acceptodds
Under review as a conference paper at ICLR 2027

RxnComp: Chemical Reaction Completion via Constrained Electron Redistribution Matrix Inference

Abstract

Organic reaction data extracted from patents and scientific literature underpin a wide range of reaction understanding and prediction tasks. However, many recorded reactions include only major products while omitting byproducts, leaving reaction equations incomplete and often element-imbalanced. Recovering these missing species is important for constructing chemically complete reaction records, yet manual annotation is prohibitively expensive at scale. Existing completion methods largely rely on heuristic rules, while evaluation typically depends on exact matching against curated ground-truth completions, making it difficult to assess large-scale unlabeled reaction corpora or characterize candidate quality beyond reference recovery. To address these limitations, we formulate reaction completion as a conditional generation problem and introduce **RxnComp**, a constrained generative framework that completes missing byproducts through electron redistribution matrix modeling while explicitly enforcing chemical conservation during generation. We further develop a reference-free evaluation framework with complementary proxy metrics to assess the chemical validity and overall quality of generated completion candidates without requiring ground-truth annotations. Across benchmark datasets, RxnComp consistently produces more chemically consistent and higher-quality completion candidates than existing baselines, while maintaining competitive reference-based recovery performance. These results demonstrate the potential of constrained conditional generation for scalable and chemically grounded reaction data completion.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Chemical Reaction Completion, Constrained Generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_wQWnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses missing byproducts in public reaction data by formulating reaction completion as conditional generation over missing entries of an electron-redistribution (ER) matrix. RxnComp combines autoregressive neural scoring with finite-state constrained beam search intended to enforce double electron conservation, atom/charge conservation, and valence constraints. It reports 0.738 COV-JVR on USPTO-480K and 23.35/82.38 EM@1/EM@5 on RMechDB, outperforming rule-based and sequence baselines. However, the central physical premise of the ER matrix is underspecified and unvalidated, the claimed decoder guarantee is not proven, and the main reference-free evaluation is largely self-referential. The reference-based evaluation, weak supervision, and benchmark coverage are also insufficient, with only USPTO-480K and RMechDB as main benchmarks.

Critical strengths

1、It formulates reaction completion as conditional generation over missing ER-matrix entries, using autoregressive scoring and finite-state constrained decoding, which is more chemically informed than unconstrained SMILES generation.

2、Experiments provide promising evidence: RxnComp achieves a better COV-JVR trade-off in Table 1 and much higher EM@1/EM@5 (23.35/82.38) on RMechDB in Table 2, while ablations show constrained decoding is critical for validity.

Critical weaknesses

1、The paper imposes double conservation of the ER matrix as a hard decoding constraint, but neither the main text nor the appendix provides an exact construction from RDKit reactions to (ER), (EP), and (ΔE), especially regarding diagonal entries, formal charges, radicals, aromaticity, and hypervalent species. Although Appendix A.1 acknowledges that this assumption is idealized and may fail for radicals, charged intermediates, and delocalized systems, it does not quantify how many complete reactions in datasets such as USPTO-480K and RMechDB actually satisfy (ΔE1=0), nor does it analyze whether such violations are widespread, whether they undermine test-set quality, whether they filter out correct byproducts during hard-constrained decoding, or whether they bias metrics like ECR/JVR toward compliance with the method's own constraints rather than chemical correctness. Thus, this core physical premise is both underspecified and unvalidated, potentially affecting the validity of decoding and evaluation.

2、The claimed guarantee of the finite-state decoder is not established, and the key chemical constraint remains undefined. The paper claims that FSM-guided beam search can “guarantee” a completion of the ER matrix that satisfies double conservation, symmetry, boundedness, valence validity, and nonnegative bond order. However, Eq. (2) only provides local reachability conditions--whether the remaining budget can still be filled by the remaining entries at each step. These are necessary but not sufficient for global feasibility. Remaining free variables couple multiple rows and columns, and symmetry means an off-diagonal decision affects two row budgets and two column budgets simultaneously, so local feasibility does not guarantee a globally feasible completion. Second, the paper states that one slack entry per row and column is reserved and later “uniquely” determined by the remaining budgets, but it does not rigorously prove that this construction always preserves symmetry, entry bounds [−K,K], nonnegative bond order, and valence constraints, nor does it quantify closure failures. Moreover, the set Vijchem of chemically admissible values is only described verbally in both the main text and the appendix, without an executable definition covering aromatic bonds, formal charges, radicals, multiple fragments, or hypervalent species.

3、The reference-free evaluation is largely self-referential, and the supplementary evidence in the appendix is limited. The large-scale USPTO-480K evaluation mainly relies on proxy metrics such as ECR and JVR to support claims of “chemical plausibility” and “higher-quality completion candidates.” However, ECR is precisely the matrix property enforced by the FSM decoder, JVR includes ECR, and ACR is directly derived from reactant-minus-main-product composition; thus ECR/JVR in Table 1 mainly verify compliance with the method’s own hard constraints rather than byproduct correctness, and ESR is a weak discriminator for sparse graphs. Although Appendix A.4 adds a direct evaluation on 86 reactions with recorded coproducts, a blinded expert evaluation on 50 reactions, a stratified generalization analysis on 500 reactions, and robustness experiments, these are limited in scale, rely on a single expert without inter-rater agreement, and the generalization analysis still mainly depends on proxy metrics such as COV/JVR. Therefore, the self-referential issue is not fundamentally resolved, and the main large-scale conclusions still lack sufficient independent validation.

