acceptodds
Under review as a conference paper at ICLR 2027

CatalyState: Enzyme-Conditioned Refinement of Generic Transition-State Priors

Abstract

Recent transition-state generators have shown that reaction geometry can be learned as a transferable prior from reactant and product structures. Enzymatic reactions, however, pose a different modeling problem: the relevant transition region is embedded in a specific active site, whereas enzyme-specific transition-state structures are rarely available as supervision. We present CatalyState, which treats this setting as inference over a frozen reaction prior rather than as a new coordinate-generation problem. Given a generic transition-region proposal, CatalyState learns an endpoint-grounded enzyme–geometry preference and applies Catalytic Posterior Transport (CPT) to infer a constrained distribution over nearby reaction geometries. This formulation preserves transferable reaction chemistry in the prior while using scarce enzyme-specific information for local conditioning and enforcing mapped bond-edit, steric, and geometric validity explicitly. Across enzyme- and reaction-disjoint evaluations, CatalyState distinguishes native catalytic environments from corrupted pocket and catalytic-context controls, improves local proposals across multiple independently trained reaction priors, and shows transfer of the learned preference to an independent semiempirical reaction-coordinate evaluator. Strong local-search controls further show that enzyme compatibility is useful for within-system refinement without being equivalent to physical energy or transition-state accuracy. These results support adapting, rather than relearning, transferable reaction priors for enzyme-specific reaction geometry.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
enzyme catalysis, reaction geometry, posterior inference, AI for Science

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3qc8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

CatalyState does not retrain a generic reaction model; it specializes the geometry that model produces toward a particular active site. With no supervision available for enzyme-specific transition regions, the critic is trained on solved enzyme–ligand complexes to tell native arrangements from deliberately perturbed ones, and the search is bounded twice over: hard checks from atom-mapped bond changes and clashes discard inadmissible candidates, while a trust region and a displacement cap limit how far it can drift from the starting proposal. Three experiments follow: the critic separates controlled context perturbations, the search raises its own utility on all three upstream priors, and against an external GFN2-xTB evaluation used neither in training nor in search, two of the three priors show lower reaction-coordinate error while the third does not reach significance. The authors characterize the preference as a local ranking signal rather than a calibrated energy, and separately report wrong-pocket, foreign-pocket and uniform-sampling controls.

Critical strengths

1. There is an external evaluator, and there are proper negative controls.

A wrong-pocket search, a size-matched foreign-pocket contrast, an external evaluator used neither for training nor during search, and uniform sampling sitting in the same table as the proposed solver. Few papers in this area run all of that. Fewer report it when it comes out badly, and the appendices here do so plainly — one proxy, in the authors' own words, "does not establish an independent physical advantage for native context". The terminology is disciplined in the same way: nothing the method produces is called a transition state.

2. Specialize a frozen generator instead of retraining it.

When supervision in a domain is scarce and a strong generic generator already exists, retraining is expensive and liable to discard transferable structure. The alternative taken here is to leave the generator alone, let domain evidence define a preference over its output, and use explicit support and movement limits to bound how far adaptation can go. The discussion states this as a general principle, and it is one: nothing in it depends on enzymes or on transition states.

Critical weaknesses

1. Nothing outside the critic's own score shows that the enzyme matters.

The headline context result permutes chemistry inside a fixed pocket, which is one of the counterfactual families the critic is trained on. Appendix G supplies the contrast it was not trained on — a size-matched foreign pocket — and on the enzyme-OOD split the frozen critic sits near chance; the dual-critic variant stays near chance there as well, and the authors note that it does not update the critics actually driving search. Appendix E arrives at the same place from the other direction: against the independent evaluator, native context has no advantage over a wrong pocket and the interval crosses zero. Section 6.4 puts the critic's agreement with that evaluator a little above a coin flip, and the per-fold critics doing the refinement discriminate noticeably worse than the figure shown in the main text. The paper is called enzyme-conditioned. Once the judge is anything other than the critic itself, the identity of the enzyme stops making a measurable difference.

