acceptodds
Under review as a conference paper at ICLR 2027

ZymPro: Transition-Aware Generative Modeling of Enzymatic Reaction Processes

Abstract

Enzymatic reactions are fundamental to biological processes and underpin a wide range of biotechnological applications. During these reactions, transition states represent key molecular configurations that reflect enzyme-specific catalytic environments, which lower the activation barriers to product formation. However, despite extensive research on enzymatic reaction prediction, transition states remain underexplored in generative modeling, leaving unclear how they can support reaction prediction. Thus, we introduce ZymPro for generative modeling of enzymatic reaction processes through explicit transition-state generation. ZymPro employs energy-based process supervision from reaction paths and introduces a novel process-corrected diffusion that refines denoising predictions toward the transition state. The learned process knowledge is then transferred to product generation and enzyme compatibility ranking. Extensive experiments demonstrate that ZymPro achieves strong performance in enzyme compatibility ranking for unseen reaction rules and high accuracy in product generation for unseen products. The code is available at https://anonymous.4open.science/r/ZymPro-26E4/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
enzymatic reaction modeling, molecular representation learning, reaction prediction, molecular generation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_WwjuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ZymPro for transition-aware modeling of enzymatic reactions. The method first constructs approximate transition states and reaction paths using computational methods, and then trains a process-corrected diffusion model to generate transition-state geometries conditioned on enzymes and reactants. The learned information is further used for product generation and enzyme compatibility ranking. Experiments show improvements over several baselines on both downstream tasks, and ablation studies suggest that transition-state supervision and the proposed Corrector are useful.

Critical strengths
  1. The problem is interesting and meaningful. Incorporating reaction process information into enzyme-conditioned prediction is a reasonable direction, and could be useful for understanding enzyme-specific reactions.

  2. The proposed method has some interesting designs, especially the process-corrected diffusion with return and advance corrections. The idea of using computational reaction paths to provide additional supervision is also reasonable.

  3. The experimental results are good, with clear improvements on enzyme compatibility ranking and product generation. The paper also includes quite extensive experiments, with different baselines, ablations, and generalization settings.

Critical weaknesses
  1. I am not fully convinced by the main claim of modeling enzymatic reaction processes through transition-state generation. The TS targets are approximated from computational reaction paths constructed using known reactants and products. The paper evaluates how well the model reproduces these computed structures, but does not really validate whether they correspond to physically meaningful transition states or reaction mechanisms. I think the current results mainly show that computationally derived process supervision can improve downstream prediction, which is somewhat different from the stronger motivation of modeling enzyme-catalyzed reaction processes. The claim should be better clarified.

  2. It is unclear whether the improvement really comes from transition-state modeling, or simply from additional reaction supervision. The TS targets are constructed using product information, and the training also includes direct supervision on bond changes. Such information is already useful for reaction prediction, even without explicitly modeling transition states. Although the ablation results show that TS-related features are useful, they do not clearly distinguish the contribution of TS-specific information from other reaction information. For example, ZymPro\T changes both the prediction target and the process supervision, making it difficult to isolate the effect of TS geometry. I think a simpler reaction-center or bond-change supervised baseline would be important here.

  3. The connection between the generated transition states and the downstream tasks is not fully demonstrated. In particular, enzyme compatibility ranking does not use the complete generated TS, but only extracts features from one denoising and correction step. Therefore, the improvement in this task may mainly come from a better pretrained representation, rather than explicit TS generation. It would be useful to better explain what information is actually transferred to the downstream tasks, and whether accurate TS geometry is necessary for the improvements.

  4. The enzyme-specific catalytic information is not sufficiently validated. The method emphasizes modeling enzyme-dependent reaction processes, but the model does not use explicit enzyme-reactant relative geometry during inference. Also, the enzyme-free ablation changes both the enzyme conditioning and the computational supervision, so it does not fully isolate the contribution of enzyme-specific information. I would like to see more direct analysis of whether changing the enzyme leads to meaningful differences in the predicted TS, and whether these differences agree with known enzyme-specific reaction behavior.

Questions for authors
  1. Can the authors clarify whether the main contribution is learning physically meaningful transition-state processes, or using computationally derived TS supervision to improve downstream prediction? What evidence supports the former interpretation?

  2. Have the authors considered a reaction-center or bond-change supervised baseline, without explicitly generating TS geometry? Since the current training already includes bond-change supervision, I think this comparison would help clarify the necessity of TS modeling.