4、The paper does not fully report the RMechDB train/validation/test split sizes, splitting criteria, whether splits are separated by reaction family, random seeds, or confidence intervals, nor does it check whether the RMechDB test set overlaps with heuristic training data or whether data leakage exists. The training set is built from over 30k heuristically generated examples, while evaluation relies heavily on proxy metrics such as COV/JVR that are shared with the generation constraints, raising a template-distillation risk. Although the 500-reaction stratified analysis in Appendix A.4 partially addresses out-of-template generalization, it still mainly depends on proxy metrics, is limited in scale, and lacks independent reference validation. The main benchmarks are only USPTO-480K and RMechDB, lacking independent, more reference-complete reaction-completion benchmarks such as CompleteRXN, which makes it difficult to support the stronger claim of a broadly deployable data-curation pipeline.

5、The paper cites CompleteRXN, Zhang et al. (2024), and electron-redistribution related work but does not directly compare with them experimentally or systematically discuss their differences in related work.

Questions for authors

1、Please give a formal, implementable definition of E, including diagonal entries, formal charges, radicals, aromaticity, hypervalent species, and multiple fragments, and explain how ΔE is reconstructed from an RDKit reaction. Before any filtering, what fraction of verified complete reactions actually satisfy ΔE1=0? Appendix A.1 acknowledges that this assumption may fail for radicals, charged intermediates, and delocalized systems, but does not quantify its prevalence or impact in USPTO-480K or RMechDB.

2、Can the authors provide a proof or precise algorithmic argument that Eq. (2) plus the slack-entry closure guarantees a globally feasible symmetric completion under all remaining budget, bound, valence, and nonnegative-bond-order constraints?

3、Vijchem is only described verbally in both the main text and the appendix. Please provide an executable definition covering permitted elements, valence states, radicals, ionic structures, aromatic bonds, and diagonal entries, and explain how these rules handle RMechDB radical reactions.

4、Please report the exact RMechDB train/validation/test split sizes, split criterion, separation by reaction family if any, overlap checks against all heuristic and self-supervised training data, random seeds, and confidence intervals.

5、Can the authors provide independent evidence that high JVR corresponds to chemically correct completion, rather than merely compliance with the method’s own hard constraints? 

6、The paper mainly evaluates on USPTO-480K and RMechDB but claims broad deployability for reaction-data curation. Could the authors add evaluation on CompleteRXN or another independent, more reference-complete reaction-completion benchmark?

8 comments
AuthorsOpenReview ↗

Replying to Reviewer wQWn

Thank you for your positive feedback and valuable suggestions. We believe that all the issues raised can be effectively addressed. Our detailed point-by-point responses are provided below.

R3-1. Formal and Implementable Definition of the ER Representation

We thank the reviewer for this comment. Our ER representation uses integer bond orders, with diagonal entries defined through electron-budget bookkeeping rather than RDKit electron annotations.

For a mapped reaction with NN atoms (including explicit hydrogens), let bijRb_{ij}^{R} and bijPb_{ij}^{P} denote reactant and product bond orders, respectively (zero for unbonded pairs). We define

\begin{cases} b_{ij}^{P}-b_{ij}^{R}, & i\neq j, \cr -\sum_{k\neq i}\Delta E_{ik}, & i=j. \end{cases}$$ Both ER matrices encode bond orders off-diagonal and atom-wise electron budgets on the diagonal, satisfying $E_P=E_R+\Delta E$. **Special chemical cases are handled as follows:** - **Charges and radicals:** Charge conservation is checked separately through CCR; radical states are not explicitly tracked. - **Aromaticity:** Aromatic bond orders are rounded to integers. Unchanged bonds cancel in $\Delta E$, while aromaticity changes may introduce approximation errors. - **Multiple fragments:** All fragments share one atom-index space, with zero bonds between disconnected atoms. - **Hypervalent species:** Approximated through element-specific valence limits (e.g., P and S). - **RDKit reconstruction:** Add explicit hydrogens, align atoms using mapping numbers (with structural-matching fallback), construct bond-order matrices, subtract, and close the diagonal. **Our representation provides a tractable framework for conservation-aware completion, rather than a complete electronic description.** We have clarified these details and limitations in the revised manuscript. ## R3-2. Validity and Limitations of the Double-Conservation Assumption We appreciate the reviewer's concern about the physical validity of double conservation and its potential exclusion of correct coproducts. **First, double conservation is guaranteed by construction.** Diagonal closure ensures $\Delta E\mathbf{1}=0$, while matrix symmetry guarantees zero column sums. Thus, all successfully reconstructed ER matrices satisfy this property, and counting nonzero row sums would not reveal the representation's chemical limitations. **The more relevant question is whether these constraints exclude correct completions.** We address this through reference-based evaluations: - **RMechDB (radical reactions):** 91.16% coverage and 82.38% EM@5 among covered reactions, recovering 187/249 references. - **GT84 (annotated USPTO):** 95% coverage and 72.62% EM@5 among covered reactions. **These results demonstrate substantial compatibility with reference completions, including radical reactions**, despite the representation's approximations. Remaining failures may reflect search, ranking, or representation limitations. **Second, ECR/JVR measure constraint satisfaction, not independent chemical correctness.** We therefore complement these proxy metrics with reference-based recovery and human expert evaluation. Unsuccessful generations count against coverage (COV = 0.841), rather than being treated as valid predictions. Overall, **we do not assume that our simplified ER representation captures every chemical mechanism**. Instead, we assess its practical applicability through coverage, reference recovery, and independent chemical evaluation, as clarified in the revised manuscript.
AuthorsOpenReview ↗

