Investigation of classification methods for the prediction of activity in diverse chemical libraries

Classification methods based on linear discriminant analysis, recursive partitioning, and hierarchical agglomerative clustering are examined for their ability to separate active and inactive compounds in a diverse chemical database. Topology-based descriptions of chemical structure from the Molconn-X and ISIS programs are used in conjunction with these classification techniques to identify ACE inhibitors, beta-adrenergic antagonists, and H2 receptor antagonists. Overall, discriminant analysis misclassifies the smallest number of active compounds, while recursive partitioning yields the lowest rate of misclassification among inactives. Binary structural keys from the ISIS package are found to generally outperform the whole-molecule Molconn-X descriptors, especially for identification of inactive compounds. For all targets and classification methods, sensitivity toward active compounds is increased by making repetitive classification using training sets that contain equal numbers of actives and inactives. These balanced training sets provide an average numerical class membership score which may be used to select subsets of compounds that are enriched in actives.