From the journal Digital Discovery Peer review history

10-May-2024

Dear Dr Bhowmik:

Manuscript ID: DD-ART-04-2024-000099
TITLE: OM-Diff: Inverse-design of organometallic catalysts with guided equivariant denoising diffusion

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.

I have carefully evaluated your manuscript and the reviewers’ reports, and the reports indicate that major revisions are necessary.

Please submit a revised manuscript which addresses all of the reviewers’ comments. Further peer review of your revised manuscript may be needed. 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 on 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

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

Reviewer 1

The split ratio of the dataset is unclear. How many data points are used for training, validation, and test? How was the dataset split, randomly or else?

Reviewer 2

The manuscript “OM-Diff: Inverse-design of organometallic catalysts with guided equivariant denoising diffusion” by Cornet, Benediktsson, Hastrup, Schmidt and Bhowmik describes the design and implementation of an inverse design framework for in-silico design of organometallic complexes. Specifically, the authors focus on complexes of the type L1-M-L2, which may serve as homogeneous catalysts for cross-coupling reactions. They use the binding energy as a target for optimization (based on the volcano plot paradigm of catalytic mechanisms, which has been popularized by Corminbeouf and coworkers in recent years).
The model proposed by the authors, OM-Diff, is based on a diffusion generative model and uses an all-atom point cloud representation of the molecules. The data used for this work comes from the DFT-level subset of the C-C cross-coupling database and contains approx. 7,000 data points. The M center can be Ni, Pd, Pt, Cu, Ag, or Au and there are 72 unique ligands. The majority of the manuscript deals with detailing the diffusion concept and algorithm, as well as explaining the underlying model design. In addition, the authors detail an analysis of the impact of different aspects on the performance of the model. In this regard, the authors investigate the effect of treating the metal center as contextual information and they test the impact of varying the expressivity of the denoiser architecture. The construction of the described model is rigorous and well-explained. The authors are very thorough in detailing their considerations and how these were implemented in the model.
Finally, the authors apply their model to the task of inverse-designing Pt- and Pd-based catalysts for Suzuki coupling and then to the task of inverse-designing a Cu-based catalyst for the same reaction. In terms of the chemical outcome, I think OM-Diff is underwhelming. Not only is the chemical space investigated extremely limited (being limited to 2 ligands on the metal centers is a big shortcoming), but the success rate in terms of validity, novelty and uniqueness are relatively low. They are, of course, higher than the off-the-shelf model, but it is not clear whether such an advanced model is justified by the relatively modest gains in molecular discovery. Furthermore, the Cu-catalyst task did not succeed at all, which points to a lack of applicability of OM-Diff as a useful tool.
Overall, in my view, the paper describes an interesting approach (though not an entirely novel one) and a thorough implementation, with a less-than-spectacular result. I think the manuscript would benefit from some additional chemical investigation (e.g., interpretation of the successful/failed catalysts). Nonetheless, I agree with the authors that this work provides a foundation for future work, and therefore I do believe the manuscript can be published in Digital Discovery.
Some general comments that should be modified in the text:
1. The authors use the term “a more expressive geometric neural network” several times in the text. What do they mean by “more expressive”? Is there a definition or quantitative measure of expressivity?
2. The authors model the metal centers and the ligands separately, because “the geometry of the center often has to follow strict (known) rules.” However, the data that they train on and the type of molecules they generate are all L1-M-L2 type. Thus, it is not demonstrated when OM-Diff learns how to create different types of geometries around the M center.
3. The validity of the model is not very high, even though the complexity of the chemical space is relatively small. Do the authors have any hypothesis as to why so many of the generated molecules are invalid? What kinds of invalid structures are obtained? Are there specific recurring issues that might point to where the model may be improved? It is likely that the validity suffers because of the all-atom approach. Perhaps a more practical approach would be to remove Hs or to group atoms together. Especially is some of them have very conserved geometries, e.g., Me, Ph, iPr groups. There is no need to represent each of the atoms in the group separately.
4. The current version of OM-Diff can be tasked with a single, point-wise target. To what extent can the model deal with multiple objectives and open-ended targets? Is the model generalizable to other types of tasks, that are not binding-energy based?
5. Can the authors say anything about the interpretability of the model, other than commenting that the similarity of the generated ligands demonstrates that the model has learned the underlying distribution? Is there any relationship between specific ligands and better catalysts? Are any of the generated catalysts more promising than other, previously known complexes?
6. Some additional previous works that may be cited: DiffDock (Jaakkola and coworkers), ZeoDiff (Kim and coworkers), denoising diffusion for inverse design of microstructures (Sun and coworkers),

Reviewer 3

This manuscript introduces OM-Diff, an inverse design framework that integrates an equivariant diffusion model and property predictor and is applied to a specific application taken from the literature involving the prediction of the thermodynamic volcano plot descriptors of a Suzuki cross-coupling reaction.

