3) Data generation

• Established workflows, such as optimizing geometries with cheap semi-empirical methods, followed by single-point DFT for energies and gradients, are entirely overlooked in generating the dataset.

  • We point out that the workflow we employed as detailed in Section 4 is optimizing geometries with B3LYP/def2-SVP, followed by single-point calculation with B3LYP/def2-SVP. Our workflow is more accurate than optimizing structures with cheap semi-empirical methods.

• “Labeled as important phosphate groups by our chemistry expert” (p. 5) What is the rationale behind labeling phosphate groups (which are one specific out of many relevant functional groups)?

  • By labeling some functional groups that play important roles in molecules, we can prevent accidental deletion during the screening process. For example, phosphate groups often influence various properties of molecules.

• “while the significance of other conformations is also noteworthy” (p. 6) This statement is highly ambiguous. What is the relevance of further (non-minimum) conformations?

  • That's a great point! One of the biggest features of our dataset is the relatively high sampling density of conformations for the same molecule. On average, there are dozens of conformations for each molecule in our dataset, while other datasets typically have only a few structures per molecule (usually only local minima are included). It is well known that neural network potentials are reliable within the data region but exhibit instability outside of the sampling area. Therefore, expanding the sampling area can help achieve a larger confidence interval and build a more stable model. Consequently, conformations in non-equilibrium regions are also important. Especially when you want to do some long-time MD experiments.

• The paper provides insufficient detail on the fragmentation / recombination strategy to construct the molecule library. What constitutes a fragment? How were the fragments generated and selected?

  • The fragment algorithm is not the main contribution of this work and, therefore, has not been elaborated upon in detail here. A comprehensive discussion of the fragment algorithm will be presented in an independent publication later. For simplicity, we employed a technique similar to [1], where a set of rules is defined to retain meaningful chemical substructures and recombine chemical motifs.

• An estimate of the computational resources required for dataset generation (in terms of CPU hours), would provide readers with a realistic perspective on the cost, and feasibility of expanding such datasets.

  • We added a new subsection in Section 4 to describe the computation cost.

[1] J. Degen, C. Wegscheid‐Gerlach, A. Zaliani, and M. Rarey, “On the art of compiling and using ‘drug‐like’ chemical fragment spaces,” ChemMedChem, vol. 3, no. 10, pp. 1503–1507, Oct. 2008, doi: 10.1002/cmdc.200800178.

4) Writing and Figures

• The paper would strongly benefit from repositioning the use cases earlier in the paper to align the introduction with actual practical applications. A number of vague statements regarding applicability should be removed.

  • We have moved the relevant content regarding potential application scenarios to the beginning of Section 2. And deleted some of the statements.

• The paper uses overly promotional language, particularly in the abstract and introduction. Examples include “unprecedented innovation and efficiency”, “professional and transformative research”, “high-precision […] quantum mechanical level”, “meticulously computed”, “comprehensive studies of structure–property relationships”. These sections should be thoroughly revised.

  • We have adjusted some adjectives and adverbs and removed certain phrases to make the descriptions more precise.

• A better distinction between molecules and conformations would be beneficial for readability of the paper, and for understanding the structure of the dataset. E.g. the analyses in Fig. 5a and Fig. 6 would be more informative on the molecule level (rather than on the conformation level). A plot of average conformation number against heavy atom count, for example, would help contextualize the dataset structure. Similarly, plots of further simple-to-compute topological molecular descriptors (molecular weight, ring counts, …) would add further information.

  • We have added a table with respect to average conformation number and other molecular descriptors in the main text.
  • We believe the heavy atom numbers are also informative for some tasks (e.g. transfer learning tasks) and so we keep the figure.

• The purpose of Fig. 5b, showing energy as a function of dihedral angle without specifying the molecule(s) is unclear. From my perspective, this plot, as it stands, is elementary, and provides no insights into the dataset.

  • Thank you for your feedback. Our intention of Fig. 5b was to convey the comprehensiveness of our sampling method, meaning that the entire space around a flexible bond is evenly scanned in nearly all 360 degrees. And we did the same rotational scan to every flexible bond of every molecule in the dataset. It's possible that our statement was not sufficient to convey this point. We will revise some of the sentences.

---

Once again, we want to raise our sincere gratitude to you!

We have absorbed many of your professional and detailed suggestions.

These valuable insights have truly enhanced the quality of the paper. Many Thanks!

Dear Reviewe rvr1N:

Highly constructive feedback we would like to express our deep gratitude to！ Since your feedback is very detailed, please don't mind if we address it section by section.

