G Moroy, V.T. Martini, P. Vayer, B.O. Villoutreix and M.A. Miteva, Drug Disc. Today 2012 Jan 17 (1-2) 44-55
For a drug to be successful in treating a condition, it must not only modulate the condition’s underlying mechanism, but also have a suitable absorption, distribution, metabolism, excretion and toxicology (ADMET) profile. Accurate prediction of this profile is a key way of reducing costs, animal studies and other resources for molecules that are destined to fall at ADMET hurdles. Accurate prediction is also, unfortunately, somewhat elusive.
Traditionally, computationally predicted ADMET relied on 2D and 3D QSAR/QSPR (Quantitative Structure-Activity / Structure-Property Relationships) or knowledge-based / expert systems as its preferred method of model development. QSAR comes with the problem of requiring high quality (and usually large) data sets of materials that have been tested biologically. The authors of this paper have noted recent changes to this historical approach and have detailed current movement away from using QSAR/QSPR on its own, using systems that consider the 3D nature of the interacting proteins, rather than solely a set of ligands.
There has been change in pace in isolation and production of high quality crystal structures, and the authors note that for a large number of ADMET-involved proteins, there are crystal structures in the Protein DataBank (PDB) which have allowed them to investigate the concept of using flexible docking for finding potential pitfalls in their compound development – a list of common Human CYP450 and Human sulfontransferases (SULTS), along with their PDB codes are given, with insights into the state of the art over a range of areas associated with ADMET (for example, plasma-binding protiens, hERG, ABC transporters and so on). They then go to demonstrate that for one particular family of proteins, the SULTS, it was possible to develop a flexible docking model to support ADMET prediction.
The authors conclude that whilst flexible docking is amenable to some ADMET-involved protein families it is not so straightforward: these proteins are designed to be a promiscuous (some with multiple binding sites), may not have the whole picture with regards to water, and require flexible modelling. They propose that a series of other emerging techniques such as MD / MM and proteochemometrics may be useful in addition to ligand based methods.
There is no doubt that in silico ADMET prediction is a big challenge in computational chemistry – the aim here is not to simply replace in vivo and in vitro tests, but to improve the knowledge-base and reduce those materials that are very likely to be liabilities as early as possible – preferably before synthesis: Fail early, fail fast, fail cheap.
It may be a mistake to attempt to tackle the problem using only one tool in the toolbox (e.g. topological analysis systems) – A mixture of QSAR, toxicophore / pharmacophore, flexible docking and expert systems together could open up an avenue to much more accurate desktop prediction. This paper goes a distance in explaining not only where the successes lie, but more importantly, which issues still challenge computational chemists working toward in silico structure-based ADMET prediction in drug discovery.