Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Authors developed a novel framework utilizing deep learning techniques to perform prediction of enzymatic kinetic parameters. Compared to other platforms where protein and substrate are modeled individually, the method is taking both protein and substrate together in 3D spatial orientation modeling. The current framework outperformed other methods with better prediction precision. 

The manuscript is very interesting and comprehensive for non-experts. I propose the acceptance of the manuscript. 

Below are my additional comments:

1: enzyme kinetics is pH and temperature dependent. Can the author explain if the software and predict kcat and Km in temperature or pH dependent manner? If not, could the author explain which pH and temperature that author used for collecting experimental values?

2: In drug discovery, kinetic parameters calculations on drug candidates (e.g. kon, koff, kinact/Ki) are more useful than kcat/Km calculation. Can the author comment on the prediction of drug candidates' kinetic parameters?

3: As suggested in my previous report, there are lots of long sentences. It is suggested that the author should consult on professional language editing by native speakers.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Review of the article. The article is titled: "KineticGraph: Unified-coordinate geometric graphs enable robust enzyme kinetic prediction". Using a carefully selected data set of 15,591 wild-type enzyme-substrate pairs, the authors trained the KineticGraph neural network with high accuracy in estimating a number of kinetic parameters of enzymatic catalysis.

The method is based on the application of docking, KineticGraph creates a geometry aware prediction pipeline which is created to be able to predict the enzymes kinetics through geometry aware message passing. The framework depends on the fidelity of docking poses and currently uses static conformations; moreover, assay conditions such as pH and temperature were not modeled due to inconsistent reporting across data sources. Future work may incorporate improved complex geometry estimation, conformational ensembles, and complementary energetic descriptors to better capture catalysis relevant effects.

Unfortunately, with all the advantages of the proposed method, the authors do not mention such an important drawback as the lack of solvent. As is known, water is explicitly involved in the active cites of many enzymes, acting, for example, as an intermediary in the transfer of protons in enzyme-substrate complexes. The authors will probably take this into account in the next versions of the product, and provide it for extended testing to all interested teams.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript is good but needs revision before it can be accepted. 

@ The manuscript relies heavily on docking derived enzyme substrate geometries, yet the robustness of the proposed method with respect to docking uncertainty is not adequately addressed. Only a single top ranked AutoDock Vina pose is used for each enzyme substrate pair, and no analysis is provided to evaluate how sensitive the predictions are to pose variability or docking errors. The absence of comparisons against experimentally resolved enzyme substrate complexes, where available, further limits confidence that the learned geometric representations are not biased by docking artifacts.

@ Dynamic effects relevant to enzyme catalysis, such as conformational flexibility, induced fit, and transition-state stabilization, are not modeled or discussed in the context of kcat prediction.

@ Experimental assay conditions (e.g., pH, temperature, cofactors) are not incorporated, and the impact of heterogeneous measurement conditions on kinetic labels is not analyzed.

@ No external, independent dataset or prospective validation is used to assess generalization beyond the curated benchmark.

@ The authors are encouraged to include and discuss additional recent and relevant studies that apply machine learning and graph neural network approaches to biochemical and enzymatic prediction tasks, including doi:10.1016/j.comtox.2025.100344, doi:10.1016/j.compbiomed.2024.108729, doi:10.1038/s41467-025-57215-9, and doi:10.1093/bib/bbaf187, to better contextualize the proposed framework within the current state of the field.

@  Confidence metrics of AlphaFold-predicted structures (e.g., pLDDT, PAE) are not considered, and structural uncertainty is not propagated into the modeling pipeline.

@  No interpretability or attribution analysis is provided to relate learned geometric features to known catalytic residues or mechanistic interactions.

@ Computational cost, training and inference time, and scalability of the framework are not reported.

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

I must confess that I have enjoyed reviewing this manuscript.

The authors of the manuscript “KineticGraph: Unified-coordinate geometric graphs enable robust enzyme kinetic prediction” present an innovative work that integrates explicit spatial geometry into kinetic prediction, overcoming the limitations of models based purely on sequence or 2D graphs. I find their work extremely interesting for the journal's guidelines. However, I would like the authors to comply with the following observations in order for this article to be published.

1. Provide a GitHub download link or website where users can access the code. This will be useful to various authors and encourage citation of the article.
2. While searching the web, I found a JavaScript engine with the same name, KineticGraphs. Since the authors named their algorithm with an almost identical name, I suggest they use an acronym to distinguish it from other programs.

 

Below are some minor revisions.

1. The manuscript mentions that the top-scoring pose was selected. However, Vina sometimes scores biologically irrelevant poses well. Therefore, I suggest that the authors be explicit about the potential disadvantages.
2. Three layers of GVPConv are mentioned. How was this number determined? Was over-smoothing observed with more layers? Thank you to the authors for the brief note on hyperparameter optimization in the supplementary material, which would make the paper more robust.
3. Given the dataset used by the authors and the statistical results shown in the manuscript, it would be helpful to mention which enzymes KineticGraph has difficulty with, such as metalloenzymes and highly flexible enzymes.
4. Carefully review and correct typos in the text, such as “benchmarkerik,” “determintaon,” “Kineticgraph,” among others.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have revised the manuscript in accordance with my comments. I recommend the article for publication.

Reviewer 3 Report

Comments and Suggestions for Authors

The comments are addressed.