Further Replies to Reviewer wQWn

R3-3. Scope of the Finite-State Decoder Guarantee

We thank the reviewer for this detailed observation. We agree that Eq. (2) provides necessary but not sufficient conditions for global feasibility. However, this does not invalidate our guarantee, because global feasibility is enforced by terminal verification (Algorithm 2, Lines 20–23), rather than Eq. (2) alone.

Our decoder separates two mechanisms:

  1. Local pruning: Eq. (2) eliminates partial assignments whose residual electron budgets cannot be satisfied, improving search efficiency.
  2. Terminal acceptance: A completed matrix is returned only after verifying double conservation, symmetry, entry bounds, nonnegative bond orders, and the implemented valence constraints. Otherwise, it is rejected.

Thus, our guarantee concerns the feasibility of returned completions, not the existence of a feasible completion for every input. Our experiments support this distinction: no retained candidate in the evaluated artifacts violates the terminal matrix constraints. Moreover, JVR@1 improves from 1.8% to 87.7% with electron-budget constraints, while greedy decoding already achieves 62.4% coverage, increasing to 84.1% with beam width 10.

We have clarified the respective roles of local pruning and global verification in the revised manuscript.

R3-4. Correctness of the Slack-Entry Closure Mechanism

We clarify that slack-entry determination and global feasibility verification are separate operations.

Once the remaining entries are fixed, the reserved slack entries are determined by the residual electron budgets. Symmetric assignments are enforced, and the resulting matrix is subsequently checked against all global constraints, including entry bounds, nonnegative bond orders, and valence limits.

We do not claim that every locally feasible state necessarily admits a valid closure. Any completion violating these constraints is rejected by the terminal check. If no valid candidate remains, the reaction is counted as uncovered rather than returning an invalid completion.

Consequently, the reaction-level frequency of closure-related failures is bounded above by the uncovered fraction: 15.9% on USPTO at beam width 10. This bound also includes other search failures and should not be interpreted as the exact closure-failure rate. Importantly, no retained candidate in the evaluated artifacts violates the implemented terminal constraints.

We have clarified this closure-and-verification procedure in the revised appendix.

R3-5. Executable Definition of the Chemical Feasibility Mask

We clarify that Vijchem\mathcal{V}_{ij}^{\mathrm{chem}} is an explicit bond-order and valence feasibility mask, rather than an unspecified chemical-validity oracle.

A candidate bond-order change vv is permitted only when:

  1. v∈[−K,K]v\in[-K,K], and the resulting bond order is nonnegative and within the permitted range.
  2. Bond-order increases respect the element-specific valence limits.
  3. The remaining electron budgets satisfy the reachability conditions in Eq. (2).

We use fixed integer bond-order-sum limits: 1 for H/halogens, 2 for O, 3 for N, 4 for C, 5 for P, and 6 for S, with 4 as the fallback for other elements. All input elements are permitted.

Regarding special chemical cases, formal charges are checked separately through charge conservation and RDKit sanitization; radical states are not explicitly tracked. Aromatic bonds use integer bond orders, multiple fragments share a common atom-index space, and diagonal entries follow the electron-budget definition in R3-1. These approximations may reject certain charged, radical, or heteroaromatic configurations.

Importantly, we acknowledge these limitations rather than claiming universal chemical validity. RMechDB directly tests the framework on radical reactions, where RxnComp achieves 91.16% coverage and 82.38% EM@5 among covered reactions. We have added the executable constraints and their limitations to the appendix.

AuthorsOpenReview ↗

Further Replies to Reviewer wQWn

R3-6. Clarification of Reference-Free Evaluation Metrics

We appreciate the concern that ECR/JVR partly measure constraints already enforced by our decoder. We agree that constraint satisfaction should not be conflated with chemical correctness.

However, these metrics represent chemically motivated feasibility requirements, rather than arbitrary properties introduced to favor RxnComp. Atom conservation, charge conservation, and valid bond-order assignments are necessary consistency requirements for the reaction-completion task. They are evaluated identically across all methods.

These checks are not trivially satisfied by learned baselines. For example, ACR@1 is only 0.020 for Molecular Transformer, 0.019 for T5Chem, and 0.118 for Differential-AR, despite training on the same completion data. This demonstrates that conventional sequence generation does not automatically preserve the missing atom inventory.

Nevertheless, we explicitly position ECR/JVR as reference-free feasibility metrics, not measures of experimentally verified coproduct correctness. We provide independent reference-based and expert evaluations in R3-7 to assess the latter. ESR is likewise treated as a descriptive sparsity criterion rather than independent correctness evidence.

R3-7. Independent Validation of Chemical Correctness

To address the concern regarding limited independent validation, we substantially expanded our reference-based and human expert evaluations.