  3. For enzyme compatibility ranking, why is only one denoising step used instead of the complete TS generation? Would using the full generated TS improve the results? Also, is there any relationship between TS prediction accuracy and downstream performance?

  4. Can the authors provide more evidence that the generated TS captures enzyme-specific catalytic effects? For example, how does the predicted TS change when different enzymes are provided for the same reactants?

1 comment

Dear Reviewer Wwju,

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

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

Reviewer_gFNvOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes ZymPro, which generates a transition-state representation for an enzymatic reaction given only the enzyme and the reactants, and then reuses that representation to improve two downstream tasks. The authors motivate the work with a diagnostic experiment showing that several published enzyme-conditioned product predictors leave their predictions essentially unchanged when the input enzyme is replaced, suggesting that these models do not actually use the enzyme. Because experimental transition-state structures are largely unavailable at scale, supervision is instead constructed computationally: both ends of the reaction are docked into the enzyme pocket, a reaction path is grown between them, and the highest-energy frame along that path is taken as an approximate transition state, with the remaining frames supervising a module that corrects predictions back toward the path. The transition state is represented as an atom-pair distance matrix and generated by a diffusion model conditioned on the enzyme and the reactants; the generator is then frozen, and its output serves both as a geometric prior for product generation and as additional evidence for ranking candidate enzymes. The method is evaluated on EnzymeMap under reaction-rule-disjoint and product-disjoint splits, with all baselines retrained on the same splits, and the authors report improvements over the baselines on enzyme ranking, product prediction, and geometric agreement with the computed references, together with ablations intended to isolate the contributions of the transition-state conditioning and the correction module.

Critical strengths

1. The enzyme-swap diagnostic in Section 2.2 stands on its own.

Replace the input enzyme and see whether the predicted product changes. Several published enzyme-conditioned predictors return the same product, one of them in most cases. The accounting behind that holds up: score coverage is kept separate from score-only comparison, missing scores go into one explicitly stated bottom tie block, the random-guess reference is derived rather than assumed, and the bootstrap is clustered by canonical reactant. That result does not depend on the rest of the paper being right, and it is worth having in the literature on its own.

2. The way geometry gets into the SMILES model is reusable on its own.

Pairwise atom distances are expanded on a Gaussian basis and pooled into token pairs through a normalized token-to-atom alignment read straight off the reactant tokenization, then added to the encoder's self-attention logits. Absolute distances and the change from the reactant enter as separate channels. ZymPro\G keeps the enzyme tokens, the site tokens, the pretraining and the training protocol and removes only this bias, so its effect is cleanly isolated. Nothing about the recipe depends on the geometry being a transition state; anyone with another source of pairwise structure could drop it into a sequence model the same way.

Critical weaknesses

1. The paper rests on the transition state, but nothing in it shows that this object gives you more than the reactant does.

Transition states carry the argument from the introduction onward, Problem 1 is stated as pϕ(T∣E,R)p_\phi(T \mid E, R), and the title commits to "Transition-Aware." What actually gets modelled is the highest-energy frame on a path drawn between two separately docked endpoints. Appendix A.2 is careful and calls this an operational reference. The main text never does.

The model does not predict DTSD^{\mathrm{TS}} directly. It predicts R0=(DTS−DR)/sΔR_0 = (D^{\mathrm{TS}} - D^{\mathrm{R}})/s_\Delta, and the reason given is that most atom-pair distances barely change during a reaction. If that is right, then D^TS=DR\hat{D}^{\mathrm{TS}} = D^{\mathrm{R}} belongs in Table 9. It costs nothing to compute. It is not there, so I have no way to tell whether the reported errors are good or bad. Nor could I find any number saying how far the target sits from the reactant to begin with.

BRDC only looks at direction, and the pairs it scores are already separated by the reactant: the ones that form are unbonded there, the ones that break are bonded. Shorten every unbonded pair, lengthen every bonded one, and you score perfectly without a model — above what is reported here. Section 4.5 and Figure 7(c) hang the entire case for the Corrector on this number. The localization AUROC in the same table survives, and that is where the argument belongs.

So two things are missing: how far this object is from the reactant, and whether it beats leaving the reactant alone.

2. The downstream numbers are real. What I cannot see is why they should be credited to the transition state.

ZymPro\T is the only control that swaps the transition state for the reaction endpoint. On ranking it comes in under ZymPro\R, and on product generation under ZymPro\G — both of which delete that information outright. The two branch scores are added and the transition-state readout starts at zero, so a branch carrying nothing should be pushed toward zero and land back on ZymPro\R. That is not what happens, and the paper does not explain it.

