Geometric Representation Condition Improves Equivariant Molecule Generation
摘要
评审与讨论
This paper introduces a novel framework to improve 3D molecular generation by integrating certain geometric representation conditions into the molecular generation process, and then conditioning on these representations to generate a molecule.
优点
Good background, appreciate the previous work review in the appendix.
Appreciate the heat maps in Figure 4, good demonstration of these parameters allowing a scientist to specify the tradeoff between properties depending on application.
缺点
There are no glaring weaknesses of this paper. All choices made have been clearly communicated and motivated. Important limitations of the method are clearly outlined in the Conclusions section of the paper.
Small change, but please change the colors used in the tables for red/green colorblind readers.
问题
- Please define EDM in your abstract.
- The writers vaguely refer to the improved “quality” of the generated molecules as the first achievement of this method. This is too vague—what does quality mean here?
- Again, the second bullet talks about model performance, citing a 31% increase. What is this “performance”? Please cite the metric and the benchmark here.
- Last bullet—reduce the number of diffusion steps by what percent on average? Please be specific here at this point in the paper.
Thank you so much for taking the time to review our work and for recognizing its contributions! We truly appreciate your feedback. To that end, we have carefully reviewed your comments and have submitted a revised version of the manuscript that directly addresses the questions you raised. We hope the revisions clarify any concerns.
We are grateful for your engagement with our work and look forward to hearing your further thoughts. Your feedback is invaluable in helping us improve, and we deeply appreciate your efforts in evaluating our submission.
The authors introduce GeoRCG, a general framework to enhance conditional molecule generative models by including geometric representations. GeoRCG splits the generation process into two parts the first being to create an informative geometric representation followed by generating a molecule conditioned on this representation. Using GeoRCG they can improve upon base EDM for QM9 and GEOM-DRUGs datasets, reducing the number of inference steps from 1000 to 100.
优点
- Overall GeoRCG demonstrates how leveraging powerful pretrained molecule representations to condition generative models can improve results for QM9 molecules.
- Strong evidence shows that guidance mixed with low-temperature sampling improves QM9 results in Figure 4.
- By conditioning the cheap representation generator on the property of interest it pushes the complexity of conditional generation to be more managible.
缺点
- There is an overemphasis on QM9 in the benchmarks, with little attention to the more challenging GEOM DRUGS which more accurately represents drug-like molecules. Molecule stability, connectivity, and other metrics are included in MIDI but have not been reported. This is important, especially for a method that demonstrates improvement over base EDM for QM9, since for Drugs EDM obtains only 5.5% molecule stability and 40.3% after OpenBabel (numbers taken from MIDI). Connectivity is not reported for any method, which is important for 3D molecule generation. If the molecule is not connected, RDKit can often still parse it, but it is not a single-molecule structure. From MIDI EDM + OpenBabel, the result is only 41.4%, which is quite low.
Overall, it's hard to understand what is translatable as a general method for other biological tasks since QM9 is a toy task.
问题
- How does low-temperature sampling impact prior DRUGS baselines, as it is used in Table 1? Methods mentioned, such as Chroma, suggest that it can have a significant impact.
- How does this same method apply to other models beyond EDM given it is a general framework?
- How does the performance vary as a function of sampling steps for DRUGs and their respective benchmarks?
Thank you for your thoughtful feedback and for pointing out these key areas for improvement in our work. We truly appreciate the time and effort you have invested in providing us with valuable insights. Below, we address your concerns in detail.
There is an overemphasis on QM9 in the benchmarks, with little attention to the more challenging GEOM DRUGS which more accurately represents drug-like molecules. Molecule stability, connectivity, and other metrics are included in MIDI but have not been reported.
Response:
We greatly appreciate your observation regarding the benchmarking emphasis. We fully agree that a broader evaluation, particularly on challenging datasets like GEOM-Drugs, is important. In our rebuttal, we have included additional evidence to highlight GeoRCG’s improvements over EDM across various molecular metrics for the GEOM-Drugs dataset. We respectfully refer you to our response to Reviewer WWP2 of question “The only reported 3D metric for the more realistic and larger GEOM-Drugs dataset is atom stability.”
