3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery

Thank you for your valuable feedback on our paper. We appreciate your insightful comments and have revised the manuscript accordingly. Below, we provide our responses to your concerns point-by-point:

As discussed in Section 5, the proposed method does not consider SE(3) symmetry explicitly but instead relies on data augmentation techniques. I think this aspect warrants further discussion and consideration.

We appreciate your attention to this aspect. While our method does not explicitly enforce SE(3) symmetry, it achieves implicit equivariance through data normalization and augmentation, such as random rotations, which provide sufficient coverage of the rotational and translational transformations. This approach aligns with recent successful methods in the field, including AlphaFold3, which also does not rely on explicit SE(3) equivariance. We have expanded the discussion in Section 5 to highlight this rationale.

For docking task, as far as I know, the more advanced approach Unimol-docking v2 is not included in the baselines.

Thank you for this suggestion. We have added Unimol-docking v2 as a baseline in our docking task experiments on the PoseBuster dataset, aligning with the setup in the Unimol-docking v2 paper. These results can be found in Appendix C of the revised paper.

It is important to note, however, that Unimol-docking v2 is primarily designed for blind docking, while our method, 3DMolFormer, is aimed at pocket-aware docking. We have clarified this scope difference in Section 4.1 of the revised paper.

For sbdd task, I noticed that the article directly employs evaluation metrics (Vina QED SA) as the reward function and utilizes reinforcement learning for fine-tuning. This approach may be considered unfair to other methods. It may be necessary to incorporate ablation study as well as additional evaluation methods to comprehensively validate the effectiveness of the approach, for example, the delta score proposed in paper [1] may serve as a metric for assessing whether the method is overfitting to the Vina evaluation.

You are correct in noting that we directly use the evaluation metrics to formulate the RL reward function. This reflects the flexibility of our framework, as it allows for targeted optimization towards specific objectives (i.e.., desirable molecular properties). In contrast, prior methods often only learn these properties implicitly from their training data.

We agree that using the delta score as proposed in [1, 2] is a meaningful way to assess specificity in SBDD and check for potential overfitting. We have incorporated the delta score in our evaluation, and the results are detailed in Appendix D, which further reflect the advantages of our framework in SBDD.

We hope these updates address your concerns and contribute to a clearer understanding of our work.

1. Concerns about the SE(3):

I also notice that AlphaFold3 does not employ SE(3)-equivariant architectures. In your opinion, is it possible that, in cases where the dataset is sufficiently large, simple data augmentation alone is adequate to address the SE(3)-equivariant properties? Additionally, what do the authors think about the potential pros and cons of using SE(3)-equivariant architectures compared to architectures that do not consider such properties?

2. Concerns about the docking:

Thanks for the clarification. However, I also noticed that the results for the compared methods, such as AlphaFold3 and UniMol DockingV2, were obtained under the blind docking setup. Would providing more accurate pocket location information for these methods result in better performance?

3. Concerns about the SBDD:

Thank you for the additional experiments, however, I have another concern. As I understand it, one of the key innovations of the paper is the use of a single model to unify the docking and SBDD tasks. However, from the ablation study in Appendix D, Table 5, I observed that the method seems to be highly dependent on reinforcement learning. In this case, what do the authors consider to be the benefit of unifying these two tasks for the SBDD task specifically? Or, could there be some RL-based methods that could be used for comparison?

Thank you to the authors for their efforts. In general, I am open to increasing the score after receiving additional clarifications.

Thank you for your thoughtful follow-up questions and additional concerns. We are grateful for the opportunity to provide further clarifications and have included our detailed responses below.

1. Concerns about SE(3):

   This is indeed a meaningful and interesting question. SE(3)-equivariance can be viewed as a form of "domain knowledge." Incorporating this property into model design reduces the need for the model to learn these equivariances directly from the data, which is especially beneficial when the training dataset is small or lacks diversity. However, when datasets are sufficiently large, SE(3)-equivariance can also be learned through standard architectures with data augmentation techniques like random rotation, which intuitively injects SE(3)-equivariant properties into the training data.

   Notably, AlphaFold3 [1] highlights that simplifying the architecture of AlphaFold2 by removing SE(3)-equivariant processing has only a modest impact on accuracy. This also leads to the "relatively standard architecture" of AlphaFold3. A recent study [2] draws the similar conclusion that "equivariance improves data efficiency, but training non-equivariant models with data augmentation can close this gap given sufficient epochs." Although this study focuses on rigid-body interactions, the findings suggest that SE(3)-equivariance can be compensated for by large datasets and longer training times.

