Integrating dMaSIF about holoprot

[vsomnath]

Hi Min Htoo,

Thank you for your kind words, and sorry for the delayed response!

Right now, there are no immediate plans as to integrate dMaSIF into this work.

In my opinion, the following steps are needed to accomplish this integration:

1. 3D surface generation - dMaSIF uses atomic point clouds as inputs to generate surfaces (their SDF function uses atomic radii), but HoloProt directly uses the C_alpha coordinates in the 3D structure, so the generation would have to account for this fact, by modifying the attributes and coordinates stored in the backbone object.

2. Mapping between surface and structure nodes - When using MSMS, one of the returned outputs is the residue each surface node is connected to. To integrate into dMaSIF, one would require an additional tensor that tracks the membership between each atom in the protein and corresponding residue, and uses this to assign the corresponding mapping between surface and structure.

3. dMaSIF convolutions - Right now, the APIs for surface and structure MPNs are consistent and so are the attribute tensors (like x, edge_index, edge_attr etc). This would change a bit with using dMaSIF convolutions, and one would have to account for it. It should also be feasible to create a more hacky solution, but I am usually not in favor of it.

I'll populate my thoughts here if I realize some other issues or considerations with the integration.

Best,
Vignesh

[vsomnath]

Closing this for now, feel free to reopen if necessary!