1) The goal and potential application scenario

• The paper does not clearly articulate the goal of developing the dataset, or potential use cases. Section 2.2 should be re-written to clarify these use cases, and clarify the unclear “conformer generation” task.
  • We have revised the paragraph describing the potential application scenarios in introduction in the revised version;
  • We have clarified the task of conformation generation with more precise language. Additionally, we have included relevant references in the revised version;

2) Basic concepts introduction

• Figure 2 and the accompanying discussion makes strong statements about computation accuracy using different basis sets.

  • In fact, the original figure simply illustrated the classification of different methods and basis sets. To avoid any misleading, we have removed Figure 2 and adjusted the statements related to accuracy.

• Force fields and potential energy surface are often discussed as two disjoint entities – rather than explaining that force fields are formally the derivative of the potential energy surface.

  • We have revised the relevant text to better align with reasonableness.

• The high-level introduction to DFT misses a number of key points: The main goal is to approximate solutions to the electronic Schrödinger equation for obtaining energies and, possibly, estimate further molecular property labels from the computed solution.

  • We have added this point to the main text in the revised version.

• SMILES and InChI representations are introduced without clarity on their primary function as molecular graph notations.

  • We have added their encoding method of molecular graphs in the revised version.

• “Heavy atom is any atom other than hydrogen, typically used in molecular studies to focus on more complex atomic interactions” (p. 3) This statement fails to address why distinguishing heavy atoms is relevant to real-world applications.

  • We have added why distinguishing heavy atoms is important in the revised version.

• Statements like “B3LYP/def2-SVP as one of the highest precision levels achievable“ are strongly misleading. In fact, in computational chemistry, double-zeta basis sets are considered a rather low-level method.

  • Actually, the complete sentence of the original text is "one of the highest precision levels achievable within an acceptable computational cost range for large-scale calculations of organic molecular systems. "
  • We believe this description is accurate. B3LYP is indeed one of the highest accuracy choices widely used in practice for datasets with tens of millions of samples.

4) regarding DFT level of theory problems

• In Figure 2, the precision levels are just categories, not linear scaled.

• Are there any dispersion corrections (as done in GEOM, and QCDGE)?

  • Today, the most commonly used dispersion corrections are DFT/D3[1] and DFT/D4[2] coming from Professor Grimme, the honored academic luminary in this field. We directly quote the original text from Page 2 in [1] by him: it has been designed from the very beginning as a correction for common functionals such as B3LYP, PBE, or TPSS that may be not optimal for noncovalent interactions. For simplicity, it may not be the optimal choice to employ dispersion correction on B3LYP level of theory. Given the very low cost of D3/D4 calculations, users are free to decide whether to apply it during the data preprocessing stage.

• Are the DFT level of theory and basis sets appropriately chosen? Is the choice of DFT level of theory appropriate for properties that are used as targets?

  • Yes. B3LYP, the level of theory we choose, is one of the most commonly used for small organic molecules. We refer to a voting result upon the question Which theory you used most in DFT calculating from a famous Chinese community of computational chemistry conducted on year 2024, which gather the votes of hundreds of researches , see Figure1 on http://sobereva.com/706 . Or directly the figure link.

[1] S. Grimme, J. Antony, S. Ehrlich, and H. Krieg, “A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu,” The Journal of Chemical Physics, vol. 132, no. 15, p. 154104, Apr. 2010, doi: 10.1063/1.3382344.

[2] E. Caldeweyher et al., “A generally applicable atomic-charge dependent London dispersion correction,” The Journal of Chemical Physics, vol. 150, no. 15, p. 154122, Apr. 2019, doi: 10.1063/1.5090222.

5) regarding Data Generation Questions

• There are no details given on the fragmentation algorithm used. The authors claim that this approach would give "a diverse range of significant and complex chemical space," with no evidence of this.

  • This paper focuses on datasets and benchmarks. The fragment algorithm is not the main contribution of this work and, therefore, has not been elaborated upon in detail here. A comprehensive discussion of the fragment algorithm will be presented in an independent publication. For simplicity, we employed a technique similar to [3], where a set of rules is defined to retain meaningful chemical substructures and recombine chemical motifs. In fact, we have provided evidence by detailed comparison in Appendix C, showcasing the diversity of our generated conformational space compared to QM9.

• The conformation generation through cheminformatics software like RDKit seems low-precision, given the computational cost required for DFT at B3LYP/def2-SVP. GEOM for example uses CREST, which is based on semi-empirical tight-binding model. Is RDKit sufficient?

  • Actually RDKit is only used for an initial structure generation, and subsequently we conduct a Conformation Search Process with B3LYP/def2-SVP, which is more expensive than semi-empirical methods employed by CREST. Refer to Line 301 and Line 316 for more details.

• What are convergence criteria typically used for optimization process of these structures?

  • The four criteria detailed in Line 304 are exactly convergence criteria typically used for optimization process, as is also employed in GaUssian series[4].

