From the journal Digital Discovery Peer review history
A transferable double exponential potential for condensed phase simulations of small molecules
Round 1
Manuscript submitted on 20 Apr 2023![](https://www.rsc-cdn.org/pubs-core/2022.0.146/content/NewImages/icon-letter.png)
![](https://www.rsc-cdn.org/pubs-core/2022.0.146/content/NewImages/chevron-down.png)
15-May-2023
Dear Dr Cole:
Manuscript ID: DD-ART-04-2023-000070
TITLE: A transferable double exponential potential for condensed phase simulations of small molecules
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 revisions.
Please revise your manuscript and associated data and/or code 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,
Linda Hung
Associate Editor
Digital Discovery
Royal Society of Chemistry
************
![](https://www.rsc-cdn.org/pubs-core/2022.0.146/content/NewImages/icon-comments.png)
![](https://www.rsc-cdn.org/pubs-core/2022.0.146/content/NewImages/chevron-down.png)
Reviewer 1
Presented is the application of the OpenFF software suite to optimize
a small molecule force field based on a new model to treat the vdW
interactions. The results show the utility of OpenFF in facilitating
the automated development of a new force field and is consistent with
the audience of Digital Discovery. Issues that need to be addressed by
the authors follow.
A general issue is that the model is being trained on all the
available QM or experimental data for some of the properties being
presented. While some separate test sets based on transfer free
energies and QM data are present it is limited and doesn't test the
full range of parameters being developed. This limitation should be
clearly stated in the conclusion.
Similar to the previous comment, the authors need to also state if the
force fields with which comparisons are being made were optimized using
the same training sets.
Information on the number of iterations/timing of the parameter
optimization itself need to be presented. This should include the
computer resources being used.
In the manuscript the authors comment about the need for far fewer
parameters due to the use of chemical perception-based atom types
versus the example of OPLS3e presented in the conclusion. However,
this will come at a significant decrease in accuracy over a large
number of molecules, including larger drug-like molecules versus the
small molecules included in the present study. This needs to be noted
in the concluding section.
A final issue that needs to be addressed is the overall limitation
associated with available experimental data as required to perform the
optimization of a wider range of molecules/atoms. For example, for the
additional 20 SMIRKS types (including S,P,I,F) is there available
experimental data to apply the same optimization approach.
Reviewer 2
The manuscript “A transferable double exponential potential for condensed phase simulations of small molecules” is a welcome contribution to the simulation community. It is also critically important to a larger audience as the ubiquity of simulation used to support experimental data without much care as to whether the underlying molecular model is a robust physical representation of the system of interest. Simulated properties that agree or support a hypothesis are often presented without careful scrutiny of potential insufficiencies (pun intended).
The authors not only recognized a long-standing problem – they have made productive progress to help the community to a brighter future. Atomistic simulation, e.g. MD and Monte Carlo, has been stuck in a historical molasses using very bad old technology in the form of force fields. One giant weakness of these functional forms, namely Lennard-Jones forms for repulsion / dispersion interactions, have been known to be inadequate for decades. Worse, far better alternatives exist that are computationally competitive. Even while force fields and computation have become more important with huge investments in person and computer hours, the quality of simulations in the literature is spotty at best. I typically do not put much trust in most simulated results I see published.
Potentials like Lennard-Jones were used / useful when only a limited data set of bulk thermodynamics was available for fitting molecular potential energy surfaces. Today, nearly exact electronic structure is ubiquitous and old practices need to change. We often know what the chemically accurate intermolecular potential energy surface is -- and they are not what is represented by the functional forms that are just used out of habit. This is without considering the regular lack of including large many body effects that are conveniently ignored. It is frankly an embarrassment what passes for molecular simulation at the “highest levels” today especially in computationally biology. There are so many fancy simulation applications, yet they are based on poor and outdated force fields used to feed grants and high-profile papers without being fundamentally sound science.
That long introduction is to say that the computational infrastructure detailed in the manuscript is a path out of these dark ages of intermolecular potentials. The community needs tools to develop, evaluate and apply better technologies. Only when the inadequacy of current practices becomes embarrassingly obvious will the requisite change / improvements be adopted / required in the molecular simulation community. It is not good enough to say we use famous giant force fields for our problem because all my friends do and here are a bunch of references that use them too. It is time for change. The authors have provided a practical road forward.