2. On the one metric it does not optimize, the method does not beat uniform sampling.

Table 2 reports the quantity the search optimizes, so it can only establish that the objective is optimizable, and even there a fixed-temperature baseline sits almost level with CPT-CEM. On the independent evaluator the picture changes: REPS is indistinguishable from uniform sampling, and a single projected gradient step is clearly better. Appendix F is more direct still — at the matched downstream initializer stage, the xTB path that serves as the evaluator is itself a better initializer than the refined output. The same appendix reports that the material-gain fraction falls sharply in the largest substrate bin, where the sample is small.

Questions for authors
  1. Section 6.1 concludes that endpoint supervision carries a transferable enzyme-context signal. But the headline context result permutes chemistry inside the same pocket, which is one of the counterfactual families the critic is trained on, and Appendix G reports that the untrained foreign-pocket contrast sits near chance on the enzyme-OOD split. Do the critics that actually drive refinement separate native from foreign pockets any better, and if not, what is being conditioned on?

  2. Section 6.2 reads the wrong-pocket control as evidence that refinement stays sensitive to the enzyme environment. That control is scored on the learned utility. Appendix E runs the same control against the independent evaluator and finds no advantage for native context. Which of the two supports the claim?

  3. Table 3 shows REPS indistinguishable from uniform sampling and worse than a projected gradient step on the independent evaluator, while Table 2 reports the quantity the search optimizes. What result establishes that the enzyme preference, rather than bounded perturbation under hard constraints, produces the reaction-coordinate improvement? Reporting the independent error with the critic removed — support and trust region kept, selection on feasibility alone — would settle this directly; please state the selection rule so the full model cannot re-enter through a final rerank.

  4. The independent evaluator is an endpoint-restrained scan, the utility carries an endpoint actor term, and the feasible set is built from endpoint-derived intervals. In what sense is the evaluation independent of the supervision?

  5. The discussion argues that upstream generation and enzyme-context modelling can be improved separately and composed at inference. Appendix F reports that the xTB-path initializer outperforms CPT-REPS at the matched downstream stage. What does the composition add once that path is available?

6 comments

We thank the reviewer for the careful assessment and for highlighting the importance of establishing whether enzyme conditioning contributes to physical refinement. We have extended our evaluation with matched critic ablations, native-versus-foreign-pocket comparisons, and additional physical validation across three reaction priors.

  1. Enzyme-specific information beyond synthetic perturbations

We evaluated the contribution of enzyme conditioning on 59 systems with catalytic-actor annotations and eligible physical references, using matched proposals, feasibility constraints, and refinement budgets.

Across React-OT, TS-DFM, and MolGEN, CatalyState reduces mean reaction-coordinate MAE relative to Geometry-only by 0.026, 0.015, and 0.021 Å, respectively, with 41/59, 40/59, and 40/59 system wins. The lower bounds of the corresponding 95% confidence intervals are +0.008, +0.003, and +0.005 Å.

Replacing the native enzyme pocket with a matched foreign pocket increases mean MAE by 0.025, 0.018, and 0.018 Å across the same three priors. Native conditioning achieves 41/59, 40/59, and 40/59 wins, with positive lower confidence bounds in all three comparisons.

We also evaluated native-versus-foreign-pocket discrimination on enzyme-, substrate-, and double-OOD endpoint views, obtaining native preferences in 202/287, 193/274, and 211/302 cases, respectively. This follow-up is evaluated separately from the original frozen-checkpoint experiment.

These physical comparisons provide additional evidence that catalytic-environment information contributes to refinement beyond the geometric preference learned from endpoint structures.

  1. Native versus wrong-pocket refinement under the physical evaluator

The original 104-system React-OT/CPT-REPS comparison showed no significant native-pocket advantage. That evaluation covered a broader cohort and used a different refinement configuration. The new comparisons isolate the effect of correct enzyme conditioning within the actor-annotated physical-reference cohort and extend the evaluation to three reaction priors.