[3] J. Degen, C. Wegscheid‐Gerlach, A. Zaliani, and M. Rarey, “On the art of compiling and using ‘drug‐like’ chemical fragment spaces,” ChemMedChem, vol. 3, no. 10, pp. 1503–1507, Oct. 2008, doi: 10.1002/cmdc.200800178.

[4] Frisch, M. J. et al. Gaussian~16 Revision C.01 (2016). Gaussian Inc. Wallingford CT.

6) regarding writing and figure problems

We sincerely thank you for your valuable feedback.

These detailed errors may not be noticeable without a thorough reading.

We truly appreciate the time and effort you have dedicated to carefully reviewing our manuscript !

In the revised version of PDF, we will address the issues related to the title and grammatical errors as suggested.

3) regarding the proposed relative works

Below, we one by one discuss the 5 "potential related" works proposed by the reviewer.

[1] Khan, Danish, et al. "Towards comprehensive coverage of chemical space: Quantum mechanical properties of 836k constitutional and conformational closed shell neutral isomers consisting of HCNOFSiPSClBr." arXiv preprint arXiv:2405.05961 (2024).

[2]Zhu, Yifei, et al. "QCDGE database, Quantum Chemistry Database with Ground-and Excited-state Properties of 450 Kilo Molecules." arXiv preprint arXiv:2406.02341 (2024).

[3] Axelrod, S., Gómez-Bombarelli, R. GEOM, energy-annotated molecular conformations for property prediction and molecular generation. Sci Data 9, 185 (2022). https://doi.org/10.1038/s41597-022-01288-4

[4] Khrabrov, Kuzma, et al. "nabladft: Large-scale conformational energy and hamiltonian prediction benchmark and dataset." Physical Chemistry Chemical Physics 24.42 (2022): 25853-25863. https://doi.org/10.1039/D2CP03966D

[5] Khrabrov, Kuzma, et al. "$\nabla^ 2$ DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules and a Benchmark for Neural Network Potentials." arXiv preprint arXiv:2406.14347 (2024).

First of all, as you may have noticed, $1$

$2$ [5] are not "published in chemistry related venues", and rather preprints that have not been peer-reviewed.

As for $3$ , we quote Line 215 in our paper: Remark: We also acknowledge the existence of several other notable datasets in the field such as ... these datasets focus on different domains and are not directly designed for the study of small organic molecules.

$3$ are not designed for small organic molecules as you can read from its abstract that The Geometric Ensemble Of Molecules (GEOM) dataset contains ... data related to biophysics, physiology, and physical chemistry. And you can read from **Table3 in $3$ ** for more detailed information.

As for [4], we list the main characteristics of the datasets.

References [4] and [5] emphasize the simplicity and general applicability of their datasets. However, as shown in the table, the average number of conformations sampled per trajectory in [4] is only 5. This is insufficient for constructing a high-precision potential energy surface (PES).

Perhaps you may not be familiar with this specific area, but as intuition would suggest, higher sampling density and a greater number of conformations lead to more accurate reconstructions of the potential energy surface $6-7$ . From a practical perspective, an average of 5 samples is inadequate for this purpose.

We will add $3$ and $4$ into the references and discuss the difference.

[6]Sicong Ma, Cheng Shang, Zhi-Pan Liu. Heterogeneous catalysis from structure to activity via SSW-NN method. J. Chem. Phys. 7 August 2019; 151 (5): 050901. https://doi.org/10.1063/1.5113673

[7] P.-L. Kang, C. Shang, and Z.-P. Liu. “Glucose to 5-Hydroxymethylfurfural: Origin of Site-Selectivity Resolved by Machine Learning Based Reaction Sampling”. J. Am. Chem. Soc., vol. 141, no. 51, pp. 20525–20536, Dec. 2019.http://dx.doi.org/10.1021/jacs.9b11535

Dear Reviewer MtoR,

We sincerely appreciate the time and effort you have invested in reviewing our paper. We have carefully considered each of your comments and below we provide a point-by-point response to your feedback.

1) regarding the appropriateness of the submission topic

We would like to first address the key concern you have raised.

It is important to emphasize that the name of this track is "Datasets and Benchmarks". As is widely recognized in the field of deep learning, the quality of data sources is a critical issue. Our work is specifically designed to contribute to the development of higher-quality AI resources for the community. This dataset is tailored for training AI models, intended for use with AI models, and designed for fair comparisons of the performance of AI models. These characteristics clearly align with the objectives of this track.

We kindly ask the reviewers to adopt a more inclusive perspective toward the concept of AI datasets. AI-related fields such as organic molecular and protein design are also advancing rapidly recent years. The well-known AlphaFold has recently been awarded a Nobel Prize.

Thus, we kindly request the reviewers' understanding once again, molecular datasets are an integral part of AI data resources, and the concept of AI datasets should not be disriminatively limited to only "image datasets" and "text corpus datasets".