How does low-temperature sampling impact prior DRUGS baselines, as it is used in Table 1? Methods mentioned, such as Chroma, suggest that it can have a significant impact.
Response:
Thank you for this insightful question. We have carefully examined this aspect and found that low-temperature sampling has small impact on the Drugs dataset. The results below demonstrate that the observed variations in metrics remain within the statistical variance:
| inv-temp | 0.3 | 0.5 | 1 | 1.5 | 2 |
|---|---|---|---|---|---|
| Atom Stability | 84.6 | 84.3 | 84.7 | 84.5 | 84.3 |
| Validity | 97.7 | 98.5 | 97.8 | 98.2 | 97.6 |
We attribute this small variation to the following factors: During the training of GeoRCG on GEOM-Drugs, we incorporated relatively larger random noise perturbations (i.e., adding random noise to representation conditions during training, details in lines 797–804 and 938 of our manuscript). This helped prevent overfitting, especially given the sparsity of samples in representation space for large and complex datasets like GEOM-Drugs. As a result, in these cases, exact representation values are less critical compared to the broader clusters of representations that capture abstract molecular concepts. Consequently, low-temperature sampling has limited influence on the results.
How does the performance vary as a function of sampling steps for DRUGs and their respective benchmarks?
We thank you for raising this important question! Below, we provide results for few-step sampling to illustrate how performance varies as a function of sampling steps for DRUGs.
| # Steps | 50 | 100 | 500 | 1000 | |
|---|---|---|---|---|---|
| Atom Stability | GeoBFN | 75.11 | 78.89 | 81.39 | 85.6 |
| GeoRCG | 81.44(0.10) | 83.02(0.06) | 84.03(0.37) | 84.3(0.12) | |
| Validity | GeoBFN | 91.66 | 93.05 | 93.47 | 92.08 |
| GeoRCG | 95.70(0.70) | 96.30(0.70) | 97.567(0.90) | 98.5(0.12) |
As shown in the table, GeoRCG achieves much better result than the advanced BFN-based generative model, achieving better results than 500-step GeoBFN in just 50 steps. In the most computationally intensive case (i.e., 1000-step settings), GeoRCG exhibits slightly lower Atom Stability (84.3 vs. 85.6) but significantly higher Validity (98.5 vs. 92.08). Notably, similar to the QM9 settings, GeoRCG achieves near-optimal performance with approximately 100 steps
These findings underscore GeoRCG’s robustness and efficiency, making it a compelling choice for generative tasks even with reduced sampling steps.
How does this same method apply to other models beyond EDM given it is a general framework?
Thank you for raising this question. As a model-agnostic framework, GeoRCG can be seamlessly extended to other models, such as GCDM, to potentially achieve even better results. However, the large-scale GEOM-Drugs dataset requires approximately 20 days for full training, making such improvements challenging within the rebuttal period. Nevertheless, we are actively working on further training and will update the results with any significant results we obtain.
Once again, we sincerely thank you for your thoughtful comments and constructive suggestions. We look forward to any additional suggestions you might have, and we would greatly appreciate it if you could raise your score if your concerns are partly resolved.
Thank you for taking the time to give additional benchmarks.
The DRUGS results are quite poor overall, and many prior recent methods are left out in the response to WWP2. Furthermore, the validity benchmark used by EDM overstates the meaningfulness of the performance. With such poor connectivity, one must generate 2-3x the desired number of molecules to get molecules. Given this, it's really hard to understand the value of this method since EDM does not perform well. The two metrics used in the ablations in the above response are the simplest benchmarks missing significant aspects of the desirable traits of the generated molecules. Molecule stability and connectivity should not be ignored when contextualing performance. Furthermore, these 2D metrics do not measure the accuracy of the 3D structure, which is altered by reliance on software like OpenBabel. Given the focus of 3D molecule generation is the structure, this aspect should be captured in the analysis.
Testing this method on other models besides EDM is critical to demonstrate the generalizability and measure the practical gain of this method. There is no doubt that this work makes nontrivial improvements over EDM. Still, even more performance gains can be made by just switching the architecture and/or sampling procedure as done in MiDi and SemlaFlow. This line of experimentation is needed as the method is claimed to be model-agnostic.
