Supplying alternative model weights about colabfold HOT 7 CLOSED

OpenFold weights require a conversion step to work with ColabFold. We were waiting for the OF team to release an updated set of weights, as the initial monomer weights didn't perform as well as hoped.

In the meantime we have integrated the DeepFold weights.

alphafold_finetune looks interesting, I was not aware of this project before.

from colabfold.

Ah, that's unfortunate. I am just gearing up to benchmark their multimer and protein LLM monomer implementations. Hopefully they work well.

Are scripts to enact the weights conversation available? I'm asking mostly because my existing pipelines leverage local ColabFold and it's been by far the easiest to install and most stable solution.

from colabfold.

Openfold provides a script for converting to jax weights:
https://github.com/aqlaboratory/openfold/blob/main/scripts/convert_of_weights_to_jax.py

from colabfold.

Must have missed that. Thank you for highlighting it for me.

from colabfold.

Where the multimer weights released? I might have missed that

from colabfold.

Not the OpenFold trained ones. They are using the "AlphaFold-Multimer v3" weights, but I still want to see how it performs with my systems. But I should be able to try and fine-tune those weights with their training scripts, which was the goal. Unless their weights perform significantly better, of course.

EDIT: They have pushed a multimer branch to main though, and added the protein LLM called Soloseq last month. So there are some major updates their end, but I suspect you've seen that.

from colabfold.

Yes, I've seen that but I hadn't seen any weights.

from colabfold.