Unified Guidance for Geometry-Conditioned Molecular Generation

Effectively designing molecular geometries is essential to advancing pharmaceutical innovations, a domain, which has experienced great attention through the success of generative models and, in particular, diffusion models. However, current molecular diffusion models are tailored towards a specific downstream task and lack adaptability. We introduce UniGuide, a framework for controlled geometric guidance of unconditional diffusion models that allows flexible conditioning during inference without the requirement of extra training or networks. We show how applications such as structure-based, fragment-based, and ligand-based drug design are formulated in the UniGuide framework and demonstrate on-par or superior performance compared to specialised models. Offering a more versatile approach, UniGuide has the potential to streamline the development of molecular generative models, allowing them to be readily used in diverse application scenarios.

Paper

References (84)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer D4nw7/10 · confidence 4/52024-07-02

Summary

The paper introduces UniGuide, a unified framework for geometry-conditioned molecular generation using unconditional diffusion models. UniGuide is designed to address the adaptability issues in current molecular diffusion models by providing a general training-free approach via a condition map that transforms complex geometric conditions to match the diffusion model’s configuration space, allowing for self-guidance during the generation process. UniGuide is demonstrated to be effective in various drug discovery tasks, including structure-based, fragment-based, and ligand-based drug design. The framework shows either on-par or superior performance compared to specialized models, highlighting its potential to streamline the development of molecular generative models.

Strengths

1. The paper introduces a novel, training-free method to guide diffusion models based on expected geometry conditions, enhancing adaptability without additional training overhead. 2. The framework is thoroughly evaluated across multiple drug discovery tasks, demonstrating its effectiveness.

Weaknesses

1. The contribution and the generalizability of UniGuide are somewhat overstated. In the Introduction and Figure 1, the available conditions are described in a quite general way, including not only structures and surfaces, but also densities. As this paper mainly focuses on geometry-aware conditions, clarifying the current scope in the introduction and main figures would enhance the paper's accuracy.

Questions

1. In the FBDD experiments, can the UniGuide framework also be applied to the DiffLinker model, considering it is diffusion-based backbone? 2. Based on W1, it is noticed that there is a brief exploration on the density condition in App. G. Given its potential to enhance the model's capabilities, it is suggested to provide additional details about this setting (e.g., task definition, condition mapping) and to reference App. G in the main body of the paper.

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

The limitation of this paper is well discussed

Reviewer YA7u4/10 · confidence 3/52024-07-09

Summary

This paper proposed a training-free framework for guided diffusions in unconditional molecular generation. UniGuide applies to a wide range of design tasks such as SBDD, FBDD and LBDD, unified by the proposed condition map, which projects from the product space of conditional input that lies in general geometric space to some favorable datapoint in the data space that diffusion operates within.

Strengths

- The idea of condition map that unifies different downstream tasks is novel. - This paper is generally easy to follow.

Weaknesses

- In all tables, the authors claimed to highlight the best "diffusion-based approach" in bold, which seems misleading since there are also non-diffusion baselines. I would recommend the authors to reconsider this style of presentation in order to faithfully reveal the general performance. - After checking other baseline results, UniGuide seems not so competitive in downstream tasks such as LBDD and linker design in FBDD. This casts doubt on its effectiveness. - There are a number of important baselines missing, e.g. DecompDiff [1] and IPDiff [2] for SBDD tasks, and LinkerNet [3] for LBDD. [1] DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design [2] Protein-Ligand Interaction Prior for Binding-aware 3D Molecule Diffusion Models [3] LinkerNet: Fragment Poses and Linker Co-Design with 3D Equivariant Diffusion

Questions

- The condition map works similarly to molecular translation or some form of retrieval given conditional inputs. How would the authors compare to those methods? - For SBDD, the authors only reported QVina Dock scores in the main results. However, SBDD models have been criticized for inaccurate structure modeling [1]. It seems to me that Vina Score and Vina Minimize used in [2] would serve as a better indicator for the pose quality. Can the authors also include these metrics in their results? [1] Benchmarking Generated Poses: How Rational is Structure-based Drug Design with Generative Models? [2] DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design

Rating

4

Confidence

3

Soundness

2

Presentation

2

Contribution

2

Limitations

See above.

