CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph

Thanks for your recognition of our contribution, and here we will answer your questions to eliminate your concerns one by one:

Reply to Originality:

We have carefully reviewed the article by Zheng et al. [1] and found that it was published on June 4, 2024, contemporaneously with our work. Thus, we believe that concerns regarding the novelty of our research are unfounded. Additionally, we would like to highlight several distinctions and advantages of our study in comparison to this work:

• 1. Zheng et al. primarily aim to demonstrate that 1D/2D molecular generation methods remain highly competitive compared to 3D approaches. In contrast, our work focuses solely on generative 3D SBDD tasks. Moreover, we provide a more comprehensive evaluation of 3D molecular generation methods, incorporating over 10 3D-based methods compared to the five included in Zheng et al.
• 2. In terms of evaluation metrics, beyond Vina Dock Energy, SA, and QED for drug properties, we include additional assessments such as PLIP interaction analysis, clash analysis, and geometric fidelity. Our 3D-based comparisons are therefore broader and more detailed.
• 3. We have incorporated four additional lead optimization tasks, including linker, fragment, scaffold, and side-chain design. These aspects are rarely covered in previous benchmark or SBDD task-oriented papers.
• 4. Our work unifies the code framework, whereas Zheng et al.’s published GitHub repository does not include training or generation code for the methods discussed. This difference implies a significantly greater engineering effort on our part, potentially making a substantial contribution to the community. Please refer to the '.zip' file in the supplementary material for further details.
• 5. We conducted experiments on real disease targets, ADRB1 and DRD3, which are rarely explored in previous benchmarking and SBDD task-oriented studies.

Therefore, we believe that (1) questioning the novelty of a contemporaneous study is not entirely fair in this context; and (2) CBGBench possesses distinct differences and advantages compared to the work you referenced. Besides, since Zheng et al. also gives lots of insights in SBDD tasks, we have added it as a reference in the newly updated Line. 539 for the reader to understand the field easily.

Reply to Quality:

• We choose 5 as the threshold for considering the group of atoms because too few atoms cannot form appropriate fragments. Criteria of `5 atoms' primarily references the atomic number constraints used in the DeLinker[2], where the design likely focuses on common pharmacophoric substructures like five-membered rings for fragment truncation. In practice, you can freely define these based on your own requirements, and the relevant code has been uploaded.
• Thanks for you advice on the confidence intervals computing method. We think it is really valuable, since in the macro-biomolecule design, the task of `quality identification' is very important, to tell the biologist whether the generated biomolecules are reliable. We will take the advice as future work, to establish an SBDD-version generation quality identification model, to facilitate the judgment of whether one of the tested models is significantly better than the others.

Reply to Clarity:

• We aggree with you that the phrase is confusing, so we change it into disease target, since DRD3 is a type of dopamine receptor primarily found in the brain, related to Schizophrenia and Parkinson’s disease and ADRB1 is a β1-type adrenergic receptor related to hypertension, heart failure and Arrhythmias. In comparison, in CrossDocked, the protein pockets are derived from structures in PDBbind, forming protein-ligand complexes that exist in the real world. However, while these receptors are indeed real-world entities, they are not necessarily disease targets. Therefore, in the newly uploaded version, we have revised references to “real-world target” to “real-world disease target” or “real-world target related to disease” to avoid any potential ambiguity.
• We have incorporated your suggestion and revised the text to: “the established evaluation protocols exhibit consistency and generalizability on real-world target data in evaluating the Vina Dock Energy.”
• We have changed the font size in the Figure in the latest updated version.

Dear Reviewer ftZU:

Thank you very much for recognizing our work. We have provided responses to the fragment growing and other molecule splitting rules you raised. Additionally, the revised manuscript we uploaded has incorporated your suggestions, including: replacing the less appropriate term “real-world target” with “real-world disease target”; citing and discussing the references you suggested (Appendix Line937 - 943); and updating the newly uploaded code version to include the bootstrap method for confidence interval estimation as per your recommendations.

