Full-Atom Peptide Design with Geometric Latent Diffusion

Summary

The paper presents a new diffusion model for generating peptide binders given protein pockets. It also presents a new benchmark dataset, created by selecting examples from PDB and ensuring sequence dissimilarity between training and test data.

Strengths

Thank you for constructing a benchmark dataset for this task with a proper train/test split. This is overlooked in some related work. Is the dataset available to download? Notation is clearly set out and there is a good survey of related work. There are sensible ablations and comparisons to baselines, and the presented model performs well relative to these.

Weaknesses

In the introduction and abstract, please state clearly that PepGLAD needs to be given the binding site (not the whole protein structure). Line 24 is hard to understand: ‘The key of peptide design …’. It seems to say that strong binding requires the peptide to adopt a compact, inflexible shape, but I do not understand why. The argument for the affine transformation is unclear. I could not decipher line 48 ‘These variances define divergent target distributions of Gaussian…’. Plenty of methods (AlphaFold, RFDiffusion, GeoDiff, …) are able to predict or design elongated protein and molecule shapes without any such transform, and here it should be easy because the binding site shape is given as a conditioning input. Surely a model can learn that the binder it generates should approximately match the shape of the binding site, even if the binding site is far from spherical? I think that proposition 3.1 could be stronger. Please check, but I think that if $x$ is rotated, then $L$ undergoes the same rotation, and hence $F_g(g.x)=F(x)$. Thus, $f(h, F(x))$ is invariant and $F^{-1}\overrightarrow{f}(h, F(x))$ is equivariant even if $f$ and $\overrightarrow{f}$ are not. I am surprised that this transformation improves performance. The score model in the transformed space now has to learn that bonds in different orientations should have different lengths, inferring from the inter-residue distances of the binding site how much each dimension is compressed or stretched.

Questions

Why do you need to generate all-atom designs? Do you think it could work just as well to generate C-alpha coordinates and discrete residue types? Mixed discrete/continuous diffusion models have shown some success for small-molecule drug design (DiffSBDD) and crystal design (MatterGen). Line 223: does ‘filter out’ mean ‘keep’ or ‘discard’? The dataset seems small for deep learning and for the complexity of the task. How could it be augmented? Table 3: what counts as a good RMSD, for practical purposes? Are all the methods given the same inputs, or are some doing blind docking? Equation (3): how is the MSE on full-atom structures defined when the residue type is wrong?

Rating

6

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

Like all peptide design in silico, this is hamstrung by lack of good in silico evaluation: dG values from Rosetta are not reliable estimates of real binding affinity. Ideally there would be some wet-lab evaluation of the generated designs. The method requires binding site to be identified before peptide binders are designed.

Summary

This paper explores the structure-based peptide design problem. A benchmark on this task and a powerful diffusion-based model for full-atom peptide design, named PepGLAD, are proposed. PepGLAD explores the geometric latent diffusion, where the sequence and the full-atom structure are jointly encoded by a variational autoencoder. The proposed method outperforms the existing models on sequence-structure co-design and binding conformation generation.

Strengths

1. This paper tackles an under-explored problem and constructs a new benchmark which is beneficial for this field. 2. The novel Geometric Latent Space can maintain the advantages of Latent Diffusion for efficiency and normalization and Geometric Diffusion for explicit interaction modeling. 3. PepGLAD can achieve full-atom generation compared with backbone atom generation in previous methods, which is truly an improvement. 4. PepGLAD utilizes a novel Receptor-Specific Affine Transformation to map the pockect-specific distribution to the standard distribution

Weaknesses

1. This paper failed to further explore the explicit interaction modeling in geometric latent space. I think it’s the key advantage compared to standard latent space (without the geometric latent). 2. The introduction to Receptor-Specific Affine Transformation can be more clear. I kindly suggest the authors can add one sentence or two summarizing the first paragraph in section 3.3 into the section 1 of Introduction. In such way readers may be more clear that this Affine Transformation is applied to the peptides and the pockets (i.e. binding site distribution in the introduction). Also the point 3 in line 62-65 has a mistake: according to the method section the affine transformation F is mapping binding site distribution (of peptides) into standard Gaussian (not the reverse) 3. The conditioning method of pockets is not explored. The denoiser takes the geometric latents of peptides and the coordinates of pockets as input, but the latents are coordinates are different levels of abstraction and this may cause some underlying issues.