Reviewer ZcJd6/10 · confidence 4/52024-07-15

Summary

This paper proposed a method named UniGuide for conditional molecular generation with unconditional diffusion models, without the need of additional training and parameters. The proposed framework is an extension of self-guided diffusion models on the conditional molecular generation task. The authors designed different condition maps C: S x Z $\rightarrow$ Z for ligand-based generation and structure-based / fragment-based generation, enabling guidance from conditions in a unified fashion.

Strengths

* The paper is generally easy to follow and well-written * The theoretical justification is provided and oblation studies / visualization results are sufficient

Weaknesses

* Technical contribution * My biggest concern is on the superiority of the proposed method over other conditional sampling methods. E.g. the special case of S = Z can be implemented with inpainting technique; the shape-based generation can also be implemented with some technique similar to classifier guidance (e.g. the validity guidance used in [1][2]) by designing a loss function based on denoised datapoint. However, the authors didn't discuss about the superiority of their method with sufficient experimental supports. * Limitation: The performance is strongly relied on the based unconditional model. * Clarification * I don’t think the approximation of Eq 10 makes much sense to me: f approximates the clean data point, then how could the condition c be a Gaussian distribution taking this clean data point as the mean? I think the concept of condition c $\in$ Z in this paper is closer to datapoint satisfying condition c, or there should be another mapping from the data space to the condition space. * Eq 17 doesn’t strictly follow the definition of the condition map: c and z should have the dimension of n x (3 + d) — Eq (18) should have corresponding index selection / mask operations. * Experiments * In 5.1 Ligand-based drug design, why don’t the authors compare their model with conditional EDM? How could UniGuide (shapeMol), a conditional model with soft constraints, outperform Cond-Shape2Mol? * In 5.3 Fragment-based drug design, I think it’d better also compare UniGuide with an inpainting version of other baselines. References: [1] Peng, X., Guan, J., Liu, Q., & Ma, J. (2023). Moldiff: Addressing the atom-bond inconsistency problem in 3d molecule diffusion generation. arXiv preprint arXiv:2305.07508. [2] Guan, J., Zhou, X., Yang, Y., Bao, Y., Peng, J., Ma, J., ... & Gu, Q. (2023). DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design. In International Conference on Machine Learning.

Questions

See the weaknesses above.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

N/A

Reviewer cfg35/10 · confidence 2/52024-07-15

Summary

The paper introduces UniGuide which is a general framework for conducting conditioning over the unconditional molecule diffusion models during inference. To achieve this, Uniguide introduced a concept called a condition map for different applications. With a condition map,it could control the score function by adding a task-related gradient term for generating samples with desired properties. The experiment has been conducted over 3 settings of SBDD, FBDD, LBDD to demonstrate the effectiveness of the proposed method.

Strengths

1. The paper is generally well-written. The motivation is clear and important to the broad audience for taking usage of unconditional models to different complex scenarios. 2. The paper utilizes extensive experiment settings. I appreciate the efforts to conduct experiments in all the SBDD, FBDD, and LBDD settings.

