From the journal Digital Discovery Peer review history

04-Mar-2024

Dear Mrs Volokhova:

Manuscript ID: DD-ART-01-2024-000023
TITLE: Towards equilibrium molecular conformation generation with GFlowNets

Thank you for your submission to Digital Discovery, published by the Royal Society of Chemistry. I sent your manuscript to reviewers and I have now received their reports which are copied below.

After careful evaluation of your manuscript and the reviewers’ reports, I will be pleased to accept your manuscript for publication after minor revisions.

Please revise your manuscript to fully address the reviewers’ comments. When you submit your revised manuscript please include a point by point response to the reviewers’ comments and highlight the changes you have made. Full details of the files you need to submit are listed at the end of this email.

Digital Discovery strongly encourages authors of research articles to include an ‘Author contributions’ section in their manuscript, for publication in the final article. This should appear immediately above the ‘Conflict of interest’ and ‘Acknowledgement’ sections. I strongly recommend you use CRediT (the Contributor Roles Taxonomy, https://credit.niso.org/) for standardised contribution descriptions. All authors should have agreed to their individual contributions ahead of submission and these should accurately reflect contributions to the work. Please refer to our general author guidelines https://www.rsc.org/journals-books-databases/author-and-reviewer-hub/authors-information/responsibilities/ for more information.

Please submit your revised manuscript as soon as possible using this link :

*** PLEASE NOTE: This is a two-step process. After clicking on the link, you will be directed to a webpage to confirm. ***

https://mc.manuscriptcentral.com/dd?link_removed

(This link goes straight to your account, without the need to log in to the system. For your account security you should not share this link with others.)

Alternatively, you can login to your account (https://mc.manuscriptcentral.com/dd) where you will need your case-sensitive USER ID and password.

You should submit your revised manuscript as soon as possible; please note you will receive a series of automatic reminders. If your revisions will take a significant length of time, please contact me. If I do not hear from you, I may withdraw your manuscript from consideration and you will have to resubmit. Any resubmission will receive a new submission date.

The Royal Society of Chemistry requires all submitting authors to provide their ORCID iD when they submit a revised manuscript. This is quick and easy to do as part of the revised manuscript submission process.   We will publish this information with the article, and you may choose to have your ORCID record updated automatically with details of the publication.

Please also encourage your co-authors to sign up for their own ORCID account and associate it with their account on our manuscript submission system. For further information see: https://www.rsc.org/journals-books-databases/journal-authors-reviewers/processes-policies/#attribution-id

I look forward to receiving your revised manuscript.

Yours sincerely,
Dr Joshua Schrier
Associate Editor, Digital Discovery

************
EDITOR'S COMMENT:

Editor provided data/code review (attached) and the article meets our guidelines.

No additional calculations are required, but please address reviewer comments below to the extent that they are representative of questions that might be asked by readers.

************

Reviewer 1

The paper applies GFlowNets to sample torsion angles in drug-like molecules such as alanine dipeptide, ibuprofen, and ketorolac. It offers a detailed analysis of molecular systems with only two torsion angles and investigates GFlowNets' performance depending on the number of torsion angles. The manuscript is well-written, explaining the method and experiments in great detail. Considering the above and the fact that exploring generative models for sampling molecular conformation is an intriguing research area, I am inclined to recommend this manuscript for publication, however, after addressing minor comments.

Some more detailed comments:

* The work focuses only on torsion angles, saying that some rule-based methods can generate bond lengths and angles. I believe this narrative should be changed as it is rather a current limitation of the proposed method, but not the must-have for generative models.

* Related work seems to miss conformer ensembles when thinking about property prediction from phase-space averages. Some recent references could be: 10.1088/2632-2153/acefa7, https://openreview.net/forum?id=kFiMXnLH9x, and https://arxiv.org/abs/2402.01975.

* GFlowNets, for me, are reminiscent of Hamiltonian Monte Carlo (HMC), employing a sequential update rule and a reward function. For HMC these would be velocity Verlet integration and accept/reject step. There is also much work on optimizing its parameters (integration time step and the number of integration steps), with the most recent utilizing automatic differentiation: https://doi.org/10.1063/5.0177738. Could the authors briefly mention the HMC method in their related work section? Also, while running additional experiments is unnecessary, a comparison similar to the one performed for MCMC in the Appendix could be interesting.

* Could the authors briefly discuss the practical advantages of their method compared to MCMC, even if only potential ones?

* In the discussion, the authors mention that the presented work is a 'stepping stone' toward a generalized model. Could the authors briefly explain how the reader can see it, given that in the presented work, a separate model is trained for each molecule? What are the criteria to reach this goal?

* Could the authors compare the computational cost of their method and MCMC? Reporting runtimes for both methods would be sufficient.