Questions

1. What’s the motivation for using Latent Diffusion instead of standard diffusion? Just because it’s better in image generation domain? 2. How to incorporate geometric pocket condition (e.g. coordinates) into geometric latent diffusion process? 3. Is the geometric latent are just the Ca atoms of each residues? Will using all four backbone atoms as latent improve the perforemance？ 4. Is the adaptive multi-channel equivariant encoder in dyMEAN a scalarization-based equivariant GNNs? It’s better to add some explainations in this paper. 5. Does this Receptor-Specific Affine Transformation beat simple gaussian normalization, common trick in Structure-based Drug Design field?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors didn’t discuss about the limitations of this method. This paper failed to further explore the explicit interaction modeling in geometric latent space. I think it’s the key advantage compared to standard latent space (without the geometric latent).

Summary

This paper introduces a novel latent diffusion model, Peptide design with Geometric Latent Diffusion (PepGLAD), for the task of peptide design. The authors propose an affine transformation to project the raw Euclidean space into a standardized one, ensuring physical symmetry. Overall, I think it is a good paper, with solid experimental results and novelty in methods.

Strengths

This paper proposes a latent diffusion model, Peptite design with Geometric LAtent Diffusion (PepGLAD) , applied to the peptide design task. This paper also proposes an an affine transformation to project the raw Euclidean space into a standardized one, so that the physics symmetry can be ensured. The main contributions of PepGLAD over GeoLDM seem to be its application to full-atom peptide generation and the introduction of a receptor-specific affine transformation. While GeoLDM¹¹ was designed for 3D molecule generation, PepGLAD extends this to the more complex scenario of full-atom peptide generation. The receptor-specific affine transformation is a significant contribution, as it projects the geometry to approximately N(0,I) derived from the binding site, which appears to be a novel technique in the field of latent diffusion models for molecules and proteins.

Weaknesses

There are several doubts on method contributions, datasets, as shown in Questions.

Questions

1. What are the key differentiating factors of PepGLAD compared to GeoLDM [1]? While I recognize the shift from latent molecular generation to full-atom peptide generation, are there any specific model design elements (apart from the affine transformation) that set it apart from GeoLDM? Furthermore, does the model incorporate any strategies to reduce the costs associated with generating full-atom peptide scenarios? 2. The most significant contribution made by the authors in this paper is the Receptor-Specific Affine Transformation. This transformation projects the geometry to approximately N(0,I), derived from the binding site. This approach appears to be the first of its kind applied to latent diffusion models for molecules and proteins. I am curious if there are any other works that have used this technique to ensure equivariance in their generation process, such as frame-averaging [6] [7]. 3. Compared to previously established databases [2] [3], the datasets utilized in your study appear relatively modest in size. Could you comment on the unique advantages of your proposed datasets over these larger ones? Furthermore, have you considered applying your model to existing datasets to reinforce the validity of your results? 4. I suggest to add more citations on recently proposed related works on peptide design [3] [4] [5]. [1] M Xu, et al. Geometric Latent Diffusion Models for 3D Molecule Generation [2] Z Wen, et al. Pepbdb: a comprehensive structural database of biological peptide-protein interactions. [3] L Lin, et al. PPFlow: Target-Aware Peptide Design with Torsional Flow Matching [4] Osama Abdin, et al. PepFlow: direct conformational sampling from peptide energy landscapes through hypernetwork-conditioned diffusion [5] Colin A Grambow, et al. RINGER: Conformer Ensemble Generation of Macrocyclic Peptides with Sequence-Conditioned Internal Coordinate Diffusion [6] W Jing, et al. DSMBind: SE(3) denoising score matching for unsupervised binding energy prediction and nanobody design [7] Omri Puny, et al. Frame Averaging for Invariant and Equivariant Network Design

