Surface-based Peptide Design with Multi-modal Flow Matching

1. The technical contribution is limited as PepFlow and PpFlow already apply multi-modality flow matching to peptide generation.
2. The description of how to model and leverage surface property is quite limited. Therefore, the reviewer fails to observe the key difference with previous methods.
3. The authors should put the ablation studies into the main paper to see which component contribute to the performance improvement.

• For surface generation, SurfFlow seems to combine several existing flow-matching algorithms.
• Ablation studies are needed to verify the design choice. The main contribution is to incorporate surface. So, what's the difference between other atom-based models? Have you tried other ways to process the surface?
• Why co-design methods always underperform RFDiffusion on Energy and Design. Can you think of ways to improve on this?

Insufficient Detail on Peptide Amino Acid Determination

1. The process of obtaining the final peptide amino acids requires more detailed explanation.
2. The method for determining peptide sequence length is not clearly described.

Lack of Computational Efficiency Comparison

1. The paper does not provide a comparison of overall processing times between methods.
2. Surface feature calculation may introduce significant computational overhead. A time comparison with methods like RFDiffusion, which doesn't require such CPU-intensive workloads, would better illustrate the trade-offs.

Surface Consistency Concerns

1. The model seems to generate surface positions and corresponding physico-chemical features simultaneously. While structure consistency is considered in the metrics, it's unclear how the generated surface of generated protein-peptide structures is similar with the calculated surface of the generated protein-peptide structures.
2. If significant discrepancies exist, the necessity of surface generation in the model should be justified.

Limited Ablation Study

1. The purpose of the ablation study (whether it focuses on generation loss or input features) is not clearly specified.
2. If the impact of surface generation is minimal, the primary contribution of this research may be limited to the addition of surface features.

Recovery vs Stability

1. Because surface is mostly determined the functional groups, sequence recovery would be higher and corresponding geometric metrics would be also increased. Beside geometric features, the model is not pretty strong compared to RFDiffusion, also, unstable peptides are useless because peptides are already unstable, so, lower stability arises the questions of use on real world discovery.

1. The key novelty of this work, beyond the usage of surface should be pointed out in the introduction. A few sections (such as Euclidean CFN for postion and physiochemical features susection - more on this to follow) appear to be a straight-foward adaptation of earlier works. Hence, the contributions of this work may be missed by the readers if not explicity pointed out. I had to extensively read through the appendix to appreciate the technical contributions of this work as it was not very evident from the main text. For instance how is the physiochemical features different from Campbell et. al 2024?
2. L$_{con}$ in eq 17 should be clarified. It is not mentioned explicity in the paper.
3. Which physiochemical propeties for surface are used should be mentioned in main text or the reference to it in appendix should be provided.
4. A key consideration of protein generative models are their training and sampling time performance. It would be beneficial to see wall-times of the proposed method and competing state-of-the-art methods.
5. Since the key contribution is the addition of peptide surface, justifying its usage is critical and to back it up with experimental results on its utility is a must. I couldn't find experiments to ablate all peptide surface component in its entirety and fair comparision to other seq-structure codesign methods. To my understanding, the ablation results in appendix remove 1 sub-component within "surface design" component but not the entire "surface design" component.
6. Minor:

• L59; lock-and-key and induced fit can use references.
• details of v$^{pos}$, v$^{ori}$, v$^{cat}$ networks should referenced in main text as these annotations are not directly used in the appendix.

My primary concern is the lack of biological and chemical motivation behind many of the design choices:

• The central novelty of the paper seems to be the consideration of the peptide surface in the generation process. However, the surface of a peptide is determined entirely by its amino acid sequence. Therefore, what is the motivation behind jointly generating both the amino acid sequence and surface shape? It is not possible to modify the surface shape of a peptide beyond what is determined by the amino acid sequence, so what is the purpose of generating a surface if it cannot be used?
• Similarly to above, the physicochemical properties of the surface are entirely determined by the amino acid sequence. Therefore, why model the amino acid sequence and physicochemical properties separately? Additionally, the surface properties of all amino acids are already known (such as charge, hydrophobicity, etc.), so why generate these properties instead of just using their known values?
• What is the meaning of the vector orientations on the surface? The “surface” of a peptide is really just an arbitrary boundary determined by electron density, and so there is no such thing as a directionality to the surface. While the authors distinguish between “concave” and “convex” parts of the surface, I’m not sure how that is relevant or worth modeling. And again, it is impossible for a biologist to somehow synthesize a peptide with a different surface orientation than what is determined by the amino acid sequence.
• I disagree that the physicochemical characteristics considered in the paper should be represented with categorical values. Specifically, the authors treat hydrogen bonding as a categorical variable (acceptor/donor/neutral), but I think it would be more suitable to consider hydrogen bonding potential as a continuous variable. Hydrogen bonds are a special type of dipole-dipole interaction, and even among hydrogen bond donors there can be different strengths. Therefore, why not consider hydrogen bonding as a continuous variable?

Taken together, it is hard for me to understand why the proposed model achieves superior results compared to models that only consider the sequence. If the authors successfully motivate their design choices from a biochemical viewpoint, I would be happy to raise my score. However, as it stands, it is hard to see how any output of the model besides the amino acid sequence is useful for biologists. Why not just model the amino acid sequence only?