acceptodds
Under review as a conference paper at ICLR 2027

Learning Beyond the Endpoints: Benchmarking and Learning Transition-State-Informed Reaction Representations

Abstract

Representation learning is central to machine learning for chemistry, yet learning from reactant and product structures alone leaves information about the intervening transformation underexplored. We investigate how much transition-state (TS) information benefits reaction representations and how these benefits can be transferred to predictors that use only reactants and products at deployment. We introduce Transition-State-Informed Representation Learning (TSIRL), a benchmark and learning framework that evaluates the value of direct TS access and its transfer to endpoint-only prediction. The benchmark comprises four audited reaction datasets with aligned reactant, product, and TS structures, fixed partitions, and five evaluation metrics for reaction barrier prediction. We further propose a simple and effective strategy: pretrain a shared molecular encoder on reactants, products, and TS structures with barrier supervision, then reuse the encoder in a standard endpoint-only predictor. We evaluate direct TS use and TS-informed pretraining across four datasets and six encoders, test reaction-scaffold generalization, extend TS exposure to five additional pretraining strategies, and compare against six established barrier-prediction baselines. TS information is useful in every evaluated task, while TS-informed pretraining consistently improves endpoint-only prediction, transfers to unseen reaction-scaffold families, generalizes across other pretraining strategies, and achieves the best performance among the evaluated baseline methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Representation Learning, Transition-State-Aware Pretraining, Activation Barrier Prediction, Geometric Graph Neural Networks, Chemical Reaction Benchmarking

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tEzQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether transition-state (TS) structures available during training improve reaction-barrier prediction from reactants and products alone. It curates four existing datasets and evaluates supervised RPTS pretraining followed by encoder transfer and endpoint-only fine-tuning across six encoders.

Critical strengths

The distinction between direct TS access and benefits retained after removing TS inputs is useful. The evaluation across four datasets and six encoders, together with paired uncertainty estimates and dataset-specific barrier definitions, helps substantiate the findings.

Critical weaknesses
  1. Limited incremental contribution and modest performance gains. The method adds TS inputs to barrier-supervised training and subsequently transfers a shared encoder. This is a straightforward and intuitive extension, and its methodological contribution appears limited. Tables 2 and 3 show substantial gains with direct TS access, but only a modest benefit is retained through RPTS pretraining. The limited endpoint-only improvements leave the practical effectiveness of the proposed approach insufficiently established.

  2. Insufficient access to supporting materials and documentation. Page 15 refers to exact identifiers and integrity metadata in a machine-readable release without providing an access location. Pages 14–15 similarly defer construction scripts, and p.22 defers checkpoint lineage. I could not locate these resources. Machine-readable resources should be accompanied by clear, reviewer-accessible documentation.

Questions for authors
  1. Where can reviewers access the release referenced on pp.15 and 22?

  2. Could the authors explain why training-time TS access yields only modest endpoint-only gains, and discuss how more of this benefit might be retained at inference?

4 comments
AuthorsOpenReview ↗

Response: Benchmark Release, Transfer Effectiveness, and Further TS Recovery

We sincerely appreciate your valuable feedback. We have conducted additional experiments in response to your comments and will incorporate the corresponding results into a future version of the paper.

Weakness 2 & Question 1 — Access to materials

We have updated the materials in our anonymous repository. They are available now.

Reproduction instructions and checkpoint details are provided in docs/REPRODUCIBILITY.md and REVIEWER_DEMO.md.

Weakness 1 & Question 2 — Regarding TS information gains and further TS recovery

To understand why encoder transfer yields such gains and how more of the TS benefit might be recovered, we therefore explore an extension of our current transfer strategy: supplying generated TS structures to the existing RPTS encoder and prediction head, without additional training or fine-tuning.

We evaluated this extension on Grambow using React-OT and TS-DFM to generate TS structures from R/P.

All columns use the same 1,159 Grambow validation reactions; MAE (kcal/mol) is mean ± SD over three downstream seeds. A0 (RP-pretrained) and A1 (TS-informed encoder transfer) use only R/P.

