I performed virtual screening of some molecules in Auto Dock Vina and GOLD and obtained different results for the best ligand. I'm not sure if I did something wrong, or the docking program is not the most appropriate for my protein and ligands, or the difference is due to the different interaction analysis algorithm.
The protein does not have a co-crystallized ligand, so I did not validate the program by redocking. I'm following the active site predicted by COACH-D. I don't know if this is relevant when comparing results from two different programs.
How different they are?
If you are focusing on a single ligand, it is not a good approach as you starting point is weak. Just look into the top 10/20 or 50 molecules from each docking program and compare them. How many of them are similar/same?
Just a suggestion, take a similar protein (same size or similar binding site) with known ligand (co-crystal structure) and do the docking exercise with them and see which program gives you better result. This can partially justify your results as well.
Hello. The thing here is that the underlying energy function from VINA and GOLD are different both in terms of philosophy and parameters. Essentially, GOLD is a "physics-based" scoring function which takes into account intra and intermolecular hydrogen bonds, strain energy penalty and a Van der Waals term. AutoDock Vina is based on empirical potentials. Thus, if you are looking solely and the top ranked ligand, due to the differences in the scoring function, the results are bound to be different. There is also the issue of sampling the phase-space within the docking algorithm (i.e did you run for enough steps that the search space is fully and comprehensibly covered). Apart from these two concerns, other things can plague your docking experiment, and there are wonderful reviews that discuss this issue.
As a suggestion, I find that looking at a fraction of the top-scored ligands (like the person above commented) is indeed a good idea. However, doing this blindly is not really useful. When carrying out virtual screening, it is often more interesting (in the first stage) to identify chemical scaffolds that may be active, and not necessarily will you immediately find a potent binder to your protein. Hence, it is common that you take your ligand library and cluster it using molecular fingerprints or any other clustering technique you like, to reduce the library to a sub-set of molecules which are representative of the molecular families present (i.e molecular scaffolds). This gives you two advantages:
1) You get to cut your computation time massively
2) The ranking will be more robust - While the order may vary, the most interesting chemical scaffolds will be better ranked. If you don't do this and the dataset is dominated by a family of compounds, it is highly likely that most of the top-ranked compounds will be from this family due to the dataset imbalance, meaning that you may miss out on potentially interesting chemical scaffolds.
