[chimerax-users] alphafold predict multiuser syntax

[Tom Goddard]

Hi Jerry,

  It is unlikely you will be able to get AlphaFold to predict multimers composed of multiple 1000 amino acid proteins even with the most powerful GPUs available.

  The EBI AlphaFold database produced by Google predicted structures of up to 2700 amino acids.  From the paper that was done by combining the memory of 4 GPUs (which cannot be done on Colab), that might be 64 Gbytes.   The GPU memory use of AlphaFold is said to be quadratic in number of residues.  When I last tested with Google Colab GPUs with 16 Gbytes a few months ago it failed with 1000 amino acids.  But that was AlphaFold 2.0 and now we are using AlphaFold 2.1.  It would require testing, but it is very likely that 1500 total amino acids in a complex will run out of memory, and quite possibly 1000 will run out of memory when your GPU has only 16 Gbytes.

  This is one of the limitations of AlphaFold, and of the old Google Colab GPUs with only 16 Gbytes.

  The AlphaFold database has predictions for human proteins larger than 2700 amino acids made by breaking into 1400 amino acid chunks and computed at intervals of 200 amino acids.  Then you can attempt to piece together those chunks.  I give an example of that piecing together in ChimeraX here

	https://rbvi.github.io/chimerax-recipes/big_alphafold/bigalpha.html

  Tom


> On Dec 9, 2021, at 10:12 AM, Jerry Honts <jerry.honts at drake.edu> wrote:
> 
> Unfortunately several of the proteins I’m studying are ~1000 amino acids and the evidence is they form multimers.  So I was wondering if Colab Pro + would actually succeed, or whether this is currently not a feasible solution.
> 
> I should mention that I have students use Chimera in several of my courses to solve problems (and they like it), and I hope to move students to ChimeraX this year. So thanks for maintaining and refining this wonderful software.
> 
> Jerry