EncoderRPTS: reference TS│A0: RP-pretrainedA1: TS-pretrained → R/P│RPTS: React-OT-generated TSRPTS: TS-DFM-generated TS
SchNet-1M4.490 ± 0.029│11.843 ± 0.30511.015 ± 0.264│5.552 ± 0.138 (85.6%)5.343 ± 0.049 (88.4%)
PaiNN3.900 ± 0.012│11.369 ± 0.18110.449 ± 0.164│5.289 ± 0.100 (81.4%)5.062 ± 0.116 (84.4%)
DimeNet++3.736 ± 0.167│11.701 ± 0.15311.731 ± 0.693│5.420 ± 0.228 (78.9%)5.172 ± 0.275 (82.0%)
GemNet-T3.929 ± 0.078│12.218 ± 0.07511.757 ± 0.095│5.286 ± 0.048 (83.6%)5.139 ± 0.051 (85.4%)
ClofNet3.525 ± 0.067│8.785 ± 0.1127.881 ± 0.259│7.744 ± 0.328 (19.8%)7.118 ± 0.295 (31.7%)
MACE3.330 ± 0.088│9.264 ± 0.0678.624 ± 0.135│8.310 ± 0.691 (16.1%)7.566 ± 0.588 (28.6%)

Percentages indicate the fraction of the A0-to-reference-TS MAE gap recovered.

Both generation routes yield lower mean MAE than A1 across all six encoders. TS-DFM closes 82.0–88.4% of the MAE gap for four encoders and 28.6–31.7% for ClofNet and MACE.

These results suggest that the transfer gains may partly reflect how TS information is used. Encoder transfer carries training-time TS knowledge into an endpoint-only model through its parameters, whereas generation supplies an explicit TS structure for each reaction. The larger gains through generation are consistent with useful TS-related information not being fully captured by encoder transfer alone. Our existing RPTS encoder and head can exploit this additional route without retraining, extending the framework beyond endpoint-only encoder transfer.

Encoder transfer retains its practical advantage: it improves endpoint-only prediction without running a TS generator. While the approach is simple, our experiments demonstrate its effectiveness across multiple datasets and encoders, as well as its compatibility with different pretraining objectives. Our aim is to establish compatibility with different pretraining objectives, rather than claim that barrier-supervised pretraining is the best objective. This will be explained in the following part.

Weakness 1 — Methodological contribution

We appreciate you raising this important concern. We agree that the approach is relatively simple. However, our experiments demonstrate its effectiveness across multiple datasets and encoders, and its compatibility with different pretraining objectives. Our contribution is a general TS-informed pretrain-and-transfer framework, evaluated across different encoders and pretraining objectives. The results are summarized below:

The following two tables report test MAE in kcal/mol.

Endpoint-only transfer. Across four datasets and six encoders, all 24 pairs show lower mean MAE after TS-informed transfer. Two Grambow examples are:

EncoderRP-pretrained MAERPTS-pretrained MAERelative MAE reduction
ClofNet8.6148.0506.5%
SchNet-1M11.56910.6747.7%

Compatibility with different pretraining objectives. On Grambow–ClofNet, every evaluated objective has lower mean endpoint-only MAE when TS is included in pretraining:

Pretraining objectiveRP MAERPTS MAE
Barrier-supervised8.6148.050
CoordDenoise7.8717.711
GeoSSL-DDM7.8207.705
EMPP-l17.7927.753
C-FREE-Norm7.9827.978
Energy-Referenced7.6657.594

The Table 5 comparison is within each objective: these supervised and self-supervised alternatives are components of our strategy. Together, the results of Table 5 support applicability beyond one encoder or objective.

AuthorsOpenReview ↗

Brief Follow-up on Our Rebuttal and Additional Results

Dear Reviewer tEzQ,

Thank you again for your thoughtful feedback. We would like to briefly follow up on our responses regarding the practical benefits of TS-informed learning, the scope of our contribution, and access to supporting materials.

Our responses provide the anonymous repository and reproduction instructions, clarify the framework’s applicability across encoders and pretraining objectives, and present additional generated-TS experiments using our existing frozen predictors without further training or fine-tuning.

If your schedule permits, we would greatly appreciate a brief indication of whether these additions address your concerns, or which points would benefit from further clarification.

Thank you for your time and consideration.

Best regards, The Authors

Thank you for the response. The additional materials and experiments address some of my concerns. However, my major concern remains. The general benefit of pretraining for downstream prediction is well established in the community, while the incremental gains from TS-informed encoder transfer remain small.

Overall, the revisions do not sufficiently change my assessment of the contribution, and I am inclined to maintain my rating.