2) regarding the main content about data generation

Given that this manuscript is a dataset and benchmark paper, which is submitted to the "Datasets and Benchmarks" track, we believe it is our responsibility to provide a detailed explanation of the dataset generation process.

In fact, many dataset papers fail to fulfill this obligation, resulting in a lack of transparency in the data production process, which creates significant challenges for users. A clear and explicit description of the production process is an essential feature of a high-quality benchmark dataset.

Recently, AAAI 2025 hosted a "Good DATA" workshop, which elaborated on the issue of unclear data provenance in many fields. Perhaps visiting their homepage could help you understand the significant value of the detailed explanation we provided regarding our data production process.

Furthermore, we would like to emphasize that the evaluation criteria for data-focused papers should differ from those applied to model-focused papers. The standards and writing conventions of model-focused papers should not be imposed on data-focused papers. We hope this perspective clarifies the unique considerations necessary for evaluating work in this category.

Thank you for the response, although the belligerent and condescending tone is not appreciated.

Appropriateness of submission:

• The track is "Dataset and Benchmarks", however the authors have neglected the benchmarking part. There is very little detail in machine learning in this paper, and more importantly, no insights into what this new dataset/benchmark probes in the off-the-shelf models they have implemented.
• The well-known AlphaFold was published in Nature, another peer-reviewed scientific journal. Furthermore, AlphaFold was trained on CASP, which are carefully curated data with experimental X-ray crystallography and NMR spectroscopy. This is quite different from a dataset generated with some invented fragmentation/conformer generation method, followed by slightly improved DFT calculations.

Main content about data generation:

• I don't understand the authors reponse to this. This reviewer (and also the other reviewers) have pointed out the arbitrary method of data/conformer generation that the authors have chosen, which is precisely the problem with works that lack "transparency" or "clear and explicit description of the production process".

Recently, AAAI 2025 hosted a "Good DATA" workshop, which elaborated on the issue of unclear data provenance in many fields. Perhaps visiting their homepage could help you understand the significant value of the detailed explanation we provided regarding our data production process.

• I do not understand why the authors have directed me to a conference workshop that has not happened yet. Perhaps visiting their hompage would help you understand that this conference will take place in March 2025, and that it is not a proper source for citation.
• It is insufficient to generate such a domain-specific dataset (for a "data-focused paper") with such scant detail provided on data generation, and vague, jargony, and inaccurate statements (as pointed out by reivewer vr1N) about quantum chemistry.

Relative works:

First of all, as you may have noticed [...] [3] are not designed for small organic molecules as you can read from its abstract that The Geometric Ensemble Of Molecules (GEOM) dataset contains ... data related to biophysics, physiology, and physical chemistry.

• You may have noticed, that GEOM is focused on small organic molecules. The dataset includes molecules from QM9. The "biophysics, physiology, and physical chemistry datasets" are all small organic molecule datasets. I recommend reading the paper, rather than making conclusions from the abstract of a paper.

Perhaps you may not be familiar with this specific area, but as intuition would suggest, higher sampling density and a greater number of conformations lead to more accurate reconstructions of the potential energy surface[6-7].

• Yes. I am unfamiliar in this specific area, so I will defer to the critical responses from reviewers vr1N and KztH. This statement is neither intuitive or trivial. This is a machine learning conference.
• Every reviewer agrees that this work overlaps with existing work from nabla2. Nabla2 even provides benchmarking tasks based on their dataset, with specialized splits, and results of 10 different models on the task.
• Thank you for adding the references to the related works.

Regarding DFT level of theory

• Figure 2, which the authors have removed, is not plotting the categories because there was an axis with ordinality. This also does not answer the question about the simplification of Jacob's ladder of DFT.
• A citation to a blogpost with no peer-review or any publishing guidelines is unconvincing. Regardless, even if B3LYP is the most popular level of theory, this does not answer the question: is it appropriate for the properties that you are using it for?

Data generation:

• The provided evidence of the fragmentation method, which the authors have still refused to elaborate on in the main text, is compared to QM9, which is an enumeration of all structures of 9 heavy atoms. The structures are usually non-physical, very small, and, in the words of the authors: "severely limited in the number of molecular structures, making them grossly inadequate for training large-scale models." This comparison is unconvincing that the fragmentation method is successful.
• According to your procedures, you do not perform conformer search with B3LYP/def2svp. The search is performed with low-level RDKit, and the structure is optimized with B3LYP/def2svp. Please do not make dishonest claims.

---

The author responses are insufficient. It is my opinion that the authors are quantum chemistry experts, using domain-specific jargon in hopes that reviewers who are not domain experts, at a machine learning conference, would believe the work to be more innovative and interesting than it is.

As such I will further lower my score to 1.