We further tested 30 independent enzyme–reaction systems. CatalyState achieves a mean MAE reduction of 0.025 Å over Geometry-only, with 23/30 system wins, and 0.020 Å over Wrong-pocket, with 22/30 wins.

The results indicate that enzyme-specific refinement provides measurable physical benefits when catalytic-site information is available. The 104-system analysis and the new matched evaluations are reported with their respective populations and protocols.

  1. Contribution of the learned preference and constrained optimization

To isolate the learned preference, we performed a Critic-free ablation retaining the same initial proposals, hard constraints, displacement limits, and search budget. Feasible candidates are sampled and selected uniformly, with no learned-critic evaluations or final learned reranking.

Relative to Critic-free, Full CatalyState reduces mean reaction-coordinate MAE by 0.08366, 0.03627, and 0.03995 Å on React-OT, TS-DFM, and MolGEN, respectively. The corresponding 95% confidence intervals are [0.06257, 0.10470], [0.01954, 0.05334], and [0.01798, 0.06333] Å. System-level win rates are 79.7%, 66.1%, and 67.8%.

We also evaluated CPT-CEM directly against raw reaction proposals using the independent physical metric. The median MAE reductions are 0.04956 Å (React-OT, n=111), 0.02466 Å (TS-DFM, n=76), and 0.02452 Å (MolGEN, n=86), with positive bootstrap confidence intervals. CEM additionally improves upon Uniform and REPS on React-OT and TS-DFM; the corresponding differences on MolGEN remain statistically uncertain.

For the projected-gradient comparison, we corrected the displacement-budget mismatch and repeated the evaluation with the same 0.25 Å reacting-atom RMS limit. At nearly matched critic-operation budgets (514 versus 513 units), CEM is 8.28 times faster than Gradient128 (0.079 versus 0.653 s/system) and achieves higher feasible-selection coverage (77.91% versus 72.71%). Gradient128 achieves lower reaction-coordinate MAE on React-OT by 0.0209 Å.

These experiments identify complementary contributions from the learned preference, chemical feasibility constraints, and local search. The optimizer comparison also characterizes the trade-off between physical accuracy, feasible coverage, and computational cost.

  1. Independence of the physical evaluation

The GFN2-xTB evaluator provides a separate physical objective based on the interior maximum of a restrained reaction-coordinate scan. Its energies and reaction-coordinate errors are not used to train the critic or score candidates during inference.

The evaluation shares the mapped reaction, endpoint structures, and molecular environment used to construct the refinement problem. This is necessary to compare the predicted reaction geometries against a common reaction-coordinate reference. The resulting metric measures agreement with an energy-derived semiempirical reference and is interpreted accordingly.

We additionally conducted a quantum-chemical assessment on five enzyme systems, examining stationary-point convergence, the expected imaginary reaction mode, and bidirectional connectivity. CatalyState satisfies these criteria in 3/5 systems, compared with 1/5 for Geometry-only and 0/5 for the raw proposals.

This assessment provides complementary evidence concerning the physical relevance of the refined geometries.

  1. Value of composition when a physical reaction path is available

The xTB-path initializer achieves stronger performance at the matched downstream initialization stage, as reported in Appendix F. We retain this comparison.

CatalyState is designed for enzyme–reaction systems where an upstream generator is available but a reaction-specific physical path has not been computed. Its endpoint-trained preference can be applied to React-OT, TS-DFM, and MolGEN without retraining the generators or computing a new xTB reaction path for each inference.

The practical contribution is the ability to specialize transferable reaction geometries using enzyme-context information under explicit chemical and geometric constraints. The cross-prior physical improvements and matched ablations establish the utility of this approach in the evaluated setting. When a reliable physical path is already available, it remains a strong initialization option.

Thank you for the detailed response. I consider Q4 and Q5 settled.

