I’m currently working on in silico design of antibody CDRs.
I generated random CDR sequences, performed docking using HADDOCK to predict antibody-antigen complexes, and estimated the binding affinity using PRODIGY.
Interestingly, several trials consistently yielded PRODIGY ΔG values around -10 kcal/mol, which is close to the experimentally determined ΔG of known complexes.I’m wondering why this might be the case.
Could it be that PRODIGY is not suitable for evaluating binding affinities of complexes generated by docking interfaces that are unlikely to interact in reality?
In my case, I intentionally specified non-native interface residues to force HADDOCK to generate binding poses.
PRODIGY is a supervised contact-based bayesian model, trained on a set >100 complexes to predict they ΔG values.
It has not been trained to differentiate true interface from wrong ones.
Also, it was not trained to predict the ΔG of antibody-antigen complexes.
Apologies for the additional question, but as far as I understand, the PRODIGY score is calculated based on the types of residue-residue contacts (e.g., polar/polar, charged/polar, etc.).
So intuitively, I would expect that if a complex is artificially forced into an incorrect binding mode, the resulting contacts would involve a lot of steric clashes or unfavorable interactions, leading to a worse (i.e., less negative) predicted affinity.