The ranking split is disjoint by reaction rule and nothing else. Enzyme homology is not controlled, reactant-side overlap is not controlled, and I found no mention of MMseqs2, Foldseek, sequence identity or fingerprint similarity anywhere in the paper. Candidates are then drawn after every known positive for that reactant has been removed. Put those together and memorizing which enzyme goes with which reactant is enough to rank correctly. The authors clearly know this control exists: the product-disjoint section assigns identical canonical reactant–enzyme inputs to the same split. The rule-disjoint section does not.

ZymPro\R, the direct enzyme–reactant branch on its own, already beats the strongest baseline, so some of the margin was there before any transition state entered. And no baseline gets the enzyme structure or the site annotations that ZymPro uses.

I believe this system beats these baselines. I don't yet believe it beats them because of the transition state.

Questions for authors
  1. What does the predictor D^TS=DR\hat{D}^{\mathrm{TS}} = D^{\mathrm{R}} score in Table 9, under the same records, references, selector and error definitions? And how far does the path-derived target actually sit from the reactant in the first place?

  2. What does the following rule score on the 658-reaction BRDC subset: shorten every atom pair that is unbonded in the reactant, lengthen every pair that is bonded? If it reaches the ceiling, what does Table 8 establish about the Corrector that the localization AUROC does not?

  3. Table 5 applies the enzyme swap to the predicted product. Hold the reactants fixed instead and swap in the EC-stratified alternatives already present in the ranking benchmark: how much does D^TS\hat{D}^{\mathrm{TS}} itself move? If the answer is "barely", the generated object is close to a function of the reactant alone.

  4. ZymPro\T puts an endpoint branch on top of the direct scorer, yet it ranks below ZymPro\R, which has no second branch at all. The two branch scores are summed and the transition-state readout starts at zero. How does training arrive there rather than at ZymPro\R?

  5. How many of the ranking test groups share their canonical reactant side with a training group, and how many share the exact reactant–enzyme pair? What is Top-1 on the subset that shares neither, for ZymPro and the two strongest baselines?

9 comments

Response to Reviewer gFNv — Part 1: W1/Q1 and Q2

Response to the TS reference and the reactant-copy baseline (W1, Q1)

What does the predictor D^TS=DR\widehat D^{TS}=D^R score in Table 9, under the same records, references, selector and error definitions? And how far does the path-derived target actually sit from the reactant in the first place?

We thank the reviewer for suggesting this comparison. Our additional results show that the computational references differ from the reactant geometries, and ZymPro predicts these references more accurately than the reactant-copy baseline. We have added this baseline to Table 9 and reported the comparison in the corresponding discussion.

We evaluated the reactant-copy predictor on the same 3,141 finite-reference reactions used in Table 9, with identical atom masks, references, and error definitions. Because this baseline retains the reactant geometry unchanged, its errors directly quantify how far the computational reference lies from the reactant.

The reactant-copy baseline yields a median distance RMSE of 2.538 Å and a median coordinate RMSD of 3.221 Å. ZymPro improves on both metrics, reducing median distance RMSE and median coordinate RMSD by 30.5% and 39.8%, respectively. These results show that ZymPro captures geometric changes toward the computational reference beyond retaining the input reactant geometry.

Table 1. Reactant-copy, baseline, and ZymPro geometric errors under the Table 9 evaluation protocol.

PredictorMedian distance RMSE (Å) ↓Median coordinate RMSD (Å) ↓
Reactant-copy: D^TS=DR\widehat D^{TS}=D^R2.5383.221
GoFlow [1]2.6242.981
TSDiff [2]4.6355.209
ZymPro\N2.2142.403
ZymPro1.7631.938

Manuscript revisions: We clarified the computational-reference interpretation in the abstract, Section 1 (including the contribution statements), Section 2.3, and Section 3. We added reactant-copy to Table 9 in Appendix A.9 and introduced this comparison in Section 4.4; Appendix A.11 states the physical-validation limitation.

Response to the evaluation of the Corrector (W1, Q2)

What does the following rule score on the 658-reaction BRDC subset: shorten every atom pair that is unbonded in the reactant, lengthen every pair that is bonded? If it reaches the ceiling, what does Table 8 establish about the Corrector that the localization AUROC does not?

