Alternatives To High Throughput Docking

Despite the ever-increasing number of protein kinase crystal structures available in the public domain, still only a small percentage (< 10%) of the human kinome is represented. Homology modelling is discussed below as a route to generating a 3D structure of a protein kinase, where a crystal structure is unknown. There are, however, numerous well-tested methods for screening that rely on information regarding known inhibitors of the target protein as an alternative to the receptor structure. There are many reviews discussing the methodologies of ligand-based virtual screening [109-111] and so only examples, where protein kinases are the focuses of attention are presented here.

The most popular methods for ligand-based virtual screening can be classified as substructure searching, similarity-based methods, QSAR modelling and pharmacophore methods. Each method requires that the molecules are characterised using a set of descriptors that are generally calculated computationally. Many descriptors have been proposed based on 1D, 2D and 3D characteristics of the molecules. 3D descriptors can be subdivided into those that are dependent on special alignment and those that are independent of the coordinate space. Pirard and Pickett [112] compared the ability of BCUTs, a 2D fingerprint and multi-pharmacophore descriptors, to classify EGFR tyrosine kinase inhibitors. BCUTs are derived from molecular graphs and have been developed to describe the bonding patterns and properties of molecules pertinent to ligand-receptor interactions including charges, polarizabilities, H-bond characteristics and surface areas. First BCUTs were calculated for a total of 770 inhibitors from five different protein kinases. The first six principal components (PCs) from a PC analysis could, for the most part, discriminate between inhibitors of the different enzymes, though overlap was found between p38 and spleen tyrosine kinase (SYK) inhibitors. Clustering using the first six PCs verified the relative separation of the different inhibitor sets. A success rate greater than 70% was obtained for all enzymes except Cdk1, in classifying the inhibitors using a five-component partial least squares discriminate analysis model (PLS DA) . On a separate test set of 52 endothelial growth factor receptor tyrosine kinase (EGFR TK) inhibitors, the partial least squares discriminant analysis (PLS DA) model successfully classified 48 compounds. How then did the performance of the BCUTs compare to the Daylight fingerprint and the pharmacophore analysis? The success rate in classifying the 52 novel EGFR inhibitors was low for the Daylight fingerprint, whereas the multi-pharmacophore analysis gave comparable classification rates to the BCUTs. The BCUTs are thought to encode more information than the 2D descriptors and are also fast to calculate, not requiring the generation of 3D databases. One downside of BCUTs is in interpretation, as for many descriptors it is not easy to place a chemical rationale behind the classification of compounds.

Methods such as pharmacophore models and 3D comparative molecular field analysis (COMFA) [113] require the generation and superposition of 3D representations of molecules. Putta et al. [114] have formulated an algorithm for efficiently mining the 3D conformational space of compounds and finding alignments likely to be biologically relevant. Molecules are first superposed using subshape matching, which allows the alignment of multiple conformers of each molecule and the matching of different sized molecules. Feature maps are used to score the alignment where a high score would result from the close overlay of many, equivalent features in an alignment; typical features being hydrophobic centres and hydrogen bonding groups. The matching algorithm was validated using a set of CDK2 compounds, comparing predicted conformations of each ligand with the crystallographic observations. One of the alignments closely matched the observed binding modes, correctly identifying the standard donor-acceptor sites of interaction. The method was also able to map out other regions of the ATP-binding site, though some error was noted with regard to functionality lying near solvent space. The program CATALYST (Accelrys) generates pharmacophore models based on the superposition of inhibitors molecules. Bhattacharjee et al. [115] used CATALYST to build 3D QSAR pharmacophore models capable of identifying inhibitors of the cyclin-dependent kinase Pfmrk. The models were constructed by overlaying 15 structurally diverse kinase inhibitors having a range of activities. A model comprising two hydrogen bond acceptor points, an aliphatic and an aromatic pharmacophore point gave a correlation of 0.9 for the training set when calculating the activities. Cross validation using an additional 15 compounds saw a reduction in the correlation to 0.7. A further test was performed using the model to select potential inhibitors from a database of 29,000 compounds. Steric factors for the binding site were introduced through a shape-based restriction on the compounds. There are several methods for restricting the molecules in a pharmacophore screen based on sterics. Exclusion volumes prohibit any atom in the specified volume of space, inclusion volumes state that one atom must occupy the specified space and shape constraints state that all the atoms must fall within the defined volume. The pharmacophore screen helped to generate a shortlist of 16 compounds subsequently found to have an IC50 <25 mM.

0 0

Post a comment