acceptodds
Under review as a conference paper at ICLR 2027

EnergyRetro: Bridging Bond Energetics and Reaction Knowledge for Multi-Step Retrosynthesis

11 views

Abstract

Retrosynthesis planning aims to identify synthetic routes by recursively transforming a target molecule into accessible starting materials. Although single-step prediction and multi-step search have advanced considerably, bond energetics and prior reaction knowledge are rarely integrated to guide both local disconnection decisions and complete-route planning. To address this gap, we propose **EnergyRetro**, a multi-agent framework that bridges bond dissociation energy (BDE)-derived information with reaction knowledge for multi-step retrosynthesis. EnergyRetro coordinates specialized agents to combine two complementary sources of information: bond energetics and reaction knowledge. Reaction knowledge retrieved from the reaction knowledge graph (RKG) provides precedents for plausible disconnections, while BDE-derived energy profiles assess their energy risk. The two signals are then integrated into reaction costs that guide multi-step planning toward complete synthetic routes. Round-trip validation provides an additional consistency check, while memory enables validated reactions and completed subroutes to be reused. EnergyRetro solves 185 of 190 USPTO-190 targets; removing energy information or RKG retrieval reduces the solve rate by 15.29 and 34.24 percentage points, respectively. It further achieves solve rates of 97.36% and 97.18% on PaRoutes n1 and n5. Finally, we apply EnergyRetro to synthesis planning for a drug candidate targeting SARS-CoV-2 Mpro, demonstrating its applicability to structure-based drug design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Drug Discovery, Knowledge Graphs, Multi-Agent Systems, Retrosynthesis Planning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7TjWOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This work proposes EnergyRetro, which feed Reaction knowledge graph (RKG) and bond dissociation energy (BDE) into an agentic single-step model and uses Retro* like algorithms for multistep retrosynthesis search. It further utilized round-trip consistency as well as RKG support to ensure the proposed reaction is chemically valid.

Critical strengths
  1. Integration of Bond Dissociation Energy (BDE) into Retrosynthesis Planning This work proposes to use BDE to ensure proposed reaction in retrosynthesis planning is chemically valid, which is quite interesting and it did improve the performance as the ablation study shows.
Critical weaknesses
  1. The experimental design does not isolate the claimed contribution. EnergyRetro simultaneously changes single-step candidate generation, reaction scoring, retrieval, memory reuse and search during retrosynthesis planning. However, the experiments do not establish which component accounts for the reported improvement. In particular, the paper does not specify which single-step model was used by the conventional baselines or compare those using the same scoring function (energy risk, round-trip, RKG support). Consequently, the results support the performance of the entire system as a whole, but do not establish the benefit of energy guidance or the proposed constrained multistep planning algorithm.

  2. Key implementation and evaluation choices are insufficiently specified for a controlled comparison. Although sec 3.4 and appendix B.4 describe and A*-like partial route search, its implementation relationship to Retro* and the substantive differences from existing planners remain unclear. Those baseline model configurations are also underspecified. Moreover, the reported planning budget uses model-specific units, where EnergyRetro may invoke LLM multiple times for one valid expansion and others may just expand a molecule or an entire route. SR@NLLMSR@N_LLM need to explicitly state such difference. The provided repository in anonymous.4open.science is empty with only a readme, makes it harder to reproduce and interpret.

Questions for authors
  1. What is the single-step model used for other baselines and how does the "Agentic Single-step Expansion" module perform compared to other single-step models?
  2. RKG construction is filtered by matching either reactants or products against reactions in USPTO-190 and Chembl-1000. However, as pointed by prior RAG-driven single-step or multi-step methods [1][2][3], similar molecules may still serve as important reference to the current molecule. Could the authors provide a similarity analysis or reactions in the RKG against test set molecules?
  3. What makes the constrained multi-step search algorithm differ from Retro*? Is it possible to provide an ablation study on energy scoring, RKG retrieval and route memory while keeping single-step model, planner, and budget fixed?
  4. What is the actual value of depth limit dretd_ret and does NLLMN_LLM accounts those multiple LLM calls during one "Agentic Single-step Expansion"? What is the intuition behind current λ\lambda settings for the reaction cost?
  5. Is it possible to provide a self-contained repo instead of a single readme file?

[1] Single-step retrosynthesis prediction by leveraging commonly preserved substructures [2] How Well Can LLMs Synthesize Molecules? An LLM-Powered Framework for Multi-Step Retrosynthesis [3] LLM-Augmented Chemical Synthesis and Design Decision Programs

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors(Part 1)

We thank the reviewer for the detailed comments and for recognizing the interest and potential of incorporating BDE information into retrosynthesis planning. We agree that the original manuscript could better separate system-level gains from the effects of individual components, and we appreciate the opportunity to clarify the experimental settings and code availability.

W1. The experimental design does not isolate the claimed contribution

We agree that EnergyRetro combines candidate generation, reaction assessment, retrieval, memory, and search. The reported full-system results should therefore be interpreted as the performance of the integrated framework. To provide more granular evidence, we conducted the following component ablations on USPTO-190:

SettingSR@500
Full EnergyRetro97.40%
Without energy guidance82.11%
Without RKG63.16%
Without Route Memory73.68%

These ablations indicate that each component contributes within the EnergyRetro implementation. We agree they do not, on their own, isolate the benefit of the planner under a fixed set of single-step proposals and identical reaction scores. We will narrow our claims accordingly and distinguish the full energy-guidance pathway from the isolated effect of the ΔBDE descriptor.

Q1. Single-step baseline configurations and Agentic Single-step Expansion

We follow the single-step model configurations benchmarked by Retro-R1: Greedy DFS and Retro* use the template-based MLP with V1 weights; EG-MCTS uses V2; PDVN uses V4; Retro-R1 uses V1 for candidate generation and V4 for reranking; and RetroAgent uses pretrained DESP weights. V1–V4 refer to pretrained variants of the same MLP architecture.

For additional context, on USPTO-50K without reaction-class labels, our energy-guided single-step module with GPT-5.5 obtains Top-1/3/5/10 accuracies of 65.7/74.1/78.2/82.8%. Published Top-1 results for LocalRetro, Graph2Edits, and EditRetro are 53.4%, 55.1%, and 60.8%, respectively. These comparisons are informative but not fully controlled, since model pretraining and retrieval resources differ. We will clarify these configurations and limitations in the paper.

Top-k exact-match accuracy (%) on USPTO-50K without reaction-class labels.

MethodTop-1Top-3Top-5Top-10
LocalRetro53.477.585.992.4
RetroComposer54.577.283.287.7
RetroKNN57.278.986.492.7
GraphRetro53.768.372.275.5
G2Retro54.174.181.286.7
MEGAN48.170.778.486.1
MARS54.676.483.388.5
DualTF53.670.774.677.0
Retroformer53.271.176.682.1
G2GT54.169.974.577.7
RetroBridge50.874.180.685.6
RetroMoE54.876.783.489.3
ChemDual46.360.167.077.4
RetroDFM-R65.083.788.992.5
RSGPT63.484.289.293.0
EnergyRetro65.774.178.282.8

Q2. Similarity between RKG reactions/products and test molecules

We agree that reaction-level filtering does not remove all structurally related precedents. Our additional canonical-product audit found target overlaps of 5.26%, 5.60%, and 3.59% and unique-intermediate overlaps of 6.35%, 9.56%, and 7.83% on USPTO-190, PaRoutes-n1, and PaRoutes-n5, respectively.

Excluding identical molecules, the median maximum Tanimoto similarity between each test target and retrieved RKG products is 0.713/0.673/0.676, respectively. The fractions with a retrieved non-identical product of similarity ≥ 0.9 are 4.21/4.31/3.99%. These analyses quantify the retrieved structural context but do not prove complete independence from benchmark-related knowledge. We will add the statistics and explain the remaining limitation.

AuthorsOpenReview ↗

Rebuttal by Authors(Part 2)

Q3. Difference from Retro* and fixed-configuration component ablations

Like Retro*, EnergyRetro uses best-first search over an AND-OR representation. Our contribution is not a claim that the underlying AND-OR search principle is new. Rather, the implementation connects BDE-derived energy risk, RKG support, and round-trip consistency to reaction costs, applies molecular validity/depth/ancestor-cycle constraints, and reuses completed subroutes through Route Memory. The reaction cost is

c(r)=1+λEe(r)+λG[1−s(r)]+λRT[1−sRT(r)].c(r)=1+\lambda_E e(r)+\lambda_G[1-s(r)]+\lambda_{RT}[1-s_{RT}(r)].

The component ablations above use the same EnergyRetro model configuration and search budget and quantify the impact of removing energy guidance, RKG, or Route Memory. A planner-only comparison using identical single-step candidates, identical scores, and identical budgets would isolate the contribution of the search procedure relative to Retro*. Given the limited rebuttal period, this comparison has not yet been completed. We will complete it as soon as possible and update the manuscript with the results.

W2/Q4. Search depth, expansion units, LLM calls, and reaction-cost weights

We agree that expansion counts must be distinguished from external LLM calls. In Table 1, N denotes method-specific planning expansions (which may be single-step or route-level); for EnergyRetro, a complete Agentic Single-step Expansion may involve multiple LLM calls. N_LLM separately counts external calls and is therefore the more direct measure for comparing agentic computation budgets.

On USPTO-190, under 100 and 300 external LLM-call limits, EnergyRetro achieves 84.21% and 88.95% solve rates; AOT* achieves 83.16% and 90.53%. EnergyRetro averages 132.6 calls across all targets (68.4 for solved targets) and consumes 48.3 million tokens. We will show both expansion and call budgets clearly, without suggesting that EnergyRetro wins at every budget.

In our experiments, we set the maximum route depth to d_max = 15 for USPTO-190, ChEMBL-1000, PaRoutes-n1, and PaRoutes-n5. Route depth counts reaction steps along a synthesis branch; it is not the number of internal LLM calls. Each invocation of the complete single-step module counts toward E_max, while its external calls contribute to N_LLM.

We use fixed weights λ_E = 0.3, λ_G = 0.5, and λ_RT = 0.5. The base cost of 1 penalizes extra reaction steps, while the additional terms incorporate assessed energy risk and lack of RKG/round-trip support. These are fixed heuristic weights; we do not claim they have been shown to be globally optimal.

W2/Q5. Code repository availability — clarification

We appreciate the reviewer's attention to reproducibility and apologize for any confusion caused by the repository state at the time of inspection. We had submitted the code repository with the paper, but the repository was still being updated, and the version visible when the reviewer accessed it may not have reflected the complete implementation. We understand why a README-only snapshot would make reproduction difficult, and we should have communicated the update status more clearly.

The anonymous repository is available at here. We will provide a concise inventory of the available agent workflow, planner, energy-tool interfaces, RKG scripts, prompts, evaluation code, environment instructions, and runnable examples. We appreciate the opportunity to clarify this point and will ensure the review-accessible snapshot is complete and clearly documented.

Thanks for the rebuttal. However, this still does not isolate different components in multistep retrosynthesis, and the provided ablation is only for its single-step part. Atop of many other ambiguity in the manuscript, I am keeping my original score. I would suggest a much more detailed and controlled ablation.

Thank you for your follow-up and valuable suggestions. We would like to clarify two aspects of our methodology and experimental evaluation.

  1. Multistep Component Ablations.

The component ablations reported in our previous rebuttal are conducted on end-to-end multistep retrosynthesis planning (SR@500), rather than single-step prediction tasks. Under the same EnergyRetro framework and search expansion budget, the full system achieves 97.40%, while removing the energy-cost contribution, RKG, and RouteMemory yields 82.11%, 63.16%, and 73.68%, respectively. These experiments evaluate the contributions of the corresponding components to multistep route-solving performance.

The previously reported λ_E = 0 setting removes the energy term from the search cost while retaining energy information in agentic candidate generation and refinement. We are conducting an additional w/o Energy experiment that removes energy information from both stages to provide a more fine-grained controlled comparison.

  1. Relationship to Retro and Planner Contributions.

EnergyRetro adopts a Retro*-like AND–OR best-first search backbone. Our contribution is not a fundamentally new search algorithm, but the integration of BDE-informed agentic reasoning, RKG-supported reaction evaluation, and RouteMemory into multistep retrosynthesis planning. Accordingly, our component ablations are performed within this common search framework to evaluate the contributions of the proposed mechanisms.

We will further clarify the ablation configurations, computational budgets, and methodological distinctions in the revised manuscript.

Reviewer_pqe4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EnergyRetro combines BDE-derived energy information, reaction knowledge retrieval, and multiple agents for retrosynthesis planning. Reaction assessments guide AND-OR search with memory reuse. Experiments report strong route coverage and improvements from both energy information and retrieval.

I like the motivation of this paper, it takes an interesting angle and has potential, but it would benefit from further refinement.

Critical strengths
  1. Connecting reaction-level energy and knowledge assessments with route-level search is a reasonable and useful design.
  2. The 97.4% solve rate on USPTO-190 is encouraging. The large performance drops after removing energy or RKG also support their usefulness.
Critical weaknesses
  1. The central energy component needs clearer specification. Eq. (1) includes bond-order changes, while the cited ALFABET model was developed for single, non-ring bonds. Handling unsupported bonds is not explained, and Appendix B.1 only qualitatively describes the mapping from energy profiles to e(r)e(r). This makes the component difficult to reproduce and assess.

  2. Reaction Memory is constructed during training, but its training data and split are not specified (Sec. 3.1, Appendix B.3). Appendix A.2 documents overlap filtering for ORD records, without clearly covering this additional memory. Since it supplies reaction precedents during evaluation, this uncertainty affects the generalization claims.

Questions for authors

See weaknesses

1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for recognizing the motivation of combining reaction-level assessments with route-level search and the encouraging benchmark performance. We appreciate the two specific concerns, which help us improve the technical precision and reproducibility of the manuscript.

W1. Specification of the energy predictor and e(r)

We agree that citing the original ALFABET model without describing our adaptation creates ambiguity. We retain its architecture and training hyperparameters but train the BDE predictor on BDE-db2, extending the reported elemental coverage to C, H, N, O, S, F, Cl, P, Br, and I. We will state this explicitly in Sec. 3.1 and Appendix B.1.

RXNMapper provides atom correspondences used to identify broken, formed, and bond-order-changed bonds. The reaction-center descriptor is computed as the sum of +D_R for broken bonds, −D_P for formed bonds, and D_R − D_P for bond-order changes. The Evaluation Agent then uses this profile, together with relation-specific RKG evidence, to assign e(r) ∈ [0, 1] and a short rationale. The executor incorporates it into the reaction cost with λ_E = 0.3. If BDE prediction is unsupported, the absence is explicitly passed to the agent rather than replaced with an unsupported value.

We agree that elemental coverage alone does not establish support for every bond type. We will specify the actual handling of bond-order changes, ring-forming/breaking edits, and unsupported structures, and provide the full scoring prompts, model settings, and input/output schemas. We will also make clear that e(r) is a heuristic score informed by bond energetics, not a directly calibrated reaction-energy or barrier prediction.

W2. Reaction Memory provenance and evaluation-time isolation

We appreciate the request for a precise description of the memory construction. Reaction Memory is built only from the Retro training split* through label-guided reflection: predicted precursors are compared against training labels, and reusable records are extracted from successful transformations. The memory is then frozen before test evaluation. Thus, no test labels or test outcomes are used to update this memory during evaluation.

Route Memory is distinct: it starts empty for each target and caches completed subroutes discovered within that target's search. The ORD-derived RKG is a third, separate source of reaction precedents, subject to its own filtering and overlap audit. We will distinguish these three resources explicitly and report the training split, record construction, and freeze protocol in Appendix B.3.

We separately audited the ORD-derived RKG after reaction-level filtering using canonical product SMILES. The following table reports benchmark targets and unique intermediates appearing as RKG products.

BenchmarkTargets appearing as RKG productsUnique intermediates appearing as RKG products
USPTO-19010/190 (5.26%)32/504 (6.35%)
PaRoutes-n1560/10,000 (5.60%)1,804/18,872 (9.56%)
PaRoutes-n5359/10,000 (3.59%)2,025/25,871 (7.83%)
ChEMBL-100021/1,000 (2.10%)N/A

ChEMBL-1000 contains target molecules without reference synthetic routes, so reference-intermediate overlap is not applicable. Excluding identical products, we also computed the maximum Tanimoto similarity between each test target and its retrieved RKG products. The table below reports the median of these per-target maxima and the fractions of targets whose maximum similarity reaches each threshold.