We now report results on three complementary benchmarks with recorded or curated coproduct references:

BenchmarkNCOVTop-1Top-5
GT84 (annotated USPTO)8495%EM 60.71%EM 72.62%
RMechDB (zero-shot)24991.16%EM 23.35%EM 82.38%
CompleteRXN (zero-shot)30,46692.45%Exact 84.31% / DeepEquiv 89.27%Exact 87.75% / DeepEquiv 91.84%

GT84 excludes two training-overlapping reactions from the original 86 annotated reactions. RMechDB evaluates transfer to radical chemistry, while CompleteRXN provides large-scale independent reference-based evaluation on 30,466 USPTO-derived reactions, without training or adaptation on this benchmark. EM@k on RMechDB is computed over covered reactions.

Furthermore, we expanded the blinded expert study from one to five chemists, obtaining an average acceptance rate of 67%, with Fleiss' κ=0.92\kappa=0.92, indicating very strong inter-rater agreement.

These results provide independent evidence beyond proxy constraint satisfaction. Reference-based evaluation measures recovery of known coproducts, while expert assessment evaluates chemical plausibility. Together, they substantially strengthen the reliability of our evaluation framework.

R3-8. RMechDB Protocol and Data Leakage

We clarify that RMechDB is exclusively used for zero-shot evaluation, with no training or validation split.

  • Dataset: 249 evaluation reactions, obtained from 517 native test records through fixed parsing, product-selection, and task-alignment filters.
  • Training separation: No RMechDB reactions are used for training, fine-tuning, checkpoint selection, or hyperparameter tuning. Therefore, family-wise train/test splitting within RMechDB is not applicable.
  • Overlap verification: We found zero overlaps between the 249 evaluation reactions and our USPTO-derived heuristic/self-supervised training data under complete-reaction, reactants-plus-main-product, canonical-reactant, and full-product matching.
  • Structural similarity: Nearest-neighbor Morgan fingerprint Tanimoto similarity averages approximately 0.25, with a maximum of approximately 0.63.
  • Reproducibility: Training seed 42; deterministic beam-search inference.

We additionally computed 95% Wilson confidence intervals: COV 91.16% [87.0%, 94.1%], EM@1 23.35% [18.3%, 29.3%], and EM@5 82.38% [76.9%, 86.8%], with EM calculated over covered reactions.

These analyses provide explicit evidence against direct training–evaluation overlap and clarify the statistical uncertainty of our results.

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

This paper addresses the widespread omission of byproducts in public reaction datasets . RxnComp formulates reaction completion as conditional generation over the missing entries of an electron redistribution (ER) matrix. The core problem is that missing byproducts leave the ER matrix incomplete, weakening electron-level supervision for mechanism modeling and reaction prediction; existing rule/template-based methods and atom-balance-constrained decoding approaches (Molecular Transformer, T5Chem) rely on hand-crafted templates or heuristics and do not guarantee electron-level physical consistency.

Critical strengths

1.Clear problem framing and formalization: the link between missing byproducts and incomplete ER matrices is made explicit (Figure 1), and reaction completion is unified as conditional generation over missing ER entries (Section 4.1), which is physically better motivated than template matching or pure SMILES sequence generation.

2.Targeted constraint design with ablation support: double conservation, symmetry, and sparsity priors are operationalized as FSM budget tracking and feasibility masking, with a slack-entry closure that avoids infeasible terminal states; the ablations in Table 3 give a reasonably clear attribution (closure constraints mainly determine JVR; capacity and supervision mainly determine COV).

3.Thoughtful evaluation protocol: reference-free proxy metrics (SVR/ACR/CCR/ECR/ESR/JVR and COV-JVR) are proposed for the large unlabeled setting, complemented by EM@k on RMechDB where references exist; the gains in Table 2 (EM@5 = 82.38, JVR@5 = 98.24) are substantial relative to baselines.

Critical weaknesses

The weak-supervision source may introduce bias, and its relation to the evaluation set needs more transparency. The heuristic template library covers only five high-frequency byproduct families (Appendix A.2); although the stratification in Table 10 partially mitigates the distillation concern, Table 4 shows that the mixed train/validation splits union ground-truth and heuristically completed reactions while the test split removes all ground-truth reactions. The overlap between training and inference distributions, and the effect of template-library size on the COV ceiling, should be stated more clearly.

Questions for authors
  1. Statistical robustness: how many raters were involved in the Table 9 assessment, what were the criteria, and what was the inter-rater agreement? What are the sample sizes and sampling procedure for the stratified subsets in Table 10?

  2. Baseline fairness: are Differential-AR and RxnComp matched in parameter count, training epochs, and beam width? What explains the COV discrepancy for SynRBL/Molecular Transformer between Tables 1 and 2? Are candidate counts and filtering protocols unified across methods?

4 comments
AuthorsOpenReview ↗

Replying to Reviewer PLfi

We sincerely thank the reviewer for the positive feedback and constructive comments. We believe that all the concerns raised can be readily addressed. Please find our detailed point-by-point responses below:

R2-1. Bias induced by heuristic weak supervision

As Appendix A.7 notes, heuristic data construction can introduce bias. We quantify the overlap as follows and will add it to Appendix A.2.

How large the overlap is. Every training label comes either from the template library (33,122 reactions) or from a recorded coproduct (824 reactions). At inference, only 3,133 of the 40,000 cleaned test reactions (7.8%) are completable by the template library (Table 4). In the 500-reaction sample of Table 10, 32 reactions (6.4%) are rule-covered and 34 (6.8%) have a residual formula that appears in the weak-label training data. Training therefore lies inside template-or-recorded support, while more than 90% of the inference distribution lies outside it.

