I'm running virtual screening tests with AutoDock GPU and am trying to validate that my configuration (parameter file, search box settings based on the target, search parameters, etc.) is correct, or at least reasonable.
For a few targets I've been able to do small scale validation, e.g. by taking protein-ligand structures from Protein Data Bank, redocking, and comparing AutoDock's predicted poses to the crystollography. I feel that approach does not scale very well.
The DUDE database seems like an obvious choice, and I tried two targets from there so far: AA2AR and ADA, with several different AutoDock configurations, all with very poor results (ROC area-under-the-curve pretty close to 0.5/random guessing in all cases). And that makes me think there really is something very wrong with my setup.
What I'm looking for is any reference target + ligands set, in the style of DUDE with known actives and decoys, or without decoys but with known/expected AutoDock affinity scores or a known AutoDock-predicted ranking, that I could use as a kind of unit test for my setup. I want to eliminate, for the moment, all the hard chemistry variables and reduce this to a debugging exercise: "If the software is configured correctly the output will definitely be close to X. If I'm not getting X then I'm Doing it Wrong™."
Is there any reference/validation set of that sort for AutoDock GPU?
I think you need to do it by yourself and it would be an interesting project. I came across the dataset with free energy calculation, however, I never had the opportunity to see this kind of dataset/database.
- Badges
- Science topic
- Similar topics
- Filing