Notes:
The authors might mention that it has long been known that the typical Lennard-Jones mixing rules used, with geometric mixing of the epsilon coefficient isn’t optimal even for Lennard-Jones and the current practice is not based on any theoretical justification that I am aware of. On the other hand mixing the C6 coefficient of the dispersion geometrically is empirically and theoretically better motivated with better results. An empirical investigation of this is nicely demonstrated in a paper (one of many) by Waldman and Hagler
https://doi.org/10.1002/jcc.540140909; the theory of mixing can be referenced adequately via https://doi.org/10.1039/TF9605600761.
In Fig. 1 I am not sure that it is enough to go to about 150K in the repulsions to compare the potentials. What is the motivation to show only the attractive region ? The repulsive parts are important and different too at relevant energies not shown in Fig. 1. Please comment and thank you. Fig. 1 does do a good job showing the attractive parts so maybe a second panel?
I cannot stress enough how important the infrastructure to run and test new potentials is that is described at the end of Section 1. OpenFF and OpenMM have done an amazing job in this regard with excellent GPU support making training / testing / evaluating against meaningful data bases practical and indeed something that should be required. While I do not share the philosophy of building force fields with condensed phase information as a primary source of intermolecular potential energy information, the framework presenting lets the community readily test what actually works best and will ultimately result in optimal strategies in building both ad hoc and more transferable force fields. In the end it is not a philosophical but rather an empirically testable question.
Overall, this is a very interesting paper and a launching point for further development. Using the described software and hardware infrastructure, an entirely novel force field was developed and validated in short order. This is terrific and of great value at minimum as a demonstration project. The practical utility will be found as it get’s use.
I would encourage the author’s to also evaluate their potentials vs chemically accurate electronic structure calculations. My group’s experience is that, carefully parametrized force fields e.g. TraPPE CO2 can reproduce CO2 dimer energies for a variety of relative orientations – literally sitting on top of CCSD(T) large basis set calculations. On a cautionary note, there were other orientations, presumably not well represented in the condensed phase fitting, that were systematically far from correct. Given that one seeks transferability to new environments, it is worth checking force fields for some representative species vs the true Born-Oppenheimer PES.
Another suggestion is presenting some additional insight into the functional form chosen here. It can be convincingly argued reasonably that the repulsions should be exponential for an empirical potential as is discussed in Section 1. Note, further it is obvious the asymptotically dispersion is inverse sixth power with distance. However, in the van der Waals contact distances the dispersion is well represented by a rough 70-30 / 60-40 mixture of inverse 6 and inverse 8,10 distance potentials – e.g. the Buckingham has some appropriately damped 1/r8 dependence. An argument can be made that having the correct functional form and reasonable, tested mixing rules will lead to transferability. Still, the empirical results presented indeed speak for themselves – they are excellent.
However, while the author’s comment on the exponential nature of the repulsions, explaining the DE form in the context of the power law expansion in the bonding distance region would help me to understand the choice of functional form. This issue is alluded to in References (10,11) where the from was introduced but more focus was put on fit and calculation speed than whether the correct functional behavior is present. My thinking is that if we want detailed accurate contact geometries like in ligand binding e.g., in such cases, as opposed to perhaps bulk liquid structure and thermodynamics, the 1/r6, 1/r8, 1/r10 dependence is potentially critical; in these situations small geometry changes can have significant influence on observable phenomena like binding, inhibition or catalytic ability. This is a hypothesis of course, and I think some context along these lines and the mixing rules would make this truly excellent manuscript even better.
I leave it to the authors to decide in what manner to address my comments. I am grateful to see the systematic activities presented in the manuscript and such powerful tools made accessible to the simulation community. The paper is an extremely valuable addition to the literature that should be of interest to a very broad audience – albeit so many simulators are neglecting to provide sufficient care in fidelity of their force field choices vs their speed and handiness for a given application.
Reviewer 3
Details regarding the "Data Reviewer Checklist" are shown as follows.
6c. Have the authors clearly specified which versions of the software libraries they depend upon were used in the course of the work?