For these reasons, I recommend the authors resubmit and take the time to perform a deeper benchmarking effort to understand the practical impact and provide evidence that this is a general method. For the current results, there are many simpler approaches that do much better across all benchmarks. There would be great value added to the community if GeoRCG could be shown to improve more recent methods.
This paper introduces an intermediate representation approach to find better property conditioning for molecular generation. The authors show success for properties such as polarizability (\alpha) over the QM9 data set, but other state-of-the-art models outperform the model on the more extensive dataset DRUG.
优点
This paper is well-written and easy to read. The authors present their results clearly and clearly, and the method is also well described.
缺点
The main weakness of this approach is that the performance in the DRUG dataset does not beat SOTA. QM9 is a great research-level dataset for small molecules, but DRUG is industrially relevant and has more realistic molecules.
问题
Is it possible to get (during revisions) a more extensive effort to match or exceed SOTA for the DRUG dataset? This would raise my reviewer score. At the moment, I think this paper is a technical but not revolutionary improvement and may belong to a journal or another venue rather than ICLR. If the improvement over larger molecules is demonstrated, I think it may be more of interest to this community.
We sincerely thank you for your detailed review and constructive feedback. Below, we provide a detailed response to your comments.
Is it possible to get (during revisions) a more extensive effort to match or exceed SOTA for the DRUG dataset?
Thank you for bringing up this important point. We understand the value of achieving SOTA performance on the GEOM-Drugs dataset and appreciate your suggestion. While GeoRCG does not currently reach SOTA results on this dataset, we believe this is due to its reliance on EDM, which is a relatively weaker generator. A promising way to enhance performance would be to replace the base generator with a stronger model, such as GCDM [1].
However, the large-scale GEOM-Drugs dataset requires approximately 20 days for full training, making such improvements challenging within the rebuttal period. Nevertheless, we are actively working on further training and will update the results with any significant results we obtain.
In the meantime, we are pleased to share several new promising results obtained during rebuttal based on our current models:
- Improved performance on GEOM-Drugs across more metrics: We observed notable improvements in molecule stability, 3D metrics (e.g., angle distribution), and connectivity, measured with OpenBabel tools, compared to the base EDM model. (Please see our detailed response to Reviewer WWP2 regarding the question on 3D metrics for GEOM-Drugs.)
- Superior results with few-step sampling: GeoRCG achieves better results than the advanced 500-step GeoBFN model with only 50 steps, and nearly optimal performance with just 100 steps. This is a significant improvement over the 1000 steps required by the EDM baseline, showcasing the efficiency and potential of GeoRCG. (Please see our detailed response to Reviewer 2uMF regarding the question "How does the performance vary as a function of sampling steps for DRUGs and their respective benchmarks?")
We believe these results underscore the promise of GeoRCG, even with the current models.
In the meantime, we welcome any further suggestions or ideas you may have and will gladly incorporate them as we continue refining our work. Your insights have been instrumental in helping us improve the paper, and we sincerely thank you again for your feedback.
[1] Geometry-Complete Diffusion for 3D Molecule Generation and Optimization.
The paper presents a theoretical foundation and specific implementation of latent diffusion for molecular generation. Instead of directly operating with molecular graphs (e.g., continuous Euclidean coordinates and categorical atom types), the method proposes projecting molecules into a latent space and using latent representations that are O(3) and SO(3) invariant. Diffusion is performed in this simplified latent space, and then the molecules are reprojected to predict their structures. This approach reduces computational costs, resulting in a denoising neural network that is much smaller and simpler, as well as reducing the number of steps required for diffusion.
优点
Improved Speed through Latent Diffusion: Latent diffusion has great potential to enhance the speed of 3D molecule generation. In this paper, the authors demonstrated that they were able to reduce the number of diffusion steps from the commonly used 1000 or 500 to just 100 without any performance drop. Performing diffusion in a simpler latent representation appears to be a highly effective approach.
Better Bond Length and Angle Distributions on QM9 Dataset: The proposed model generates bond length and bond angle distributions that more closely resemble those of the QM9 dataset. This is a reasonable metric for assessing the quality of generated 3D structures and indicates an improvement over previous methods.
