Hi,
So I was trying to model the interaction between two proteins (known to physically interact) using AlphaFold 3 server. https://alphafoldserver.com/
The prediction fails, however, with most of predicted folding having very low accuracy as well as the interface prediction failing (way below the ipTM threshold of 0.8).
When I try to model the individual proteins separately, the prediction fails similarly for both of them, with very low accuracy throughout most of the structure.
The original AlphaFold 3 paper states that proteins with low MSA (multiple-sequence alignment) depth are predicted with low accuracy. https://www.nature.com/articles/s41586-024-07487-w
I wanted to explore if shallow MSA depth would be the explanation for my failing predictions, so I was wondering how would I be able to check the MSA of my proteins.
A second, unrelated question, would be if adding ions or a prosthetic group to the modeling would improve the results (I'm starting to search the literature to know if either of these play a role in the interaction between these structures), or if the only solution would be not to model the full-length proteins but only the domains that interface each other.
Unfortunately, I don't think they have released a mechanism to get your hands on the MSA to inspect it, nor have they provided colabfold-esque plots. You could try and replicate their MSA protocol from the paper, but it could be a lot of work - I've not gone through it in detail.
They still appear to rely on the same protein sequence databases as AF2 (Uniref30/90, BFD, Mgnify, etc...). I suspect you'll get a good idea of what the AF3 MSA looks like using the AF2 MSA step.