In its current form, the manuscript is intelligible to readers interested in catalyst optimization and probably not understandable either to people unfamiliar with catalysis. The technical sections (section methods) are written for an audience of non-chemists. The language, terminology and even emphasis is not placed on aspects that will match the interest and knowledge of chemists capable of assessing the relevance and value of the actual chemical predictions. Alternatively, the "chemical section" (section experiment and results) is presented with so little relevant chemical context (taken straight from existing literature but at the same time omitting a considerable amount of relevant literature on the topics) that it is impossible to understand the chemical relevance of the inverse design objective and task. With this two major problems, we get lost in trying to assess the value of the proposed approach.

One can also strongly doubt of the relevance of equivariance and the authors should present the invariant analogue model for comparison.

Overall, the content of the manuscript in its current form is not adequately elaborated and can thus not be understood and properly evaluated by the audience it does target. If the authors do not want to modify their technical part to make it more intelligible, they should at least considerably improve the chemical section and also demonstrate that equivariance is necessary to learn this target property.

This text has been copied from the PDF response to reviewers and does not include any figures, images or special characters.

Dear Dr. Schrier,

On behalf of all co-authors, I would like to resubmit our manuscript entitled “OM-Diff: Inverse-design of organometallic catalysts with guided equivariant denoising diffusion” in Digital Discovery.
Thank you for giving us the opportunity to improve our manuscript based on the reviewers’ comments. To facilitate further review, we quote review comments point by point in the reply document and respond to each of them. We also attached the manuscript with marked changes and a clean version. The changes are marked with red and blue in the marked manuscript. We hope that our work is now suitable for publication in your highly esteemed journal.
Thank you in advance for your time in handling the manuscript.

Review 1
Comment The split ratio of the dataset is unclear. How many data points are used for training,
validation, and test? How was the dataset split, randomly or else ?

Answer We thank the reviewer for these important remarks. We have added Section S1.1 in
Appendix detailing how the data was split for the different results reported in the manuscript.
Review 2

The manuscript “OM-Diff: Inverse-design of organometallic catalysts with guided equivariant denoising diffusion” by Cornet, Benediktsson, Hastrup, Schmidt and Bhowmik
describes the design and implementation of an inverse design framework for in-silico design of organometallic complexes. Specifically, the authors focus on complexes of the type
L1-M-L2, which may serve as homogeneous catalysts for cross-coupling reactions. They
use the binding energy as a target for optimization (based on the volcano plot paradigm
of catalytic mechanisms, which has been popularized by Corminbeouf and coworkers in
recent years). The model proposed by the authors, OM-Diff, is based on a diffusion generative model and uses an all-atom point cloud representation of the molecules. The data
used for this work comes from the DFT-level subset of the C-C cross-coupling database
and contains approx. 7,000 data points. The M center can be Ni, Pd, Pt, Cu, Ag, or
Au and there are 72 unique ligands. The majority of the manuscript deals with detailing
the diffusion concept and algorithm, as well as explaining the underlying model design.
In addition, the authors detail an analysis of the impact of different aspects on the performance of the model. In this regard, the authors investigate the effect of treating the
metal center as contextual information and they test the impact of varying the expressivity of the denoiser architecture. The construction of the described model is rigorous
and well-explained. The authors are very thorough in detailing their considerations and
how these were implemented in the model. Finally, the authors apply their model to
the task of inverse-designing Pt- and Pd-based catalysts for Suzuki coupling and then
to the task of inverse-designing a Cu-based catalyst for the same reaction. In terms
of the chemical outcome, I think OM-Diff is underwhelming. Not only is the chemical
space investigated extremely limited (being limited to 2 ligands on the metal centers is a
big shortcoming), but the success rate in terms of validity, novelty and uniqueness are
relatively low. They are, of course, higher than the off-the-shelf model, but it is not clear
whether such an advanced model is justified by the relatively modest gains in molecular
discovery. Furthermore, the Cu-catalyst task did not succeed at all, which points to a
lack of applicability of OM-Diff as a useful tool. Overall, in my view, the paper describes
an interesting approach (though not an entirely novel one) and a thorough implementation, with a less-than-spectacular result. I think the manuscript would benefit from some
additional chemical investigation (e.g., interpretation of the successful/failed catalysts).
Nonetheless, I agree with the authors that this work provides a foundation for future
work, and therefore I do believe the manuscript can be published in Digital Discovery.