We thank the reviewer for this suggestion. The proposed rule achieves 100% BRDC, without identifying which atom pairs actually undergo bond changes. We have therefore replaced BRDC with localization AUPR to further evaluate the Corrector’s ability to identify changing atom pairs. This complements localization AUROC by measuring precision–recall performance when changing pairs constitute only a small fraction of all atom pairs. AUPR summarizes the trade-off between retrieving changing pairs and avoiding false positives. Higher AUPR indicates more accurate identification of changing atom pairs across recall levels. We evaluate it on the same 658-reaction subset, using the same atom-pair scores and labels as localization AUROC.

We compare the full model with three ablations: ZymPro\A removes Advance Correction, ZymPro\B removes Return Correction, and ZymPro\C removes the entire Corrector. ZymPro achieves higher AUPR than all three variants, reaching 37.84, compared with 30.68 without the Corrector—a 7.16-point improvement. Together with the localization AUROC results, these findings support the Corrector’s contribution to identifying reaction-center atom pairs.

Table 2. Localization AUROC and AUPR on the 658-reaction evaluation subset. Values are multiplied by 100; higher is better.

Method / variantAUROC ↑AUPR ↑
Random Guess50.006.27
GoFlow [1]65.20 ± 1.2129.88 ± 0.83
TSDiff [2]72.93 ± 0.8527.13 ± 0.97
ZymPro79.81 ± 1.1437.84 ± 0.64
ZymPro\A77.42 ± 1.2633.76 ± 0.69
ZymPro\B75.19 ± 1.3832.33 ± 0.43
ZymPro\C71.24 ± 1.4830.68 ± 0.52

Manuscript revisions: We replaced BRDC with localization AUPR in Section 4.5, Figure 7(c), and Table 8 in Appendix A.5. Appendix A.5 defines the AUROC/AUPR evaluation, and Figure 7(c) includes the 6.27% Random Guess AUPR line. Figure 7(a–c) now displays standard-deviation error bars.

References

[1] Galustian et al. GoFlow: Efficient transition state geometry prediction with flow matching and E(3)-equivariant neural networks. Digital Discovery, 2025.

[2] Kim et al. Diffusion-based generative AI for exploring transition states from 2D molecular graphs. Nature Communications, 2024.

Dear Reviewer gFNv,

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

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

Thank you for the detailed response and the new experiments. I consider Q1 and Q2 resolved.

Three follow up questions:

  • [W2 / Q3] The direction result is convincing. But 0.053 Å is about 2% of the reaction-center change you are predicting, and the 8.63× is measured against seed noise rather than against that change. Could you regenerate the TS with your L0 EC-unmatched enzymes and report product Top-1?

  • [W2 / Q4] That joint optimization does not guarantee recovering the direct-only model describes the behaviour rather than explaining it, and leaves the other gaps in Table 6 hard to read as information content. Could you run a test-time ablation — setting uTSu^{TS} to zero in the trained full model, no retraining — to separate the branch's information contribution from training effects?

  • [W2 / Q5] Exact disjointness does not stop homology transfer — a close relative of a test enzyme can sit in training with a different reactant. Could you stratify ranking Top-1 by the observed enzyme's maximum sequence identity to the training enzymes, for ZymPro and the two strongest baselines? It needs no retraining, and if accuracy holds at low identity the concern goes away.

Reviewer_pYc8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a generative model for enzymatic reaction processes. Enzymes are important because they stabilize reactions and can also make them more energy efficient or faster. The main contrast of this approach with prior work is the focus on modelling the transition states, not just the final reaction outcome. The paper proposes process-corrected residual diffusion as a way to train both a transition-state predictor and a corrector. It also uses energy-based process supervision to get around the problem that there is not enough transition annotation data. Experiments are provided on EnzymeMap, where comparisons with several existing models are favourable to the proposed method.

Critical strengths

The paper addresses an important practical problem and focuses on handling the lack of annotated data, for which it proposes a reasonable methodology (which combines ideas that have been used in modern generative model literature, but in a useful way for this problem, and with appropriate leverage of domain knowledge). The empirical analysis is well done in terms of both comparing with other methods but especially in terms of the ablations, which elucidate the impact of different parts of the method.

Critical weaknesses

The method is relatively complex, although the ablations show that the different elements all contribute to the results. While the results are positive, accuracies are still low (eg see Table 3) and likely insufficient for practical use.

Questions for authors

Please discuss the computational cost of both training and inference in the proposed model. Also, please comment on whether the variability in the predictions of the proposed methods could be reduced, and if so how.

