Receptor-Specific Diffusion Model: Towards Generating Protein-Protein Structures with Customized Perturbing and Sampling

• The paper would benefit from a clearer explanation of why an Equivariant Graph Neural Network (EGNN) is particularly well-suited for addressing the protein-protein docking problem. Providing more background on EGNN’s advantages in capturing molecular interactions could better highlight its relevance and strengthen the motivation for its inclusion in this context.
• The process of identifying binding sites and integrating this information into the diffusion model is not fully detailed. It would be more helpful to provide a more comprehensive description of the methods used to determine binding sites and their role in the generation process.

The key aspects of the diffusion model method are questionable and confusing:

1, For the “receptor-specific” section: The paper achieves this "receptor" specification by adding the mean position of the binding pocket to the mean of the Gaussian in the prior distribution and gradually decreasing its magnitude during sampling. To me, this approach seems like it is steering the center of the sampled ligand protein to remain in the pocket center. However, in most molecular protein docking and protein-protein docking problems, this can be done by simply removing the Center of Mass (CoM) for the ligand during sampling, which makes it confusing why it is necessary to complicate the process by adding this mean into the prior distribution and modifying the training and sampling processes. Also, Figure 2 shows that incorporating a personalized mean into the sampling distribution reduces the RMSD of alpha carbon considerably. However, I did not see any CoM removal in the training algorithm, so the difference between RSDM and w/o PSD groups in RMSD could simply be because the mean of the ligand is not re-centered.

2, For “step-by-step data purification”: In Equation 8, $\mathcal{N}(x_{t-1}|\mu_{\theta}(x_t^{(l)},t)-\frac{\gamma_t}{T}x^{(r)},\Sigma_{\theta}(x_t^{(l)},t))$ indicates that $x_{t-1}$ is sampled from this parameterized normal distribution. However, prior to this equation, the authors mention that $x_{t-1}$ is directly predicted by $f_{\theta}(x_t^{(l)},t)$. These two expressions on how to obtain $x_{t-1}$ are contradictory; if you directly predict $x_{t-1}$ from the previous time step, how could it be equivalent to sampling from a parameterized distribution as described in Equation 8? There is no sampling involved if you are directly predicting. I would appreciate if the authors could clarify whether they used sampling to obtain $x_{t-1}$ or if it is deterministically obtained by network prediction.

Furthermore, the authors introduce this “step-by-step data purification” by starting with the criticism, “Such the reverse process (predict $x_0$ with the score network) poses a challenge to the model’s predictive ability and complicates the training process.” I don’t see why this challenges the model’s predictive ability. In diffusion or score-based models, predicting $x_0$, $\epsilon_t$, or $v_t = \alpha_t x_0 + \sigma_t \epsilon$ are three commonly used spaces to parameterize diffusion models. If you directly predict $x_{t-1}$ from $x_{t}$, then your network is no longer directly or indirectly parameterizing the score $\nabla_{x_t} \log P(x_t)$; it is directly predicting a sample instead of the gradient of the probability distribution, so it is no longer a score-based or diffusion model.

Taking a step back, regardless of the confusion discussed above, even if the paper claims that directly predicting $x_t$ is better than $x_0$, I did not see any ablation study showing that predicting $x_t$ improves performance over predicting $x_0$, given the authors' criticism that predicting $x_0$ challenges the model’s predictive ability.

• The idea of shifting the mean in the personalized sampling distribution has been explored in prior works, such as DecompDiff [1], potentially making the technical contribution a bit weak.
• Missing important baselines, e.g., RFdiffusion [2].

[1] Guan, Jiaqi, et al. "DecompDiff: diffusion models with decomposed priors for structure-based drug design." arXiv preprint arXiv:2403.07902 (2024).

[2] Watson, Joseph L., et al. "De novo design of protein structure and function with RFdiffusion." Nature 620.7976 (2023): 1089-1100.

The presentation of this work could be improved as it contains some typographical errors; for instance, 'sets' on line 513 may be a typo.

The method itself is relatively straightforward, with similar approaches employed in DecompDiff [1]. Informative priors are used to refine the diffusion and reverse processes, enhancing the quality of generated samples, although some important references are missing.

Furthermore, since the binding site is unknown during inference, it would be beneficial to investigate how the quality of the predicted binding site affects performance.

The proposed method underperforms on antibody structure prediction compared to several baselines in terms of IRMSD.

The generalizability of this approach could be further validated by extending the framework to protein-ligand (small molecule) complex structure prediction tasks.

Reference:

[1] Guan, J., Zhou, X., Yang, Y., Bao, Y., Peng, J., Ma, J., Liu, Q., Wang, L. and Gu, Q., 2024. DecompDiff: diffusion models with decomposed priors for structure-based drug design. ICML 2023.