AuthorsOpenReview ↗

Clarifying the Benchmark and Framework Contributions

Thank you for following up and for acknowledging that the additional materials and experiments address some of your concerns. We would like to clarify the overall scope of our contribution: we formulate the question of how TS information can benefit endpoint-conditioned prediction, establish a diagnostic benchmark and extensible framework, and develop two deployment strategies, with generated-TS prediction delivering substantial gains on both Grambow test splits. Encoder pretraining and transfer constitute one component of this broader contribution.

The benchmark asks a more specific question. It separates the direct predictive value of TS structures from the benefit accessible when only R/P structures are supplied at deployment. Across four audited datasets and six encoders, we evaluate these questions with leakage controls, alongside complementary grouped-random and unseen-family evaluations on Grambow. The gap between direct TS access and encoder transfer is itself an informative finding: it motivates examining how different strategies make TS-related predictive benefit accessible.

The framework accommodates different TS-information transfer strategies. Encoder transfer is one such strategy and can incorporate different supervised and self-supervised pretraining objectives. Our endpoint-only comparison uses an RP-pretrained control, so it examines the incremental benefit of TS-informed initialization over RP pretraining. Table 5 extends this comparison to five additional objectives, showing compatibility with both supervised and self-supervised pretraining. These objectives are alternative components of the framework, rather than competitors to a single prescribed pretraining method.

We also introduce a generation-based prediction strategy within this framework. It combines an existing TS generator with the trained RPTS encoder and prediction head, keeping all components frozen. This strategy requires no additional training or fine-tuning and is instantiated with both React-OT and TS-DFM. Both variants yield lower mean MAE than encoder transfer across all six encoders on the Grambow grouped-random and unseen-family test splits (Tables 3–4). This strategy also outperforms all six established baselines evaluated on the Grambow grouped-random test set, using only R/P structures as external inputs at deployment. With ClofNet and TS-DFM, it achieves an MAE of 6.912 kcal/mol, compared with 10.870 kcal/mol for the strongest external baseline, a 36.4% reduction (Table 16). Thus, the framework supports both encoder transfer without TS generation and explicit structural recovery using existing generators.

Our contribution therefore combines a diagnostic benchmark, an extensible framework, and two practical strategies for using TS-informed learning. We do not position the work as another standalone 3D predictor whose contribution rests solely on state-of-the-art MAE, nor as a demonstration that pretraining generally helps. We would appreciate your reconsideration of the work based on these contributions together, rather than the encoder-transfer gains alone.

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

The paper studies how much transition-state (TS) structure helps reaction-barrier prediction and whether that benefit survives when the TS is available only during training. It curates four existing reactant/product/TS datasets with endpoint-pair grouped splits, a reaction-family split and several metrics, and proposes barrier-supervised RPTS pretraining of a shared 3D encoder that is then transferred into an endpoint-only predictor. Direct TS input lowers MAE substantially, and RPTS pretraining gives positive mean MAE gains for endpoint-only models across six encoders and four datasets.

Critical strengths

Clear benchmark question. Separating the direct value of TS input from the value retained after TS removal is a precise and useful framing for deployment without TS.

Barrier definitions, TS validation standards and QC exclusions are documented per dataset, and grouping by endpoint pair prevents alternative-TS leakage.

Critical weaknesses
  1. The transfer gain is not attributed under matched controls.

(i) The RP-pretrained control continues the already converged RP parent, while the RPTS branch receives an encoder from a second supervised run that also processes an additional geometric state, and pretraining compute and learning rates are not matched; one disclosed adaptation cell also changes the learning rate between branches.

(ii) The transfer gains are small, from a few hundredths to under 1 kcal/mol, and several pairs (including MACE on unseen families) have intervals crossing zero. Improves in all 24 pairs is therefore not a statistical result

  1. The method has limited novelty, its gains are not shown to come from TS information, and the benchmark is not yet reproducible.

The author honestly admit this point. However, as chemistry papers are already hard to publish on ICLR, without technical novelty it would be worse. Also, the experiments also do not show a prominent gain. In the pretraining-objective study, every RP-only self-supervised objective reaches lower endpoint-only MAE than the proposed RPTS pretraining, and adding TS within those objectives changes MAE only marginally. Besides, since the audited benchmark is a central contribution, an accessible anonymous artifact or complete reconstruction materials should be provided during review.