   In summary, SE(3)-equivariant architectures have the advantage of explicitly ensuring equivariance, particularly in data-limited scenarios, but come at the cost of increased architectural complexity. In contrast, simpler architectures paired with data augmentation offer scalability and efficiency, albeit with a higher dependency on large and diverse datasets. We appreciate this thought-provoking discussion and have included these points in Section 5 of the revised paper.

2. Concerns about the docking:

   Thank you for this question! Blind docking and pocket-aware docking are fundamentally different tasks with distinct inputs and training data, making direct comparisons difficult. Blind docking methods cannot leverage pocket information directly, as they are not trained or designed to incorporate it.

   Specifically:

   • AlphaFold3, which was recently open-sourced, does not include a channel for binding pocket input, making it infeasible to provide pocket information.
   • Uni-Mol Docking V2 is a variant of Uni-Mol designed explicitly for blind docking, transferring knowledge from pocket-aware docking (by using the weights of Uni-Mol) but trained for the blind docking task. Meanwhile, the original Uni-Mol has been included as a baseline for pocket-aware docking in Section 4.1.

3. Concerns about SBDD:

   As we state in the abstract of the paper, our approach leverages the duality between docking and structure-based drug design (SBDD) by "utilizing docking functionalities within the drug design process." Specifically, the reinforcement learning (RL) process in 3DMolFormer is applied solely to the generation of ligand SMILES, while the protein-ligand docking ability, obtained from supervised fine-tuning, remains frozen during RL fine-tuning. This integration demonstrates the synergy of unifying these two tasks.

   Regarding your observation on the dependency of the method on RL, this reflects the importance of RL in optimizing ligand generation within the unified framework. Before 3DMolFormer, structure-based 3D drug design was dominated by diffusion-based methods, where RL was not applicable. Our method introduces a distinct paradigm by employing a "pre-training + fine-tuning" approach tailored for transformer models, with RL fine-tuning being a widely adopted strategy in this context.

   While RL-based methods for comparison are scarce, we acknowledge DeepLigBuilder [3], which employs Monte Carlo tree search (MCTS) for SBDD. However, the RL technique of MCTS differs significantly from our approach. Moreover, it designs drugs for only one protein target, whereas our model is evaluated across 100 targets. And DeepLigBuilder is not open-sourced, making it unavailable for direct comparison in our experiments.

Reference:

[1] Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 2024: 1–3.

[2] Brehmer J, Behrends S, de Haan P, et al. Does equivariance matter at scale? arXiv preprint arXiv:2410.23179, 2024.

[3] Li Y, Pei J, Lai L. Structure-based de novo drug design using 3D deep generative models. Chemical Science, 2021, 12(41): 13664–13675.

Thank you again for your insightful questions. We hope these clarifications address your concerns and further highlight the contributions and strengths of our work. We greatly appreciate your openness to revisiting your evaluation and look forward to any additional feedback.

Thank you to the authors for the detailed responses and discussions, which have largely resolved my concerns. My only remaining doubt is whether it is fair to directly use evaluation metrics (such as VINA) as the optimization objective for RL, since there have been previous approaches based on optimization theory that achieve better results than data-driven methods (such as diffusion-based methods), at least on the surface. However, this controversy remains an open question, and I will leave it to the AC to make the final decision. Additionally, I have raised the score to 6 for their efforts and improvements.

We appreciate your feedback and have addressed all of your concerns thoroughly, and we sincerely hope that you will reconsider your evaluation on our paper:

This paper mention figure 1 multiple times when introducing model structure, however there is no figure 1 in the preprint.

Figure 1 is located at the top of page 3, just above Figure 2. To enhance clarity, we have highlighted this in the revised version of the manuscript to make it more prominent and easy to find.

For the pose evaluation, the model only check for RMSD. Could you also report other pose-relate metrics like steric clashes and strain-energy?

Thank you for this insightful suggestion. We reviewed the methodology in PoseCheck [1], which introduced clash scores and strain energy as metrics for assessing the rationality of structure-based drug design, rather than protein-ligand docking. We have integrated these metrics into our evaluation and reported the results in Appendix D of the revised manuscript, which further confirm the advantages of our 3DMolFormer.