缺点
Limited 3D Metrics on Larger Datasets: The only reported 3D metric for the more realistic and larger GEOM Drugs dataset is atom stability. Since the atom stability is only 0.86 for GEOM Drugs itself, this raises questions about the reliability of the metric. A more comprehensive and accurate comparison is required to fully assess the model’s performance on larger datasets.
Overlooking Models That Do Not Rely on External Software: The paper states that models such as MiDi and LDM3DG use domain knowledge through Open Babel, which gives them advantages. However, there are models like JODO, EQGATDiff, and SemlaFlow that directly predict bonds without relying on external software. Including these models in comparisons would provide a more comprehensive evaluation and highlight the strengths and weaknesses of the proposed method.
Questionable Reliance on Lookup Tables: The reliance on a lookup table for bond lengths is questionable. Depending on the molecular configuration and the specific energy calculation method used (which is GFN2-xTB for GEOM Drugs), bond lengths can vary within a 10% interval. This variability suggests that a static lookup table may not accurately capture the nuances of bond lengths across different molecules.
Lack of Comparison with Faster Models: While the method aims for faster molecule generation by reducing the number of diffusion steps, previous models have already utilized flow matching to reduce steps to 200 (EquiFM) or even 20 (SemlaFlow). Including these models in the comparison would strengthen the research by providing context for the improvements and demonstrating how the proposed method stands relative to existing fast-generation techniques.
问题
Could you add bond length and bond angle metric comparisons for the GEOM Drugs dataset? Providing these metrics would offer a more complete evaluation of the model’s performance on larger and more complex datasets.
Would you consider comparing your model with others like JODO, EQGATDiff, or SemlaFlow, which generate the full graph including bonds? A comparison with these models, especially SemlaFlow due to its speed and reduced number of steps, could be particularly beneficial in highlighting the advantages and limitations of your approach.
In the paper, there is no reference to the GEOM paper, and there is a claim: “Crucially, many structures in GEOM-DRUG lack the equilibrium conditions necessary for pre-training methods that enable effective learning of force fields,” which could be misleading. All GEOM Drugs molecules have optimized geometries with respect to GFN2-xTB energy calculations. Could you reconsider this statement? Clarifying this point and accurately referencing the GEOM dataset would enhance the credibility and accuracy of your pape
We sincerely thank you for your detailed and constructive feedback. Your insights have been invaluable in helping us improve the clarity and rigor of our work. Below, we address each of your points in detail.
The only reported 3D metric for the more realistic and larger GEOM Drugs dataset is atom stability. Since the atom stability is only 0.86 for GEOM Drugs itself, this raises questions about the reliability of the metric. A more comprehensive and accurate comparison is required to fully assess the model’s performance on larger datasets.
Response:
Thank you for pointing out this limitation! We recognize the restricted scope of 3D metrics reported for GEOM-Drugs in our original submission. In response, we have expanded our evaluation to include additional metrics, such as bond lengths, angles, connectivity, MiDi-based stabilities, and metrics derived using Open Babel. The results of these analyses are summarized below:
| Mol stable | Atom stable | Validity | Uniqueness | Angles | Bond Length | Connectivity | |
|---|---|---|---|---|---|---|---|
| Data | 99.9 | 99.9 | 99.8 | 100 | 0.05 | ~0 | 100 |
| EDM | 5.5 | 92.9 | 97.5 | 99.9 | 6.23 | 0.2 | 35.6 |
| GeoRCG | 10.8(0.5) | 94.61(0.036) | 98.40(0.07) | 100(0) | 2.87(0.02) | 0.46(0.10) | 55.1(0.01) |
| EDM + OBabel | 40.3 | 97.8 | 87.8 | 99.9 | 6.42 | 0.2 | 41.4 |
| GeoRCG + OBabel | 67.9(2.0) | 98.97(0.047) | 92.37(0.60) | 100(0) | 3.08(0.04) | 0.29(0.11) | 58.1(0.21) |
Results demonstrate that GeoRCG shows consistent improvements over its base model, EDM, across nearly all metrics, except for bond lengths. Notably, GeoRCG achieves significant gains for the “Angles” metric, which is a more complex 3D statistic involving 3-tuples rather than the simpler 2-tuples of bond lengths. This improvement likely results from the representation condition, which imposes a more global structural constraint, encouraging the model to capture higher-level geometrical relationships. This enhanced representation leads to better performance on metrics that reflect intricate structural properties.
