Generate structures from intermediate layers for OpenFold

[girip]

Overview
OpenFold's structure module (similar to AlphaFold structure module) could be used to generate structures from the intermediate layers. We aim to offer a flexible approach for viewing these structures during inference.

High-Level Goals

• Develop functions to generate structures from representations using the structure prediction module from OpenFold.
• Integrate the structure prediction approaches within the web interface.
• Use structure evaluation methods to provide measures for comparing the generated structure with the reference structure.

Deliverables

• Test functions for generating structure from intermediate representations
• Web prototype for user selecting a layer to generate a structure using the structure module. Allow for changing the structure module to different networks.
• Impliment TM-score, lDDT measures for evaluation of structure.

[AlexisF101]

I am interested in working on this issue. I will develop a test script to extract the representation from the layers inside the Evoformer with Python and PyTorch. Then I will input those to create structures in the structure module. I can develop the web interface through Streamlit.

[archbilesherman]

I am interested on working on this issue and my plan is to extract intermediate representations from the Evoformer using Pytorch hooks and feed them into the StuctureModule to generate structures. Then I'll build a simple Streamlit interface that lets a user select a layer, visualize the structure from that layer, and compare it with the reference.

[ssagili3]

I’d like to work on this issue. I plan to write a test script in Python/PyTorch to extract Evoformer representations and feed them into the structure module to generate structures. I also like the idea others mentioned about using Streamlit as the interface, and I’d be happy to build on that to make the outputs easy to explore.

[sherrylicodes]

Hi, I’d love to contribute on the evaluation and visualization components. I can implement TM-score and lDDT metrics for comparing generated vs. reference structures, and add these into the Streamlit interface so users can see quantitative feedback for each layer’s prediction. I can also help create a reproducible CLI/Notebook pipeline that automates extraction → structure generation → evaluation.