[1] Harris, Charles, et al. "Posecheck: Generative models for 3d structure-based drug design produce unrealistic poses." NeurIPS 2023 Generative AI and Biology (GenBio) Workshop. 2023.

We greatly appreciate your thorough review and valuable feedback. Below, we provide detailed responses to each of your concerns and hope this will lead you to reconsider your evaluation.

Response to Weaknesses

1. Docking:
   • Thank you for suggesting the inclusion of PoseBusters as a benchmark. We have conducted experiments and included results in Appendix C of our revised paper. However, it is important to note that PoseBusters is designed primarily for blind docking, whereas our 3DMolFormer focuses on pocket-aware docking. To bridge this difference, we evaluated 3DMolFormer on PoseBusters with provided pocket information, representing a different setup from typical PoseBusters evaluations. We have discussed this further in Section 4.1.
   • Speed and Accuracy Trade-offs: While 3DMolFormer excels in scalability due to parallel inference, direct speed-accuracy comparisons with models like AlphaFold3 and Chai-1 are not meaningful, as they target different tasks. Only for context, 3DMolFormer predicts a binding pose in an average of 0.8 seconds using an A100 80G GPU, whereas AF3 takes over 60 seconds. We have argued this in Section 4.2 - Results.
2. Synthetic Accessibility (SA) Score:
   • The original SA score ranges from [1, 10], where lower values indicate better synthetic accessibility. For consistency with previous studies such as Pocket2Mol and DecompDiff, we applied a negative linear transformation to scale the SA score to [0, 1], where higher values are preferable. The threshold of 0.59 follows DecompDiff, where scores above this threshold signify good synthetic accessibility. We have clarified this in Section 4.2 - Footnote and Appendix D.
3. Multi-Score Thresholds:
   • The thresholds for Vina Dock < −8.18, QED > 0.25, and SA > 0.59 were derived from DecompDiff to ensure comparability with established baselines. Moreover, QED and SA scores above their thresholds are assigned a reward value of 1 since they are auxiliary to the main docking score. We have included further clarification in Section 4.2 - Reward Function of the revised manuscript.
4. Docking Formula:
   • You correctly noted that we did not use a fixed threshold for the Vina Dock score in our reward function. Instead, we aimed to minimize this score as much as possible. We used a reverse sigmoid function, as discussed in Section 4.2 - Reward Function, to convert the Vina Dock score into a [0, 1] range where higher values are preferred.
5. Molecule Distribution:
   • We appreciate your suggestion regarding molecule distribution. We now include the distributions of molecular weights, logP values, and the number of rotatable bonds in Appendix D. Although these distributions are not used for direct comparison, their contributions are captured within the QED metric, which reflects drug-likeness.
6. Benchmarks for 3D Generation:
   • We appreciate your recommendation to include benchmarks such as PoseCheck and DrugPose. We have incorporated clash scores and strain energy assessments from PoseCheck to evaluate the structural rationality of generated molecules. Results are included in Appendix D, which further validate the advantages of our 3DMolFormer.
   • For DrugPose, while it offers various metrics for 3D drug discovery, its overlap with PoseCheck in the context of structure-based drug design makes PoseCheck a sufficient benchmark for our evaluation. We have included a discussion on this in Appendix D.

Response to Questions

• PoseBusters Performance: We present our new evaluation on PoseBusters in Appendix C, highlighting the difference in application context between blind and pocket-aware docking. (Also refer to Response to Weaknesses 1)
• Speed and Accuracy Comparisons: We provide speed and performance insights in Section 4.1, noting that 3DMolFormer is highly scalable and efficient. However, due to different scope of tasks, the comparison of speed-accuracy trade-off between 3DMolFormer and AF3 or Chai-1 is meaningless. (Also refer to Response to Weaknesses 1)
• Threshold Selection Explanation: The selection of thresholds—QED > 0.25, SA > 0.59, and Vina Dock < −8.18—follows standard benchmarks from DecompDiff for consistency, detailed in Section 4.2. (Also refer to Response to Weaknesses 3)
• Protein Selection: As mentioned in Section 4.2 - Data, for the protein pockets for 3D drug design , we selected 100 diverse pockets from the CrossDocked2020 dataset, ensuring they had <30% sequence similarity to pre-training data to evaluate the generalizability of our model.