Yes, however, as the installation process can be complex and may cause issues with the local system, I strongly recommend the inclusion of instructions for creating a virtual environment. Using a virtual environment prevents potential conflicts with the local system by isolating package installations. I recommend adding the following to the README.md in the https://github.com/jthorton/de-forcefields repository:
- from now on, everything contained between the opening and closing tags <to be written in README.md> is what I suggest it should be written in the README.md
<to be written in README.md> ++++++++++++++++++
# Header: Installation
## subheader: Installation of virtual environment
We first create a virtual environment to avoid dealing with other packages installed in the system:
Test folder:
`$ mkdir Test && cd Test`
We create a virtual environment:
`$ conda create -n de-forcefields_env`
It will be installed in:
`~/.conda/envs/de-forcefields_env`
To activate this environment, use
`$ conda activate Site-Net_env`
To deactivate the environment:
`$ conda deactivate`
To further remove it:
`$ conda remove -n de-forcefields_env --all`
## subheader: Installation of dependencies and packages:
In this order, install
`(Site-Net_env) $ conda install -n de-forcefields_env -c conda-forge de-forcefields`
</to be written in README.md> ++++++++++++++++++
2a. Are the data cleaning steps clearly and fully described, either in text
or as a code pipeline?
6b. Are scripts to reproduce the findings in the paper provided?
Both, not entirely.
The main issue lies in the difficulty of reproducing the data curation and training and reproducing the optimizations and benchmarks as described in the manuscript due to the absence of adequate documentation in the accompanying GitHub repository. The repository's file structure, as found at https://github.com/jthorton/double-exp-vdw, is complex.
For example, the scripts to reproduce the training for the qm-benchmarks are located in https://github.com/jthorton/double-exp-vdw/tree/main/inputs-and-results/benchmarks/qm-benchmarks. However, this main page only displays a list of Python scripts without clear instructions about what each script is doing at each time, their function, the order of execution, or what output to expect at each stage.
I recommend creating a Jupyter notebook in the main directory of the double-exp-vdw repository to guide users through the training process for the optimizations and benchmarks. Furthermore, a separate Jupyter notebook should be created to explain the usage of scripts for the curation and creation of the datasets. Another notebook is needed for the ancillary data analysis.
I would gladly continue reviewing the code from this point onward once this is fixed.
4a. Is a software implementation of the model provided such that it can be trained and tested with new data?
Not in its current version. As written in the answer to question 6a, the lack of documentation
impedes the straightforward reproduction of this code using the provided test cases. Once this is solved, it will allow to be tested with new data.
![](https://www.rsc-cdn.org/pubs-core/2022.0.146/content/NewImages/icon-response.png)
![](https://www.rsc-cdn.org/pubs-core/2022.0.146/content/NewImages/chevron-down.png)
Dear Dr Hung,
We are grateful to the three Reviewers for their helpful and very supportive comments. We have revised the manuscript in response to their suggestions, as detailed below:
Referee: 1
A general issue is that the model is being trained on all the available QM or experimental data for some of the properties being presented. While some separate test sets based on transfer free energies and QM data are present it is limited and doesn't test the full range of parameters being developed. This limitation should be clearly stated in the conclusion.
Response: Thank you, this is a fair point. We have chosen to use the published, standard OpenFF training and test sets for direct comparison with the Sage force field. As highlighted in Sections 4.2 and 4.3, these do not currently exercise all force field parameters. For quantum mechanical data, it will be relatively straightforward to fill in the gaps in future datasets with additional quantum calculations on relevant chemistries. For condensed phase properties, it may be possible to add extra training data by being less stringent in the curation from ThermoML (e.g. functional groups required 5 measurements before being accepted for inclusion: https://doi.org/10.1021/acs.jctc.3c00039). Additional sources of data could also be included through OpenFF-Evaluator such as speed of sound, dielectric constants and crystallographic densities and sublimation enthalpies. We have expanded our discussion of this limitation and future directions in the revised Conclusions.
Similar to the previous comment, the authors need to also state if the force fields with which comparisons are being made were optimized using the same training sets.
Response: This information was presented for the most part in the Methods section, but we have now further emphasised/clarified this information in the main text:
• The caption to Fig 2 makes it clear which properties were used for training the three compared force fields.
• Section 4.2 (Non-bonded Parameters): DE-FF and Sage were trained on exactly the same condensed phase set, except for the addition of six pure water densities.
• Section 4.2 (Bonded Parameters): DE-FF is trained on a subset of the full Sage valence training set (filtering out any molecules that would be parameterised with unoptimised DE-FF nonbonded parameters). In addition, we did re-train a Sage-style force field to this subset of the training set. The results are reported in Fig S12, and show little difference between any of the three force fields. We therefore compared directly DE-FF and the standard Sage force field for the remainder of the tests.