Questions for authors

see weakness

5 comments
AuthorsOpenReview ↗

Response: Training Controls, TS Recovery, and Reproducibility

Thank you for your careful review and constructive comments. We have run additional experiments to address your concerns and will include the results in the revised paper.

The materials are now available in our anonymous repository.

Reproduction instructions and checkpoint details are provided in docs/REPRODUCIBILITY.md and REVIEWER_DEMO.md.

Weakness 1 (i) — Learning rate and training controls

Thank you for carefully checking the training settings. We apologize for not explaining the learning-rate adjustment clearly and will add this explanation to the main text.

In our Cyclo setting, DimeNet++ encountered non-finite losses at 1e-3. We lowered the rate to 5e-5 for failed runs and dependent tasks, retaining successful runs and other settings. Training instability has also been reported in prior work: PhAST (Table 6) reports NaN training losses for DimeNet++ with supernode extensions across a wide range of learning rates. This explains the seed-2 exception (A0: 5e-5; A1: 1e-3).

Weakness 1 (ii) — Encoder transfer gains and TS attribution.

We appreciate this concern and agree that the transfer gains warrant further investigation.

To understand such transfer gains, we hypothesize that explicit TS structures may help activate learned TS-related information that is not fully utilized in endpoint-only prediction. We therefore explore an extension of our current transfer strategy: supplying generated TS structures to the existing RPTS encoder and prediction head, without additional training or fine-tuning.

We evaluated this extension on Grambow using React-OT and TS-DFM to generate TS structures from R/P.

EncoderRPTS: reference TS│A0: RP-pretrainedA1: TS-pretrained → R/P│RPTS: React-OT-generated TSRPTS: TS-DFM-generated TS
SchNet-1M4.490 ± 0.029│11.843 ± 0.30511.015 ± 0.264│5.552 ± 0.138 (85.6%)5.343 ± 0.049 (88.4%)
PaiNN3.900 ± 0.012│11.369 ± 0.18110.449 ± 0.164│5.289 ± 0.100 (81.4%)5.062 ± 0.116 (84.4%)
DimeNet++3.736 ± 0.167│11.701 ± 0.15311.731 ± 0.693│5.420 ± 0.228 (78.9%)5.172 ± 0.275 (82.0%)
GemNet-T3.929 ± 0.078│12.218 ± 0.07511.757 ± 0.095│5.286 ± 0.048 (83.6%)5.139 ± 0.051 (85.4%)
ClofNet3.525 ± 0.067│8.785 ± 0.1127.881 ± 0.259│7.744 ± 0.328 (19.8%)7.118 ± 0.295 (31.7%)
MACE3.330 ± 0.088│9.264 ± 0.0678.624 ± 0.135│8.310 ± 0.691 (16.1%)7.566 ± 0.588 (28.6%)

Both generation routes yield lower mean MAE than A1 across all six encoders. TS-DFM closes 82.0–88.4% of the MAE gap for four encoders and 28.6–31.7% for ClofNet and MACE.

These results suggest that the modest transfer gains may partly reflect how TS information is used. Encoder transfer carries training-time TS knowledge into an endpoint-only model through its parameters, whereas generation supplies an explicit TS structure for each reaction. The larger gains through generation are consistent with useful TS-related information not being fully captured by encoder transfer alone. Our existing RPTS encoder and head can exploit this additional route without retraining, extending the framework beyond endpoint-only encoder transfer.

Weakness 2 — Methodological contribution

Thank you for raising this concern. We agree that the scope of our methodological contribution needs to be clear.

While the approach is simple, our experiments demonstrate its effectiveness across multiple datasets and encoders, as well as its compatibility with different pretraining objectives. The results are summarized below.

Endpoint-only transfer (Table 3). The evaluation covers four datasets and six encoders, alongside a separate Grambow unseen-family evaluation. Two Grambow examples illustrate improvements without a TS generator or changes to the downstream architecture:

EncoderRP-pretrained MAERPTS-pretrained MAERelative MAE reduction
ClofNet8.6148.0506.5%
SchNet-1M11.56910.6747.7%

Compatibility with different pretraining objectives (Table 5). On Grambow–ClofNet, each evaluated objective has lower mean endpoint-only MAE when TS is included in pretraining:

Pretraining objectiveRP MAERPTS MAE
Barrier-supervised8.6148.050
CoordDenoise7.8717.711
GeoSSL-DDM7.8207.705
EMPP-l17.7927.753
C-FREE-Norm7.9827.978
Energy-Referenced7.6657.594

The comparison is within each objective: these supervised and self-supervised alternatives are pretraining components of our strategy. Together, Table 5 shows compatibility with different objectives and an extension to generated TS inputs without additional training.