Given the extended discussion period, we sincerely ask if there are any unresolved concerns or whether our current responses have addressed all your doubts about the manuscript. Considering that all reviewers with positive opinions have provided satisfactory feedback and stronger recommendations after our responses, we hope that if you have any remaining questions, you feel free to raise them.

We sincerely look forward to your response.

Best regards,

Authors.

IV. Reply to Data

Question 5:

The dataset are all constructed based on CrossDocked2020. The four task aims to generate different part of a molecule, so the preprocessing of the fragmentation of the molecules in CrossDocked2020 is different, leading to different 4 subtasks and the corresponding 4 datasets. You can refer to Sec. 3 for details.

V. Reply to Training and Sampling Protocol

Question 15:

In fact, every paper presents various examples highlighting its strengths and weaknesses. We plan to include generated molecular case studies in the appendix in future versions to provide more intuitive visualizations.

Question 16:

In fact, we reviewed the standard training protocols currently in use and found that 5M iterations represent the maximum used. Therefore, we standardized on 5M iterations for all models to ensure fairness, avoiding cases where some models might fail to converge with fewer iterations. The final model is selected based on the lowest validation loss, which we believe makes this unified protocol fair.

The metrics we reported are highly consistent with those in published models, although their reports often lack comprehensiveness. To address this, we adopted our proposed protocol, which is currently the most comprehensive available. Additionally, we attribute some discrepancies to the inherent randomness of generative models and different evaluation aspects.

Question 17:

As described in te introduction, we aim to make comparisons under the same network architecture as much as possible. Based on previous studies [2], [12], we can infer that Pocket2Mol achieves state-of-the-art performance in AR, while others, like GraphBP, perform less favorably. This led us to explore whether the lack of network representational capacity or the difference in generation strategies is the main cause of the performance disparity. Therefore, we adopted the GVP architecture used in Pocket2Mol to reach conclusions like those in Line 82-83 and Line 401-408. Otherwise, it would be meaningless to draw these conclusions regarding the strengths and weaknesses of different generation strategies. Similarly, existing state-of-the-art one-shot SBDD models, such as MolCraft, employ a Transformer + EGNN network architecture. This naturally raises the question of whether their superiority over earlier models like DiffBP/DiffSBDD stems from their generative strategy or the stronger representational power of their network architecture. We believe that comparisons should be made under consistent architectures to fairly evaluate the relative strengths of different generative strategies and to highlight potential directions for future development. Therefore, we do not view modifying the model architecture as an issue; rather, we see it as an advantage that ensures fair comparisons.

Question 18:

All the code has been organized into a unified repository and refactored. If you are interested, please download the supplementary.zip file and compare it with the code repositories of previous methods. This will allow you to assess whether our engineering efforts are sufficiently comprehensive.

Question 19:

This is a protocol shared by all previous methods: generating 100 molecules for each pocket in the protein test set, resulting in 10,000 molecules across 100 pockets. In Table 2, the number 100 refers to the 100 pocket-ligand pairs in the test set. Such protocols typically do not require mathematical notation, just as one wouldn’t use $N_{\text{class}} = 100$ to explain that CIFAR-100 has 100 classes or $N_{\text{class}} = 1000$ to indicate that ImageNet-1k has 1000 classes. I sincerely recommend that you review relevant prior research to familiarize yourself with the common consensus in this field (´Д｀).

Question 21:

As stated in the paper, the generation target for each subtask corresponds to a different part of the molecule. Therefore, it is necessary to finetune the aforementioned models to better adapt them to the new downstream tasks. Additionally, for diffusion models, certain preprocessing techniques, such as center-of-mass adjustment, require re-centering the coordinate system’s origin, which necessitates retraining the model.

Question 24:

The choice of 100 is based on the aforementioned protocol, which involves generating 100 molecules for a single pocket. Additionally, for the two targets in question, the number of known chemically active molecules we collected is also around 100. Therefore, comparing the distribution of 100 generated molecules versus 100 reference molecules is reasonable.