Information on the number of iterations/timing of the parameter optimization itself need to be presented. This should include the computer resources being used.
Response: We agree that this is a useful addition. We have added timings, number of iterations and compute resource information for both valence and non-bonded training in the section “A transferable DE-based force field for small molecules” of the revised manuscript.
In the manuscript the authors comment about the need for far fewer parameters due to the use of chemical perception-based atom types versus the example of OPLS3e presented in the conclusion. However, this will come at a significant decrease in accuracy over a large number of molecules, including larger drug-like molecules versus the small molecules included in the present study. This needs to be noted in the concluding section.
Response: Apologies for being unclear on this point. Indeed the Sage force field has been directly benchmarked against OPLS4 to test this very question [https://doi.org/10.1021/acs.jcim.2c01185]. As the reviewer points out, OPLS4 does perform better on this benchmark, but also has the disadvantages of having many more parameters and being proprietary. We have noted this in the revised Conclusions, and also pointed the reader to the OpenFF-BespokeFit package (for fitting bespoke dihedral parameters against QM reference data for further accuracy improvements).
A final issue that needs to be addressed is the overall limitation associated with available experimental data as required to perform the optimization of a wider range of molecules/atoms. For example, for the additional 20 SMIRKS types (including S,P,I,F) is there available experimental data to apply the same optimization approach.
Response: Thank you, we believe we have now responded to this point in our answer to comment 1, and in the revised Conclusions.
=====================
Referee: 2
The authors might mention that it has long been known that the typical Lennard-Jones mixing rules used, with geometric mixing of the epsilon coefficient isn’t optimal even for Lennard-Jones and the current practice is not based on any theoretical justification that I am aware of. On the other hand mixing the C6 coefficient of the dispersion geometrically is empirically and theoretically better motivated with better results. An empirical investigation of this is nicely demonstrated in a paper (one of many) by Waldman and Hagler
https://doi.org/10.1002/jcc.540140909; the theory of mixing can be referenced adequately via https://doi.org/10.1039/TF9605600761
Response: We fully agree. Another study has looked at this recently, with alternative mixing rules performing similarly to the Lorentz-Berthelot mixing rules used here [https://doi.org/10.1021/acs.jctc.2c01170]. However, it would be interesting to re-investigate this in the context of DE-FF. We have added a small discussion to the revised Conclusions.
In Fig. 1 I am not sure that it is enough to go to about 150K in the repulsions to compare the potentials. What is the motivation to show only the attractive region ? The repulsive parts are important and different too at relevant energies not shown in Fig. 1. Please comment and thank you. Fig. 1 does do a good job showing the attractive parts so maybe a second panel?
Response: Thank you, we have added a second panel showing the repulsive region (up to around 1000K).
I cannot stress enough how important the infrastructure to run and test new potentials is that is described at the end of Section 1. OpenFF and OpenMM have done an amazing job in this regard with excellent GPU support making training / testing / evaluating against meaningful data bases practical and indeed something that should be required. While I do not share the philosophy of building force fields with condensed phase information as a primary source of intermolecular potential energy information, the framework presenting lets the community readily test what actually works best and will ultimately result in optimal strategies in building both ad hoc and more transferable force fields. In the end it is not a philosophical but rather an empirically testable question.
Overall, this is a very interesting paper and a launching point for further development. Using the described software and hardware infrastructure, an entirely novel force field was developed and validated in short order. This is terrific and of great value at minimum as a demonstration project. The practical utility will be found as it get’s use.
I would encourage the author’s to also evaluate their potentials vs chemically accurate electronic structure calculations. My group’s experience is that, carefully parametrized force fields e.g. TraPPE CO2 can reproduce CO2 dimer energies for a variety of relative orientations – literally sitting on top of CCSD(T) large basis set calculations. On a cautionary note, there were other orientations, presumably not well represented in the condensed phase fitting, that were systematically far from correct. Given that one seeks transferability to new environments, it is worth checking force fields for some representative species vs the true Born-Oppenheimer PES.