Two things still open:

  • [W2 / Q3] Critic-free removes critic-based selection entirely, while Uniform in Table 3 keeps it and already ties REPS. So the 79.7% win rate does not separate CPT from reranking uniform samples with the same critic, and Appendix E rules out reconstructing that contrast from the existing rows. Could you report CEM against Uniform directly on the independent metric?

  • [W1 / Q2] Swapping the pocket also changes the steric support, not only the critic score. Could you run the foreign-pocket comparison under the critic-free configuration?

I noticed the posted PDF is still the original submission. Could you upload the revised version so I can go through the changes in context?

Reviewer_ZG8ZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the problem of predicting the 3D arrangement of atoms near the transition state of an enzyme-catalyzed chemical reaction, when chemical bonds are breaking and forming. The authors propose CatalyState, a framework that refines predictions from pretrained reaction models using information about the enzyme's catalytic environment. The method learns an enzyme-conditioned geometric preference function from observed enzyme–substrate complexes, using synthetic geometric and chemical perturbations as negative examples. To refine the initial predictions, the authors introduce Catalytic Posterior Transport (CPT), which combines techniques for constrained probabilistic optimization, including relative-entropy reweighting (REPS), the cross-entropy method (CEM), and trust-region constraints. CPT searches for geometries that better satisfy the learned enzyme preference while preserving chemical feasibility and remaining close to the original prediction. Experiments show that the learned preference distinguishes native enzyme environments from synthetic perturbations, improves compatibility scores across three pretrained reaction generators, and yields modest improvements in reaction-coordinate accuracy under an independent semiempirical evaluator (i.e., a computational model that approximates quantum-mechanical calculations and is calibrated to empirical measurements).

Critical strengths

Clean and modular formulation of enzyme-conditioned reaction-geometry prediction The paper proposes a useful decomposition that separates transferable reaction knowledge from enzyme-specific geometric preferences. By learning the latter from endpoint structures rather than requiring transition-state labels, the approach addresses an important supervision bottleneck. The ability to adapt different pretrained reaction generators without retraining them is particularly appealing. Although the underlying optimization techniques are established, their combination with endpoint-supervised preferences provides a potentially useful framework for enzyme-specific refinement.

Multiple complementary evaluations, including an independent physical proxy The experiments assess several aspects of the proposed approach, including enzyme-context discrimination under out-of-distribution splits and refinement across three pretrained reaction generators. Importantly, the authors also evaluate the refined geometries using an independent semiempirical reaction-coordinate metric, rather than relying solely on improvements in the learned preference score. While the physical improvements are modest and vary across generators, they provide encouraging evidence that the learned preferences capture information relevant to reaction geometry.

Critical weaknesses

Limited evidence that the physical improvements arise from enzyme-specific conditioning It is unclear whether the performance improvements actually depend on using information about the correct enzyme, or whether the method primarily learns general properties of plausible molecular geometries that would improve predictions regardless of the enzyme. Most notably, Appendix E reports that refinement using the correct enzyme pocket does not outperform refinement using a mismatched pocket under the independent physical metric (49% win rate, with a median difference near zero). This result raises a fundamental question about the role of enzyme-specific conditioning in the proposed approach. Additional ablations, such as evaluating a geometry-only critic under the same physical metric, could help isolate the contribution of enzyme-specific information. It would also be useful to understand whether the negative result reflects limitations of the learned preference or the physical evaluator's sensitivity to enzyme-specific effects.

Limited isolation of the methodological contribution relative to prior work The proposed framework combines several established ideas, including learned protein–ligand compatibility scoring (e.g., EquiScore by Cao et al., 2024), pretrained reaction-geometry generators, and constrained optimization techniques such as REPS and CEM. This in itself is not a huge issue, and the application to enzyme-specific reaction refinement is interesting. However, the experiments leave some uncertainty about which aspects of this combination are responsible for the improvements. In particular, fixed-temperature weighting performs competitively with CPT-REPS, while CEM and projected gradient achieve better reaction-coordinate accuracy in the reported solver comparisons. Since the authors appropriately position their contribution around the overall problem decomposition rather than a novel optimization algorithm, I would be interested in more targeted analysis of what the enzyme-conditioned preference and constrained transport each contribute beyond simpler alternatives.

