This is very arbitrary unfortunately as docking programs are based on empirical (simplified) scoring functions. One way to check the results is to:

1. carefully inspect the structures (I mean all of them)

2. run extra calculations on the most promising complexes using atomistic models (eg charmm, gromacs)

I know, it is not very satisfying

Yikes, I usually don't even bother with using web servers. I always used my local supercomputing cluster to do much larger sample spaces. I'd agree with Francesco's approach, but I'd also recommend writing a proposal for CPU-time to run a much larger set if you have the opportunity to.

As for 'determining the best of the best' it really depends on the size of the protein subunits though. Can you give us some physical characteristics of the system? You'd have to factor in AA-size and the types of residues there are. Use your physical intuition & reasoning.

First, try to validate your docking protocol on a ligand for which you have the result (e.g. the "good" X-ray position). After that, you should predict the position of other ligands (if not completely different in structure).

Be carefull with the scoring function : is it able to predict the pose you know from X-ray ? if yes, is it the best predicted pose (e.g. best score) ? The problem of course with unknown ligands, is that you will (probably) not look at the poses that are not predicted as the best scores...

Wait, Chhavi: is your system based on any crystallography/NMR or is the two protein subunit model completely ab-inito?

first, in your results, eliminate all the structure with wrong ramachandran, if you works in flexible mode...

if not the better way is to find the better electrostatic interaction or hydrophobic zone...

the major part of the time protein/protein interaction are well describe, alpha-helix in hydrophobic cleft or or different pi-stacking as trypto-trypto, arg-tryp etc...

you must if you have the possibility to obtain the PDB of your prot to analyse the water position with cristallographic importance ( bfactor near the bfactor of backbone), it could reveals some particular zone of interaction for ASP GLU,

and following all the good advice given previously...

the prot-prot interaction is probably the most difficult exercice in docking...

thanks everybody,i think i can put together things and work now and ya Jonathan the two subunits i homology modeled from a template taken from RSCB. so they arent ab initio models!

I run my own molecular dynamics simulations on the best ~10 ranking models from the docking procedure and subsequently I choose 4-5 models with minimum interaction energy. Then I utilize, if available, biochemical information from the literature as clews to decide which model is more compatible.

I second georgios

generally scores and energies are compared for their best values but extra precesion, QM/MM docking results and MD simulations confirm the best ones.

I got great results using FireDock and making the ACE (since this is calculated from solvation Delta Gs) my distinction criterion between them. Obviously you need to check the biochemical background from the literature for you to have a better picture to pick the best complex(es).