BenchmarkMedian maximum similarityFraction ≥ 0.8Fraction ≥ 0.9
USPTO-1900.71332.11%4.21%
PaRoutes-n10.67318.88%4.31%
PaRoutes-n50.67621.39%3.99%
ChEMBL-10000.4715.30% (53/1,000)0.70% (7/1,000)

These statistics describe the retained ORD-derived reaction precedents, while the training-only construction and pre-evaluation freeze specify the scope of Reaction Memory.

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

The paper proposes EnergyRetro, an LLM-based multi-agent framework for multi-step retrosynthesis that combines two signals. The first is reaction precedents retrieved from a Reaction Knowledge Graph (RKG) built from about 2.35M Open Reaction Database (ORD) records, using exact SMILES, canonical SMILES, fingerprint similarity and fragment matching. The second is a bond-energetics signal, ΔBDE, computed from RXNMapper atom maps and ALFABET bond-dissociation-energy predictions. Solver, Critic, Rewriter/Selector and Evaluation agents use these signals to propose and refine disconnections. An LLM then assigns an energy-risk score, an RKG-support score and a round-trip (Molecular Transformer) consistency score, which are combined into a reaction cost used by an A*-style AND-OR search with route memory. The authors report a 97.4% solve rate on USPTO-190, 84.0% on ChEMBL-1000, and about 97% solve rates with the best top-10 route recall on PaRoutes n1/n5. In the ablations, removing energy information lowers the solve rate by 15.29 points and removing RKG retrieval lowers it by 34.24 points. A SARS-CoV-2 Mpro inhibitor case study is also included.

Critical strengths
  1. The retrieval component is specified with unusual transparency, which makes it auditable. Appendix A gives the RKG schema and statistics (Tables 3 and 4), the exact Cypher queries (Listings 1 to 6), retrieval latencies (Table 6), and an explicit benchmark-overlap analysis (Table 5). Most LLM-planner papers do not provide this level of detail.

  2. The evaluation is large in scale and uses recent comparison methods. The system is tested on USPTO-190, ChEMBL-1000 and the full PaRoutes sets (two sets of 10,000 targets each), against both search-based planners and recent LLM-based ones (Retro-R1, RetroAgent, AOT*, LLM-Syn-Planner), and Table 1 reports LLM-call and token costs, which comparable work often omits. This shows the system runs at a realistic scale and provides a basis for cost comparison.

Critical weaknesses
  1. The RKG may leak test information, and the knowledge available to the compared methods is not matched. On the evidence as it stands, I consider this decisive. ORD contains the USPTO grant reactions (1976 to 2016) and USPTO-480K, which are the very sources from which USPTO-190 (Retro* protocol) and the PaRoutes reference routes were built. The authors' own analysis confirms this: before filtering, 104 of the 190 USPTO-190 routes contain at least one reaction that exactly matches an ORD record, and for PaRoutes the figure is about 36 to 39 percent (Table 5). The filter in Appendix A.2 removes a record only when the atom-map-stripped, canonicalized reactant and product signatures match exactly, and inputs with an unspecified role on the ORD side are counted as reactants, so records that list such a reagent can easily slip through. More importantly, the filter removes reactions, not products. Exact- and canonical-SMILES retrieval (Listings 1 and 2) and the path query (Listing 6) can still return other recorded syntheses of the target molecule or its route intermediates. The authors report this kind of residual product overlap on USPTO-50K (243 test products, 4.85%), but give no such analysis for USPTO-190 or PaRoutes, and ChEMBL-1000 is not filtered at all. Meanwhile, removing the RKG drops the solve rate from 97.4% to 63.16%, and the PaRoutes R@1 and R@10 metrics measure exact recovery of patent routes, which is especially sensitive to such retrieval. For the reproduced LLM baselines, Appendix B.5 does not say whether they had access to the same or a comparable reaction corpus. As a result, even if no test reaction survives verbatim, the current comparison cannot separate better planning from access to a larger, test-adjacent knowledge base. Resolving this requires rebuilding the RKG from the benchmark-sanctioned training reactions only (or excluding all test-route molecules at the product level) and rerunning every benchmark, which is beyond what the rebuttal period allows.