We thank you again for your comprehensive review. We hope these clarifications address your concerns of our work.

We sincerely appreciate your positive feedback on the novelty, clarity, and strong performance demonstrated by 3DMolFormer. Below, we address your concerns point-by-point.

Response to Weaknesses

The multi-objective optimization of the RL seems to be overly simplistic, including a reward function that assigns a constraint-based reward for QED and SA. Literature on multi-obj DRL shows using more sophisticated reward functions and multi-objective optimization techniques greatly improve agent performance and stability.

Thanks for your suggestion! We understand and acknowledge that the reinforcement learning (RL) reward function in our work could be seen as simplistic compared to more sophisticated multi-objective optimization approaches in the literature of general machine learning. However, our design was intentionally straightforward to maintain focus on demonstrating the core innovations of our model, namely the dual-channel transformer framework and its applicability to both docking and drug design tasks. We have included an expanded discussion in Section 5 of our revised paper to provide context for this design decision and outline potential extensions using advanced RL techniques.

Minimal ablation studies are conducted, and all results are based on one run. More runs should be conducted to demonstrate the soundness of the model.

Thank you for this suggestion. In response, we have taken two steps to strengthen our experimental evidence:

1. Reproducibility Across Runs: We conducted additional experiments, running 3DMolFormer five times on both protein-ligand docking and pocket-aware 3D drug design tasks. We now report the standard error of these results in Appendices C and D of the revised paper, demonstrating the robustness and consistency of our model's performance across multiple runs.
2. Ablation Study on RL Fine-Tuning: We performed an ablation study to analyze the impact of freezing the weights for ligand SMILES generation during RL fine-tuning. The results, included in Appendix D, show that freezing these weights significantly reduces the quality of generated ligands and hinders reward optimization. This validates the importance of allowing weight updates during RL fine-tuning, as it enables the generation of molecules with increasingly higher expected rewards.

Minor Edits: Line 482: the parameter σ in Eq. (4) was is to 100

The typo in line 482 has been corrected in the revised paper.

Response to Question

The authors state: “... the sampling of ligand SMILES utilizes the weights of the RL agent’s model, which are continuously updated during finetuning. In contrast, the generation of atomic 3D coordinates uses the weights from the model finetuned for docking, which remains unchanged during this process.” Why don’t the authors freeze the GPT weights during the sampling of ligands? If this hinders the model’s performance, why is this not shown in the ablation studies?

Thank you for this insightful question. The core purpose of RL fine-tuning in our framework is to iteratively optimize the weights for ligand SMILES generation, ensuring that the generated molecules achieve progressively higher expected rewards. Freezing these weights would undermine the RL process, preventing any improvement in the ligand generation mechanism. Conversely, the generation of atomic 3D coordinates is not the target of RL fine-tuning, which is why we use frozen weights for this part.

To substantiate this, we conducted an ablation study comparing performance with and without freezing the weights for ligand SMILES generation during RL fine-tuning. The results, detailed in Appendix D, reveal that freezing the weights significantly reduces reward optimization and degrades the quality of generated ligands. These findings underscore the necessity of weight updates in achieving the goals of RL fine-tuning and validate our design decisions.

Thank you again for your valuable feedback. We hope our responses and revisions address your concerns adequately and further highlight the strengths of our work.

I'm Xiangyu Huang from Department of Life Science, Tsinghua University. I didn't write the comment above and I doubt that the commenter stole my personal information.

Dear Nahyun,

Thank you for your interest in our work. We sincerely appreciate your efforts to reproduce our results and your thoughtful questions.

We acknowledge that there may be some typos or ambiguities in the released code, which could be misleading. We will carefully review and update the code as soon as possible to ensure clarity. Additionally, we are working to address the challenges of uploading large model checkpoints due to internet restrictions in mainland China, and we will do our best to make these available to facilitate your work.

Regarding your Point 5, the discrepancy between your results and those reported in the paper is significant—far beyond what might be expected from experimental randomness. Given that your results are even lower than the performance of "w/o pretraining" (i.e., training on PDBbind alone), there may be a misalignment in the training process, possibly due to issues in our code. We deeply apologize for any confusion or inconvenience this has caused and will prioritize investigating this further.

Thank you again for your patience and for bringing these issues to our attention. Please don’t hesitate to reach out if you have further questions in the meantime.

Best regards,
Xiuyuan Hu, on behalf of all authors