Request for missing code

[MarissaRao]

Fantastic work!

However, I’m a bit confused about the GNN encoding and 3D feature aspects of your project.

Although GNN encoding is referenced in your article, I couldn’t find the corresponding code in your project or training scripts.

What’s more, I noticed in your article that you compared 3D features with 2D features (like MAT-2D and MAT-3D), but there are only specific Python files for generating 2D-MAT or 2D-UniMol features. How can I obtain the 3D molecular features?

Do you happen to have any additional code that incorporates GNN encoding and 3D feature during the training process? If so, I would greatly appreciate it if you could share it.

Thank you so much!

[WillHua127]

ReactZyme/process_mat.py

Line 57 in c4d1554

# AllChem.EmbedMolecule(mol, maxAttempts=500)

This line allows you to generate 3d conformers using MAT. As for UniMol, you may refer to their official repository to generate 3D conformers.

[MarissaRao]

Thank you for your response.
And could you please explain how to incorporate GNN encoding during the training process? I would greatly appreciate any guidance you could provide.
Looking forward to your reply. Thank you once again!

[WillHua127]

You can either follow simple graph convolution, where you propagate features before doing the convolution. Or you can use pre-trained EGNN to encode molecules (like using ESM to encode proteins), there are multiple ones available on Github. Or you can pre-train GNN on your own, and use it for encoding.