Hello,
I’m looking into the feasibility of using AlphaFold Multimer (more specifically the ColabFold implementation) to generate PDB complex predictions, then using these complexes as inputs to PRODIGY to predict the binding free energy. I’m looking at complexes with 2-6 chains, for which I know the binding order. That is, I might have a 2-mer binding a 3-mer, forming a 5-mer.
AlphaFold outputs 5 PDB predictions, each based on a different training seed.
On a few test cases, I’ve gotten huge (50 orders of magnitude) differences in PRODIGY Kd predictions between the 5 PDBs.
I’ve tested another Kd predictor, and it doesn’t have such high variances between these 5 PDBs.
I was wondering if this has been tried before, and if their are any ideas as to what would lead to such high variance in Kd prediction. For example, is there a processing step I might be missing on the alpha-fold generated PDB files? I will try and comment two visualizations of the stimulated PDBs with corresponding Kd predictions for reference.
Thanks!
Holly