AuthorsOpenReview ↗

Brief Follow-up on Our Rebuttal and Additional Results

Dear Reviewer FfDn,

Thank you again for your careful review. We would like to briefly follow up on our responses concerning training controls, the interpretation of transfer gains, and the methodological contribution.

We have clarified the learning-rate exception and the scope of the comparisons, distinguished improvements in mean MAE from statistical evidence, and explained the within-objective RP/RPTS comparisons. We also provide additional generated-TS experiments with frozen downstream predictors and access to the supporting materials.

If your schedule permits, we would appreciate hearing whether these responses address your concerns and whether any specific issue remains unresolved. We would be happy to clarify further.

Best regards, The Authors

Thank you for the clarification and additional experiments. The generated-TS results are useful, and I appreciate the release of reproduction materials. However, my main attribution concern remains. Explaining the learning-rate exception does not remove the mismatch, and the response does not report controls matching the additional training and pretraining compute.

The generated-TS experiment evaluates a different inference setting by supplying explicit TS structures at test time. It therefore does not isolate the source of the original endpoint-only transfer gains. Several within-objective improvements are marginal, and the RP-only self-supervised alternatives still outperform barrier-supervised RPTS pretraining. At present, I do not find sufficient evidence to change my assessment.

However, as I said in W2, if AC treats novelty level is sufficient in the track, you can treat my score as 2.5~3. My rating is based on my previous view of AI4S papers at ICLR.

Reviewer_1XLEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses an under-examined information-use question in reaction barrier prediction: transition-state (TS) structures are often available during training, yet deployment typically provides only reactants (R) and products (P). The authors note that existing work either generates TS from RP endpoints (TSNet, React-OT, etc.) or generates a TS on the fly at inference, but lacks a systematic evaluation of how much value TS information provides and whether that benefit transfers to endpoint-only models.

Critical strengths
  1. Clear problem decomposition with methodological value: the paper explicitly separates the "direct predictive value" of TS (true TS at inference) from its "transferable value" (TS only during pretraining) (Sections 3 and 5.1), and measures both systematically with a four-way RP/RPTS/RP-pretrained/RPTS-pretrained design across four datasets × 6 encoders. This decomposition of "information value vs. deployment feasibility" is more useful as a benchmark than simply reporting another barrier predictor.

  2. Auditing and leakage control are reported honestly: Appendix A details each dataset's target definition (electronic barrier, ZPE-inclusive activation energy, free-energy barrier) and TS validation standard (whether IRC was run, whether exactly one imaginary frequency exists), and explicitly states that "absolute errors across datasets do not rank chemical difficulty or label accuracy" (end of Section 4). Table 7 reports crossing counts for the grouped splits, and the paper proactively discloses that the Cyclo audit finds 651 reactant-pair groups spanning partitions, so "the split is therefore not an OOD evaluation" (Appendix A.6). This explicit delimitation of the evidence boundary is commendable.

Critical weaknesses

The gap in magnitude between transfer and direct gains suggests the "transferable value" may be quite limited, yet no attribution analysis is provided. On Grambow, for instance, the direct TS gain for ClofNet is 5.115 while the transfer gain is only 0.564, about 11% of the former; the limitations section acknowledges that "positive matched gains demonstrate useful changes in the learned representation without identifying the specific geometric or electronic features transferred," but offers no further experiment to explain why most TS information is not retained. This leaves the conclusion that "TS structures are a resource for reaction representation learning" largely qualitative.

Questions for authors

Why does only about 11% of the direct TS gain on Grambow (about 5.1) transfer to the endpoint model? Can the authors use TS-encoder probes, retention analysis of the t−r feature, or stratification by reaction type / barrier range to explain which TS information is retained and which is lost?

2 comments
AuthorsOpenReview ↗

Response: Further TS Recovery and Structural Error Analysis

Thank you for recognizing the value of separating direct TS use from endpoint-only transfer, and for your suggestions on analyzing the gap. We have conducted additional experiments and will include the results in the revised paper.