Weaknesses

1. However, the paper's contribution is hard to evaluate. Though the author claims a new general framework for Uniguide, the form takes exactly as a gradient guidance form for diffusion models which has been explored in the previous literature and following ups [1]. Hence, the key contribution, given in the previous works, limits to a direct application of gradient-based guidances of molecule diffusion to different scenarios. I would like the author to clarify more about the contribution. 2. The important related works are missing such as [1,2]. I suggest the authors do a comprehensive review over the relevant literature. [1] Equivariant Energy-Guided SDE for Inverse Molecular Design. ICLR 2023 [2]Training-free Multi-objective Diffusion Model for 3D Molecule Generation ICLR 2024

Questions

refer to above

Rating

5

Confidence

2

Soundness

3

Presentation

3

Contribution

2

Limitations

No

Reviewer YdVq6/10 · confidence 2/52024-07-22

Summary

This paper presents a framework for geometric guidance of diffusion models to enable flexible generation for protein and small molecule tasks. Their method is based on self-guidance from geometric conditions. They propose a condition map to map geometric conditions to the latent condition space to guide diffusion. They demonstrate experiments on three main tasks: ligand-based, structure-based and fragment-based drug design.

Strengths

The paper addresses a relevant problem in drug design effectively. There is a need for a general framework for guiding diffusion models for the multitude of protein and small molecule design tasks. This work proposes a novel solution by focusing on geometry-based conditioning, and using a condition map to map different conditions to a common space. They provide extensive experimental evaluation on three major tasks, achieving comparable performance, or outperforming recent methods.

Weaknesses

A key benefit of this approach is the generalizability of the method. The main concerns I have are in relation to this. Regarding training the model, can the authors provide additional details on the training? For example, are all three tasks trained with different conditioning to guide the model, together? Or is one model trained separately for each task? What is the effect of training on multiple tasks together? For example, if one model is trained per task, then it is difficult to see the benefit of this approach, other than in the conditioning map. In the baseline experiments, I feel that there are some related work missing on works that condition diffusion models. For example, how would a simple conditioning mechanism, such as the ones used in the image-text generation literature, fare against this approach? How would this method compare to other latent diffusion models, such as [1-2]? It seems there are a multitude of latent diffusion models for generation out there, I find it difficult to compare those with this. The related work section also does not contrast this work with others in great detail. ### References [1] McPartlon, Matt, et al. "LATENTDOCK: Protein-Protein Docking with Latent Diffusion." [2] Watson, Joseph L., et al. "De novo design of protein structure and function with RFdiffusion." Nature 620.7976 (2023): 1089-1100.

Questions

- What is the relation with this work and latent diffusion? Could one similarly encode the geometric conditions using a latent encoder? - What is the computational complexity in training these models? - What other design tasks could this extend to?