Answer We thank the reviewer for taking the time to review our manuscript so thoroughly.
We address each of the the reviewer’s comments here below, and we additionally highlight the
implemented modifications in red in the amended manuscript.

Comment The authors use the term “a more expressive geometric neural network” several times
in the text. What do they mean by “more expressive”? Is there a definition or quantitative measure
of expressivity?

Answer We thank the reviewer for this important remark. We have amended the initial explanation to make it clearer. It now reads,

Along with the usual scalar hidden states, a set of (equivariant) vectorial hidden states is
also maintained and updated for each atom through message-passing [1]. The messages
exchanged in our εθ are constructed using local descriptors based on pairwise distances
and angles, whereas EGNN [2] only leverages pairwise distances to construct messages.
This leads to a more expressive architecture, as local angular information can now be
resolved [3], while remaining cheap to evaluate compared to architectures leveraging
higher-order tensors

Comment The authors model the metal centers and the ligands separately, because “the geometry
of the center often has to follow strict (known) rules.” However, the data that they train on and the
type of molecules they generate are all L1-M-L2 type. Thus, it is not demonstrated when OM-Diff
learns how to create different types of geometries around the M center.

Answer We thank the reviewer for their remark, and acknowledge that the initial formulation
was approximate. We intended to express that fixing the geometry around the center, and providing
it as context (instead of learning it) could be beneficial. The experiments only demonstrate that the
model can generate one type of coordination, i.e. L1-M-L2. However, provided data with multiple
metal centers featuring different coordination patterns, OM-Diff can handle those out-of-the-box.
In the manuscript, we have amended the paragraph titled Organometallic complexes in Section 2.1 Data representation. In particular, we have added the following closing sub-paragraph
contrasting the generality of the introduced formulation with the presented experiments

In our experiments in Section 3, we consider a simple case where only the metal center
is fixed and only one coordination pattern is present in the training data. However, the
formulation in Eq. (1) entails more complicated cases such as centers featuring different
coordination patterns, or scaffold-based design where some of the ligands can be also be
considered part of the center subset.

Comment The validity of the model is not very high, even though the complexity of the chemical
space is relatively small. Do the authors have any hypothesis as to why so many of the generated
molecules are invalid? What kinds of invalid structures are obtained? Are there specific recurring
issues that might point to where the model may be improved?

Answer We agree with the reviewer that validity is limited, and that highlighting the limitations
of the model more clearly can be beneficial to the community.
We first note that it is difficult to put the obtained values in perspective as (1) no previous work
has directly operated in a similar setup (i.e. generation in 3D of relatively large organometallic
complexes from scratch), (2) there exists no unique definition of validity – e.g. in related work on
organic molecules, charged complexes are often deemed valid, or disconnected fragments are filtered
out, thereby leading to validity metrics that can not compared. In this paper, as the ultimate goal
was to present a closed-loop framework and run DFT calculations with the generated complexes,
we designed a strict validity metric as to ensure that the samples could effectively be used for
calculations. In parallel, we also wanted to see whether the model could be of practical interest
without resorting to pre-processing nor post-processing of the generated samples.
As for the kind of invalid structures, Table S2 provides an element of response. We can observe
that around 50% of the generated structures are already deemed invalid due to unphysical pairwise
distances: either because of disconnected fragments or clashing atoms. Among those passing the
distance check, the bonding structure can be inferred by RDKit for around 66% of the samples
without the addition of charges nor implicit hydrogen atoms. In summary, the most recurring
failure modes are: disconnected/clashing atoms, and missing protons leading to charged complexes.
Regarding directions for improvement, while a post-processing procedure would increase the
effective validity it would not fix the model itself. However, additionally modelling the bonding
information in the diffusion process has recently been shown to mitigate some of the connectivity,
and valence issues. This could constitute an interesting avenue for future work.
In the manuscript, we have amended the paragraph title Validity in 3.2 Unconditional
Generation where we added the following closing sub-paragraph

In Table S2, , we provide more details regarding the validity results. We observe that
already 50% of the generated structures are deemed invalid due to abnormal pairwise
distances, i.e. isolated or clashing atoms. Among the structures passing the distance
check, only around 60% can be properly parsed by RDKit [4], i.e. the bonding structure
can be inferred without charge and the resulting object contains exactly two fragments
corresponding to L1 and L2. This highlights the most recurring failure modes that were
observed: disconnected/clashing atoms, and missing protons leading to charged complexes. While an appropriate post-processing procedure, e.g. removing the disconnected
fragments or fixing valences, could further increase the effective validity, modelling the
bonding information [5, 6] could help solve connectivity and valence issues at the source.

