Where are the Pockets?

(February 8th, 2016) Allosteric sites on proteins have enormous potential for the development of new and better drugs. Finding such sites is, however, very difficult. But that is changing thanks to AlloPred, an online utility from the laboratory of Michael Sternberg at Imperial College, London.

But first of all, why is there such an interest in finding new allosteric sites? Most drugs on the market today target an orthosteric site - they attach to the same site as a natural ligand that they are mimicking. Nicotine, for example, does its thing on the brain by binding to the site normally targetted by acetylcholine, the natural brain transmitter. Finding drugs that hit the orthosteric site is relatively (ah, big word that - "relatively") easy. Given that the natural ligand has a well-defined biological response, screening candidate ligands is a matter of monitoring for that response in a model system. The key thing here is that adding the natural ligand, or its pharmacomimetic, always elicits a response. Watch out for that response and you'll spot the mimic.
But drugs that work by hitting the orthosteric site present some very real problems. The target site may be conserved across other proteins, raising the risk side-effects. You are pretty much limited to one direction of response - you can switch things on, but not off. Finally, by clubbing the orthosteric site you are probably interfering with some delicate signalling. It is a bit like trying to overcome someone's deafness by screaming at them hysterically.
An allosteric site is one that will not evoke a response on its own, but instead modulates the effectiveness of the main signalling ligand. Less like screaming hysterically and more like just talking louder. And there is another plus: allosteric sites are also less conserved than orthosteric sites, so drugs targetting them are more likely to be highly specific. Needless to say, because allosteric sites do not evoke a response on their own, screening for them can be tricky. Finding them in the first place is hard - you can't just label a ligand and see where it sticks. That is why the race is on to perfect computational and predictive approaches.
Michael Sternberg's lab is in just such a race. They use a two-step approach to predicting allosteric sites. First, a standard computational approach called Fpocket is used to find pockets in the protein. The next step is to predict whether anything binding to those pockets will affect the orthosteric site. This is done using a trick called Normal Mode Analysis (NMA). The idea behind this is that although there is a huge number of states a molecule can be in, you can get away with reducing them to a relatively set of orthogonal harmonic modes. This makes it computationally tractable to treat the molecule as a set of springs and predict the shapes it will adopt should a molecule occupy a pocket. As a final step, machine learning (Support Vector Machines) is then used to predict the allosteric pockets. 
As Lab Times so often says, "never mind the maths, just operate the program". Sternberg's expertise has been transferred from his brain and uploaded to the internet. The AlloPred program is freely available, web-based and incredibly easy to use. To use it, the only bits of information you need are the PDB identifier of the molecule you are interested in, and the location of the active site. We tapped in the example parameters suggested at the website - the catalytic domain of a receptor-type adenlyate cyclase from Trypanosoma brucei - and got our results in just a few minutes. And the results couldn't be easier to interpret: you are presented with a split screen, with the details of the located pockets on the left and a 3D model on the right. In our case, eight pockets were found and had been ranked by AlloPred for their likelihood as being allosteric pockets. On the right screen, a Jmol applet displays the protein, which you can move around with your mouse. If you select any of the pockets shown in the left screen, a corresponding filling molecule is drawn in the Jmol applet occupying the pocket.
Sternberg tested the algorithm on a set of 40 known allosteric proteins and found that in 23 out of 40 cases, the highest ranked pocket in each protein turned out to be a known allosteric pocket and in 28 out of 40 cases an allosteric pocket was ranked first or second. Sternberg's results, along with an explanation of how AlloPred works, are laid out in a recent publication.

Steven Buckingham

Picture: ProteinDataBase

Last Changes: 03.02.2016