What the model does outside that support. The template library is used only to construct training labels. It is not queried at inference and is not a candidate inventory, so the model is not restricted to the five families. On the 468 rule-uncovered reactions, RxnComp returns a candidate for 81.6% and a jointly valid Top-1 candidate for 70.7%, against 6.8% for the rule-based baseline; on the 466 reactions with unseen residual formulas the figures are 81.5% and 70.6% (Table 10).

What this shows. The model does not perform template lookup. Table 10 uses proxy metrics, so correctness outside template support is assessed separately with the reference-based evaluations on GT84, RMechDB, and CompleteRXN.


R2-2. Effect of template-library size on the COV ceiling

A direct ablation over library size is not part of the current experiments. The existing results bear on the question as follows.

  • The library does not impose a hard ceiling. COV counts reactions for which the decoder returns at least one candidate, and the decoder searches the ER space without consulting templates. Overall COV is 0.841 although only about 8% of test reactions are template-completable, and COV is 0.825 for reactions with more than 10 residual heavy atoms, where the rule-based baseline covers none (Table 10).

  • Template support still helps. COV is 1.000 on rule-covered reactions and 0.816 on rule-uncovered ones, and it decreases from 0.966 to 0.766 as similarity to the training reactions decreases (Table 10).

  • The self-supervised objective is the main driver of coverage beyond the templates. It reconstructs masked product entries and does not depend on template labels; removing it lowers COV from 0.841 to 0.445 (Table 3).


R2-3. Statistical robustness of the expert evaluation (Table 9)

As described in Appendix A.4, the assessment in Table 9 involved a single chemist, so no inter-rater agreement could be computed for it. The stated criterion is binary: a Top-1 prediction is accepted if the predicted coproduct set is judged chemically plausible and acceptable without further correction, and invalid or unparsable predictions count as unsuccessful. The appendix accordingly presents the study as preliminary supporting evidence.

We extended the expert study with four additional chemists. Together with the original rater, the five raters give RxnComp an average acceptance rate of 67%, with Fleiss' κ = 0.92 across all five raters, which indicates very strong inter-rater agreement.


R2-4. Sampling protocol for the stratified analysis (Table 10)

The 500 reactions were drawn uniformly at random from the inference-only test set in 10 rounds of 50, one round per random seed (seeds 42–51), and pooled; the pooled set was then frozen. Selection did not depend on any model's output.

The strata are assigned after sampling, so their sizes reflect the test distribution:

StratificationSubsets (N)
Template coverageRule-covered (32), rule-uncovered (468)
Residual formulaSeen in weak-label training data (34), unseen (466)
Similarity to training reactions[0, 0.25) (231), [0.25, 0.5) (182), [0.5, 0.75) (58), [0.75, 1] (29)
Residual heavy atoms0–2 (52), 3–5 (45), 6–10 (135), >10 (268)

Similarity to training reactions is the maximum Tanimoto similarity between a test reaction and the training set, using 2,048-bit radius-2 Morgan fingerprints computed on the disconnected reactant-plus-observed-major-product graph. It is a post-hoc diagnostic used only for this stratification and is not a model input or a training signal. All sampled reactions stay in the denominator, including generation and sample-construction failures, so both numbers in each cell of Table 10 are computed over all N reactions of the stratum.

AuthorsOpenReview ↗

Further Replies to Reviewer PLfi

R2-5. Fairness of the Differential-AR comparison

They share the training-data source, the evaluation pipeline, and the evaluation beam width. They differ in parameter count and schedule:

Differential-ARRxnComp
Parameters4,844,8771,535,681
Epochs40 epochs; evaluated checkpoint from epoch 19, the best on validation4
Beam width1010
Training reactions33,94633,946
Optimizer, learning rate, batchAdamW, 4e-4, 24AdamW, 1e-3, 16
Seed4242

Differential-AR has 3.15 times as many parameters and a longer schedule, so its low coverage of valid completions is not explained by smaller capacity or shorter training. Because the two models also differ in output representation and optimization settings, we treat Differential-AR as a design-level comparator. The strictly controlled evidence for the electron-budget constraint is the ablation "w/o electron budget constraint" in Table 3, which changes only the constraint within the same model. We will add the table above to the appendix.


R2-6. COV discrepancy across Tables 1 and 2

Two things differ between the tables:

Datasets. Table 1 is the USPTO inference-only test set (39,687 patent reactions) and Table 2 is RMechDB (249 elementary radical steps). COV is the fraction of reactions in a dataset for which a method returns at least one candidate, so it depends on the dataset.

Models. SynRBL is the same tool in both tables. The Molecular Transformer rows are two different checkpoints. In Table 1 it is the model we trained on the same completion data as RxnComp, with code adapted from the original implementation. In Table 2 it is the frozen CompleteRXN checkpoint of Vogel et al., so that the Molecular Transformer and CRB rows differ only in the decoding constraint.

  • SynRBL's COV goes from 0.708 to 37.35%, consistent with a rule library built for patent-style reactions matching far fewer radical steps.

  • Molecular Transformer's COV is 0.983 on USPTO and 100% on RMechDB. In both cases an unconstrained sequence model emits an output for nearly every input.

