The word "pharmacophore" was coined by Paul Ehrlich (a great word coiner; he also coined "magic bullet") as a reference to the essential chemical functionalities in a drug that makes it active. In the time since then, that word has been modified to mean the essential set of features in a drug, including their geometry and 3D orientation, that makes the drug active.
Pharmacophore modeling has become a stable of drug discovery and virtual screening, especially in the context of Computer Aided Drug Design (CADD). But even the average medicinal chemist is always interested in pharmacophores because he wants to know what's common between diverse structures that bind to a receptor that makes them active. Thus arises the goal of finding the "common pharmacophore" for different structures that bind to a site.
Sadly, the methods used for achieving this goal often fail because of some fundamental mistaken assumptions. The first assumption is that there is one, single pharmacophore. But ask anyone who uses a decent pharmacophore modeling program, and he or she will tell you that the program generates many pharmacophores, and it is not easy to decide which one among them is the "correct" one. The usual flaw in trying to come up with a common pharmacophore is to overlap compounds and look at common functionalities (hydrophobic, aromatic, charged etc.) in the same places. While this is a justifiable technique, the simple fact as noted above is that there are many ways to do this. One may overlay compounds so that say aromatic portions match with aromatic portions. But what about the other portions? How do we know that they don't constitute good binding features? Sometimes such trends or the lack thereof can be gleaned through traditional med chem and SAR, but more often than not they don't work.
In pharmacophore based virtual screening (VS), the technique is to come up with a pharmacophore and then compare it to a 3D library of possibly hundreds of thousands of compounds to find a lead that matches the pharmacophore in its 3D conformation. Since the pharmacophore essentially includes 3D information, the compounds in the databases have many conformers. In most cases, conformational searching for these compounds cannot be very exhaustive, but in most cases drug-like compounds have few rotatable bonds and so exhaustive conformation generation is not required.
Such simplistic techniques are seen in some early pharmacophore models. Say you have a peptide with a positive and negative end (NH2 and COOH) and you want to develop a good small molecule analog. You look at the minimum energy conformation of that peptide, or even the bioactive conformation if you know it. Then you note that the distance between the positive amino and negative carboxylate ends is 10 A. So with this pharmacophore, just an amino-carboxylate distance of 10 A, what you do is simply run this feature against a large 3D database of molecules hoping that you will come up with something that will also have the same distance between such two groups and therefore be active. This is admittedly the most simplistic pharmacophore ever, but Merck made it work for some fibrinogen agonists in 1992, which was an early example of pharmacophore-based virtual screening. The problem inherent in this protocol is that you are obviously neglecting other parts of the molecules which will dictate binding. Binding is a multipronged phenomenon and does not always depend on just two or three strong interactions.
This problem with investigating pharmacophores points to a larger truth; that ligands which bind to a common site in a protein may not bind the same way in many cases. For example, several people tried to come up with a "common pharmacophore" hypothesis for compounds binding to tubulin, such as taxol, epothilone and discodermolide. But a 2004 paper indicated that the binding site was promiscuous, that each ligand exploited the site in a unique way, and so trying to overlap them and come up with a common pharmacophore hypothesis was largely a futile effort.
If you have ever built a pharmacophore, you will also realise that it depends crucially on user input. Chemical intuition and knowledge of binding as well as known SAR data is very important for making decisions along the way. Pharmacophore modeling even more than other CADD protocols is not a black box.
In the future, pharmacophore models will continue to be used. Again, we must be realistic about their utility, which is seldom to come up with an ultra-potent lead, but to help in lead discovery. Some people think that because of this multiple-solution scenario, pharmacophore development is inherently flawed. I am more optimistic because first of all it is now possible to generate many pharmacophores and assess them against multiconformer libraries of a million compounds if necessary in a relatively short period of time. Secondly, model generation combined with careful analysis and astute chemical intuition has been known to pay dividends in the past, and there is no reason why it should not pay dividends in the future. At some point in the future, I am planning a post on some more details of pharmacophore modeling.
Of interest: A pharmacophore for kinase frequent hitters