Docking despair

I work on a number of drug design projects where the number of compounds being produced by the chemists is far greater than I could possibly hope to co-crystallise with the protein and solve the subsequent structures. My hope was that I could use a docking program to quickly get an idea of how compounds might bind which could perhaps help me decide which compounds are interesting enough to put into crystal screens. I have three completely separate projects, and libraries of compounds for all three projects where I know both the experimental binding affinity and structure. Thus I figured the best thing to do was to run these libraries through a number of docking programs and see which predicted my binding conformations best. I have no hope of predicting affinity data because docking algorithms are not up to this yet, but a hint at the order of affinities would be an extra bonus. If this all worked successfully I would then try looking at libraries such as ZINC to try and find totally new compounds.

After about a year, and a couple expensive software licenses, I have come to the conclusion that no current docking programs will be able to help me at this point in time. I have evaluated GOLD, FlexX and Autodock 4 with libraries from all three projects, played around with lots of parameters, and it seems none of these programs can accurately predict binding orientations, and certainly cannot rank compounds in any experimentally meaningful way. Meanwhile the couple ZINC sub-libraries I did look at with both Autodock and FlexX did not manage to dock many compounds in a meaningful way.

I think the reason for this failure is down to the protein structures themselves. Looking at the literature and at the situations where docking has worked well for people, it seems that proteins with very well defined binding sites and obvious hydrogen bond donors/acceptors work well with docking prgrams. Although two of my proteins contain strongly charged positive ions, the rest of the binding pockets rely on hydrophobic interactions which seem to be difficult to model (and hence recognise) by the docking programs. Meanwhile my third protein is a promiscuous aromatic binder whose binding site is necessarily quite unspecific, something that again is probably not good for the success of the docking programs. Thus I think at the moment the docking programs are not yet able to model my systems, and cannot really help with quickly screening my compounds.

However I do have some ideas for using the programs to simulate molecular dynamics...