Evaluate Docking Result

[Anonymous]

Hi,
I have a set of outputs conformation from rosetta after hotspot placement. It’s a docking of antigen and antibody. So, how should i evaluate the results? I filtered out the top100 lowest ddG, but, when proceeding to SASA, shape complementarity score and number of hydrogen bond, the results varied. Some having more H-bond with lower Sc score ( around 0.3-0.4). Some with no H-bond but better Sc score.

For example:
A: -31 ddG, 3 H-bond, 2641 SASA, 0.565 Sc
B: -29 ddG, 0 H-bond, 2050 SASA, 0.610 Sc
C: -20 ddG, 5 H-bond, 2039 SASA, 0.323 Sc

How can I determine which one is a better docking conformation?

Thank you.

[Anonymous]

I asked around to some of the people who do antibody docking, and their main recommendation in evaluation is to take a look at the output structures in PyMol or the like, and see if they look reasonable and look like native antibody-antigen complexes. In addition to the metrics you’re looking at, they also suggest looking at packing (either visually or with “packstat” metrics), and the number of buried unsatisfied hydrogen bonds, as opposed to simply raw numbers of hydrogen bonds. (Buried unsatisfied hydrogen bonds are atoms which could be making a hydrogen bond but aren’t, and which can’t make one to the implicit solvent. You’ll want to keep this number low.) Having good solvation energies (numerically low fa_sol scores) for the complex was also suggested. The use of the InterfaceAnalyzer application (https://www.rosettacommons.org/docs/latest/interface-analyzer.html) was recommended for generating relevant metrics. Comments were that Sc tends to be low in antibody interfaces, so that’s not always a reliable discriminant, and that hydrophobic surface area burial tends to be more reliable than just total surface area burial.

One thing you might want to do is to run some known antibody/antigen interfaces which are similar to your system through the InterfaceAnalyzer. This way you can get a sense of where the metrics should be for “native-like” interfaces, and thus can discard models which are much worse than they “should” be.

[Author]

Posts