Thank you for taking time to review our work and providing your feedback. Here is our response to the review comments:

> The work focuses only on torsion angles, saying that some rule-based methods can generate bond lengths and angles. I believe this narrative should be changed as it is rather a current limitation of the proposed method, but not the must-have for generative models.

Our focus on sampling the torsion angles is based on prior work in the chemistry domain that applied similar parameterization to the problem, e.g. https://arxiv.org/abs/2206.01729 and https://arxiv.org/abs/2306.07472

> Related work seems to miss conformer ensembles when thinking about property prediction from phase-space averages. Some recent references could be: 10.1088/2632-2153/acefa7, https://openreview.net/forum?id=kFiMXnLH9x, and https://arxiv.org/abs/2402.01975.

Both works mentioned by the reviewer are either concurrent work that published either at the same time as the paper (AI4Science workshop at NeurIPS – OpenReview link) or after the paper was submitted (arXiv link Feb 3, 2024). As such, they were not included in the related works section.

> GFlowNets, for me, are reminiscent of Hamiltonian Monte Carlo (HMC), employing a sequential update rule and a reward function. For HMC these would be velocity Verlet integration and accept/reject step. There is also much work on optimizing its parameters (integration time step and the number of integration steps), with the most recent utilizing automatic differentiation: https://doi.org/10.1063/5.0177738. Could the authors briefly mention the HMC method in their related work section? Also, while running additional experiments is unnecessary, a comparison similar to the one performed for MCMC in the Appendix could be interesting.

GFlowNets indeed have similarities with other sequential sampling methods, such as Hamiltonian MC, hierarchical variational inference, etc. However, these theoretical links are not the focus of our work and therefore we don't include it in our manuscript. Section 3 of our paper gives the necessary background on GFlowNets and one can follow the cited works e.g. https://arxiv.org/abs/2301.12594 to learn more about the theoretical aspects.

> Could the authors briefly discuss the practical advantages of their method compared to MCMC, even if only potential ones?* In the discussion, the authors mention that the presented work is a 'stepping stone' toward a generalized model. Could the authors briefly explain how the reader can see it, given that in the presented work, a separate model is trained for each molecule? What are the criteria to reach this goal?

As detailed in the discussion section, future work will extent the current model in a way that it will be able to generate conformers for the given molecular graph which was not seen during the training. One promising approach is to make the gflownet model conditional on the input molecular graph.

> Could the authors compare the computational cost of their method and MCMC? Reporting runtimes for both methods would be sufficient.

Given the similarities to MCMC don't expect our method to be more computationally efficient than MCMC in the current implementation, as it requires training the model for each molecule.


Thank you,
Alexandra

22-Mar-2024

Dear Mrs Volokhova:

Manuscript ID: DD-ART-01-2024-000023.R1
TITLE: Towards equilibrium molecular conformation generation with GFlowNets

Thank you for submitting your revised manuscript to Digital Discovery. I am pleased to accept your manuscript for publication in its current form. I have copied any final comments from the reviewer(s) below. You may incorporate revisions into your page proofs, if desired.

You will shortly receive a separate email from us requesting you to submit a licence to publish for your article, so that we can proceed with the preparation and publication of your manuscript.

You can highlight your article and the work of your group on the back cover of Digital Discovery. If you are interested in this opportunity please contact the editorial office for more information.

Promote your research, accelerate its impact – find out more about our article promotion services here: https://rsc.li/promoteyourresearch.

If you would like us to promote your article on our LinkedIn account [https://rsc.li/Digital_showcase] please fill out this form: https://form.jotform.com/213544038469056.

We are offering all corresponding authors on publications in gold open access RSC journals who are not already members of the Royal Society of Chemistry one year’s Affiliate membership. If you would like to find out more please email membership@rsc.org, including the promo code OA100 in your message. Learn all about our member benefits at https://www.rsc.org/membership-and-community/join/#benefit

By publishing your article in Digital Discovery, you are supporting the Royal Society of Chemistry to help the chemical science community make the world a better place.

With best wishes,

Dr Joshua Schrier
Associate Editor, Digital Discovery

Reviewer 1

The authors tended to disregard my comments rather than incorporate them into their manuscripts.

For instance, mentioning the use of torsional angles in other works to assess model performance might be seen as a weak argument because it may stem from those methods' limitations rather than strengths (similar to many classical approaches, though). Such a discussion would encourage researchers to focus on these issues (e.g., mode collapse is related) but was not meant to criticize the presented approach.

Furthermore, dismissing a body of work because it is concurrent (10.1088/2632-2153/acefa7 has been on arXiv since 2020) and is not worth referencing seems odd. It would be beneficial to contextualize their generative model and discuss its suitability for specific tasks, particularly for practitioners in the chemistry community. However, I apologize for not checking the dates of other proposed works submitted to arXiv.

However, despite these points, I find the work compelling and worthy of publication in Digital Discovery. I recommend this manuscript for publication, leaving it to the authors to decide whether to discuss the limitations of their method.