google-deepmind / alphafold

Open source code for AlphaFold.
Apache License 2.0
12.29k stars 2.2k forks source link

Generating multiple (>10) models from one sequence #664

Open cgovaert opened 1 year ago

cgovaert commented 1 year ago

Hi all

I'm new to the field. I would like to generate between 100 and 50 models for my target sequence to explore conformal flexibility of a specific region. This region is often (comparing prediction on my target and various orthologs) predicted with lower confidence, and sometimes with different conformation. We have experimental evidence that this domain may be mobile.

Thus I want to generate many models from the same target sequence to identify possible conformational diversity.

We can run Alphafold locally but I'm not sure how to paramtyrize such a run.

Can anybody help here ?

Thank a bunch Cedric

— Prof. Cedric Govaerts, Ph.D. Universite Libre de Bruxelles Campus Plaine. Phone :+32 2 650 53 77 Building BC, Room 1C4 203 Boulevard du Triomphe, Acces 2 1050 Brussels Belgium

tcoates5 commented 1 year ago

While you certainly can do this with AlphaFold, you will probably only explore "near-native" conformations with it, because of what it is trained to produce (crystal structures). If you want to explore conformational space more thoroughly, I would recommend using molecular dynamics. If you would like to use AlphaFold to investigate where the flexible region likely is, some studies have indicated that segments with lower pLDDT scores tend to be more flexible, and at least with 25 structures (typical output for multimers), this has matched where the variation occurs for me. For doing this with AlphaFold, the easiest way would be to run in multimer mode (which, according to the latest update notes, is best to do even for monomers if you know the stoichiometry) and change the num_multimer_predictions_per_model value. As there are 5 models, pass the number of structures you want / 5 as the value to this flag in the command line. In other words, for 100 output structures, add --num_multimer_predictions_per_model=20 to your command line arguments.

cgovaert commented 1 year ago

Great idea. will do. To be more precise, I do want to run MD, I want to see if there is/are other conformation of a subdomain that we believe changes upon ligand binding. Thanks