the effectiveness of virtual synthons, and constructive suggestions on the experiments and method description. We have revised the manuscript to address your comments and marked the changes in blue. Unless quoting the review, all section, figure, table, and equation references below refer to the revised manuscript.

Critical Weakness 1

The confidence-based selection rule is introduced from ML implementation perspective, and experiments in Appendix B support the cutoff. However, it does not directly assess whether retained hypotheses provide complementary chemical alternatives or redundant guidance. Evidence on this distinction would strengthen the selection rationale.

Thank you for raising this question. The original experiments in Appendices B and C examined annotated-center coverage, which products should receive additional candidates, and the effect of candidate count on generation. These results did not directly establish whether additional centers provide useful information beyond the first candidate for the same product. We therefore added two experiments. The first changes additional center positions while preserving candidate counts and confidence values. The second generates samples separately under different candidate centers and tests whether additional centers recover correct reactants missed under the first candidate, including a comparison with the same total number of samples.

1. Additional candidates cover annotated centers missed by the first candidate

Appendix B, Figure 5(b), shows that increasing the candidate count improves annotated-center coverage more strongly in the low-confidence group. The confidence threshold is τ=0.5\tau=0.5. Here, mm is the number of retained candidates for each center type, atom or bond.

GroupCenter coverage, m=1m=1Center coverage, m=3m=3Increase (percentage points)
Low confidence50.67%88.00%+37.33
High confidence91.52%99.02%+7.50

The detailed coverage in the low-confidence group is as follows.

Retained candidatesCenter coverage
Top 1 of each type50.67%
Top 2 of each type77.33%
Top 3 of each type88.00%

Expansion from m=1m=1 to m=3m=3 increases coverage by 37.33 percentage points. The annotated center follows the single-center supervision label used by the center predictor. Additional candidates thus contain correct center information missed by the first candidate, although coverage alone does not establish whether the generator uses that information.

2. Prioritizing low-confidence products is more effective than selecting products randomly

Appendix C.3, Table 10 and Figure 6, compare expanding low-confidence products with expanding the same number of randomly selected products. Both expand selected products from m=1m=1 to m=3m=3. The following changes, reported in Table 10, are in percentage points.

Product selectionChange in center coverageChange in Top-10 accuracy
Random selection, mean ± SD over five seeds+7.73 ± 2.77+0.40 ± 1.46
Low-confidence selection+37.33+3.33

Appendix C.4, Table 11, also reports generation results with fixed candidate counts on the low-confidence subset.

StrategyTop-10 graph-match accuracy
Fixed m=1m=155.33%
Fixed m=3m=358.67%
Fixed m=5m=560.67%

These results support increasing candidate counts and prioritizing low-confidence products. However, the original random control randomizes which products receive additional centers; their center positions are still chosen by the predictor. It therefore does not isolate whether gains depend on the predicted positions. The new position controls and separate-candidate generation experiments are presented in Part 2/2.

AuthorsOpenReview ↗

Response to Reviewer m6VQ — Critical Weakness 1: Candidate complementarity (Part 2/2)

This comment continues our response to Critical Weakness 1 on the complementarity of additional candidate centers.

3. Predicted additional positions outperform random or nearby positions

We conducted controls on low-confidence test reactions without updating model parameters. Settings and results are provided in Appendix K.1, Table 29.

SettingConfiguration
Samples per reaction and condition100
Generation steps per sample50
Random-position controlsThree predefined random seeds
Matched settingsSame model, sampling parameters, and per-reaction sampling seeds
RankingDescending generation frequency; ties broken by lexicographic order

Both random-position and local-neighbor expansion retain the first-ranked candidate and preserve the candidate count and set of confidence values used by predicted Top-3 fusion, replacing only the additional positions. The local control preferentially selects positions sharing an endpoint or close in graph distance, testing whether nearby positions also improve generation.

ConditionTop-10 graph-match accuracyDifference from first-ranked condition (percentage points)
First-ranked candidate only52.00%—
Predicted Top-3 fusion58.00%+6.00
Random positions, seed 202646.00%−6.00
Random positions, seed 202744.67%−7.33
Random positions, seed 202847.33%−4.67
Local-neighbor positions44.00%−8.00

Predicted Top-3 fusion improves Top-10 accuracy by 6.00 percentage points over the first-ranked condition and outperforms the random-position and local-neighbor controls. These results support the usefulness of the predicted additional positions; arbitrary added positions do not yield the same improvement.

