Support for protein structures and residue features?

[Garhorne0813]

Hi, thanks for releasing Equiformer V2!

I have two quick questions:

1. Can it handle protein coordinates (e.g., PDB data)?
2. How to add amino acid properties (e.g., hydrophobicity, charge) or precomputed embeddings? Any tips for modifying the code?

Thanks!

[yilunliao]

Hi @Garhorne0813

Thanks for your interest, and my responses are as follows.

1. Yes. EquiFold and FlowPacker are two examples of applying Equiformer series to protein-related tasks.
2. I think you can treat some additional properties as extra type-0 vectors (or scalars). One similar example is how we encode the atom type -- we encode them as long int and then convert to a feature vector with torch.nn.Embedding as here. Depending on whether the properties are discrete (e.g., atom type) or continuous, you might need different ways to convert the properties to features (e.g., discrete can use the same way as atom type and continuous ones can maybe use a linear layer to expand the dimension).

I am not very familiar with protein-related tasks, so feel free to correct my understanding or let me know if something is unclear.