Help Wanted: Implementation of various upgrades to Alphafold

Hi my name is Michael, I’m a structural biologist and unfortunately don’t have the coding expertise to fix the issue I would like to solve. In short, last year deepmind came out with a protein prediction software that revolutionised the structural biology field. The biggest limitation of this platform is the number of residues that it can calculate (currently ~2000aa), as I work on a large protein complex every residue we can get in the model is super helpful. To help this limitation a number of groups have produced various ways to optimise the platform and I would like to integrate them into an easy to use Colab like this one:

and here is the list of papers that have provided optimisations:

Protein complex prediction with AlphaFold-Multimer | bioRxiv (Source code for the latest version of alphafold specific to protein complexes)

To start with if someone could give me an idea of the feasibility of integrating these adaptations that would be excellent. Secondly if someone would be keen to either create or help create such a colab notebook that would be even more excellent.

Cheers
Michael Healy
Michael.healy@uq.edu.au

3 Likes

Hi there

I do remember this has been a pretty big deal in the whole field. Quite a bit of shakeup there. I’m very much interested in pursuing this idea, but it’ll most likely have to wait until the end of the course.
Let’s get back on this once this course ends.

2 Likes

Wow, that sounds like a real challenge! I guess Alpha-Fold has a rather involved training process, and also not that simple architecture (some kind of transformer?)

However, I guess that you can also check HF Transformers hub to see if any other variant of Alpha-Fold is available.

Still, would be interesting to see how this idea evolves!