Comment It is likely that the validity suffers because of the all-atom approach. Perhaps a more
practical approach would be to remove Hs or to group atoms together. Especially is some of them
have very conserved geometries, e.g., Me, Ph, iPr groups. There is no need to represent each of the
atoms in the group separately.

Answer We thank the reviewer for these relevant suggestions. We generally agree that removing
hydrogen atoms is an interesting option. While a simple pre-processing step removing all hydrogen
atoms from the training database is straightforward to implement, an hydrogenation algorithm
would be required to append the Hs to the generated complexes before running DFT calculations.
Grouping atoms together is another promising approach, whilst being slightly more involved. It
would require to define (1) a coarse-graining procedure mapping all-atom structures to a grouped
representation, and (2) a post-processing procedure to replace the group nodes by their actual
composition. The latter would additionally require to decide on the orientation and/or attachment
point if multiple are possible. Both approaches would lead to smaller point clouds, and hence
faster training/sampling along with (most likely) increased validity. As mentioned in the previous
answer, in this paper we aimed at minimizing the amount of pre-/post-processing, to illustrate the
practicality of the framework out-of-the-box.
We have however added the suggestions in a paragraph in the discussion section. We copy the
relevant excerpt below for reference,

To address the limited effective validity of the generated complexes, a possible avenue
would be to only consider heavy atoms in the generative process, and adding an hydrogenation post-processing step. This would also make the training and sampling faster by
reducing the size of generated point clouds. Another related approach would consist in
grouping atoms together, i.e. representing fragments as coarse-grained nodes [7]. Such
procedure would require to define (1) a mapping transforming all-atom structures to
coarse-grained representations, and (2) a post-processing step replacing the group nodes
in the generated structures by their actual composition. The latter would additionally
require to decide on the orientation and/or attachment point if multiple possibilities exist [8]. A third approach could leverage the guidance setup described from Section 2.3,
where a target function could be designed to include the feedback of a classifier trained
to distinguish between valid and invalid compounds. [...] Other methodological improvements in OM-Diff could include modelling of atom and bond types as categorical
variables [5, 6] (instead of the continuous relaxation used in this work) and predicting
the denoised structures directly, as it has been shown to work better for atomistic data
[6].

Comment The current version of OM-Diff can be tasked with a single, point-wise target. To what
extent can the model deal with multiple objectives and open-ended targets? Is the model generalizable
to other types of tasks, that are not binding-energy based?

Answer We thank the reviewer for these relevant questions. In principle, multi-conditional generation is possible in the OM-Diff framework, and can be performed by composing individual
energy functions for each condition. The same holds for open-ended targets. We have amended the
manuscript and added a paragraph entitled Advanced conditioning in Section 2.3 Regressor
guidance. It reads,

While the exposition of the guidance mechanism was particularised to a single target
scalar value, it can readily be extended to more advanced conditioning schemes. To
condition on multiple properties for instance, an energy function that rely on several
property predictors, {yϕi
}
I
i=1, can be designed. An example of such function could write,
fϕ

In terms of generalizability, the guidance procedure is agnostic of the nature of y and is therefore
not limited to binding-energy.

Comment Can the authors say anything about the interpretability of the model, other than commenting that the similarity of the generated ligands demonstrates that the model has learned the
underlying distribution? Is there any relationship between specific ligands and better catalysts? Are
any of the generated catalysts more promising than other, previously known complexes?

Answer We thank the reviewer for raising a relevant question. If we focus on the 15 complexes
that were validated with DFT, we can see that the model prefers phosphine ligands, with 13 out
of 15 complexes containing at the least one. Phosphines are a popular choice in cross coupling
reactions due to their σ-donating and π-accepting character, as well as being spectator ligands,
which can be fine tuned for a specific reaction. Therefore, we think it is likely that the model has
learned the underlying chemistry and chooses those ligands that fit for this energy range. In the
case of Pd-based complexes, a CO is a ligand in 4 out of 5 complexes, which is also a σ-donating and
π-accepting ligand. For the other ligands present in the 15 complexes, there 4 complexes are with
pyridine-based ligands, 3 complexes with imidazole-based ligands, 3 complexes with imidazolidinebased ligands. The changes there are more nuanced and they effect the reaction energies through
by modifying the electron density of the conjugated part of the ligand.
We note that these complexes do not improve the reaction energetics compared to the ones in
the original database [9] article, as it already contained Pd and Pt catalysts lying in the energetic
region of interest. We do not consider this as being an issue, since the manuscript focuses on the
description, functionalities and capabilities of OM-Diff. We think that further analysis, such as
bond order analysis or population analysis, is out of scope of this article.
We have amended the manuscript and added a sub-paragraph under Inverse-designing optimized catalysts for the Suzuki reaction that briefly touches upon this topic. It reads,

Interestingly, we note that although the training dataset contains various ligand types,
eight out of ten Pd/Pt complexes contain phosphine-type ligands, that are σ-donors and
π-acceptors. Furthermore, four out of five Pd-complexes contain a CO ligand, a σ-donor
and π-acceptor. These ligand properties are known to be important for cross-coupling
reactions. Although the model doesn’t explicitly learn the electronic structure properties,
its choice of ligands can be rationalized.