  2. The bond-energetics contribution emphasized in the title is not supported. Fig. 1(c) and (d) only show that ΔBDE is broadly distributed over reference reactions; they do not show that it separates plausible from implausible disconnections, which would require a comparison against negative or decoy reactions (rejected LLM proposals, incorrect template applications). Eq. 1 sums homolytic BDEs over mapped bonds only, ignoring leaving groups and by-products, and the resulting values (a median of -83.5 kcal/mol per reaction) are far from typical reaction enthalpies, so the treatment of +139.45 kcal/mol as an extreme energy risk in Sec. 4.5 is uncalibrated. The cited ALFABET model predicts BDEs only for non-ring single bonds in molecules limited to C, H, N and O, and the paper does not explain how bond-order changes, ring-forming or ring-breaking bonds, or bonds involving halogens, S, B or metals are handled, although these are common in the case-study molecule and in coupling chemistry. The ablation is also confounded: energy information enters in five places (the fifth Solver strategy, the Critic, the Rewriter, the Selector and e(r)), so the "without energy" variant simultaneously changes the number of proposals and the prompt content, while the energy term contributes at most 0.3 per step against a base cost of 1. It is therefore impossible to tell whether the 15.29-point drop comes from ΔBDE itself or from the extra proposal strategy and context. A placebo control (shuffled or random BDE values with the pipeline unchanged) would answer this directly.

Questions for authors
  1. After filtering, what fraction of USPTO-190 and PaRoutes targets and route intermediates still have exact or canonical product matches in the RKG? What fraction of reactions in your solved routes (and in R@1-matched routes) correspond to a retrieved RKG record? Can you report results with an RKG restricted to the benchmark-sanctioned training reactions, or with product-level exclusion of all test-route molecules? This is the primary question for my rating.

  2. Can you report a placebo ablation (shuffled or random ΔBDE with the pipeline unchanged) and a "λ_E = 0 only" ablation, each with mean and standard deviation over at least three runs? These would show whether the ΔBDE information itself, rather than the extra agent or context, accounts for the gain.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors(Part 1)

We thank the reviewer for the detailed and constructive assessment. We appreciate the recognition of our transparent RKG implementation, extensive evaluation, and resource accounting. We agree that benchmark overlap and the independent contribution of bond energetics deserve more careful analysis. We address each concern below.

W1 & Q1. Potential test leakage through the RKG and unmatched knowledge resources

We agree that removing exactly matching reaction records is not sufficient to exclude alternative syntheses of test products or intermediates. To quantify this residual overlap, we performed an additional audit using canonical product SMILES after the reaction-level filtering in Appendix A.2:

BenchmarkTargets appearing as RKG productsUnique intermediates appearing as RKG products
USPTO-19010/190 (5.26%)32/504 (6.35%)
PaRoutes-n1560/10,000 (5.60%)1,804/18,872 (9.56%)
PaRoutes-n5359/10,000 (3.59%)2,025/25,871 (7.83%)
ChEMBL-100021/1,000 (2.10%)N/A

ChEMBL-1000 contains target molecules without reference synthetic routes, so reference-intermediate overlap is not applicable. These results make the residual overlap explicit: most targets and intermediates do not have exact canonical product matches, but some overlap remains. We do not claim that these statistics alone rule out information advantages from structurally related reactions or alternative routes.

We also audited the maximum Tanimoto similarity between each test target and retrieved RKG products, excluding identical molecules:

BenchmarkMedian maximum similarityFraction ≥ 0.8Fraction ≥ 0.9
USPTO-1900.71332.11%4.21%
PaRoutes-n10.67318.88%4.31%
PaRoutes-n50.67621.39%3.99%
ChEMBL-10000.4715.30% (53/1,000)0.70% (7/1,000)

These results characterize the degree of structural similarity available to the retriever, which is intended to use related reaction precedents. However, similarity statistics are not a substitute for a benchmark-training-only RKG evaluation. We will distinguish system-level gains from claims about a planner operating with matched knowledge access, and clarify that the reproduced baselines do not necessarily use identical external reaction corpora.

Product-level exclusion control on USPTO-190. We excluded RKG reaction records whose canonical products match any of the 190 test targets or their 504 unique reference-route intermediates, applying the same exclusion rule to Reaction Memory. Route Memory started empty for each target. The model configuration, planner, building-block stock, and search budget were held fixed. The results are reported below.