4. Additional candidates generate correct reactants missed under the first candidate

We generated samples separately under the first, second, and third candidate centers to test whether the latter two supplement the first candidate's outputs. Here, success means that correct reactants appear anywhere in the sample set, without requiring a Top-10 rank. Results for equal samples per candidate and equal total sampling counts are provided in Appendix K.2, Table 30.

SettingMetricResult
100 samples for each of the first three candidatesFirst candidate fails, but second or third generates correct reactants20%
Same settingAbove successes whose additional candidate contains the annotated center missed by the first candidate14%
Fixed total samples, with some allocated to second and third candidatesNewly successful samples relative to using only the first candidate16%
Same settingPreviously successful samples that no longer generate correct reactants1.3%
Same settingNet increase in the proportion of successful samples14.7 percentage points

All percentages use all low-confidence samples as the denominator. The 14% group is a subset of the 20% group.

With 100 samples per candidate, 20% of all low-confidence samples fail under the first candidate but succeed under the second or third. In samples accounting for 14% of the full low-confidence subset, the additional candidate contains the annotated center missed by the first candidate.

To test whether the gain comes only from increased sampling, we also held the total number of samples fixed and allocated some samples to the second and third candidates. Compared with allocating all samples to the first candidate, 16% of samples become successful and 1.3% lose a previously correct result, giving a net gain of 14.7 percentage points.

Thus, multiple candidate centers help generate correct reactants even when the total sampling count is unchanged, supporting the complementary information provided by additional candidates.

AuthorsOpenReview ↗

Response to Reviewer m6VQ — Critical Weakness 2: Training predictions and distribution shift

Critical Weakness 2

Appendix F addresses ground-truth-versus-predicted conditioning by training with predicted centers. Please clarify whether these training predictions come from examples used to fit the center predictor and how their error distributions compare with those at inference. Separate training does not itself demonstrate or resolve the concern on training-testing distribution-shift problem.

Thank you for highlighting the source of the training predictions, their error distributions, and the implications for distribution shift.

1. Source of predicted centers used for training

In settings B and C, the predicted centers used for generator training are obtained from samples outside the center predictor's own training set. The center predictor is fitted on a separately partitioned dataset. We then freeze it and apply it to the molecules in the retrosynthesis training set; these molecules were not included in the center predictor's training set. The generator therefore receives out-of-sample center predictions rather than predictions on examples used to fit the center predictor, avoiding leakage from such overlap.

2. The same prediction procedure does not guarantee identical error distributions

We acknowledge that the error distribution of the predicted training centers is not guaranteed to match that at inference. Both stages use the same frozen center predictor, preprocessing, and prediction procedure. In contrast, setting A uses ground-truth centers for training and imperfect predicted centers at inference, so the source of its conditioning differs between the two stages.

Using predicted centers during training does not eliminate all distribution shift. The generator is trained with center predictions on the retrosynthesis training set, whereas inference uses predictions on unseen test molecules; differences between these sets may remain.

The A/B comparison in Appendix F primarily examines the effect of training with ground-truth rather than predicted centers when inference uses predicted centers. Settings B and C use predicted centers in both stages, removing this particular difference in the source of conditioning. This does not establish that their training and inference error distributions are identical.

3. Clarifications in Appendix F

We have renamed Appendix F “Comparison of Training with Ground-Truth and Predicted Reaction Centers” and clarified that settings B and C use a separately trained, frozen center predictor. Its center scores are precomputed and cached for generator training, and the generation loss does not update the center predictor.

We have also clarified the interpretation of the comparisons in Table 16. A versus B examines the effect of using ground-truth or predicted centers during training. B versus C compares hard and virtual synthons when predicted centers are used. A versus C changes both the source of training centers and the way synthons are constructed, so the improvement cannot be attributed solely to either factor. We further explain that using the same prediction procedure during training and inference does not guarantee identical center-error distributions, and these results do not establish that all forms of distribution shift have been eliminated.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.