Comment Some additional previous works that may be cited: DiffDock (Jaakkola and coworkers),
ZeoDiff (Kim and coworkers), denoising diffusion for inverse design of microstructures (Sun and
coworkers)
Answer We thank the reviewer for pinpointing these relevant papers. We have added them in
the introduction section.
Review 3
This manuscript introduces OM-Diff, an inverse design framework that integrates an
equivariant diffusion model and property predictor and is applied to a specific application taken from the literature involving the prediction of the thermodynamic volcano plot
descriptors of a Suzuki cross-coupling reaction. In its current form, the manuscript is
intelligible to readers interested in catalyst optimization and probably not understandable
either to people unfamiliar with catalysis. The technical sections (section methods) are
written for an audience of non-chemists. The language, terminology and even emphasis
is not placed on aspects that will match the interest and knowledge of chemists capable of
assessing the relevance and value of the actual chemical predictions. Alternatively, the
”chemical section” (section experiment and results) is presented with so little relevant
chemical context (taken straight from existing literature but at the same time omitting
a considerable amount of relevant literature on the topics) that it is impossible to understand the chemical relevance of the inverse design objective and task. With this two
major problems, we get lost in trying to assess the value of the proposed approach.
Answer We thank the reviewer for taking the time to review our manuscript, and raising concerns
about its overall readability. To address the reviewer’s concerns, we have considerably amended
the submitted manuscript, and added detailed explanations to the method sections to make them
more accessible. We have marked all changes in blue.
We feel important to restate that the main contribution of our paper is not a discovery campaign
per se, but rather the introduction of a generic framework (based on state-of-the-art machine
learning methods) enabling the discovery of organometallic complexes. To illustrate what the
framework is capable of, we select a well-studied yet practically relevant problem, and show that
OM-Diff is a tool that can be leveraged in a practical discovery pipeline.
Comment One can also strongly doubt of the relevance of equivariance and the authors should
present the invariant analogue model for comparison.
Answer We thank the reviewer for this remark.
Regarding the denoising neural network, as there exists no canonical representation for atomistic
point clouds, an invariant model could not be used. A simple illustration of this could be to consider
a point cloud and a rotated version thereof. As the output of an invariant model is insensitive to
rotations of its input, the two point clouds would erroneously lead to two identical predictions.
However, an alternative could be to have a model that does not account for symmetries, and sees
a point cloud and a rotated version thereof as two distinct (and unrelated) objects. This approach
has been shown to perform significantly worse than its equivariant counterpart, even when coupled
with data augmentation [10].

Regarding the surrogates used for guidance and screening, it is correct that invariant models
could be used as the prediction task is inherently invariant, i.e. energy prediction. However, leveraging equivariant features in the message-passing phase, even if the downstream task is invariant,
has been shown to be a useful inductive bias and lead to improved performance [11]. Additionally,
in our models the equivariant features are limited to vectors, and therefore do not incur a significant
overhead with respect to an invariant model.
Comment Overall, the content of the manuscript in its current form is not adequately elaborated
and can thus not be understood and properly evaluated by the audience it does target. If the authors
do not want to modify their technical part to make it more intelligible, they should at least considerably improve the chemical section and also demonstrate that equivariance is necessary to learn
this target property.
Answer We thank the reviewer for raising concerns about the readability of the manuscript. We
took the feedback seriously, and consequently amended numerous sections of the document to make
it more accessible for an audience that is less familiar with technical aspects of generative modelling
architectures. We have marked all changes in blue.
While all changes cannot be enumerated here, we highlight a concrete example at the beginning
of the methods section where we have added a signpost to help the reader navigate the section. It
reads,
Problem Statement The goal of OM-Diff is to inverse-design novel organometallic complexes, that we refer to as C, with desired properties, that we denote y. We frame
this task as a conditional sampling problem, where we aim to sample from the conditional
distribution q(C|y). We approximate the conditional distribution, q(C|y) ≈ pθ,ϕ(C|y), by
combining an unconditional diffusion model, pθ(C), with a property predictor pϕ(y|C).
Outline In Section 2.1, we introduce how organometallic complexes C are represented
in practice, i.e. the data representation that the different models operate on. In Section
2.2, we present the unconditional generative model. Using the unconditional generative
model and an auxiliary property predictor, we present in Section 2.3 the guidance procedure for performing conditional sampling. Finally in Section 2.4, we present details
about the surrogate model used to screen the complexes sampled from the generative
model.
Regarding the chemical section, we refer to our previous answer.
Concerning equivariance, the reviewer is correct that it is not strictly required to learn the
target property. Regarding the surrogate used for guidance, the two requirements are that (1)
the model is a continuously differentiable function (fulfilled by using smooth activation functions),
and (2) its output is rotationally invariant as to preserve the overall rotation-invariance of the
learned conditional distribution [12]. As mentioned in the previous answer, rotation invariance can
be achieved by a model that only leverages invariant features. However, resorting to equivariant
features has been shown to yield better performance [11]. For the screening surrogate, any model
could technically be used. For simplicity, we use an architecture similar to that of the guidance
regressor.