If this is still unclear, we can explain it this way: in the case study, there are two protein targets, each with approximately 100 known active molecules. For each target, we generated 100 \times \text{(number of methods)} molecules using different methods, and then compared these generated molecules. :)

[1] Jiaqi Guan, 3D Equivariant Diffusion for Target-Aware Molecule Generation and Affinity Prediction

[2] Xiangang Peng, Pocket2Mol: Efficient Molecular Sampling Based on 3D Protein Pockets

[3] Wang Yidong et al. Usb: A unified semi-supervised learning benchmark. ArXiv, abs/2208.07204, 2022.

[4] Raphael Townshend et al. ATOM3D: Tasks on Molecules in Three Dimensions, https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/hash/c45147dee729311ef5b5c3003946c48f-Abstract-round1.html

[5] Wangyang Wang et al, RELATION: A Deep Generative Model for Structure-Based De Novo Drug Design, https://pubs.acs.org/doi/10.1021/acs.jmedchem.2c00732

[6] Mikhail Andronov et al. Exploring Chemical Reaction Space with Reaction Difference Fingerprints and Parametric t-SNE, https://pubs.acs.org/doi/10.1021/acsomega.1c04778

[7] Haitao Lin et al. Functional-Group-Based Diffusion for Pocket-SpecifiMolecule Generation and Elaboration, https://proceedings.neurips.cc/paper_files/paper/2023/file/6cdd4ce9330025967dd1ed0bed3010f5-Paper-Conference.pdf

[8] Odin Zhang et al. ResGen is a pocket-aware 3D molecular generation model based on parallel multiscale modelling, https://www.nature.com/articles/s42256-023-00712-7

[9] Yang Yaodong et al. Enabling target-aware molecule generation to follow multi objectives with Pareto MCTS, https://www.nature.com/articles/s42003-024-06746-w

[10] Sebastian Salentin et al. PLIP: fully automated protein–ligand interaction profiler. https://pmc.ncbi.nlm.nih.gov/articles/PMC4489249/

[11] Melissa F Adasme, PLIP 2021: expanding the scope of the protein–ligand interaction profiler to DNA and RNA, https://academic.oup.com/nar/article/49/W1/W530/6266421

[12] Meng liu et al., Generating 3D Molecules for Target Protein Binding, https://arxiv.org/abs/2204.09410

Question 7:

MAE is used to compared point to point. For example, for a drug dataset which can bind to certain pocket, where the average molecule contains 2 benzene rings and 4 hydroxyl groups, its distribution can be represented as $\text{Multinomial}(1/3, 2/3)$. If, in the generated molecule set, each molecule contains 1 benzene ring and 2 hydroxyl groups, the overall distribution matches, and the JS divergence would be optimal. However, for such molecules, their pharmacophore, which interacts with the drug, may require 2 benzene rings and 4 hydroxyl groups. In this case, KL divergence fails to capture the differences in functional group distributions. To address this limitation, we introduce $\text{MAE} = 0.5 \times (|2 - 1| + |4 - 2|)$ as a point-to-point comparison metric, which is reasonable and aligns with the approach in D3FG [7].

For other metrics, a similar rationale can be applied. If you find any metrics difficult to understand, please let us know, and we are extremely pleased to provide further examples to clarify the metrics you are struggling with. :)

Question 8:

‘Atom type’ is denoted by AT, 'Functional Group' is denoted by FG and RT means 'Ring type'. They are three different and independent evaluation frameworks. For atom type, we only consider the distribution of seven atom types: C, N, O, F, Cl, P, and S (Table 15 in the latest-uploaded version). This results in a generated probability distribution $\hat{p} \in \mathbb{R}^7$ with $\sum_i^7 p_i = 1$ , which can be compared with the true distribution $p \in \mathbb{R}^7$ to compute the divergence. For functional groups (FG), $p$ is a 25-dimensional probability vector (Table 17 in the latest-uploaded version, where we show the top 10 dimensions). For rings, $p$ is a 6-dimensional vector (Table 16 in the latest-uploaded version).