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Authorsrebuttal2024-08-13

Additional Analyses for SBDD and LBDD

### Follow-Up on Comparison of UniGuide with validity guidance   We want to follow up on the “Validity Guidance” discussion (ZcJd), as mentioned in [1]. **We have conducted experiments using the surface loss computation and self-guidance for the LBDD task.** We selected the hyperparameters $\sigma$ and $\gamma$ in the surface loss computation to achieve a high DICE score between the implicitly defined surface ($S=\gamma$) and the meshes UniGuide utilises for LBDD ($\sigma=1$, $\gamma=2$, DICE > 0.8). Our surface calculations are performed using the Open Drug Discovery Toolkit (ODDT), which assigns specific radii to individual atom types and employs the marching cubes algorithm to generate meshes [2]. We performed several runs around the above hyperparameter specifications. The runs were similar in performance and we report the best result below: | |$\text{Sim}_S ~(\uparrow)$ | $\text{maxSim}_S~~(\uparrow)$ | $\text{Sim}_G ~(\downarrow)$ | $\text{maxSim}_G ~(\downarrow)$ | **Ratio** $(\uparrow)$ | Connect. $(\uparrow)$ | Unique. $(\uparrow)$ | Diversity $(\uparrow)$ | QED $(\uparrow)$ | | -------- | -------- | -------- | -------- | -------- | -------- | -------- | -------- | -------- | -------- | | Validity Guidance | 0.59 | 0.76 | 0.20 | 0.20 | **2.96** | 97%| 100% | 0.76 | 0.69 | | UniGuide (EDM) | 0.74 | 0.86 | 0.21 | 0.20 | **3.53** | 99%| 99% | 0.73 | 0.74 | **Although Validity Guidance for LBDD yields low graph similarity, the shape similarity remains suboptimal compared to UniGuide.** Additionally, we frequently encounter numerical instability for this guidance term, an issue not present with UniGuide’s formulation of LBDD. We believe this ablation will provide a valuable comparison for the final version of our manuscript and hope we have addressed the reviewer's remaining concerns.   [1] Guan, J., Zhou, X., Yang, Y., Bao, Y., Peng, J., Ma, J., ... & Gu, Q. (2023). DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design. In International Conference on Machine Learning. [2] Lorensen, William E., and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm, 1998     ### Follow-Up on SBDD   We want to follow up on the discussion regarding baselines for SBDD (ZcJd & YA7u), specifically IPDiff [1] and DecompDiff [2]. As previously mentioned, incorporating a form of "Clash Drift" into UniGuide, particularly within the condition map, is straightforward and helps prevent the generated ligand from colliding with the pocket. **We have implemented this adjustment, and our initial results indicate that similar to IPDiff and DecompDiff, it improves the overall VINA scores.** Specifically, UniGuide's VINA score, compared to the results reported in Table 1 of the rebuttal PDF, **improved significantly from −5.074 to -5.81. These initial results are promising and we are confident that adjustments of UniGuide's guidance parameters will lead to further improvments.** We again appreciate the reviewers' suggestions to include [1, 2] and believe that these additional results underscore UniGuide's flexibility to readily incorporate various conditions, enabling the comparison of underlying models across different levels. We hope that these findings address all the reviewers' remaining concerns, and we are looking forward to discussing these additional results.   [1] Huang et al. Protein-ligand interaction prior for binding-aware 3d molecule diffusion models [2] Guan et al. DecompDiff: diffusion models with decomposed priors for SBDD

Authorsrebuttal2024-08-13

Additional results posted as official comment

Dear ZcJd, As promised in our initial response, we conducted additional experiments to compare UniGuide with an adaption of validity guidance for LBDD and included a version of "clash drift" into UniGuide's condition map for SBDD. As we believe that the results are relevant to all reviewers, we kindly refer you to our official comment. Thank you again for your valuable feedback.   Best regards, The Authors

Reviewer ZcJd2024-08-13

Thank the authors for the detailed response! Most of my concerns have been addressed and I'd love to raise my score to 5. However, the authors still didn't show the superiority and highlight the difference of their method compared to some other techniques like classifier guidance, which doesn't need additional training, can also be applied in different application scenarios, and have very similar mathematical formation. That's the main reason for not giving a higher score.