References
[1] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural
message passing for quantum chemistry. In International conference on machine learning, pages
1263–1272. PMLR, 2017.
[2] Vıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural
networks. In International conference on machine learning, pages 9323–9332. PMLR, 2021.
[3] Chaitanya K. Joshi, Cristian Bodnar, Simon V Mathis, Taco Cohen, and Pietro Lio. On the
expressive power of geometric graph neural networks, 2023.
[4] Greg Landrum et al. Rdkit: A software suite for cheminformatics, computational chemistry,
and predictive modeling. Greg Landrum, 8:31, 2013.
[5] Clement Vignac, Nagham Osman, Laura Toni, and Pascal Frossard. Midi: Mixed graph and
3d denoising diffusion for molecule generation. In Joint European Conference on Machine
Learning and Knowledge Discovery in Databases, pages 560–576. Springer, 2023.
[6] Tuan Le, Julian Cremer, Frank Noe, Djork-Arn´e Clevert, and Kristof T Sch¨utt. Navigating
the design space of equivariant diffusion-based generative models for de novo 3d molecule
generation. In The Twelfth International Conference on Learning Representations, 2024.
[7] Tomer Weiss, Eduardo Mayo Yanes, Sabyasachi Chakraborty, Luca Cosmo, Alex M Bronstein, and Renana Gershoni-Poranne. Guided diffusion for inverse molecular design. Nature
Computational Science, 3(10):873–882, 2023.
[8] Xiang Fu, Tian Xie, Andrew Scott Rosen, Tommi S. Jaakkola, and Jake Allen Smith. MOFDiff:
Coarse-grained diffusion for metal-organic framework design. In The Twelfth International
Conference on Learning Representations, 2024.
[9] Benjamin Meyer, Boodsarin Sawatlon, Stefan Heinen, O Anatole Von Lilienfeld, and Cl´emence
Corminboeuf. Machine learning meets volcano plots: computational discovery of cross-coupling
catalysts. Chemical science, 9(35):7069–7077, 2018.
[10] Emiel Hoogeboom, Vıctor Garcia Satorras, Cl´ement Vignac, and Max Welling. Equivariant
diffusion for molecule generation in 3d. In International conference on machine learning, pages
8867–8887. PMLR, 2022.
[11] Benjamin Kurt Miller, Mario Geiger, Tess E Smidt, and Frank No´e. Relevance of rotationally
equivariant convolutions for predicting molecular properties. arXiv preprint arXiv:2008.08461,
2020.
[12] Fan Bao, Min Zhao, Zhongkai Hao, Peiyao Li, Chongxuan Li, and Jun Zhu. Equivariant
energy-guided sde for inverse molecular design. In The eleventh international conference on
learning representations, 2022.





Best regards,
Arghya Bhowmik

26-Jun-2024

Dear Dr Bhowmik:

Manuscript ID: DD-ART-04-2024-000099.R1
TITLE: OM-Diff: Inverse-design of organometallic catalysts with guided equivariant denoising diffusion

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 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:

Several reviewers have noted that the article does not adequately situate itself with respect to prior work. Please address the issue of thoroughly describing relevant background work in the introduction and in clearly delineating the similarities and differences of the present work from those past work for the benefit of readers.

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

Reviewer 1

Choosing an unconventional train-validation-test split, such as the 90-5-5 split in the study presented mentioned, deviates from the more standard practice of an e.g. 80-10-10 split. Could the authors give some rationale why this uncommon imbalanced split was necessary? And considering this imbalanced split, did the authors run the diffusion model multiple times to demonstrate the robustness of the model?