Question 9&10:

The SBDD task is not about generating the “correct” ligand; the task aims to explore the chemical space of ligands that can bind effectively to the target. The requirement is for the generated ligands to exhibit properties such as realistic molecular geometry, drug-like properties, pharmacophoric substructures, expected interaction types, and the ability to bind well to the target protein. From this perspective, we suggest reevaluating the protocol and metrics. “Correctness” has never been the focus of this field—if your model could only predict existing, validated drugs, the goal of “drug discovery” would no longer exist. Additionally, given that many probabilistic models cannot compute precise likelihoods, such as MolCraft using BFN and LiGAN using VAE, comparing the NLL (negative log-likelihood) on the test set becomes challenging to implement.

Question 11:

Section 4 provides clear comparisons of the metrics and explicitly explains which metrics are better when larger and which are better when smaller. For metrics like JSD, which match the true distribution, the goal is to minimize it, and this does not create any conflict. Additionally, each metric is ranked internally, so differences in metric scales do not affect the comparisons. Could you please specify which aspects made you feel that the comparisons were vague? This would help us address your concerns more effectively.

Question 12:

The three Vina Energy represents the energy calculated with intial pose, optimized pose and optimized conformations. Although the predictions made by Vina are neither fully precise nor comprehensive, truly accurate validation requires wet-lab experiments, which are neither feasible nor affordable for screening generated molecules. Furthermore, currently, no widely recognized tools superior to Vina have been adopted in the field (as evidenced by the references we cited). Therefore, our aim is to build upon Vina and address some of its limitations.

Regarding LBE, we suggest reviewing the analyses provided in Appendix B2 and B4. If you still have concerns, we welcome further discussion. Additionally, for the presentation of metrics, you may refer to the main experimental tables in previous works such as [1].

Question 13:

Please refer to Sec.4 Line 255 - 275, and besides, PLIP tool is commonly used to analyze the interaction pattern (Figure 2c in page 1024 in [8] and Figure 3c in [9]). If you are interested in PLIP tool, we here give you a detailed paper for reference [10][11]. We welcome further discussion.

Weakness 3:

The use of additional decimal places is because many values cannot be meaningfully distinguished using only two decimal points, so we have adopted four decimal places for precision. The absence of standard deviations is due to the generative nature of the task, where the comparison focuses on the differences between distributions. Sampling 10,000 molecules for distribution comparison is an established protocol followed by nearly all prior studies, and we believe that adhering to a reasonable protocol based on published works is appropriate.

The metrics we present are those typically reported in prior studies, and the benchmarking tables follow the format in [3]. If a ranking is all that’s needed, please refer to the last column in each table. In fact, the performance differences among these methods are minimal, and comparing them does indirectly serve as a form of ablation analysis. We have clearly outlined the reasons for the performance differences in the Conclusion section, so please review this in detail. Additionally, we would appreciate specific examples to clarify what kinds of insights are being sought.

Weakness 4:

We would appreciate it if specific evidence for the allegations could be provided. If the following questions are specific concerns, we would be happy to address them one by one.

Question 1 & 2:

We have added an additional table in the Appendix to clarify the supported tasks in Table 9, along with the corresponding losses in Table. 10. Besides, the architecture has already been outlined in Table 9.

Question 3:

We aim to introduce the task of de novo Complete Binding Graph to unify the lead optimization process. Subsequently, in the experiments, we first compare the performance on de novo design and then evaluate the related performance on lead optimization. Our paper's presentation follows a structured flow: introduction of tasks and corresponding data -> presentation of evaluation protocols -> experiments.

Question 4:

Many methods require modifications to be adapted for these tasks. Although the training objectives are usually the same, the conditions in conditional probabilistic models differ, necessitating various adaptations, such as zero-centering-of-molecule techniques. Of course, we would prefer there will be newly established models better suited to these tasks; however, as a benchmarking study, we primarily utilize existing models and adapt them for relevant tasks instead of proposing a new one. Moreover, models designed for de novo generation are not always effective in lead optimization. A typical example of this is autoregressive models, which may perform well in de novo generation but fall short in lead optimization.