USPTO-190 settingSR@500 (%)
Existing reaction-level filtering97.4
Strict product-level exclusion93.7
AuthorsOpenReview ↗

Rebuttal by Authors(Part 2)

W2. Physical meaning, coverage, and isolation of the bond-energetics contribution

We appreciate this concern and agree that the original description requires greater precision. Our implementation adapts the ALFABET architecture and training hyperparameters but trains on BDE-db2, extending the reported elemental coverage to C, H, N, O, S, F, Cl, P, Br, and I. This differs from directly applying the original ALFABET checkpoint. For each candidate, RXNMapper supplies mapped endpoint correspondences. Broken bonds contribute +D_R, formed bonds contribute −D_P, and bond-order changes contribute D_R − D_P; their sum defines the mapped reaction-center ΔBDE.

We emphasize that this quantity is a bond-change descriptor, not a calibrated reaction enthalpy, activation barrier, or direct proof of synthetic feasibility. Leaving groups, by-products, solvent, and other thermodynamic contributions are not fully captured. The Evaluation Agent combines the bond-level profile and RKG evidence to assign a normalized heuristic energy-risk score e(r) in [0, 1]; the reaction cost uses λ_E = 0.3. Unsupported calculations are flagged as missing rather than assigned fabricated BDE values. We will document the predictor's actual supported bond types, treatment of ring changes, and complete scoring prompts and schemas.

We also agree that the broad ΔBDE distributions in Fig. 1 are descriptive rather than a discrimination experiment against implausible disconnections. The case-study value of +139.45 kcal/mol should therefore be presented as a high value of this descriptor within the assessed examples, not as a calibrated thermodynamic risk threshold.

The λ_E = 0-only ablation removes the energy term from the reaction cost while retaining energy information in candidate generation and refinement. The three-run comparison below evaluates this cost contribution alongside a placebo control with shuffled ΔBDE.

Q2. Placebo Energy and λ_E = 0-only controls

We evaluated EnergyRetro, λ_E = 0 only, and Placebo Energy (shuffled ΔBDE) on USPTO-190, with three runs per setting. The SR@500 results (%) are reported below as individual runs and mean ± sample standard deviation.

SettingRun 1Run 2Run 3Mean ± SD
EnergyRetro98.4295.7997.3797.19 ± 1.32
λ_E = 0 only82.1178.4278.9579.83 ± 2.00
Placebo Energy68.4264.2163.1665.26 ± 2.78

EnergyRetro achieves the highest mean solve rate across these three settings. Placebo Energy also yields a lower solve rate than setting λ_E = 0, indicating that incorrect energy assignments can adversely affect planning. The placebo protocol should specify the shuffling scope and confirm that the agent pipeline, prompts, model configuration, RKG snapshot, and search budget are unchanged.

Thank you for the rebuttal. However, critical issues remain unresolved:

  1. The new evidence reinforces rather than resolves the test-leakage and unmatched-corpus concerns, and omits the requested RKG match-rate audit on solved/R@1 routes (W1 & Q1).
  2. The λ_E = 0 ablation shows no gain from energy guidance in the single-step agents over w/o energy, the Placebo shuffling protocol is left unspecified (retaining an unedited placeholder note), key bond/element handling details remain undocumented, and the manuscript PDF has not been revised (W2 & Q2). Therefore, I maintain my original rating.
AuthorsOpenReview ↗

Rebuttal by Authors

Thank you for your careful follow-up. We sincerely apologize for the remaining ambiguities and would like to clarify the following points.

  1. W1 & Q1: Additional RKG Match-Rate Audit on R@1 Routes

We conducted an additional reaction-step-level RKG overlap analysis on the R@1-correct routes from PaRoutes-n1 and PaRoutes-n5. Specifically, we examined whether each reaction step had an alternative synthetic precedent in the RKG sharing the same canonical product but involving different reactants (e.g., A + B → P versus C + D → P). Exact reaction duplicates were excluded during RKG construction.