Authorsrebuttal2024-08-14

Highlighting methodological differences and superiority of UniGuide

  Thank you for your kind answer and for increasing the score. We are glad that our clarifications and additional results have addressed most of the reviewer's concerns. Below, we would like to offer further clarifications on UniGuide's superiority and distinct advantages over alternative guidance methods that control generation during inference:   > Highlight the difference to other controlling techniques - **UniGuide maps the condition to the configuration space**: - Prior works [1,2,3], centred around classifier-guidance [1], guide towards conditions that describe global input properties. For example, [2,3] control the generation of molecules given a desired quantum property. We highlight that such prior works are limited in their applications and cannot (directly) condition on geometric conditions as introduced in our work. - [1,2,3] control the generation during inference usually by mapping the output of the diffusion model to the condition's space $(Z \to S)$, by using a classifier/regressor $f_{\theta}(z_t)$. In other words, the distribution $p_{\theta}(c |z_t)$ takes the form $p_{\theta}(c|z_t) = \mathcal{N}(c |f_{\theta}(z_t), I).$ In contrast, UniGuide maps the source condition to the configuration space $(S \to Z)$ which simplifies the gradient signal computation $\nabla_{z_t} p_{\theta}(c |z_t)$ to a squared error between the sample $z_t$ and the target condition $c = C({s}, \hat{z}_t)$ and eliminates the need for additional networks, see Eq.15. - **UniGuide is a self-guiding method**: UniGuide modifies the reverse process of the unconditional diffusion model without relying on additional networks to guide the generation, unlike [1,2,3] that use additional differentiable networks for guidance. This highlights the flexibility of UniGuide as it does not require additional training: neither training a specialised conditional diffusion model nor training additional classifiers/regressors for guidance. - **UniGuide is a gradient-based method**: In our experimental evaluations, we include other controlling techniques during inference, such as inpainting or position guidance. - For SBDD, inpainting directly inpaints the noised condition to the generation, leading to a perturbed denoising process through the used hard constraints. - For LBDD, ShapeMol+g directly leverages the positions of the reference ligands and adds additional position corrections to the conformations based on distance measures.   > Elaborate on the superiority of UniGuide - **Unification** through the condition map: UniGuide leverages a novel condition map to incorporate diverse geometric conditions, such as protein structures, molecular fragments, and molecular surfaces. This extends unconditional base models to diverse drug discovery tasks. Beyond established drug discovery tasks, UniGuide enables tackling entirely novel tasks where neither baselines nor sufficient data exists, e.g. the tasks of density-based drug design or symmetric proteins. We kindly refer the reviewer to the general response and the figures of the PDF. - **Equivariance**: Different from the text and image domains, equivariance is an important inductive bias of molecular generative modelling. In Theorem 4.1, we derive the conditions on UniGuide's condition map that enable guidance while respecting the molecular domain by maintaining equivariant updates. - Empirically, our results show that through UniGuide's soft gradient-based updates, we achieve state-of-the-art performance by providing the best trade-off between how well we incorporate the condition into the generative process (LBDD: shape similarity and graph similarity, SBDD: Vina) and the overall quality of the molecules (QED, SA, validity). We kindly point the reviewer to Tables 2 & 3 in the PDF. - Finally, we believe our follow-up response shows UniGuide's superiority over validity guidance. However, we are unsure if the reviewer has seen it. We provide results demonstrating that our condition map is more reliable and outperforms validity guidance regarding shape similarity, highlighting its effectiveness.   In summary, like classifier guidance, UniGuide controls the generation without additional training. In addition, UniGuide extends this approach [1] to the molecular domain by mapping the source condition to the configuration space. This enables direct self-guidance without additional networks while preserving equivariance. Our experimental evaluations show that UniGuide outperforms various baselines, including other techniques that control the generation solely during inference across multiple downstream tasks. We are thankful for the reviewer’s feedback and hope that our answers sufficiently clarify the remaining concern.   [1] Dhariwal et al. Diffusion models beat gans on image synthesis [2] Equivariant Energy-Guided SDE for Inverse Molecular Design [3] Training-free Multi-objective Diffusion Model for 3D Molecule Generation

Reviewer ZcJd2024-08-14

Thank you for the follow-up clarification. I have seen the new experiments about comparing UniGuide and validity guidance. However, I'm curious about what the specific difference is in terms of implementing these two methods? From my understanding, they both use the loss in the coordinate space derived from the geometric constraint to guide the diffusion sampling process.

