We ran a centroid blind docking for a published reference system. [Its docking had been calculated with BiGGER - don't know if this is relevant]. We ran 1000 structures. Among the 5 most negative scores was a structure that looked as if it could be perturbed into one that resembles the structure in the literature. We ran full atom perturbation docking with an angle of 45 deg & a 10A distance. The results look as broadly distributed as the centroid blind docking, and none of the 10 structures with the most negative scores resemble the starting structure at all, or for that matter, each other. [And they all have scores with +/- 1 of each other - it's as if these 2 proteins stick together pretty much at random]. There's a distance criterion between an atom in each docking partner that's pretty easy to measure with a script & I did that for all 1K of the full atom structures. Nothing comes close to the cutoff for viable structures used in the published work. The chances I'm using reasonably intelligent flags are miniscule, since this is my first attempt at this computation, & I'd appreciate suggestions for better choices.
If constraints are the way to go, I'd really appreciate a pointer to a URL with an example, as I've not found anything along those lines I can wrap my head around. [There are some tutorial pages that look good, but they depend on your having an account on a specific system where the example files are or were stored, & I don't, & so it goes].