BenchmarkR@1 RoutesExamined Reaction StepsSame-Product Matches
PaRoutes-n13,4119,893477 (4.82%)
PaRoutes-n53,00311,268432 (3.83%)

The results indicate that only 4.82% and 3.83% of the examined reaction steps in R@1-correct routes have same-product alternative reaction precedents in the RKG, respectively. These findings quantify the limited product-level overlap between the correctly recovered top-ranked synthesis routes and the retrieval knowledge graph.

  1. Regarding W2 & Q2 (Energy Ablations, Placebo Protocol, and BDE Implementation).

We sincerely apologize for mistakenly treating λ_E = 0 only and w/o Energy as equivalent ablation settings in our previous response. The former removes only the energy term from the search cost while retaining energy information in single-step agents, whereas the latter removes energy information from both components. We are conducting the separate w/o Energy experiments and will provide the corrected comparison as soon as possible. Regarding Placebo Energy, we randomly shuffle the numerical BDE values assigned to individual bonds and subsequently recompute ΔBDE from the reactant- and product-side bond changes, rather than directly shuffling the final ΔBDE values. This preserves reaction-center mapping and bond-change calculations while disrupting the correspondence between chemical bond environments and their BDE values. We also clarify that our BDE-db2-trained ALFABET-derived predictor uses explicit hydrogens and encodes endpoint elements, bond types (single, double, triple, and aromatic), and ring sizes capped at six. The feature vocabulary covers H, C, N, O, F, P, S, Cl, Br, and I. Ring opening/closure is handled through bond breaking/formation, while unseen environments, including boron and metals, map to an UNK token rather than validated BDE predictions. We apologize for the unedited placeholder and the delay in revising the manuscript PDF. We will incorporate these corrections and implementation details into the revised version.

We appreciate the reviewer's constructive criticism and will continue addressing these concerns with additional experimental evidence and clearer documentation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Additional Experiments and Clarifications

Dear Area Chair and Reviewers,

We sincerely thank you for your careful evaluations and constructive feedback. In response to your concerns, we have conducted additional controlled experiments and provided further methodological clarifications. The key updates are summarized below.

1. Controlled energy ablations in multistep retrosynthesis.

We conducted additional experiments on USPTO-190 to distinguish the contributions of energy-informed agentic reasoning and energy-aware search scoring. All settings were evaluated using end-to-end multistep SR@500 over three independent runs, with results reported as mean ± sample standard deviation.

SettingRun 1Run 2Run 3Mean ± SD (%)
EnergyRetro98.4295.7997.3797.19 ± 1.32
λ_E = 0 only82.1178.4278.9579.83 ± 2.00
Placebo Energy68.4264.2163.1665.26 ± 2.78
w/o Energy73.6874.7471.0573.16 ± 1.90

The λ_E = 0-only setting retains energy information within the single-step agents while removing the energy term from the search cost, whereas w/o Energy removes energy information from both stages. Their comparison shows a 6.67-percentage-point improvement associated with energy-informed agentic reasoning. Full EnergyRetro further improves upon λ_E = 0 only by approximately 17.4 percentage points, supporting the additional contribution of energy-aware multistep scoring. The lower performance of Placebo Energy further highlights the importance of chemically meaningful bond-energy assignments.

2. RKG overlap analysis and evaluation fairness.

We conducted additional RKG overlap audits and a strict product-level exclusion experiment. On USPTO-190, excluding RKG entries associated with test targets and reference intermediates yields an SR@500 of 93.7%, compared with 97.4% under the original reaction-level filtering. Additionally, our R@1 route analysis on PaRoutes-n1/n5 finds same-product alternative reaction precedents for only 4.82% and 3.83% of the examined reaction steps, respectively. These analyses further characterize the extent of benchmark-related information available through the RKG.

3. Methodological clarifications and reproducibility.

We clarified the BDE predictor's adaptation from ALFABET using BDE-db2, the handling of bond changes and unsupported environments, and the interpretation of ΔBDE as a heuristic bond-change descriptor. We also detailed the relationship between EnergyRetro and Retro*-like search, distinguished Reaction Memory from RouteMemory, and clarified baseline configurations, search budgets, and external LLM-call accounting.

We believe these additional experiments and clarifications strengthen the empirical evidence and provide a more precise understanding of EnergyRetro's contributions. We will incorporate the corresponding updates into the revised manuscript.

We sincerely appreciate the reviewers' valuable feedback and consideration.

Best regards,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.