How do you choose a good crystal structure for docking?

The first step in much of SBDD, including docking, is the selection of a good crystal structure if it exists. The crystal structure is used as the starting point for seeking new leads and optimizing them. Consider any docking method evaluation paper in J. Med. Chem. and one will come across a benchmarking set of protein structures that are used as starting models for testing the docking protocols.

Now crystal structures are frequently as close as you can get to “reality”, but even they are models and should be treated with some skepticism. But the more obvious question for such a study when multiple crystal structures of a protein are available is, which crystal structure among those should you use?

The short answer to this question is, choose one with good resolution (preferably 2.0 A or less), which does not have missing portions, and which is preferably also unencumbered by the presence of a whole lot of counterions, stabilizing molecules, and other ligands.

But is that really all? Maybe not. Recently, I was playing around with docking some molecules into kinase crystal structures. I was trying to see if docking scores can correlate with the selectivity for one related protein over the other. Usually they don’t, but I was going to look at similar proteins and similar structures, so I though it may be worth a shot. I was particularly looking at cyclin-dependent kinases (CDKs) which share a lot of homology especially in their ATP binding pocket. CDK2 is probably the most well-characterised CDK among the CDKs, and there are at least four to five different high-resolution CDK2 structures in the PDB. Also, I was more keen on using CDK2, because it was one of the proteins used for benchmarking the docking program.

So I decided upon two structures, both of high resolution. One had ATP docked into it, the other one had Staurosporine. I took an inhibitor which was known to be selective for another CDK over CDK2. First I docked it into that other CDK, and into the CDK2 structure that had ATP bound to it (without the ATP of course). I noted that the score for the other CDK was higher (which actually means more negative, since it is supposed to reflect the free energy of binding). That was consistent with the experimental data, which showed that the inhibitor was in fact more selective for the other kinase. But then, I docked it into the other CDK2 structure, and now the score was much better than for the other kinase. So the two docking runs gave two opposite results for the same protein. One predicted that the inhibitor would be less selective for CDK2, and the other one predicted that it would be more selective.

Now one of the things this says is that you cannot trust docking scores much. But this still was weird, because the question persists; which CDK2 structure should I use if I am going to do some SBDD and selectivity studies? I don’t know the answer to this question, but I took a look at the two structures to try to figure out. In the one with the ATP, the adenine region of ATP nicely made two hydrogen bonds with the hinge region of the kinase, and so did my inhibitor which was supposed to be an ATP mimic. In the other one however, the backbone carbonyl that was supposed to form the hydrogen bond to the inhibitor was rotated by almost 90 degrees upwards. It did not form a bond with stauroporine, and it did not have to, because staurosporine does not “look” like ATP. And needless to say, it could not hydrogen bond with my inhibitor too. That’s why the docking score was much worse.

What’s the solution for circumventing such a problem? One quick answer that comes to my mind is; if you are docking a ligand that is “similar” to ATP, use the protein structure that has ATP bound to it. However, “similarity” can be a tricky concept, and should be considered carefully. Also, it may be slightly easy for kinase inhibitors, because there are literally hundreds of very typical planar, heterocyclic amino-pyrimidine based kinase inhibitors that share some very obvious similarity to ATP (or not…)

But probably the best message to take home from this from a computational standpoint is that rigid protein docking not surprisingly can get you into some bad trouble. Not allowing the protein to move means that you are going to preconstrain the protein based on its preconstrained conformation in the crystal. To test this thought, I did an induced-fit docking run on both structures with the inhibitor. Gratifingly, both the runs converged on the same protein-ligand structure.

Choosing a PDB x-ray structure may not be as easy as we think, and may have to be done critically. And more importantly as usual, what we put in is what we get out. Rigid docking is ok if there’s only one crystal structure, and then only because there’s no other choice. But in other circumstances, always allow the protein to move. That’s closer to nature.

New tagline for SBDD

I have come up with a new tagline for structure based drug design:

“Semi-rational in foresight, rational in hindsight”

And in fact I think that could apply to many aspects of designing drugs.

For instance, you may think, “Hey, that group seems to hydrogen bond with an NH hydrogen; let me make it more electronegative by putting a nitro group somewhere close to it. Oops! That changed the binging mode completely. Tough luck”.

Or “Hmm…it seems that this part of the molecule that forms all these hydrogen bonds is similar to that other molecule in that other protein-ligand complex which shows swashbuckling picomolar activity. Why don’t I modify this part of the molecule to resemble the other molecule then?…Oops! Changed the entire binding mode again.”

And did I mention…the tagline actually should be “Semi-rational or irrational in foresight, almost always irrational and sometimes rational in hindsight”

And all this is only at the molecular level; I don’t even have to get started on how small changes can play havoc with things like solubility, pharmakokinetics and bioavailability, and of course, the notorious tox. Chemists are long familiar with what the addition of a single methyl group can do to binding and activity. At the same time, it’s curious how there in fact are so many inhibitors binding to a certain class of proteins that may differ radically in other parts, yet have a highly conserved set of atoms which consistently bind to the same part of the protein. Almost miraculously, all those changes in the other parts don’t seem to affect the way those few atoms bind to the protein; the best example that comes to my mind is that almost ubiquitous set of ‘hinge binding’ atoms in protein kinase inhibitors that mimics the ATP motif. Of course, this makes them wildly non-selective because almost no kinase is ‘unhinged’, but often puts you at a good starting point in terms of potency.

The fact, we still have miles to go in understanding the subtleties of protein-ligand interaction. But we have also made a mighty fine start in understanding, among other things, subtle differences in Van der Waals interactions, desolvation and entropic penalties, bioisosterism (or the lack thereof), and most recently, the role of water in ligand binding.

It’s a good day to be alive sir.