The definition of COV is the same in both tables. We will name the dataset and the checkpoint in each caption and use one number format.


R2-7. Unified candidate count and filtering protocol

The filtering and scoring protocol is identical for all methods, and we verified it in the evaluation code. All multi-candidate methods decode with beam width 10.

  • Candidate counts. RxnComp, Differential-AR, T5Chem, both Molecular Transformer checkpoints, and CRB decode with beam width 10 and are evaluated at k ∈ {1, 3, 5}. SynRBL and Rule-based Completion return a single completion, so only k = 1 applies.

  • Top-k. For a reaction with n unique candidates, Top-k uses min(k, n) of them in the method's own order. Lists are never padded or duplicated.

  • Filtering and ranking. No method's candidates are filtered or re-ranked using ground truth or ECR/JVR.

  • Denominators. Generation failures stay in the COV denominator.

  • Metric type. SVR, ACR, CCR, ECR, JVR, and EM are reaction-level existential hits. ESR is a continuous candidate-level score and is averaged over the available Top-k candidates.

An automated check confirmed Metric@1 ≤ Metric@3 ≤ Metric@5 for every binary metric and method.

AuthorsOpenReview ↗

Further Replies to Reviewer PLfi

If you have any further questions or concerns, please do not hesitate to let us know. We sincerely appreciate your positive feedback and constructive suggestions, and we would be grateful for your continued support and consideration of our paper throughout the discussion.

AuthorsOpenReview ↗

Follow-Up Before the Discussion Period Closes

Dear Reviewer PLfi,

Thank you very much for your support! We sincerely hope that our responses have satisfactorily addressed all your concerns. If so, we would be truly grateful if you could consider supporting our paper during the remaining discussion period. We believe your perspective and understanding of our clarifications could also help address similar concerns raised by other reviewers and facilitate a constructive discussion.

As the discussion period draws to a close, we would greatly appreciate it if you could let us know whether any concerns remain. Despite the limited time, we will make every effort to provide further clarifications and address any outstanding questions.

Thank you again for your time, patience, and consideration!

Best regards,
The Authors

Reviewer_HFZ1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RxnComp infers missing reaction byproducts by autoregressively completing an electron redistribution matrix conditioned on reactants and recorded major products. A finite-state decoder tracks electron budgets, masks inadmissible assignments, and uses beam search and terminal closure. Training combines a small recorded-byproduct set with heuristic completions and masked reconstruction. Evaluation includes reference-free validity and coverage on USPTO reactions, reference recovery on RMechDB and some annotated USPTO reactions, and a small blinded expert assessment.

Critical strengths
  1. The task and constrained representation are useful. Missing byproducts limit applications of electron-level reaction modeling, and generating multiple completions under atom and electron constraints offers a concrete approach. Reporting coverage plus validity helps to strengthen the results.

  2. The reported evidence extends beyond the main proxy metric. RxnComp improves COV-JVR compared to SynRBL (Table 1), while Table 3 provides a clear ablation of the effectiveness of each component. Reference-based tests and the expert study attempt to check chemical usefulness as well. These are meaningful strengths, subject to the protocol issues below.

Critical weaknesses
  1. Reliability of evaluation on unlabeled reactions. The RMechDB evaluation, the 86 annotated USPTO reactions, and the blinded expert study provide useful evidence beyond proxy validity. However, it remains unclear how well these results establish chemical plausibility across the much larger unlabeled USPTO set. The paper itself notes that the annotated USPTO subset is small and relatively simple. Please further clarify the scope of the reference benchmarks and the expert-evaluation results.

  2. Baseline comparability and candidate selection. Table 2 shows substantial gains, including over Molecular Transformer with constrained beam search, but the baseline training and decoding setup needs more detail. Were the learned baselines trained or adapted using the same completion data and input information? Please specify the constraints and decoding budgets used for the CRB baseline and explain the filtering and ranking applied to each method. The proposed method is extremely strong compared to the baselines; more details/insights would be appreciated.

Questions for authors

Main questions

Please see weakness above. Further discussion is appreciated.

Other presentation comments

  • Please explicitly define A(R) and A(P) at their first use in line 149.
  • Please remove the extra closing parenthesis in line 152.
6 comments
AuthorsOpenReview ↗

Replying to Reviewer HFZ1

Thank you for your positive feedback and valuable suggestions. We believe that all the issues raised can be effectively addressed. Our detailed point-by-point responses are provided below.

R1-1. Reliability and Scope of Evaluation on Unlabeled USPTO Reactions

We thank the reviewer for highlighting this important concern. We would like to clarify that our evaluation framework combines reference-based validation, human expert assessment, and large-scale proxy-based evaluation to provide complementary evidence of prediction correctness, chemical plausibility, and scalable hypothesis generation. Importantly, these evaluations address distinct questions rather than relying solely on proxy validity metrics.

(1) Reference-based evaluation: direct validation against known coproducts.

We evaluate RxnComp on three complementary benchmarks with recorded or curated coproduct references, providing direct evidence of prediction correctness beyond proxy metrics.