Would you consider comparing your model with others like JODO, EQGATDiff, or SemlaFlow, which generate the full graph including bonds? A comparison with these models, especially SemlaFlow due to its speed and reduced number of steps, could be particularly beneficial in highlighting the advantages and limitations of your approach.
Response:
We greatly appreciate this suggestion, and present a preliminary evaluation on the QM9 dataset here.
| QM9 | Mol stable | At stable | Validity |
|---|---|---|---|
| Data | 98.7 | 99.8 | 98.9 |
| MiDi | 97.5 | 99.8 | 97.9 |
| + JODO* | 93.4 | 99.2 | - |
| + SemlaFlow | 99.6 | 99.9 | 99.4 |
| + EQGAT-diff | 98.7 | 99.9 | 99 |
| GeoRCG | 98.2 | 99.9 | 99 |
(JODO is marked with * to indicate its use of the EDM codebase for stability evaluation, which permits less valency and may result in lower results. GeoRCG incorporates Open Babel to address EDM’s limitations in bond computation, which ensures fairness since all other models implicitly learn bond distributions which are calculated by Open Babel during the data preparation stage)
While GeoRCG demonstrates competitive performance among these newest SOTA methods, it does not outperform the best-performing models such as SemlaFlow. This is a reasonable outcome given that GeoRCG uses the weaker base generator EDM, which does not explicitly learn any bond information thus can be weak in capturing complex molecule distribution. Future improvements could involve integrating GeoRCG with stronger base generators to enhance its performance further.
Questionable Reliance on Lookup Tables: The reliance on a lookup table for bond lengths is questionable.
Response:
We appreciate your concern regarding the reliance on lookup tables for bond calculations and largely agree with it. To clarify, we adopted this convention to ensure fair comparisons: In our main tables (e.g., Table 1), EDM (our base model) and almost all other methods included in our comparisons also rely on lookup tables for bond calculations.
That said, more accurate methods for bond calculation, such as using Open Babel, can certainly be employed. As shown in Table 2, the use of Open Babel addresses the limitations of lookup tables and highlights GeoRCG’s improved performance when leveraging this more advanced approach.
… previous models have already utilized flow matching to reduce steps to 200 (EquiFM) or even 20 (SemlaFlow)
Response:
Thank you for this observation. Actually, a comparison with EquiFM is already included in Table 4, where we demonstrate that GeoRCG can outperform EquiFM using just a quarter of the number of steps. Regarding SemlaFlow, as noted earlier, GeoRCG currently does not surpass its performance, which may be attributed to the weaker base generator, EDM, used by GeoRCG.
In the paper, there is no reference to the GEOM paper, and there is a claim: “Crucially, many structures in GEOM-DRUG lack the equilibrium conditions necessary for pre-training methods that enable effective learning of force fields,” which could be misleading. All GEOM Drugs molecules have optimized geometries with respect to GFN2-xTB energy calculations. Could you reconsider this statement? Clarifying this point and accurately referencing the GEOM dataset would enhance the credibility and accuracy of your paper
Response:
Thank you for pointing this out. We apologize for the oversight in not citing the original GEOM paper but the later ones, and have included this in the revised manuscript.
Regarding the statement “lack the equilibrium conditions …,” our intent was to emphasize that not all conformations in GEOM-Drugs are in the minimum-energy state: The dataset typically includes the top five minimum-energy conformations per molecule. This may impact the efficacy of pre-training. We quite appreciate your feedback and will revise this statement to ensure greater clarity and accuracy.
Once again, we thank you for your thoughtful comments and constructive suggestions. We look forward to any additional suggestions you might have, and we would greatly appreciate it if you could raise your score if your concerns are partly resolved.
The paper, titled "Geometric Representation Condition Improves Equivariant Molecule Generation" (GeoRCG), presents a novel approach to improving molecular generative models by incorporating geometric representation conditions. The GeoRCG framework divides the molecule generation process into two stages: first, generating an informative geometric representation; second, generating a molecule conditioned on this representation.