Reviewer 2

In the revised manuscript, it is clear that the authors made a very sincere and thoughtful effort to implement the feedback from the reviewers and modify the text accordingly. I appreciate their additional work and their willingness to make changes where necessary, as well as their detailed and courteous replies to all reviewers. I was happy to see that they added a bit more chemical interpretation of the results, even if they only reiterate previously known insights and do not reveal new ones (which would have been very impressive).

As I wrote in the previous round of review, I think the OM-Diff model described herein has some specific limitations (as do all models), but that does not mean that there is no value to the work. On the contrary, the field of catalysis is enormously important and this contribution may prove to be a stepping-stone for future developments. Therefore, I support accepting the manuscript for publication.

The only two issues that I feel need to be clarified at this point are the extent of novelty of the work, and the extent of applicability/impact of the model. Regarding the first point: As I mentioned in my previous review, the approach taken here is not novel. Reference 40 in the manuscript also used guided diffusion with an external predictor network as a replacement for conditional generation. Yet, the authors do not state this explicitly (even after I noted it before). Instead, it seems to me that they actually bury this very relevant comparison by simply referring to reference 40 as a work that deals with fragments rather than atoms. This, in my view, is a disingenuous representation and must be rectified. Regarding the second point: the authors should state much more clearly that the current demonstration of OM-Diff is highly specialized and really does not provide a general solution for homogeneous catalysis. The fact that they only demonstrated L-M-L systems and only focused on Suzuki coupling means that the scope of the study is limited. This is not a criticism of the work; it is perfectly acceptable to have a small-scope study. However, it is less acceptable to frame the work as a general solution for inverse design for homogeneous catalysts, when such broad applicability has not been demonstrated whatsoever. The authors have added a very nice summary at the end, which details possible future developments. If indeed such advances are made, then perhaps the model may begin to provide a general use-case solution. In its current form, and based on the current results, it is not justified to claim so, in my view.

Reviewer 3

I would like to thank the authors for the revision. The emphasis is now more clearly placed on the introduction of the framework and less on the design and discovery. The framework is useful and interesting. Publication in Digital Discovery would be appropriate.
My only remaining comment concerns the references, which are still a bit light. In the introduction, the authors tend to miss contributions from many other groups regarding generative models (more careful work should be done here) and/or genetic algorithms in particular (e.g., Aspuru-Guzic, Kulik, Corminboeuf, .....). I made the same observation, while reading the rest of the text so perhaps a last in-depth literature check would be wise.

\title{\textbf{Review reply} \\ \textsc{OM-Diff}: Inverse-design of organometallic catalysts with guided equivariant denoising diffusion}

\section*{Editor's comments}
\begin{quote}
\textit{Several reviewers have noted that the article does not adequately situate itself with respect to prior work. Please address the issue of thoroughly describing relevant background work in the introduction and in clearly delineating the similarities and differences of the present work from those past work for the benefit of readers.}
\end{quote}

{\color{olive}
\paragraph{Answer} We thank the editor for reminding us for better literature review. We have reworked and significantly added to the introduction section, also specifically clarifying how the current work expands on previously published work.
All modifications are marked in blue.

}


\section*{Review 1}

\begin{quote}
\textit{Choosing an unconventional train-validation-test split, such as the 90-5-5 split in the study presented mentioned, deviates from the more standard practice of an e.g. 80-10-10 split. Could the authors give some rationale why this uncommon imbalanced split was necessary?}
\end{quote}


{\color{olive} \paragraph{Answer} We thank the reviewer for this important comment. Regarding the diffusion model, we do not use the validation/test partitions for anything but Fig.~S1. We instead evaluate the performance of the model based on generated samples.
As we use the same split for training the guidance model, we made the validation and test folds as small as possible, while keeping a sufficient amount of data points to allow for meaningful estimation of the error metrics (e.g. for early stopping of the model), and producing Figs.~7b and S12.

For the screening surrogate, the values reported in Table 1 were obtained through a 10-fold cross-validation procedure. For each fold, the model was trained on $80\%$ of the data, while $10\%$ was used for validation, and the last $10\%$ for testing; as per standard practice.

In \textbf{S1.1 Data splits}, we have added some clarifying sentences.
\begin{quote}
\textit{{\color{blue}}}
\end{quote}

}


\begin{quote}
\textit{And considering this imbalanced split, did the authors run the diffusion model multiple times to demonstrate the robustness of the model?}
\end{quote}
{\color{olive}
\paragraph{Answer} We only trained the diffusion model once. We note that as we include $90\%$ of the dataset in the training data, we expect less variance from multiple runs than if trained on $80\%$.
}