Response: The reason that we do not currently use dimer interaction energies to train force fields is primarily because the effective fixed charge model used is somewhat pre-polarised to account for polarisation effects in the condensed phase. The concern is therefore that electrostatics will tend to be too polar to describe dimer energetics well, even with a good LJ / DE-FF model. Nevertheless, we do agree that this is a useful check that potential energy surfaces have not strayed too far from physical reality. We have added a comparison with the DESS66x8 QM dataset in the section “A transferable DE-based force field for small molecules” in the revised manuscript, some example dimer interaction plots in the Supporting Information, and a full set of comparisons with the DESS66x8 QM dataset in the Supporting Data.
Another suggestion is presenting some additional insight into the functional form chosen here. It can be convincingly argued reasonably that the repulsions should be exponential for an empirical potential as is discussed in Section 1. Note, further it is obvious the asymptotically dispersion is inverse sixth power with distance. However, in the van der Waals contact distances the dispersion is well represented by a rough 70-30 / 60-40 mixture of inverse 6 and inverse 8,10 distance potentials – e.g. the Buckingham has some appropriately damped 1/r8 dependence. An argument can be made that having the correct functional form and reasonable, tested mixing rules will lead to transferability. Still, the empirical results presented indeed speak for themselves – they are excellent.
However, while the author’s comment on the exponential nature of the repulsions, explaining the DE form in the context of the power law expansion in the bonding distance region would help me to understand the choice of functional form. This issue is alluded to in References (10,11) where the from was introduced but more focus was put on fit and calculation speed than whether the correct functional behavior is present. My thinking is that if we want detailed accurate contact geometries like in ligand binding e.g., in such cases, as opposed to perhaps bulk liquid structure and thermodynamics, the 1/r6, 1/r8, 1/r10 dependence is potentially critical; in these situations small geometry changes can have significant influence on observable phenomena like binding, inhibition or catalytic ability. This is a hypothesis of course, and I think some context along these lines and the mixing rules would make this truly excellent manuscript even better.
Response: Thank you, we agree that ultimately improved accuracy and parameter transferability will come from physically-motivated functional forms that are capable of describing intermolecular interactions both close to the bonding region and at long range. However, as discussed in the context of the Buckingham-6-8 potential, these potentially come with speed and divergence (as r tends to zero) issues. Thus, long term it may be that effective potentials, such as DE-FF, trained on physical interaction potentials will be a good compromise (e.g. we show in Fig. 1 that the DE-B68 water model follows closely the B68 potential energy surface). In any case, our contribution here is to provide the computational infrastructure to test these hypotheses.
In the revised Introduction, we have elaborated more on the effective treatment of multipole and many-body dispersive interactions by DE-FF. And in the revised Conclusions, we have commented on the systematic next steps that will be needed for functional form exploration.
==========================
Referee: 3
Comments to the Author
Details regarding the "Data Reviewer Checklist" are shown as follows.
6c. Have the authors clearly specified which versions of the software libraries they depend upon were used in the course of the work?
Yes, however, as the installation process can be complex and may cause issues with the local system, I strongly recommend the inclusion of instructions for creating a virtual environment. Using a virtual environment prevents potential conflicts with the local system by isolating package installations. I recommend adding the following to the README.md in the https://github.com/jthorton/de-forcefields repository:
- from now on, everything contained between the opening and closing tags <to be written in README.md> is what I suggest it should be written in the README.md
<to be written in README.md> ++++++++++++++++++
# Header: Installation
## subheader: Installation of virtual environment
We first create a virtual environment to avoid dealing with other packages installed in the system:
Test folder:
`$ mkdir Test && cd Test`
We create a virtual environment:
`$ conda create -n de-forcefields_env`
It will be installed in:
`~/.conda/envs/de-forcefields_env`
To activate this environment, use
`$ conda activate Site-Net_env`
To deactivate the environment:
`$ conda deactivate`
To further remove it:
`$ conda remove -n de-forcefields_env --all`
## subheader: Installation of dependencies and packages:
In this order, install
`(Site-Net_env) $ conda install -n de-forcefields_env -c conda-forge de-forcefields`
</to be written in README.md> ++++++++++++++++++
Response: Thank you, we have edited the README as suggested: https://github.com/jthorton/de-forcefields#installation
2a. Are the data cleaning steps clearly and fully described, either in text
or as a code pipeline?
6b. Are scripts to reproduce the findings in the paper provided?
Both, not entirely.