The core idea is to first generate a compact geometric representation of a molecule using a pre-trained geometric encoder. This representation captures essential information about molecular structure without the complexity associated with 3D symmetries, making the generation task simpler and more effective. Leveraging this representation, the second stage uses a molecule generator to produce the final molecule. The framework employs EDM as the base generator and shows significant improvements in both unconditional and conditional molecule generation tasks. Specifically, GeoRCG achieves an average 31% performance gain over state-of-the-art methods on challenging conditional generation tasks.
优点
The key strength of GeoRCG lies in its innovative use of geometric representations to condition molecular generation. By transforming the generation problem into a two-stage process—first generating a geometric representation and then generating the molecule conditioned on this representation—the paper introduces an effective way to simplify the complex task of molecular generation. This approach addresses major challenges like handling 3D geometric symmetries and provides a significant improvement over existing methods that attempt to directly learn molecular distributions.
The clear and structured presentation of the methodology, supported by well-executed empirical evaluations and visual explanations, adds to the clarity and accessibility of the paper. GeoRCG's advancements could have a significant impact on drug discovery and material design, highlighting its importance for both research and practical applications.
缺点
While the paper emphasizes the use of geometric representations to simplify the generation task, there is insufficient analysis of how different pre-trained encoders impact the overall quality of the generated molecules. The choice of pre-trained encoders (UniMol and Frad) is central to the approach, but the authors do not explore how variations in the pre-training dataset or encoder architecture affect the representations. Conducting a more comprehensive analysis, such as comparing multiple pre-trained models trained on different datasets or architectures, would help clarify the impact of representation quality and improve confidence in the method’s robustness.
The representation generator aims to remove symmetries such as O(3) and S(N), but the impact of symmetry removal on downstream tasks is not thoroughly analyzed. Specifically, it would be beneficial to explore whether there are specific symmetries that contribute positively to certain molecular properties or whether removing all symmetries has unintended negative effects on some downstream applications. Conducting ablation studies that selectively preserve certain symmetry properties could offer insights into how symmetry affects molecule generation and provide a more nuanced understanding of its role.
In the conditional generation setting, the paper discusses training the representation generator on (molecule, property) pairs. However, this strategy is limited to simple properties like HOMO-LUMO gap, polarizability, etc., and there is no clear extension for complex properties such as molecular binding affinity or ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties. Such properties typically require more context or knowledge beyond geometric structure alone. Addressing this limitation, either by discussing possible extensions or incorporating more sophisticated conditioning mechanisms (e.g., using multi-modal data such as 3D structure and protein targets), would make GeoRCG more applicable to real-world drug discovery problems.
问题
See weakness
We sincerely thank you for your detailed and insightful review! We address your comments below.
Conducting a more comprehensive analysis, such as comparing multiple pre-trained models trained on different datasets or architectures, would help clarify the impact of representation quality and improve confidence in the method’s robustness.
Response:
Thank you for highlighting this important consideration! We greatly appreciate your suggestion and would like to offer some clarifications:
- Impact of different pre-trained encoder architectures
In Appendix D (lines 1012–1016), we have presented ablation studies examining the effect of different architectures (pre-trained on the same dataset), specifically Frad and Unimol, on the GEOM-Drugs dataset. This dataset is particularly challenging for the encoder.
Our findings reveal that different architectures have distinct effects. Interestingly, we observe that architectures producing representations with a clearer reliance on node numbers tend to perform slightly better on downstream tasks. Figure 6 provides an intuitive visualization of this reliance, shedding light on the characteristics of encoders that are effective for irregular data like point clouds.
- Effectiveness of pre-training.
Additionally, we evaluated the impact of pre-training (specifically on the PCQM4Mv2 dataset) using Frad on the QM9 dataset. These results, detailed in Appendix D (Table 6), underscore the clear benefits of pre-training, showing significant improvements in downstream performance.
We hope these clarifications address your concerns and demonstrate the robustness of our method.
Conducting ablation studies that selectively preserve certain symmetry properties could offer insights into how symmetry affects molecule generation and provide a more nuanced understanding of its role.