Authorsrebuttal2024-08-14

Technical difference between UniGuide and validity guidance

We thank the reviewer for engaging in the discussion and raising this important clarifying question. It is correct that both UniGuide and the validity guidance from [1] compute the loss in coordinate space. However, UniGuide's condition map explicitly maps the condition to target points related to the surface, expressed as $c = C(s, \hat{x}_t)$, where $s$ represents the surface and $\hat{x}_t = f\_\theta(x_t, t)$ is the clean data conformation approximation. In contrast, [1] leverages the shape descriptor $S(x_t)$ to decide if a point $x_t$ is inside the volume but then does not use the surface for the gradient update but regresses against all surface-generating points. Note that [1] relies on this differentiable function $S(x)$ to implicitly define a surface via $S(x)-\gamma=0$. For a point $x_t$ that satisfies $S(x_t)<\gamma$, the gradient computes as follows: $$ \begin{aligned} \nabla_{x_t} S(x_t) = \nabla_{x_t} \Big [ -\sigma \log \Big ( \sum_i^N \exp(-\|x_t-y_i \|_2^2 / \sigma ) \Big ) \Big ] \\ \end{aligned} $$ $$ \begin{aligned} \qquad \qquad & = \frac{1}{\sum \exp(\dots)} \sum\_i^N \underbrace{\exp(\dots)}\_{\omega_i} \nabla\_{x_t} \|x_t-y_i \|_2^2 \\ \end{aligned} $$ $$ \begin{aligned} \qquad \qquad & = \frac{1}{\sum \omega_i} \nabla\_{x_t} \sum_i^N \omega_i \|x_t-y_i \|_2^2 \end{aligned} $$ This gradient formulation is quite similar (up to the weighting) to UniGuide's special case $S=Z$, as it computes an L2 loss on a given conformation ($\{y_i\}$) and does not generalise to arbitrary geometric conditions. In practice, this means that [1] calculates gradients with respect to the surface-defining points of the noisy, generated ligand rather than the condition surface. We consider this suboptimal and have experienced that this formulation can lead to instabilities, as it heavily depends on the quality of the noisy samples and never explicitly defines the surface. (Note that we also had similar issues when guiding the clean data point's conformation $\hat{x}_t$ instead of $x_t$.) Within UniGuide, in contrast, the surface is explicitly defined, and the gradient updates relate directly to the surface. UniGuide's condition map provides guidance targets $c = C(s, \hat{x}_t)$ (see Eq. 20 and 21, manuscript) that are directly associated with the surface, either on the surface or projections into it. For these target conditions $c$, we compute $\nabla\_{x_t} \| \hat{x}_t - c \|_2^2$. This technical difference comes with several advantages: - The separation of the condition from the generated configuration, combined with the subsequent comparison via a direct L2 loss, shifts the focus to appropriately defining the condition. - UniGuide does not require the condition definition to be differentiable, offering greater flexibility. - The condition's explicit nature and the more flexible gradient definition enable us to derive various projections, such as setting a margin for projecting into the surface. This precision is challenging to achieve with a distance-based gradient formulation. Finally, we would like to reemphasise that UniGuide's formulation is versatile and extends to other tasks, including density-based drug discovery (as noted in our response to D4nw). In contrast, validity guidance shares limitations similar to UniGuide's special case where $S = Z$. We hope this explanation provides a clear and satisfactory understanding of the distinction between how UniGuide implements conditions compared to validity guidance. [1] DECOMPDIFF: Diffusion Models with Decomposed Priors for Structure-Based Drug Design

Reviewer YA7u2024-08-13

Thank you for the response. Given the additional results and comparison with more diffusion baselines, I'm still a bit concerned about the effectiveness and necessity of introducing geometric guidance, since it only marginally improves upon the performance of backbone models (EDM, DiffSBDD, etc). The author mentioned that these models have not been specifically engineered for certain tasks, but it seems to me that the point of adapting existing models for downstream tasks is primarily aimed to benefit from them so as to boost the task-specific performance, which UniGuide has yet to achieve. In this regard, I'm inclined to maintain the score.