The main issue lies in the difficulty of reproducing the data curation and training and reproducing the optimizations and benchmarks as described in the manuscript due to the absence of adequate documentation in the accompanying GitHub repository. The repository's file structure, as found at https://github.com/jthorton/double-exp-vdw, is complex.
For example, the scripts to reproduce the training for the qm-benchmarks are located in https://github.com/jthorton/double-exp-vdw/tree/main/inputs-and-results/benchmarks/qm-benchmarks. However, this main page only displays a list of Python scripts without clear instructions about what each script is doing at each time, their function, the order of execution, or what output to expect at each stage.
I recommend creating a Jupyter notebook in the main directory of the double-exp-vdw repository to guide users through the training process for the optimizations and benchmarks. Furthermore, a separate Jupyter notebook should be created to explain the usage of scripts for the curation and creation of the datasets. Another notebook is needed for the ancillary data analysis.
I would gladly continue reviewing the code from this point onward once this is fixed.
Response: Thank you very much for the helpful feedback. We have significantly restructured the supporting data, and added tutorials and Jupyter notebooks to illustrate the training and testing of the DE-FF. In detail:
• We do think that it is clearer to separate out the data curation, training and testing steps, rather than having a single notebook in the main directory. However, we have simplified the directory structure, with signposts to each stage.
• Jupyter notebooks have been added to explain dataset curation for both the quantum chemical and physical property datasets: https://github.com/jthorton/double-exp-vdw/data-set-curation
• We note that training against physical properties in particular is computationally expensive (since new atomistic simulations in the condensed phase must be run for each new parameter set), and so we have provided a tutorial on a reduced physical property training set to illustrate the key concepts: https://github.com/jthorton/double-exp-vdw/inputs-and-results/training Full sets of input files are additionally provided for both the non-bonded and bonded training runs that we performed in this study.
• Finally, full instructions have been added for the testing of the final force fields, including Jupyter notebooks and/or detailed instructions: https://github.com/jthorton/double-exp-vdw/inputs-and-results/testing
• Note that for reproducibility, the required dependencies have been provided in a single environment file: https://github.com/jthorton/double-exp-vdw/blob/main/environment.yaml
Please also note that the Open Force Field software stack, on which we build, is itself extensively documented. For example, OpenFF-Evaluator has tutorials on curating data and training force fields: https://docs.openforcefield.org/projects/evaluator/en/stable/
4a. Is a software implementation of the model provided such that it can be trained and tested with new data?
Not in its current version. As written in the answer to question 6a, the lack of documentation
impedes the straightforward reproduction of this code using the provided test cases. Once this is solved, it will allow to be tested with new data.
Response: We believe that our answer above, particularly the provided notebooks for the train and test stages address this comment.
Round 2
Revised manuscript submitted on 16 Jun 2023![](https://www.rsc-cdn.org/pubs-core/2022.0.146/content/NewImages/icon-letter.png)
![](https://www.rsc-cdn.org/pubs-core/2022.0.146/content/NewImages/chevron-down.png)
07-Jul-2023
Dear Dr Cole:
Manuscript ID: DD-ART-04-2023-000070.R1
TITLE: A transferable double exponential potential for condensed phase simulations of small molecules
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 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 Twitter account @digital_rsc 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,
Linda Hung
Associate Editor
Digital Discovery
Royal Society of Chemistry
![](https://www.rsc-cdn.org/pubs-core/2022.0.146/content/NewImages/icon-comments.png)
![](https://www.rsc-cdn.org/pubs-core/2022.0.146/content/NewImages/chevron-down.png)
Reviewer 2
The modifications address my concerns.
Reviewer 1
The authors have addressed my concerns.
Reviewer 3
I thank the authors for providing a revised manuscript and detailed response.
I can confirm the authors have addressed my comments and questions: The documentation has been improved in the 'de-forcefields' and 'double-exp-vdw' repositories as requested, as well as the Jupyter notebooks provided. In addition, the 'qm-benchmarks' are now well-documented and reproducibility has been tested.
I can also confirm the authors have addressed the questions raised by referees #1 and #2
Transparent peer review
To support increased transparency, we offer authors the option to publish the peer review history alongside their article. Reviewers are anonymous unless they choose to sign their report.
We are currently unable to show comments or responses that were provided as attachments. If the peer review history indicates that attachments are available, or if you find there is review content missing, you can request the full review record from our Publishing customer services team at RSC1@rsc.org.