Questions for authors

See weaknesses.

4 comments

Response to Reviewer

We thank the reviewer for the constructive assessment and for recognizing the value of separating transferable reaction priors from enzyme-specific geometric preferences. We agree that the two central questions are whether the physical improvements genuinely depend on the correct enzyme environment and what each component contributes beyond existing scoring and optimization methods. We have conducted additional controlled experiments to address both concerns.

  1. Does the correct enzyme environment improve physical refinement?

We agree that the original 104-system React-OT experiment did not establish a consistent native-pocket advantage (49.0% wins, with a median difference near zero). This result is retained. Importantly, the original evaluation includes systems without catalytic-actor annotations and uses a different critic and refinement configuration.

To directly isolate enzyme conditioning, we conducted matched physical comparisons on 59 systems with both catalytic-actor annotations and valid GFN2-xTB references. All comparisons use the same reaction proposals, geometric constraints, displacement budgets, and physical evaluator.

For React-OT, TS-DFM, and MolGEN, respectively:

  • Full versus Geometry-only: mean MAE reductions of 0.026/0.015/0.021 A˚0.026/0.015/0.021\,\mathrm{\AA}, with 41/5941/59, 40/5940/59, and 40/5940/59 system wins.
  • Native versus Wrong-pocket: mean MAE reductions of 0.025/0.018/0.018 A˚0.025/0.018/0.018\,\mathrm{\AA}, with 41/5941/59, 40/5940/59, and 40/5940/59 system wins.

The lower bounds of the 95% bootstrap confidence intervals are positive for all six comparisons. We additionally evaluated 30 independent enzyme–reaction systems, obtaining mean improvements of 0.025 A˚0.025\,\mathrm{\AA} over Geometry-only (23/30 wins) and 0.020 A˚0.020\,\mathrm{\AA} over Wrong-pocket (22/30 wins).

These results provide direct evidence that the correct enzyme environment contributes beyond generic geometric plausibility. They do not invalidate the original 104-system observation; rather, they establish a positive effect under matched conditions within the actor-supported cohort. We have clarified the cohort and protocol differences in the revised manuscript.

  1. What does the learned enzyme-conditioned preference contribute?

We performed a separate Critic-free ablation that retains the same feasibility constraints and local search protocol while removing learned preference scoring.

Across React-OT, TS-DFM, and MolGEN, Full CatalyState improves mean reaction-coordinate MAE by 0.08366/0.03627/0.03995 A˚0.08366/0.03627/0.03995\,\mathrm{\AA}, respectively, with 95% confidence intervals of [0.06257,0.10470][0.06257,0.10470], [0.01954,0.05334][0.01954,0.05334], and [0.01798,0.06333] A˚[0.01798,0.06333]\,\mathrm{\AA}.

Together, these controls distinguish three contributions: Critic-free isolates the value of learned preference, Geometry-only isolates the additional enzyme-specific information, and Wrong-pocket tests whether refinement depends on the correct catalytic environment.

We also evaluated native-versus-foreign-pocket discrimination on the enzyme-, substrate-, and double-OOD endpoint splits. Native environments are preferred on 202/287202/287, 193/274193/274, and 211/302211/302 views, respectively, providing complementary evidence beyond synthetic perturbation recognition.

  1. What is the contribution beyond existing scoring and optimization methods?

We agree that REPS, CEM, and learned protein–ligand compatibility scoring are established techniques. Our contribution is not a new generic optimizer, but their formulation and integration for a different inference problem: enzyme-conditioned refinement of frozen reaction priors using endpoint supervision without enzyme-specific transition-state labels.