\section*{Review 2}
\begin{quote}
\textit{In the revised manuscript, it is clear that the authors made a very sincere and thoughtful effort to implement the feedback from the reviewers and modify the text accordingly. I appreciate their additional work and their willingness to make changes where necessary, as well as their detailed and courteous replies to all reviewers. I was happy to see that they added a bit more chemical interpretation of the results, even if they only reiterate previously known insights and do not reveal new ones (which would have been very impressive).
As I wrote in the previous round of review, I think the OM-Diff model described herein has some specific limitations (as do all models), but that does not mean that there is no value to the work. On the contrary, the field of catalysis is enormously important and this contribution may prove to be a stepping-stone for future developments. Therefore, I support accepting the manuscript for publication.
The only two issues that I feel need to be clarified at this point are the extent of novelty of the work, and the extent of applicability/impact of the model. }
\end{quote}


{\color{olive}
\paragraph{Answer} We thank the referee for taking the time to review our manuscript.

}


\begin{quote}
\textit{Regarding the first point: As I mentioned in my previous review, the approach taken here is not novel. Reference 40 in the manuscript also used guided diffusion with an external predictor network as a replacement for conditional generation. Yet, the authors do not state this explicitly (even after I noted it before). Instead, it seems to me that they actually bury this very relevant comparison by simply referring to reference 40 as a work that deals with fragments rather than atoms. This, in my view, is a disingenuous representation and must be rectified. }
\end{quote}

{\color{olive}
\paragraph{Answer} We thank the reviewer for this important remark. We have added a separate sub-section in the introduction discussing previous work. Along with other related work, we specifically explain the similarities and differences between GAUDI and OM-DIFF in this section.

}



\begin{quote}
Regarding the second point: the authors should state much more clearly that the current demonstration of OM-Diff is highly specialized and really does not provide a general solution for homogeneous catalysis. The fact that they only demonstrated L-M-L systems and only focused on Suzuki coupling means that the scope of the study is limited. This is not a criticism of the work; it is perfectly acceptable to have a small-scope study. However, it is less acceptable to frame the work as a general solution for inverse design for homogeneous catalysts, when such broad applicability has not been demonstrated whatsoever. The authors have added a very nice summary at the end, which details possible future developments. If indeed such advances are made, then perhaps the model may begin to provide a general use-case solution. In its current form, and based on the current results, it is not justified to claim so, in my view.
\end{quote}

{\color{olive}
\paragraph{Answer} We thank the reviewer for this important remark. We have amended the discussion to reflect better the limited scope of the study.

}




\section*{Review 3}

\begin{quote}
\textit{I would like to thank the authors for the revision. The emphasis is now more clearly placed on the introduction of the framework and less on the design and discovery. The framework is useful and interesting. Publication in Digital Discovery would be appropriate. }
\end{quote}
{\color{olive}
\paragraph{Answer} We thank the reviewer for this important remark.

}

\begin{quote}
\textit{My only remaining comment concerns the references, which are still a bit light. In the introduction, the authors tend to miss contributions from many other groups regarding generative models (more careful work should be done here) and/or genetic algorithms in particular (e.g., Aspuru-Guzic, Kulik, Corminboeuf, .....). I made the same observation, while reading the rest of the text so perhaps a last in-depth literature check would be wise. }
\end{quote}

{\color{olive}
\paragraph{Answer} We thank the reviewer for pointing out this. We have added a more thorough literature review in the introduction, including a separate related work' subsection.

}

19-Jul-2024

Dear Dr Bhowmik:

Manuscript ID: DD-ART-04-2024-000099.R2
TITLE: OM-Diff: Inverse-design of organometallic catalysts with guided equivariant denoising diffusion

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 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

EDITOR'S COMMENT;

- The authors may wish to incorporate the references mentioned by Referee 1 as a "note added in proof" during the proofing stage.

Reviewer 3

The authors have addressed all the comments appropriately.

Reviewer 1

I want to clarify that in my previous comments, I hope the authors can generate samples multiple times to evaluate the performance of their model, such as the validity and uniqueness and novelty, rather than to train the model a couple of times. As far as I know, the generation of diffusion model is pretty random. From the revised manuscript, I guess the authors may only generate samples once but increases the numbers of generated samples. But it may be more persuasive to repeat the generation under a given number of samples. For example, run the generation process three times and each time generate 10000 samples to check validity and report the average. In addition, when I re-reviewed the paper, I found Jin and Merz recently published similar work (LigandDiff : https://doi.org/10.1021/acs.jctc.4c00232 and multi-LigandDiff: https://chemrxiv.org/engage/chemrxiv/article-details/6658e2eb418a5379b0b280ba ), which the authors did not mention in the introduction. The authors may consider citing their work to provide readers a comprehensive introduction because in the current introduction, the authors do not mention any diffusion models used for organometallics.

Reviewer 2

I support publication of the revised manuscript.