Question 6:

The use of this notation is to unify the models under the task of “Fill-in-the-gap in Complex Binding Graph.” If a unified notation system were to be applied to describe all models, the article would become overly verbose and lean more toward a review-style paper rather than a benchmark and dataset study. Please differentiate the purposes and writing styles of these two types of articles. For reference, you may look at Atom3D [4].

Question 20:

The context means the text in the main body, we here change the discription as main text to avoid confusion.

Question 22:

Please refer to Line 452 - 459 to figure out if you are sure that the conclusion is right or wrong. We here need more detailed problems.

Question 23:

The visualization is a very common technique in the field of Computational Chemistry, for example [5] and [6]. Please refer to them.

Question 25:

Although Vina is known to have limitations, there is currently no better computational method widely available to replace docking as an initial screening metric. Therefore, we have continued to use it. From the results in Table 7, it is evident that DiffSBDD underperforms. Furthermore, in the real-world target case study, DiffSBDD also appears to perform poorly, which demonstrates the consistency of these two conclusions.

We kindly ask the reviewer to carefully read our paper to fulfill the responsibilities of peer review. The inconsistency between 'A' and 'B' that you pointed out does not pertain to the same objects we described, and thus the claim of inconsistency does not apply.

Weakness 1:

The motivation is clearly stated in both the abstract and introduction. If there are any specific concerns, we welcome precise feedback. While de novo tasks are indeed common, lead optimization tasks have received far less attention in the literature.

Weakness 2:

Our contributions include a unified setting of tasks, a unified codebase, a unified evaluation protocol, extensive empirical study and insights derived from experimental results. The framing of tasks is not well-known; if it is, please provide specific examples. Simply asserting a lack of novelty is easy, but pointing out similar prior work would constitute a more solid argument.

Given the numerous issues raised in this review, we have broken down our responses into several sections to address each point individually. Additionally, since your questions were not numbered, we have re-labeled them for clarity. Due to space limitations, we did not copy the re-labeled questions with number, and you may re-number your questions and cross-reference them accordingly. The questions and weaknesses can be classified into the following parts:

I. Novelty& Contribution:

Weakness 1, 2

II. Presentation:

Weakness 3, 4

Question 1, 2, 3, 4, 6, 20, 22, 23 25

III. Metrics:

Question 7, 8, 9&10, 11, 12, 13, 14

IV. Data:

Question 5

V. Training and Sampling Protocol:

Question 15, 16, 17, 18, 19, 21, 24

Besides, we friendly recommend the reviewer to get more familiar with some representative prior works, such as TargetDiff[1] and Pocket2Mol[2], to avoid questions that challenge well-established concepts and common consensus in the field. :)

Dear Reviewer G2c6:

Many thanks for your raising your rating to 5. Besides, your insightful and helpful comments help us to improve our manuscript a lot. In the former discussion, we suppose we have thoroughly considered all the feedback and provided a point-to-point response to you, including the total 25 questions and 4 weakness.

Since the reviewer-author discussion phase appears to have been appropriately extended, we sincerely request that you take the time to reassess our responses and the contributions of this paper. As you may have noticed, nearly all reviewers with positive opinions have provided very strong recommendations for the paper (score 8), which we believe is worth your attention as a responsible reviewer. Given that your current rating leans somewhat negative, we earnestly hope that you could re-evaluate the paper and consider providing a relatively positive recommendation. Your recognition of our work would greatly motivate us to continue our deep exploration in the SBDD field, unify frameworks, and refine engineering efforts.

Best,

Authors

Thanks for your valuable comments and insightful questions, here we will reply to the two weakness with questions one by one.

Reply to W1:

Firstly, in our article, one-shot models are not limited to diffusion-based models alone. As shown in Table 1, one-shot models here refer to approaches that are in contrast to auto-regressive models. For instance, LiGAN uses a VAE model framework, which we also classify under one-shot models.

Secondly, we included BFN in the diffusion model category because, in diffusion processes, the forward noising step adds noise to data using $x_t = \alpha_t x_0 + \sigma_t \epsilon$ , which can be viewed as adding noise to the mean of a Gaussian distribution parameterized by $\mu( x_t) = \alpha_t x_0$. The reverse sampling process follows a similar logic. However, while discrete state diffusion models like [1][2] also add noise in parameter space, BFN’s update strategy for categorical variables is distinctly different, yet it can still be considered a nonlinear update process in parameter space, as opposed to the linear updating strategy with Markov transition probability in [1][2].