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

See Appendix. J Limitations

Summary

The authors proposed PepGLAD, a latent diffusion model for full-atom peptide design. This work mainly addressed two challenges for peptide design, (1) full-atom modeling, and (2) diverse binding geometry. Specifically, a variational auto-encoder was trained to learn latent representations of protein-peptide interactions. Then, a latent diffusion model was employed to capture the underlying distribution to sample binding site-conditioned peptide conformations. The authors curated a benchmark dataset to evaluate the model performance, and compared with existing models to showcase its effectiveness.

Strengths

- Full-atom modeling is significant to better represent protein-protein/peptide/ligand interactions. This work goes beyond backbone-only modeling and realized full-atom level generation. - Employing latent diffusion with a pre-trained VAE is a reasonable choice, which has recently become prevalent for protein generative modeling. - Provided a new benchmark dataset with careful curation. Proposed novel evaluation metrics and discussed limitations of existing metrics (e.g., AAR). - Under the proposed evaluation metrics, PepGLAD outperforms existing models.

Weaknesses

- Dataset is small (6105 non-redundant complexes). Data augmentation attempts (ProtFrag) do not seem to help (Table 4 ablation results). Although sequence-based clustering has been utilized to prevent data leakage, training/testing VAE and diffusion model on 6105/93 data points make the results less convincing, *especially considering the enormous design space for sequence-structure co-design tasks*. - Guidance with an additional classifier (Sec 3.5) makes the model more complicated (which seems a bit ad hoc to me). - From the results shown in Table1, it seems that the model is able to generate a variety of conformations (high Div.) which are reasonable ($\Delta G < 0$). Given the extremely limited size of dataset, I am not sure if the model can properly capture the *conformation space* at all. In other words, I am a bit skeptical about the proposed evaluation metrics.

Questions

- L#130. For each residue, the initial coordinates are initialized with the same $\vec{\boldsymbol{z}}_i$. How significant are these initial vectors? Can we use an SE(3)-invariant latent space and a regression model to generate atomic coordinates (similar to AlphaFold StructureModule, conditioned on the sampled residue type)? - Classifier guidance seems ad hoc and brings more complication to modeling. Would an architecture similar to that used in [ESM3](https://www.biorxiv.org/content/10.1101/2024.07.01.600583v1) or [PVQD](https://www.biorxiv.org/content/10.1101/2023.11.18.567666v1) help? - Table 1: out of 40 samples generated by PepGLAD, there are about 20 different conformations (Div.) and 20 valid peptides (Success). How should the readers interpret this? Peptides exhibit high structural flexibility upon binding? - Did I miss sequence recovery metrics in the model benchmark section (excluding Consistency, which only models the joint distribution)?

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

N/A

Reply

I am glad that my concerns are all address by the authors. And I will lift my rating.

Thank you for the clarifications

Thank you for the clarifications. Regarding proposition 3.1: you're absolutely right. Sorry, and thank you for your patience. One further detail I could not find in the paper is the exact definition of the binding site. Is this a potential source of data leakage? I am wondering if the binding site is defined in terms of the ground-truth pose of the same peptide. I keep my score as-is.

Thanks for your rebuttal!

I have read all of the author rebuttal, which addressed most of my concerns. Thus I will raise my score. Thanks for the detailed response.

Response to Rebuttal

Dear Authors, Thanks for your response. Most of my questions have been properly addressed. Although some limitations still remain, e.g., dataset size and benchmark metrics, I think researchers can refer to this work for some insights and know-how to solve their domain specific challenges. To that end, I have raised my rating and support the acceptance of this manuscript. Best,