Our anonymous repository is now available. It provides fixed partitions, filtering records, preparation code, recorded settings and archived predictions. results/README.md maps the predictions to the paper's tables, and REVIEWER_DEMO.md provides a runnable four-checkpoint Grambow example, including endpoint-only inference with TS geometry absent.

Weakness 1 & Question 1 — Regarding the recovery and what TS structure contributes

Thank you for your insightful suggestions on encoder probes, retention analysis of the t−r feature, and stratification by reaction type or barrier range. We are currently pursuing these analyses. In parallel, we examined the gap from another angle by testing whether explicit TS structure recovery could make more of the predictive benefit accessible.

We hypothesize that explicit TS structures may help activate learned TS-related information that is not fully utilized in endpoint-only prediction. We therefore explore an extension of our current transfer strategy: supplying generated TS structures to the existing RPTS encoder and prediction head, without additional training or fine-tuning.

We evaluated this extension on Grambow using React-OT and TS-DFM to generate TS structures from R/P. The results are summarized below:

All columns use the same 1,159 Grambow validation reactions; MAE (kcal/mol) is mean ± SD over three downstream seeds. A0 (RP-pretrained) and A1 (TS-informed encoder transfer) use only R/P.

EncoderRPTS: reference TS│A0: RP-pretrainedA1: TS-pretrained → R/P│RPTS: React-OT-generated TSRPTS: TS-DFM-generated TS
SchNet-1M4.490 ± 0.029│11.843 ± 0.30511.015 ± 0.264│5.552 ± 0.138 (85.6%)5.343 ± 0.049 (88.4%)
PaiNN3.900 ± 0.012│11.369 ± 0.18110.449 ± 0.164│5.289 ± 0.100 (81.4%)5.062 ± 0.116 (84.4%)
DimeNet++3.736 ± 0.167│11.701 ± 0.15311.731 ± 0.693│5.420 ± 0.228 (78.9%)5.172 ± 0.275 (82.0%)
GemNet-T3.929 ± 0.078│12.218 ± 0.07511.757 ± 0.095│5.286 ± 0.048 (83.6%)5.139 ± 0.051 (85.4%)
ClofNet3.525 ± 0.067│8.785 ± 0.1127.881 ± 0.259│7.744 ± 0.328 (19.8%)7.118 ± 0.295 (31.7%)
MACE3.330 ± 0.088│9.264 ± 0.0678.624 ± 0.135│8.310 ± 0.691 (16.1%)7.566 ± 0.588 (28.6%)

Percentages indicate the fraction of the A0-to-reference-TS MAE gap recovered.

Both generation routes yield lower mean MAE than A1 across all six encoders. TS-DFM closes 82.0–88.4% of the MAE gap for four encoders and 28.6–31.7% for ClofNet and MACE.

These results suggest that the modest transfer gains may partly reflect how TS information is used. Encoder transfer conveys training-time TS knowledge through model parameters, whereas generation supplies an explicit structure for each reaction. The larger gains are consistent with useful TS-related information not being fully captured by encoder transfer alone. Our existing RPTS encoder and head can use this additional route without retraining, extending the framework while retaining the original option of inference without TS generation.

We conducted two complementary analyses of structural recovery, measuring pair-distance error over either all atoms or atom pairs within the reaction center and its one-hop neighbors. We correlated each measure with the increase in absolute barrier error relative to reference TS, using the same frozen predictor and averaging barrier-error increases over three seeds. The results are summarized below (Pearson r; the same 1,159 Grambow validation reactions):

EncoderReact-OT: all atomsReact-OT: center + 1 hopTS-DFM: all atomsTS-DFM: center + 1 hop
SchNet-1M0.3470.3960.3650.357
PaiNN0.3660.4070.3880.409
DimeNet++0.3230.3540.3820.396
GemNet-T0.3570.4000.4050.418
ClofNet0.2400.2780.2640.256
MACE0.2030.2300.2090.189

Restricting the error to the reaction-center neighborhood increases correlation for all six React-OT predictors, while TS-DFM shows mixed changes. These preliminary associations suggest that local structural accuracy may be particularly relevant for some generation routes. Our ongoing analyses will further examine which TS-related features encoder transfer retains and how reaction-center geometry contributes to prediction.

AuthorsOpenReview ↗

Brief Follow-up on Our Rebuttal and Additional Results

Dear Reviewer 1XLE,

Thank you again for your constructive suggestions on understanding the gap between direct TS use and endpoint-only transfer.