Unlike general protein–ligand compatibility scoring, our preference operates on mapped reacting atoms, bond-edit information, and catalytic-site context. CPT then restricts refinement to chemically admissible local geometries while retaining the transferable structure supplied by the upstream generator.

The additional experiments help distinguish preference learning from optimization. Under matched physical evaluation, CEM improves median reaction-coordinate MAE over raw proposals by 0.0496/0.0247/0.0245 A˚0.0496/0.0247/0.0245\,\mathrm{\AA} across React-OT, TS-DFM, and MolGEN. Relative to Uniform and REPS, CEM provides stronger physical results on React-OT and TS-DFM, while the MolGEN differences remain statistically uncertain.

We acknowledge that fixed-temperature weighting remains competitive in learned-utility optimization. Thus, we do not attribute the contribution solely to adaptive reweighting.

We also corrected the previously identified gradient displacement-budget mismatch and repeated the comparison under the same 0.25 A˚0.25\,\mathrm{\AA} reacting-atom RMS constraint. With nearly identical critic-operation budgets (514 versus 513), CEM is 8.28×8.28\times faster than Gradient128 and achieves higher feasible-selection coverage (77.91% versus 72.71%). Gradient128 achieves lower reaction-coordinate error on React-OT, illustrating an accuracy–efficiency trade-off rather than universal CEM superiority.

The resulting evidence supports our intended methodological interpretation: CatalyState combines transferable reaction priors, enzyme-specific endpoint preferences, and chemically constrained local refinement. The practical benefit comes from their complementary roles, rather than from claiming novelty or universal superiority for an individual optimization algorithm.

We hope these results and the clarified methodological positioning resolve the remaining concerns.

Reviewer_ZG8ZOpenReview ↗

Thanks for resposne

I thank the authors for their response. I am happy with the response, and will keep my weak accept rating (and will argue for acceptance provided other reviewers do not find any major issues).

Thank you for your positive feedback and for taking the time to consider our response. We appreciate your support and your willingness to advocate for the paper's acceptance.

Reviewer_K6f9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CatalyState addresses enzyme-conditioned refinement of reaction geometries without enzyme-specific transition-state supervision. It combines a frozen reaction generator with an ensemble critic trained on endpoint complexes and synthetic counterfactuals. Catalytic Posterior Transport then searches near the generated structure using learned preferences, hard geometric constraints, and bounded updates. The paper evaluates endpoint discrimination, refinement across three reaction generators, and agreement with an independent GFN2-xTB reaction-coordinate proxy. The central contribution is a modular formulation that separates transferable reaction geometry from enzyme-specific preferences. This is well motivated, but the evidence is stronger for optimizing a learned local score than for improving geometry specifically because of the correct enzyme environment.

Critical strengths

Adapting existing reaction generators using endpoint supervision is a sensible response to the scarcity of enzyme-specific transition-state structures. The separation of preference, feasibility, and movement constraints makes the approach understandable and potentially reusable. The paper considers multiple upstream generators, grouped evaluation splits, out-of-fold critics, and several matched search methods. It reports feasible-selection rates alongside utility gains is helpful, and it distinguishes search randomness from critic-training variability. At a fixed displacement budget, REPS improves reaction-coordinate MAE over raw React-OT and MolGEN proposals, with median paired changes of −0.0246 Å and −0.0175 Å and confidence intervals excluding zero. The within-system utility-error ordering analysis also provides encouraging evidence that the critic captures some useful local geometric information. It distinguishes transition-region hypotheses from validated transition states, reports unfavorable controls, and explicitly explains that the implemented search does not certify inference from the stated continuous posterior.