Thus, both traditional diffusion models and BFN exhibit the following features: (1) unification in parameter space for noise addition and sampling; and (2) progressive updates with either a linear or nonlinear strategy. Based on these similarities, we classify BFN as a diffusion-based approach.

Reply to W2:

We apologize for the lack of clarity in our explanation. We agree with your point that calculating clashes within the molecule itself is meaningless, as in structures like C=O-C, it is very likely that the distance between the two carbon atoms will be less than their van der Waals radii overlap. In fact, the term Ratio_cm here does not refer to clashes within the molecule itself; rather, it indicates the proportion of molecules with protein-molecule clashes relative to the total number of molecules. If any atom in a molecule causes a clash, we consider that molecule to have a clash. Therefore, this metric is stricter than Ratio_ca and will yield a higher ratio. We will make this explanation clearer in the next version (updated line 285) to avoid any unnecessary misunderstandings.

Minor： Besides, we add the suggested reference in our updated paper. Thanks for your valuable advice.

[1] Emiel Hoogeboom, Argmax Flows and Multinomial Diffusion: Learning Categorical Distributions, https://arxiv.org/abs/2102.05379

[2] Andrew Campbell, A Continuous Time Framework for Discrete Denoising Models, https://arxiv.org/abs/2205.14987

We sincerely appreciate your recognition of our work. In response to the issues you raised, we provide the following replies:

Following [1], we focused on the two targets ADRB1 and DRD3, both of which play significant regulatory roles in biological systems, making case studies on these targets meaningful.

Additionally, based on your valuable feedback, we recently included experiments with the KRAS12 target (PDB code: 7O83) and used the bound structure’s protein configuration as the pocket (see Fig. 10 in the Appendix E3 in the latest article version). Using the methods incorporated in our case study, we generated 100 molecules for the target, then assessed their molecular fingerprint distribution as well as Vina Energy & LBE. The results have been added to Appendix E3 in Figures 11(a) and (b). Furthermore, we provided a summary table of Mean & Median values related to binding affinity of the generated molecules on KRAS12 for brief reading.

[1] Haotian Zhang, Delete: Deep Lead Optimization Enveloped in Protein Pocket, https://arxiv.org/pdf/2308.02172

Dear Reviwer kydR:

We sincerely and deeply appreciate your efforts and valuable advice. It is the dedication and active engagement of reviewers like you, who continuously engage in discussions with authors, that drive the development of the ML community and the AI4Science community.

As you can see, we have added extensive parts for real-world target analysis, including KRAS12 as you suggested. Since the reply has been made from about one week before, and there will be about three days left until the discussion deadline, we would like to know if we have addressed all your concerns at this stage. If you have any further related concerns, please feel free to share them with us, and we will be more than happy to address them. If you are satisfied with our responses, we kindly ask if you would consider reevaluating the manuscript and providing a stronger recommendation. If you recognize the value of our work and the efforts we have made during the rebuttal phase, we kindly ask you to consider raising the score to ‘Accept’, as two of the reviewers has raised their rating. We would be immensely grateful, as this would acknowledge the significant effort we have invested. In the future, we will expand our framework and codebase with the newest advancements, making ongoing contributions to the AI for drug design community.

Your feedback, recognition of our work, and recommendation are extremely important to us. :)

Dear Reviewer kydR:

Thank you for your insightful and helpful comments once again. We suppose we have thoroughly considered all the feedback and provided a point-to-point response to you, espicially add the target of KRAS12 as a case study.

The deadline for author-reviewer discussions is approaching. If you still need any clarification or have any other questions, please do not hesitate to let us know. We hope that our response addresses your concerns to your satisfaction and we would like to kindly request a reconsideration of a higher rating if it is possible.

Best regards,

Authors.

Thanks for your recognition of the value of our work. Here is our reply to your questions.