Our additional experiments examine a complementary route: supplying generated TS structures to the existing frozen RPTS predictors without further training or fine-tuning. We also report structural-error analyses over both the full molecule and the reaction-center neighborhood. These analyses provide evidence about sensitivity to TS geometry, while leaving the question of which features encoder transfer retains open.

If your schedule permits, we would greatly appreciate your thoughts on whether these additions help address your concern and which aspects warrant further clarification.

Thank you for your time and consideration.

Best regards, The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary for the AC

Dear AC,

Thank you for handling our submission. We appreciate the reviewers' constructive feedback and have revised the manuscript, expanded the experiments, and released supporting materials.

Our contribution. TSIRL distinguishes the direct predictive value of TS structures from their benefit when only R/P structures are externally supplied at deployment. It covers four audited datasets and six encoders, with endpoint-pair grouped-random evaluation and an unseen depth-1 reaction-family evaluation. Encoder transfer demonstrates broad applicability across splits, encoders and pretraining objectives, without requiring a TS generator at inference. The revised framework also includes generated-TS prediction using the existing frozen RPTS encoder and head, without additional training or fine-tuning.


Encouraging reviewer feedback

We are encouraged that the reviewers recognize the value of our contributions:

  • tEzQ: “The distinction between direct TS access and benefits retained after removing TS inputs is useful.” The reviewer also values the four-dataset, six-encoder evaluation.
  • FfDn: separating direct and retained TS value is “a precise and useful framing for deployment without TS.” The reviewer highlights endpoint-pair grouping against alternative-TS leakage.
  • 1XLE: the decomposition is “more useful as a benchmark than simply reporting another barrier predictor.” The reviewer also commends the audit and clear evaluation boundaries.

Their questions guided the following experiments and revisions. References below point to the revised manuscript.


Additional experiments

Further TS recovery — addressing tEzQ, FfDn and 1XLE. We extended the initial validation-set study to the Grambow grouped-random test set (1,198 reactions) and unseen-family test set (1,196 reactions). React-OT and TS-DFM supply generated structures to existing RPTS predictors, with all components frozen. Both routes yield lower mean MAE than encoder transfer across all six encoders on both splits (Tables 3–4).

With TS-DFM, SchNet-1M improves from 10.674 to 4.989 kcal/mol on the grouped-random split and from 13.158 to 7.239 on the family split. ClofNet improves from 8.050 to 6.912 and from 10.304 to 8.304, respectively. These results suggest that the achievable benefit depends partly on how TS-informed learning is used: encoder transfer uses R/P representations through transferred parameters, whereas generation supplies explicit reaction-specific geometry to the original RPTS predictor.

Practical effectiveness. Appendix B.1, Table 16 compares both strategies with six established barrier predictors on the same grouped-random test reactions. Encoder-transfer ClofNet achieves 8.050 kcal/mol, compared with 10.870 for the strongest external baseline in this comparison, MFP + RF. Generated-TS prediction further reduces MAE to 7.679 with React-OT and 6.912 with TS-DFM.

Structural analysis — addressing 1XLE. Our reply compares generated-TS geometry error with added barrier error relative to reference TS on the initial validation set. Restricting geometry error to the reaction center and its one-hop neighbors increases Pearson correlation for all six React-OT predictors; TS-DFM shows mixed changes. The React-OT results suggest that local structural accuracy may be particularly relevant to barrier prediction.


Framework scope — addressing tEzQ and FfDn

Encoder transfer lowers mean endpoint-only MAE in all 24 dataset–encoder comparisons (Appendix B) and across all six encoders on the Grambow unseen-family split (Table 4). Table 5 shows lower mean MAE for RPTS pretraining within each of five additional objectives. These supervised and self-supervised objectives are alternative components of our strategy, demonstrating applicability beyond barrier-supervised pretraining.


Training settings and reproducibility

Addressing FfDn. Appendix C.3 and Table 17 document recorded training settings, checkpoint selection, parameter inheritance and the lower learning rates used to recover failed Cyclo DimeNet++ runs.

Addressing tEzQ and FfDn. Our anonymous repository provides fixed partitions, filtering records, training metadata and archived predictions. Appendix C.1 links the materials; docs/REPRODUCIBILITY.md and REVIEWER_DEMO.md provide reproduction and checkpoint-inference instructions.