Critical weaknesses

Appendix E reports that native-context refinement beats matched wrong-pocket refinement on only 49.0% of 104 React-OT systems. The median native-minus-wrong-pocket MAE difference is +0.0007 Å, with a confidence interval spanning zero. This is directly relevant to the central claim and deserves prominence in the main results. Table 1 shows sensitivity to synthetic context perturbations, but these perturbation families are also used during training. Recognizing corrupted chemistry or reassigned actors does not by itself demonstrate a preference for the correct enzyme over a plausible alternative. Consistent with this concern, Table 6 reports only 49.48% native-over-foreign-pocket accuracy on enzyme-OOD examples for the base endpoint critic. These findings do not prove that enzyme conditioning is ineffective, but they leave its contribution to physical refinement unresolved. On React-OT, REPS does not show a detectable reaction-coordinate advantage over uniform search: the median difference is +0.0018 Å, with a 95% interval of [−0.0035, +0.0077] Å. CEM and projected gradient achieve lower reaction-coordinate error than REPS. The reported path-normal advantage over projected gradient is approximately 0.0041 Å; its practical significance is unclear relative to the 0.0593 Å reaction-coordinate disadvantage. The authors appropriately identify CEM as the stronger practical optimizer, but the complete cross-prior physical comparison is provided only for REPS. Consequently, Table 2 mainly establishes success at optimizing the critic's own utility, while the external evaluation leaves the benefit over simpler search methods uncertain. A matched physical evaluation of the recommended CEM configuration, including a critic-free or geometry-only control, would make the contribution substantially easier to assess. Appendix A describes the 171-system benchmark as a generated/aligned coordinate cache but does not sufficiently specify how endpoint structures were constructed, aligned, or validated. This matters because endpoint information supplies both the method's constraints and the external scan's starting information. Similarly, the paper refers to a fixed refinement horizon, configured distance margins, severe-clash thresholds, and a bounded restoration phase without giving all necessary values or formulas. The xTB protocol needs more detail on system composition, treatment of the enzyme environment, protonation, restraints, and scan resolution. The physical tables evaluate 76--110 of 171 systems per prior; a breakdown of exclusions is needed to understand how representative these subsets are. The Gibbs construction in Equation 3 is standard and valid under the stated assumptions. However, the reference distribution is an isotropic Gaussian around one generated structure, and the implemented updates use current-population weighting followed by highest-utility selection. Appendix B correctly acknowledges the absence of importance correction to the fixed reference and a global posterior guarantee. I therefore understand CPT primarily as constrained local optimization motivated by a posterior objective. This can still be a useful contribution, but the manuscript should align its distributional language with what is implemented and evaluated. No posterior approximation or distributional-quality assessment is presented.

Questions for authors

How should the central enzyme-conditioning claim be interpreted given the native-versus-wrong-pocket physical result in Appendix E? Does the evaluator capture enzyme-dependent effects sufficiently to test this claim? Can the recommended CEM configuration be compared directly with raw proposals, uniform search, and projected gradient on the same physical-evaluation systems across all three priors? Under identical support constraints and search budgets, how much physical improvement remains when the learned enzyme critic is removed or replaced by the geometry-only critic? Can the authors provide the coordinate-generation and xTB protocols, missing search hyperparameters, and an accounting of exclusions from the 171-system cohort?

1 comment

We thank the reviewer for the constructive review. We agree that the key question is whether correct enzyme context improves physical geometry, beyond learned utility.

  1. Native versus wrong-pocket (the 104-system concern).

We agree that the original React-OT/CPT-REPS comparison on 104 systems did not show a native-context benefit (49.0% wins; median native-minus-wrong MAE +0.0007 A˚+0.0007\,\mathrm{\AA}, CI including zero). We retain this result.

The new enzyme-context test uses the 59 systems with both catalytic-actor annotations and eligible physical references, under matched native/foreign-pocket search. For React-OT, TS-DFM, and MolGEN, native conditioning reduces mean reaction-coordinate MAE by 0.025/0.018/0.018 Å, with 41/59, 40/59, and 40/59 wins; the 95% CI lower limits are +0.007/+0.003/+0.003 Å. A separate 30-system physical evaluation gives +0.020 Å (22/30 wins).