BenchmarkNScopeCOVTop-1Top-5
GT8484USPTO reactions with recorded coproducts95%EM 60.71%EM 72.62%
RMechDB249Elementary radical reactions; no adaptation91.16%EM 23.35%EM 82.38%
CompleteRXN30,466USPTO-derived curated completions; zero-shot92.45%Exact 84.31% / DeepEquiv 89.27%Exact 87.75% / DeepEquiv 91.84%

These benchmarks serve complementary purposes. GT84 provides direct evaluation on the target USPTO distribution, with two training-overlapping reactions removed from the original annotated subset. Although relatively small and containing simpler reaction patterns, it provides valuable reference-based validation. RMechDB evaluates generalization to elementary radical chemistry without adaptation, while CompleteRXN substantially expands reference-based evaluation to 30,466 USPTO-derived reactions. Notably, RxnComp achieves 84.31% top-1 exact match and 89.27% equivalence-aware match on CompleteRXN, providing additional evidence beyond the original small-scale annotated benchmark. For CompleteRXN, DeepEquiv recognizes chemically equivalent representations, while RMechDB EM@k is computed over covered reactions, consistent with Table 2.

(2) Human expert evaluation: assessing chemical plausibility.

We further expanded our blinded expert study to five chemists, achieving an average acceptance rate of 67%, with Fleiss' $\kappa = 0.92, indicating very strong inter-rater agreement. This provides independent human assessment of chemical plausibility beyond structural validity and conservation-based proxy metrics. Importantly, expert acceptance evaluates the plausibility of proposed coproducts rather than establishing that they were experimentally observed.

(3) Large-scale proxy evaluation: assessing scalable hypothesis generation.

For the 39,687 unlabeled USPTO test reactions, where ground-truth coproducts are unavailable, we employ structural validity and conservation-based metrics (SVR, ACR, CCR, ECR, and JVR) to assess whether generated hypotheses satisfy fundamental chemical constraints.

We emphasize that these proxy metrics are not intended to establish the exact correctness of experimentally observed coproducts. Instead, they evaluate whether RxnComp can consistently generate structurally valid and chemically constrained reaction hypotheses at scale, which is an essential prerequisite for reliable reaction completion.

To further examine scalability and robustness, we evaluated RxnComp on an additional 100,000 unlabeled USPTO reactions, obtaining COV = 0.837 and COV-JVR = 0.745, closely matching the original results of 0.841 and 0.738, respectively. These findings demonstrate that constraint-consistent hypothesis generation remains reliable as the evaluation scale increases.

Overall, our evaluation framework explicitly distinguishes prediction correctness, chemical plausibility, and scalability. Small-scale reference benchmarks with recorded ground truth provide direct validation, the substantially larger CompleteRXN benchmark extends reference-based assessment, and expert evaluation offers independent chemical judgment. Meanwhile, large-scale proxy metrics assess the ability to generate chemically constrained hypotheses across extensive unlabeled reaction data. Together, these complementary evaluations provide substantially stronger evidence for RxnComp's reliability than any single evaluation protocol alone, while avoiding the assumption that proxy validity necessarily implies experimentally verified correctness.

AuthorsOpenReview ↗

Further Replies to Reviewer HFZ1

R1-2. Baseline Comparability and Candidate-Selection Protocol

We thank the reviewer for raising these important questions. We have substantially clarified the training configurations, input conditions, decoding constraints, and candidate-selection protocols of all baselines. Importantly, the substantial performance improvements of RxnComp are primarily attributable to its explicit conservation-aware generation mechanism, rather than differences in candidate filtering or evaluation protocols. We provide the details below.

(1) Baseline training and implementation details

We have ensured comparable training data and input conditions wherever applicable. Specifically, Differential-AR, T5Chem, and Molecular Transformer (Table 1) are trained or fine-tuned on the same reaction-completion data as RxnComp, using the same observed reaction context (reactants and recorded major product). Their configurations are summarized below.

SettingRxnCompDifferential-ART5Chem
ImplementationOursIn-house AR baselineOfficial implementation
InitializationFrom scratchFrom scratchOfficial simple_pretrain checkpoint
Train / Validation33,946 / 2,41733,946 / 2,41733,946 / 2,417
Parameters1.54M4.84M14.71M
Training epochs440 (best checkpoint: 19)19 (best checkpoint: 14)
OptimizerAdamWAdamWAdamW
Learning rate1e-34e-45e-4
Batch size162416 × 2 accumulation
Beam width101010
Random seed424242
Decoding constraintsElectron-budget FSM + valence maskNoneNone

Differential-AR is an in-house autoregressive comparator that predicts structural differences from the same incomplete reaction context. Unlike RxnComp, it does not explicitly maintain the residual electron/bond-order budget or enforce corresponding reachability constraints during decoding. Therefore, conservation properties must be learned implicitly. We treat this model as a design-level comparator rather than a strictly parameter-matched ablation.

T5Chem uses the official implementation and pretrained checkpoint, followed by fine-tuning on our completion data. It achieves 99.79% validation coverage and 97.77% EM@1, indicating successful fitting to the in-domain completion task. However, its coverage decreases substantially on external benchmarks (9.52% on GT84 and 14.06% on RMechDB), primarily because generated outputs fail to parse or preserve the supplied major product.

Molecular Transformer (Table 1) was trained using code adapted from the original implementation (Schwaller et al., 2019), with the same completion data and inputs as RxnComp. It contains 12.09M parameters, was trained for 20,000 optimizer steps, and uses beam width 10.

