SaProt with structures

[loodvn]

The current SaProt model was based only on VH sequence and used the model zero-shot scores.

Add SaProt model but using both sequence and predicted structure as input. Could either do the likelihood-based zero-shot predictions or use embeddings.

[apathithyan]

Hey @loodvn, I will be able to take this up. Currently setting up SaProt, should be ready with a PR by Nov. 11.

Just to clarify, the VH and VL structures are in the ABB3 and MOE folders respectively?

[loodvn]

Yep, thanks that's great! I think the MOE structures are probably more accurate so you can just use those. They contain the complex of VH and VL structures, so I'm not sure how SaProt handles multiple chains but feel free to use any approach you think makes sense.
@sritterginkgo is handling all GitHub-related things so he'll be able to coordinate further with you.

[apathithyan]

Thanks for the response,

As per my trials on https://www.rcsb.org/3d-view, I observed that:

• (GDPa1) VH sequences are stored in ./abdev-benchmark/data/structures/AntiBodyBuilder3/GDPa1/{antibody_name.pdb}
• (GDPa1) VL sequences are stored in ./abdev-benchmark/data/structures/MOE_structures/GDPa1/{antibody_name.pdb}

Could you confirm if I'm reading the folder structure correctly? If the complexed VH-VL structures are located elsewhere, I'd gladly revise the implementation to use those instead.

My current approach generates SaProt embeddings from VH and VL (structures+sequences) separately, then concatenates them, similar to ESM2's philosophy.

[sritterginkgo]

Hi,

I don't think that I follow. VL and VH sequences are present in both structure sources, see example below (MOE left, AntiBodyBuilder3 right), VH green, VL magenta. For clarity they do have different chain names, ABB3 uses H/L for heavy/light. MOE uses B/A for heavy/light.

The example uses the provided structures for abagovomab.

[apathithyan]

Thanks for the clarification. You're right, I had incorrectly checked the files earlier.

Will update the PR taking this into account

[apathithyan]

Will be closing this PR to clean up the implementation.

Will submit a corrected version soon with the right usage of the .pdb structures

[sritterginkgo]

Sounds good

[apathithyan]

Have added the corrected version here - #36