saro00 / dgn Goto Github PK

Dear authors,

Very nice work, came to know of this through Dominique's excellent talk recently at Valence Discovery's M2D2 series.

Since DGN exhibits SOTA performance over many common GNNs, I am excited to try it out on some biological tasks. However, I am already using another open-source model that relies on torch_geometric's Data interface (to batch graphs and so on), so I believe I also need to port this to torch_geometric to make it compatible as the task I'm working on requires me to sample data points simultaneously in each iteration.

Admittedly, I am really not familiar with torch_geometric, so I was wondering if you happened to have a torch_geometric version, or if you have some pointers on the easiest way to port this model to torch_geometric.

To clarify, the input will be the molecular identity (such as a SMILES string), which need to be pre-processed to the input features needed by DGN, such as the eigenvectors, the adjacency matrices and so on.

Thank you very much!

Thanks for sharing the codes. Could you please provide the best hyperparameters to reproduce the results of PNA? The original repo of PNA did not provide the hyperparameters for PATTERN.

Thanks.

The paper mentions that the first k non-zero eigenvalues of the laplacian are used for Phi(k). However, in the code, I don't see filtering the zero-eigenvalues out by setting a low threshold of 1e-6 for example. Phi(1) contains the lowest non-zero eigenvalue and Phi(k) contains the lowest K non-zero eigenvalues, right, according to the Figure 1? Here

In that case, why aren't the eigenvalues with low mangitude, eigenvalue < e-6, filtered out in get_k_lowest_eig? Here. To get the lowest k non-zero eigenvalues. If the eigenvalue is less than 1e-6 they could be considered as the zero eigenvalues and hence filtered out?

Would be great if you could share your thoughts on this