search

THUNLP-MT / dyMEAN

This repo contains the codes for our paper "End-to-End Full-Atom Antibody Design"
https://arxiv.org/abs/2302.00203
MIT License
94 stars 11 forks source link

The complex and Hard pipline #24

Open Cosmos9526 opened 3 months ago

Cosmos9526 commented 3 months ago

Dear researcher thanks for sharing this pipline, but you try to design Antibody with a lot of package, but I believe it can be easier than you made it pipeline, You try to use for example IgFold and Rosseta for what ? while you can easily put antigen as a input for diffab .... best wish

kxz18 commented 3 months ago

Hi, IgFold and Rosetta are optional. They are only included as pipeline-based baselines in our paper. Our end-to-end model can be used with an antigen and sequence of antibody frameworks. Please see our instructions here.

Cosmos9526 commented 3 months ago

Thank you for your response. After generating multiple CDR samples, how do you choose the best model? Is there any scoring method for Antigen to CDR docking, such as using Prodigy?

Additionally, once the best model is selected, how do you coordinate the construction of the full atomic structure of the antibody (considering both human and mouse models)?

kxz18 commented 3 months ago

We didn't explore the scoring paradigm much in this paper. But I think common tools like rosetta and foldx might serve the purpose of ranking the samples, though the actuall effects remained unknown.

The full-atom structure is simultaneously generated in our end-to-end model (i.e. the sequence and the full-atom structure are jointly designed). There is no need to do further side-chain packing or something like that. However, openmm relax might still be needed for solve clashes and physically invalid local geometries (e.g. bond lengths), which we have also integrated in our inference API.

Cosmos9526 commented 3 months ago

I’ve reviewed over 15 antibody construction packages, but each of them had issues that made them unusable. The package you constructed is much better than the others, but it requires a considerable amount of time to complete the full structure. best wish

kxz18 commented 3 months ago

Thanks for the appreciation! I think the most time is spent on openmm relax. Current algorithms are still struggling on generating physically valid local geometries, though the approximate spread of the side chains might be correct, therefore openmm relax might be necessary. I'm also delving into this problem and upgrading generative algorithms. Wish that in the future the deep generative models can directly generate structures that do not have obvious problems at the first glance.

Best wishes