We respectfully invite the AC and reviewers to consider these revisions alongside our individual replies and the original benchmark results. We hope they clarify the effectiveness and generality of TSIRL. Thank you again for your thoughtful consideration.

Best regards,
The authors

AuthorsOpenReview ↗

Summary of Our Revisions and Request for Feedback

Dear Area Chair and Reviewers,

We sincerely thank the AC and reviewers for their time, care and constructive suggestions. As the discussion deadline approaches, we summarize our revisions and the evidence supporting our contribution for the AC's consideration. We are grateful for the opportunity to strengthen the paper through this process and are encouraged by the strengths recognized in the reviews:

  • tEzQ recognizes the useful distinction between direct and retained TS benefits, supported by four datasets, six encoders and paired uncertainty estimates.
  • FfDn describes the benchmark question as “a precise and useful framing for deployment without TS” and highlights the endpoint-pair leakage controls.
  • 1XLE values the problem decomposition as “more useful as a benchmark than simply reporting another barrier predictor” and commends the transparent dataset audits.

We have provided point-by-point responses to all weaknesses and questions, added experiments, revised the manuscript, and released anonymous materials. Below we summarize the resulting contributions and updates for the AC's consideration. References point to the revised manuscript.

Practical effectiveness — tEzQ and FfDn. In response to the reviewers' questions about practical gains, we highlighted that encoder transfer improves mean endpoint-only MAE across four datasets and six encoders, with Grambow reductions of 6.5% for ClofNet and 7.7% for SchNet-1M, without requiring TS generation. We also expanded the external comparison: encoder-transfer ClofNet achieves 8.050 kcal/mol, compared with 10.870 for MFP + RF, the strongest of the six external predictors in this comparison. Generated-TS prediction further reduces ClofNet MAE to 6.912 with TS-DFM (Appendix B.1, Table 16).

Further TS recovery — tEzQ, FfDn and 1XLE. To investigate the gap between direct TS and encoder-transfer benefits, we tested an extension that supplies React-OT- and TS-DFM-generated structures to the existing frozen RPTS encoder and head, without additional training or fine-tuning. Both routes yield lower mean MAE than encoder transfer across all six encoders on both Grambow test splits (Tables 3–4). For example, TS-DFM reduces SchNet-1M MAE from 10.674 to 4.989 on the grouped-random split and from 13.158 to 7.239 on the unseen-family split. This suggests that the achievable benefit depends partly on how TS-informed learning is used, with explicit structural inputs providing an additional way to use the trained RPTS predictor.

Structural analysis — 1XLE. In response to the request for analysis beyond aggregate errors, our reply examines the association between generated-TS geometry error and the added barrier error relative to reference TS. We compared whole-molecule geometry with the reaction center and its one-hop neighborhood on the initial Grambow validation set. The local measure increases Pearson correlation for all six React-OT predictors, while TS-DFM shows mixed changes. These results suggest that reaction-center geometry may be particularly relevant for some generation routes.

Methodological scope — tEzQ and FfDn. We clarified that the contribution is a general TS-informed framework with applicability across splits, encoders and pretraining objectives. In particular, Table 5 compares RP and RPTS within each objective: all five additional objectives yield lower mean endpoint-only MAE when TS is included in pretraining. Supervised and self-supervised objectives can therefore serve as different components of the transfer strategy. The generated-TS extension broadens this framework, while encoder transfer retains inference without a TS generator.

Training settings and accessible materials — FfDn and tEzQ. Appendix C.3 and Table 17 document the training settings and parameter inheritance, including the lower learning rates used after non-finite losses in Cyclo DimeNet++. We distinguish the reported mean improvements from statistical significance for every pair. Our anonymous repository now provides fixed partitions, filtering records, recorded training metadata and archived predictions, with reproduction instructions and checkpoint inference described in docs/REPRODUCIBILITY.md and REVIEWER_DEMO.md (Appendix C.1).

TSIRL combines an audited benchmark with broad applicability across encoders and learning objectives, supporting both encoder transfer without TS generation and further gains from generated structures without predictor retraining.

We would welcome the reviewers' feedback on our responses and revised results, and would appreciate the AC's consideration of the full evidence in the final assessment. We are happy to provide further clarification. Thank you again for your time and thoughtful evaluation.

Cordially,
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.