Because the cohort and critic/solver settings differ, the 104-system result is not overwritten. Native is also preferred on 202/287, 193/274, and 211/302 enzyme-, substrate-, and double-OOD views.

  1. Learned critic versus Geometry-only and Critic-free.

Under matched proposals, hard support, displacement limits, and physical reference, Full improves over Geometry-only by 0.026/0.015/0.021 Å on React-OT/TS-DFM/MolGEN, with 41/59, 40/59, and 40/59 wins and 95% CI lower limits of +0.008/+0.003/+0.005 Å.

A separate matched Critic-free ablation (zero learned-critic calls) gives mean Full advantages of 0.08366/0.03627/0.03995 Å, with CIs [0.06257, 0.10470]/[0.01954, 0.05334]/[0.01798, 0.06333] Å.

These results distinguish the benefit of learned local preference from the additional contribution of enzyme-specific information. We report the two ablations separately.

  1. Recommended CEM: cross-prior physics and fair search comparisons.

The new CEM-versus-raw physical comparison yields median MAE reductions of 0.0496 Å (React-OT, n=111n=111), 0.0247 Å (TS-DFM, n=76n=76), and 0.0245 Å (MolGEN, n=86n=86), each with a positive bootstrap CI.

CEM also improves upon Uniform and REPS on React-OT and TS-DFM; MolGEN's relative differences are not conclusive.

We corrected the old gradient projection-budget mismatch and reran the comparison. At the same 0.25 A˚0.25\,\mathrm{\AA} cap and approximately equal critic-operation counts (514 vs. 513), CEM is 8.28 times faster than Gradient128 (0.079 vs. 0.653 s/system) and selects feasible candidates more often (77.91% vs. 72.71%).

Gradient128 remains 0.0209 Å better in reaction-coordinate MAE on React-OT. We therefore position CEM as offering an efficiency–feasibility trade-off, rather than universal accuracy superiority.

  1. Coordinate construction, xTB details, and the 171-system denominator.

Starting from 339 matched records, we exclude 56 by source requirements and 112 with incomplete canonical ES/EP, retaining 171 systems in five enzyme/reaction-disjoint folds.

Of 342 protonation-specific xTB paths, 318 are usable and 196 yield an eligible interior maximum across 118 systems. All 171 contribute to feasibility; the 118-system physical analysis retains raw fallback, and strict solver comparisons report pair-specific denominators.

We align EP to ES by backbone Kabsch, use a 3.5 Å residue environment with explicit H/caps and protein hydrogenation at pH 7.4, and evaluate with GFN2-xTB/ALPB-water and nine restrained images.

CPT uses 6 rounds × 64 particles + 128 final candidates, a 0.25 Å reacting-atom RMS cap, ±0.35 Å bond-edit distance margins, a 1.0 Å severe-clash threshold, and 0.01 Å edit/actor non-regression tolerances.

These settings and the explicit exclusion accounting clarify the construction and evaluation of the frozen benchmark.

  1. Posterior interpretation and physical meaning.

We agree that the Gibbs distribution is the variational reference, not a guaranteed output distribution. CPT implements posterior-motivated constrained local optimization, with a highest-utility feasible final selection; we do not claim exact posterior sampling or global approximation guarantees.

The frozen xTB path maximum is independent of critic training but remains an endpoint-informed semiempirical proxy, not an experimentally established transition state.

In a separate five-enzyme quantum-chemical assessment, stationary-point, reaction-mode, and bidirectional-connectivity criteria are met for 3/5 Full candidates, versus 1/5 Geometry-only and 0/5 raw proposals.

In summary, the additional evidence directly addresses the central questions of enzyme-context specificity, learned-critic contribution, cross-prior physical improvement, and practical optimization efficiency. Together with the clarified evaluation protocol and posterior interpretation, we hope these results resolve the reviewer's principal concerns. We welcome further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.