Reply to Q1:

We here use GVP as the backbone architectures for Auto-regressive models and EGNN as the architecture for One-shot models, for the two reasons:

1. GVP projects the vector (positions) into a higher dimensional space, and update the latent coordinates, while EGNN directly update the 3D Euclidean coordinates. From this perspective, GVP requires more memory and computational comlexity compared to EGNN. This is because GVP projects both coordinates and attributes into a high-dimensional space. This makes GVP more feasible in an auto-regression (AR) model, where the AR model only needs to model the coordinates of one atom (the next atom) at a time. However, in one-shot methods, such as diffusion, it must model the joint distribution of the coordinates and types of $N_{\text{atom}}$ atoms simultaneously, leading to potentially unaffordable computational costs. Therefore, we chose the GNN architecture with the strongest possible representational capacity within feasible conditions, adapting it to different generation methods to minimize the negative impact of insufficient network expressiveness on performance and to explore various generation strategies and methods.
2. Moreover, as described in the introduction, we aim to make comparisons under the same network architecture as much as possible. Based on previous studies [1], [2], we can infer that Pocket2Mol achieves state-of-the-art performance in AR, while others, like GraphBP, perform less favorably. This led us to explore whether the lack of network representational capacity or the difference in generation strategies is the main cause of the performance disparity. Therefore, we adopted the GVP architecture used in Pocket2Mol to reach conclusions like those in Line 82-83 and Line 401-408. Otherwise, it would be meaningless to draw these conclusions regarding the strengths and weaknesses of different generation strategies.

Reply to Q2:

We are sorry about the newest revision on the experimental results in the Appendix is not updated. We give the result in Figure 8 and 9 in the Appendix of the latest updated version. The t-SNE visualization distribution changed since the MolCraft molecule is added and making the training data for t-SNE different. Thanks a lot for your mentioning our mistake.

Reply to Q3&Weakness:

Since this paper primarily focuses on the SBDD task, we have not discussed broader biomolecular generation tasks. Here, we can provide an overview of this direction. A common consensus is that de novo biomolecule design requires models to generate both the types and structures of basic units in biomolecules.

Therefore, from a structural biology perspective, these basic units are amino acids in proteins and peptides, bases in DNA, and atoms in small molecules. Although it is possible to generate these biomolecules atom-by-atom from the bottom up, the inherent “biological language” of proteins (e.g., AFCUDNE…) and DNA (e.g., ATCG) complicates ensuring that atomic-level generation can successfully assemble corresponding amino acids or bases. Consequently, bottom-up generation is challenging for other macro biomolecules. In contrast, a top-down generation strategy that treats functional groups as fundamental units in proteins and amino acids can inspire molecular design and SBDD tasks. Examples of such approaches include [3] and [4].

Besides, from a model design perspective, as the fundamental generation units for proteins and other biomolecules are functional groups, determining the atomic coordinates within each functional group becomes an essential issue in transitioning from coarse-grained to fine-grained generation. Current models often treat the protein backbone as a rigid body and side chains as flexible elements, generating the C-alpha positions and orientations of amino acids in the backbone. This approach has inspired the motivations in SBDD tasks like [3] and the generation strategies in lead optimization like [5].

Overall, while no model currently unifies de novo design for both marco-biomolecules and small molecules, the fields are increasingly converging. For instance, AlphaFold3 focuses on protein&DNA&RNA-ligand bound structure prediction (though not yet de novo design), and research on unified approaches is expected to make steady progress in the coming years.

[1] Meng liu et al., Generating 3D Molecules for Target Protein Binding

[2] Xiangang Peng et al., Pocket2Mol: Efficient Molecular Sampling Based on 3D Protein Pockets

[3] Haitao Lin et al, Functional-Group-Based Diffusion for Pocket-Specific Molecule Generation and Elaboration

[4] Jiaqi Guan et al, DECOMPDIFF: Diffusion Models with Decomposed Priors for Structure-Based Drug Design

[5] Jiaqi Guan et al., LinkerNet: fragment poses and linker co-design with 3D equivariant diffusion