3 comments

Dear Reviewer pYc8,

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

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

Thank you for the detailed reply - it addresses my questions

Thank you very much for your supportive review and practical questions.

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

This paper explore the problem that traditional generative enzymatic reaction models often simplify the task to an end-to-end mapping of "reactant + enzyme -> product," neglecting the crucial physical intermediate state in the catalytic process—the transition state (TS). Furthermore, the difficulty in isolating and observing transition states experimentally leads to an extreme scarcity of labeled data for real transition states.

This paper proposes the ZymPro framework, which includes the following three innovations:

  • Computational path generation: Combining FlowDock molecular docking, molecular GSM growth string method, and GFN2-xTB semi-empirical quantum chemical calculations, an energy-based process supervision path is constructed.
  • Process-corrected residual diffusion model: An atom-pair Transformer denoiser and a two-headed corrector are designed to predict the "regression vector" pointing to the path and the "progress vector" pointing to the transition state, achieving precise geometric structure generation.
  • Downstream migration: The generated transition state structure is introduced as an attention bias into the Transformer encoder-decoder to predict the product, and single-step denoising is used to characterize the compatibility ranking of the auxiliary enzyme.

The experiments were divided into two categories:

  • Rule-Disjoint setting: Evaluating enzyme compatibility ranking (5,255 candidate groups) and transition state generation (3,142 reactions) under rules where no reaction was observed in the test set.
  • Product-Disjoint setting: Evaluating product generation accuracy under conditions where no product was observed in the test set (9,519 reactions, covering 34 EC families).
Critical strengths
  1. High Importance of Transition State Mechanism Research: This paper focuses on the long-neglected transition state mechanism in enzyme catalysis, revealing how the active site environment lowers the activation energy barrier and stabilizes intermediate conformations, providing solid physical mechanism support for connecting the enzyme catalytic microenvironment with chemical changes.

  2. Novel Diffusion Architecture Design: A process-corrected residual diffusion model is proposed, innovatively decoupling the correction term into two vectors: "regression path" and "path progression," effectively ensuring geometric rationality and directional accuracy during sampling and denoising.

  3. Highly Practical with No Product Dependency Setting: Unlike traditional TS generation models (such as TSDiff and GoFlow) that rely on known products as input, ZymPro can predict transition states based solely on reactants and enzymes, meeting the needs of forward prediction and enzyme screening in practical applications where no known products are available.

  4. Computational Path Addresses Data Scarcity: An automated computational pipeline (FlowDock + molecularGSM + GFN2-xTB) generates energy-based process supervision paths, successfully overcoming the bottleneck of severe scarcity of real experimental transition state data. 5. Solid Experiment and Dataset Design

  5. Solid experimental and dataset design: The datasets are rigorously divided into Rule-Disjoint and Product-Disjoint datasets, and strict leakage control is implemented during both the pre-training and fine-tuning stages. The datasets provide rich benchmarks for comparison, and the experimental conclusions are convincing.

Critical weaknesses
  1. Self-Circulating Logic in RQ3 Evaluation: In the RQ3 trajectory and transition state evaluation, both the supervisory signal learned by the model and the final Ground-Truth benchmark are derived from similar GFN2-xTB computational paths. This only proves that the model successfully fits the semi-empirical calculation results, and does not necessarily represent the true physical transition state.

  2. Negative Sample Selection Based on the Closed-World Assumption: In the enzyme compatibility ranking task, negative samples are constructed using EC stratified sampling of unlabeled enzymes. This closed-world assumption assumes that "unlabeled means no catalytic activity," but in reality, some unlabeled enzymes may simply lack experimental data, rather than being experimentally validated non-catalytic enzymes.

  3. Neglecting Multiple Pathways and Competing Reactions: The pipeline currently establishes only one deterministic computational path for each reaction. It does not consider the possibility of multiple competing reaction paths or reaction mechanisms that change with dynamic conformation in complex enzyme catalysis processes.

Questions for authors

When evaluating the downstream Product Generation task under the Product-Disjoint setting, the pre-trained Transition State (TS) generator is kept frozen while providing attention biases to the product model. Since the TS generator was separately trained on its own Rule-Disjoint split: Did any reaction rules present in the Product-Disjoint test set overlap with the reaction rules used during the TS generator's training phase?

1 comment

Dear Reviewer nBYB,

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

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

Comments on the paper

1 comment

Dear Reviewers,

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

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

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.