Response:
We truly value your suggestion regarding symmetry preservation and its potential impact. While the idea is intriguing, it actually diverges from the primary objectives of our study. Our approach intentionally removes symmetry to ensure simplicity and elegance in conditioning mechanisms. This choice not only 1) ensures compatibility with a wide range of tasks and models as a plug-and-play enhancement (since all molecule generators are able to accept scalar conditions) 2) but also simplifies the upstream generation process, making the representation generator more efficient and straightforward.
Introducing symmetry-based guidance would significantly increase complexity for both the upstream and downstream components of the pipeline. For instance, maintaining O(3) symmetry through the use of a hidden geometric object (e.g., a fragment) as guidance would require the upstream generator to learn to generate the object, which needs itself function as an equivariant generator like the downstream generator. Simultaneously, the downstream generator would necessitate intricate modifications to process these objects as conditions, diverging from the straightforward fusion of scalar representation conditions.
Nevertheless, your suggestion is both insightful and inspiring, and we are eager to explore how such directions might deepen our understanding of symmetry’s role in molecular generation in future research.
However, this strategy is limited to simple properties like HOMO-LUMO gap, polarizability, etc., and there is no clear extension for complex properties such as molecular binding affinity or ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties. Such properties typically require more context or knowledge beyond geometric structure alone. Addressing this limitation, either by discussing possible extensions or incorporating more sophisticated conditioning mechanisms (e.g., using multi-modal data such as 3D structure and protein targets), would make GeoRCG more applicable to real-world drug discovery problems.
Response:
Thank you for your insightful observation regarding complex properties and the challenges they pose. We agree that extending GeoRCG to handle such properties would broaden its applicability to real-world drug discovery. We provide some discussions regarding this.
For properties like molecular binding affinity, the key challenge lies in conditioning the representation generator to produce meaningful molecule representations sensitive to specific geometric contexts. Once this is achieved, these representations can condition downstream molecule generators designed for such tasks, such as GCDM-SBDD [1], to generate some molecules conditioning on target-sensitive representations, geometric target and values such as affinity. Since our representation generator is an MLP-based diffusion model and does not operate in 3D space, a promising direction would involve transforming geometric conditions (e.g., protein pocket structures) into symmetry-removed representations. These could then guide the representation generator to produce molecule representations sensitive to complex geometric contexts.
We believe this line of research holds great potential and would greatly enhance GeoRCG’s versatility and impact in drug discovery. Your suggestion has inspired us to explore this exciting direction in future work.
Once again, we are deeply grateful for your valuable feedback and suggestions. We hope that our responses address your concerns and provide clarity on the aspects raised. If you feel that your concerns have been resolved, we would greatly appreciate it if you could raise your score accordingly.
[1] Geometry-Complete Diffusion for 3D Molecule Generation and Optimization.
Thanks authors for the clarification. I am convinced and willing to raise my score to accept.
We are deeply grateful for your thoughtful engagement and are delighted that our responses addressed your concerns. Your updated score and support mean a great deal to us.
Please don’t hesitate to reach out if you have any further questions or if there are any remaining concerns—we would be glad to continue the discussion.
Dear Reviewers,
Thank you once again for dedicating your time to reviewing our work and for providing such detailed comments and concerns!
During the rebuttal process, we have carefully addressed each of your points and hope that our responses have resolved your concerns. If there is anything that remains unclear or if you have additional questions, we would be very happy to discuss further.
As the free discussion period is approaching its end, we kindly hope you to review our responses and let us know if they satisfactorily address your concerns.
We greatly appreciate your time and consideration. Wishing you a wonderful day!
Best regards,
The Authors
This paper proposes an approach to improve molecule generation by first generating a geometric representation and conditioning the molecule generation on the geometric representation.
The idea is interesting and shows improvement over the considered EDM architecture. However, as the reviewers pointed out, the evaluation is not sufficient to convincingly demonstrate the useful-ness of the new method. In particular, the authors should show the merits of their approach in the (hard and realistic) DRUG dataset and consider more backbone generative models / neural architectures.
Overall, I recommend rejection for this paper, but believe this would be a strong resubmission after convincingly showing the merits of their approach.
审稿人讨论附加意见
The reviewers were generally concerned about evaluation of the proposed idea, and the rebuttal did not change their opinions.
Reject