Authorsrebuttal2024-08-13

Clarification on performance and broader impact of UniGuide

Thank you very much for your response. While we consider UniGuide's performance improvements significant (details below), we want to emphasise that a performance-focused discussion overlooks UniGuide's **broader potential and impact: UniGuide's formulation enables to tackle entirely novel drug discovery tasks where no established baselines or sufficient data exists.** Specifically, our goal is not to adapt existing models to boost their performance in downstream tasks but to provide a generally applicable guidance framework that makes (unconditional) base models useful for various tasks and practical applications. This capability is well demonstrated in our experiments on density-based drug design (no data) and symmetric proteins (limited base model); see Figures in PDF. Our results show that UniGuide is versatile (EDM was applied to both LBDD & FBDD) and effective, consistently delivering performance that matches or exceeds task-specific models (which are limited to a single task and require extra training) and outperforms alternative conditioning mechanisms. This sets the basis for UniGuide's long-term objective to reliably translate novel tasks directly to a generative model by incorporating any newly developed geometric conditions through its condition map, thereby accelerating the overall drug discovery process (for real-world applications). We also want to highlight a selection of our results that demonstrate the consistency and significance of our improvements, and we kindly ask the reviewer to reassess their evaluation in light of this evidence: - **LBDD**: **UniGuide is state-of-the-art** and outperforms alternative guidance mechanisms (cf. follow-up to our general response). An improvement over the base EDM model for LBDD (as indicated by the reviewer) is not possible as **EDM can only be applied for the LBDD tasks with UniGuide** (Table 3, Rebuttal PDF). - **SBDD**: **UniGuide demonstrates superior performance over all evaluated conditioning mechanisms for diffusion models** (cf. follow-up to our general response). Specifically, we improve the base model DiffSBDD in terms VINA score of up to 1 when utilising UniGuide and up to 1.8 when additionally using UniGuide's version of Clash Drift. This results in competitive performance of the base model compared to conditional, task-specific models such as DecompDiff and IPDiff. - **FBDD**: **UniGuide enhances the VINA score by 0.5 compared to the evaluated baseline for general, pocket-conditioned FBDD tasks** and generates more valid and connected ligands (as shown in Table 13 of the manuscript). While the observed improvement initially appeared marginal to the reviewer, we believe further consideration of the experimental evidence and UniGuide's potential for novel applications and model enhancement may offer a different perspective. We appreciate your feedback and hope our explanation has been helpful in clarifying UniGuide's potential impact.

Reviewer YdVq2024-08-13

Thank you for the detailed response and improving the clarity of understanding the method. I believe that the rebuttal adds clarity to the paper. I appreciate the additional results in generating symmetric proteins. However, as other reviewers mentioned, the geometric guidance led to marginal improvements in performance. Further, the additional experiments on rebuttal Table 2 only demonstrate minimal performance improvements. As a result, I maintain my score.

Authorsrebuttal2024-08-13

Thank you and additional clarifications

&nbsp; Thank you for your response and for recognising the additional results we provided as part of our rebuttal. We are glad that our explanations have enhanced the clarity of our paper. Regarding the mention of marginal performance improvements, we have addressed this in detail in our response to YA7u, offering a broader perspective on UniGuide's potential impact and objectives. We kindly refer you to that response and hope it provides further clarity. Specifically for the FBDD experiments, while we agree that Table 2 shows UniGuide's on-par performance with a task-specific model, we would like to emphasise the significantly improved VINA scores by 0.5, as demonstrated in Table 13 of the manuscript. &nbsp; We hope these clarifications highlight UniGuide's full potential. Thank you once again for your valuable feedback and the engaging discussion.

Authorsrebuttal2024-08-13

Thank you

Thank you very much for the kind words and for increasing the score of our paper. We are pleased to hear that we were able to resolve the reviewer's concerns. Best regards, The Authors

Reviewer cfg32024-08-14

Thanks for your response

I carefully check the rebuttal. I am now convinced that Uniguide holds a difference with the direct application of gradient-guided generation and I also appreciate the novelty of geometric condition maps. I increased my scores to appreciate the efforts.

Authorsrebuttal2024-08-14

Thank you for your response

Thank you very much for the kind words; we are grateful for an increased score and are excited about the positive feedback on the novelty of the UniGuide framework. We are pleased to hear that you found our additional clarifications on gradient-guided generation satisfactory, which will be included in the camera-ready version of the paper. Best regards, The Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC