sokrypton
How to save time to predict multimer structure by reusing previous monomer structure results
Expected Behavior
I used colabfold to predict two monomer proteins cas9 and deaminase, and I wanted to continue to predict their possible complex structure by multimer method. While the cas9 protein was too large (~1400) AA), how can I adjust the code to reuse the previous monomer structure prediction results so that the prediction time of the multimer structure is greatly reduced? How, if at all, can the code be tweaked to achieve this?
Funny, we have been just discussing this to speed up AlphaPulldown ;-) In the simplest way of implementing it, you should be able to do this by providing precalculated monomeric models (e.g. from AlphaFold DB) as monomeric templates, but I don't know if ColabFold supports multiple custom templates. In AlphaPulldown it should be possible (using the mode for multimeric templates https://github.com/KosinskiLab/AlphaPulldown/blob/main/manuals/example_3.md but turning the multimeric_mode off). Perhaps that would allow reaching good multimeric predictions in less recycles. We were discussing benchmarking it over different numbers of recycles, but maybe @sokrypton, you tested that already? In addition, one could possibly use the monomeric models as ColabFold's initial guess, even if they are in the arbitrary orientation of monomers, but I'm not sure if that would work well.
Colabfold allows multiple custom templates if alignable by HHsearch.
