Discovering drug candidates in virtual chemical libraries: A novel graph-based method for virtual screening

The virtual screening of chemical libraries for drug candidates is an important technology used by researchers to find novel drugs. The chemical libraries of interest may be very large so it is important that virtual screening methods be fast. Some of the best methods for ligand-based virtual screening, pharmacophore fingerprint methods, require a conformational search for each of the molecules in the screening library. This requirement for a conformational search makes it time-consuming to screen libraries containing millions of compounds because it may take more than a minute to screen a single molecule on a single processor.; We were interested in developing a faster approach to ligand-based virtual screening that would be able to quickly screen large libraries. We developed a novel graph-based method, the subgraph fingerprint (SPrint) method, that does not require a conformational search to be performed for each molecule. In this method each molecule is represented by a feature vector where the features correspond to subgraphs present in a particular type of molecular graph. In this graph representation the vertices of the graph represent atoms and the edge between two vertices represents the maximum distance that the two atoms can be apart. Statistical models of activity built from these feature vectors are used to screen chemical libraries for drug candidates. We show that our method is accurate; fast, interpretable, and may be able to find novel structures that have different chemical scaffolds from known inhibitors. In particular, we show that the SPrint method is able to screen chemical libraries at a rate greater than 60 molecules per second on a single processor and is as accurate as a leading pharmacophore fingerprint method on a benchmark data set of estrogen receptor inhibitors.; We applied the SPrint method to the problem of finding inhibitors to the cancer target checkpoint kinase-1 (Chk1), by screening a library of 1.1 million commercially available compounds. In addition to the SPrint method we also used two external methods, a ligand-based similarity method called Rocs and a structure-based docking method called GLIDE. We discuss the results from this virtual screening study and propose several molecules for experimental testing.