Open genya opened 1 year ago
When only a single sequence is provided, alphafold is limited to what it can predict. If you have reason to believe that the region you are constraining will fold regardless if alphafold thinks it will or not, you can constrain it by adding a template.
Examples: target proteins (for designing PPI) or domains (in cases where there is an extension at the termini or loops of large protein domain)
I'm trying to use the partial_hallucination_rewire notebook to replace part of a protein with a linker-like sequence that otherwise preserves the overall predicted structure (code below). To preserve the exact sequence outside of the hallucination region, I think I have to have fix_seq=True in model.prep_inputs(), but I'm unsure whether use_templates in mk_afdesign_model() should be True or False?
The goal with these partly hallucinated proteins is to disrupt a function specific to the hallucinated region while preserving functions localized to other parts as well as overall stability/solubility. To assess which partly hallucinated variants are likely to have these properties I'm looking at how well the structure outside the hallucinated region is preserved and the pLDDT score both outside and within the hallucinated region.
These metrics look better when use_templates=True, but does this actually mean that the predicted structure is more reliable or is use_templates=True forcing the output to conform to the input by construction?
Maybe relevant is that the protein is so large (~1000AA) that my input PDB is the part of a larger alphafold model containing only the relevant domain (otherwise memory errors), and I'm unsure whether AlphaFold would predict this domain to fold on its own as it does in the larger structure (would it help to model this?).
Here's the code: