Hi Xu,

First of all thanks for your nice work. I've read your paper, and I have some questions on the proof of the equivariance of the transition kernel. In detail, suppose $\mathcal{C}^t$ is roto-translation invariant, and (thus) $\mu_{\theta}(\mathcal{C}^t, \mathcal{G}, t)$ is roto-translation equivariant with desgined GNN, we need to prove that $p(\mathcal{C}^{t-1} | \mathcal{C}^{t}, \mathcal{G}, t)$ is equivariant. I wonder if it is due to the following derivation:

$$\begin{aligned} p(R \mathcal{C}^{t-1} + g | R \mathcal{C}^{t} + g, \mathcal{G}, t) &= \mathcal{N}(R\mathcal{C}^{t-1} + g; \boldsymbol{\mu}{\theta}(R\mathcal{C}^t + g, \mathcal{G}, t), \sigma_t^2 \mathbf{I}) \ &= \frac{1}{(2 \pi)^{\frac{p}{2}}|\boldsymbol{\Sigma}|^{-\frac{1}{2}}} e^{-\frac{1}{2}(R(\mathcal{C}^{t-1}-\boldsymbol{\mu}\theta))^T \boldsymbol{\Sigma}^{-1}(R(\mathcal{C}^{t-1}-\boldsymbol{\mu}\theta))} \ &= \frac{1}{(2 \pi)^{\frac{p}{2}}|\boldsymbol{\Sigma}|^{-\frac{1}{2}}} e^{-\frac{1}{2}(\mathcal{C}^{t-1}-\boldsymbol{\mu}\theta)^T \boldsymbol{\Sigma}^{-1}(\mathcal{C}^{t-1}-\boldsymbol{\mu}_\theta)} \end{aligned}$$

where $\boldsymbol{\Sigma} = \sigma_t^2 \mathbf{I}$. I am not sure if it's correct, hope to receive your clarification. Thanks.

Could you please provide the data process script since I found that ConfGF's code does not get the features like num_graphs, atom_type, pos, bond_index , etc.?

Hi authors,

Thank you for sharing codes of your great work! According to your paper, GeoDiff has two different versions including GeoDiff-A and GeoDiff-C. But it seems only the codes of GeoDiff-C are provided. In addition, it is mentioned in the paper that graph field network (GFN) is used but I cannot find the codes of GFN, while SchNet seems to play the role of GFN in current implementations.

I am wondering whether the implementations of GFN and GeoDiff-A will be provided in the future? Thanks!

Best,

Hi, thanks for sharing this very good work!

I found the google drive link of the preprocessed dataset can not use, maybe the permission issue, could you have a double check? Thanks

Hi Minkai Xu,

I think there is a flaw in your mathematical notation for GNN, Equation 5 and 6 in particular. I found that the notations m_{ij} in equation 5 and 6 are not the same. m_{ij} in equation 5 and m_{ij} in equation 6

Dear Minkai,

I am sorry to bother you. It seems like that you did not use time step t as the parameter in the diffusion training process. Why did not you use beta or alpha with time step t in the training process?

Hi, Xu. Thanks for sharing the code. I've noticed the discussion here (https://github.com/MinkaiXu/GeoDiff/issues/11) and carefully read the code line by line. Just as what you stated in the issue 11, the "diffusion" process in the code is actually rescaled compared to the paper, i.e., $\mathcal{C}^t = \frac{1}{\sqrt{\alpha_t}}(\sqrt{\alpha_t}C^0 + \sqrt{1-\alpha_t}\epsilon)$. Based on the paper ScoreSDE (https://arxiv.org/abs/2011.13456), DDPM is a variance preserving process and DSM is a variance exploding one. I think maybe there might be some typos in your answer to issue 11 which cause contradiction: "2) use the alpha to rescale the data to achieve variation preserving" and "the problem of variation preserving is: it will change the scale of coordinates". In my perspective, after rescaling, $\mathcal{C}^t = C^0 + \frac{\sqrt{1-\alpha_t}}{\sqrt{\alpha_t}}\epsilon$ is a DSM process with variance increasing along with $t$. So I am confused about why this rescaling method will hold the scale of coordinates since in my view it seems to corrupt the scale (increase the variance) instead.

Hi Minkai Xu,

What's the motivation for that you designed two separate architectures to learn local and global structures? In loss, the loss is divided into local loss and global loss (node_eq_global - target_pos_global)**2 + (node_eq_local - target_pos_local)**2

Hi, Minkai @MinkaiXu. Rencently I am trying to do some experiments based on GeoDiff, and having some problems. Frist, I re-train the model with your original hyper-parameters. I notice that after around 800k iterations, the loss hardly decreases but oscillates wildly. I don't know whether the model truly benefit from the training in the following 2m iterations, and why the loss oscillates wildly? Second, I am trying to sample some conformations from the checkpoints getting in my training procedure, but something very strange happened. Most of the molecules during sampling occur FloatingPointError and retry with local clipping. Then I use eval_covmat.py to evaluate the quailty of conformations, I get 0.00 Cov value and thousands of Mat value. This strange thing happens all the time except one test with 700000.pt to generate molecules from 800 to 1000 (and I repeat the test again with same parameters and the strange thing occurs) I don't know why it happens because I run your code without any modification, and I wonder why you design the local clip operation and how to choose the corresponding hyper-parameter? In the end, I would very appreciate it if you could give me some explanation and suggestions.

Hi! Thanks for the nice paper. In the paper you mentioned protein conformation is a difficult task with this model. Can you explain why linearity of proteins can be an issue? What could be some improvements?

The function found here uses the full variable config which is not passed in for the type gaussian. I believe it should be cfg, but I guess it is never used.