(2) Clarification of the CRB baseline and zero-shot evaluation

We emphasize that CRB (Constrained Reaction Balancer) is a learned reaction-completion model equipped with constrained beam search, rather than a heuristic baseline. Both Molecular Transformer and CRB in Table 2 use the official pretrained checkpoint from CompleteRXN (Vogel et al., 2026), without additional training or adaptation on RMechDB.

Following the official implementation, CRB enforces atom and charge conservation during beam search, with hydrogen exempt from the hard atom-balance constraint. The end-of-sequence token is permitted only when the required balance conditions are satisfied. We use beam width 10, matching RxnComp's evaluation budget. The constrained and unconstrained variants share the same underlying learned model and differ only in their inference procedures.

Importantly, none of the evaluated models in Table 2, including RxnComp, is trained or fine-tuned on RMechDB. Thus, the comparison evaluates direct transfer to elementary radical chemistry rather than in-domain training performance.

(3) Unified candidate selection and evaluation

To eliminate potential bias from candidate selection, we clarify that all methods follow the same evaluation protocol:

  • Original candidate ranking: Each method is evaluated using its own generated candidate order, without external reranking.
  • No ground-truth filtering: Candidates are never filtered or reranked based on reference products, ECR, or JVR.
  • Consistent Top-k evaluation: For a reaction with fewer than kk unique candidates, only the available candidates are considered, without duplication or padding. Binary Top-k metrics are evaluated at the reaction level.
  • Full-set coverage: Coverage is calculated over the complete evaluation set, including failed generations.

Therefore, RxnComp does not receive any additional advantage from ground-truth-based candidate selection or post-hoc chemical filtering. Its conservation constraints are incorporated directly into the decoding process.

AuthorsOpenReview ↗

Further Replies to Reviewer HFZ1

(4) Why does RxnComp substantially outperform the baselines?

The performance gap can be explained by three important observations.

First, structural validity does not imply chemical conservation. In Table 1, Molecular Transformer achieves SVR@1 = 1.000 but only ACR@1 = 0.020, indicating that its outputs are syntactically valid yet frequently fail to account for the missing atoms. In contrast, RxnComp achieves ACR@1 = 0.971 by explicitly maintaining the residual atom inventory during electron-redistribution completion. This demonstrates the importance of incorporating conservation constraints into generation rather than learning them implicitly.

Second, electron-level constraints provide additional information beyond atom and charge conservation. CRB already incorporates atom- and charge-balance constraints, making it a particularly relevant constrained-generation baseline. However, RxnComp additionally models electron redistribution and explicitly enforces electron-budget consistency during decoding. These complementary constraints provide a mechanistic basis for generating chemically consistent completion hypotheses, particularly when transferring to elementary radical reactions.

Third, RxnComp generates a substantially stronger candidate set, rather than merely improving its first prediction. On RMechDB, RxnComp achieves 23.35% EM@1 and 81.94% EM@3 over covered reactions. This substantial improvement with multiple hypotheses demonstrates its ability to explore diverse, conservation-consistent completion candidates, which is especially valuable when several coproduct hypotheses may be plausible.

Overall, these additional clarifications demonstrate that our comparisons use transparent training and inference settings, consistent candidate-selection protocols, and no ground-truth-based filtering. More importantly, the performance improvements have a clear methodological motivation: RxnComp incorporates electron-level conservation into autoregressive generation, enabling more reliable exploration of chemically constrained reaction-completion hypotheses than conventional sequence generation or atom-balance-only constrained decoding. We have incorporated these implementation details and additional analyses into the revised manuscript.

R1-3. Definition of atom sets (line 149)

Thanks for pointing out this issue. We will define A(·) at first use as the multiset of atoms (element with multiplicity, explicit hydrogens included) of a set of molecules, so that atom conservation is an equality of multisets.

R1-4. Typographical correction (line 152)

Thanks for pointing out this issue. We will remove the extra closing parenthesis.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Discussion and Thanks to Reviewers

Dear Reviewers and AC,

As the discussion period draws to a close, we would like to briefly summarize the outcomes of our discussions and sincerely thank all reviewers for their time and constructive feedback.

To Reviewer wQWn: We are delighted that our responses and discussions have satisfactorily addressed all your concerns. We sincerely appreciate your expertise, thoughtful feedback, and active engagement throughout the rebuttal process, and we are especially grateful for your willingness to raise your rating in light of our clarifications and additional evidence.

To Reviewers wQWn and PLfi: We are truly grateful that both of you have expressed support for the acceptance of our paper following our discussions. We would sincerely appreciate your continued support during the upcoming reviewer discussion phase, where your expertise and understanding of our contributions would be invaluable in clarifying any remaining concerns.

To Reviewer HFZ1: We fully understand that this is a very busy period, with many papers to review and other commitments, and that you may not yet have had the chance to go through our responses. We have provided detailed point-by-point responses and additional evidence that we believe address the concerns raised in your initial review. Whenever your time allows, we would greatly appreciate it if you could take a look and let us know whether any concerns remain; we would be glad to clarify further. If you find your concerns satisfactorily addressed, we would be sincerely grateful if you would consider reassessing your rating. We also warmly welcome your participation in the upcoming reviewer discussion and hope that exchanging perspectives with the other reviewers will help facilitate a balanced assessment of our work.

Once again, we sincerely thank all reviewers and the AC for their time, patience, and thoughtful consideration. We believe our work makes meaningful and novel contributions to the community, and we are grateful for the opportunity to clarify its significance throughout the discussion process.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.