Linear and Nonlinear Support Vector Machine for the Classification of Human 5-HT1A Ligand Functionality

Upon binding to a receptor, agonists and antagonists can induce distinct biological functions and thus lead to significantly different pharmacological responses. Thus, in silico prediction or in vitro characterization of ligand agonistic or antagonistic functionalities is an important step toward identifying specific pharmacological therapeutics. In this study, we investigated the molecular properties of agonists and antagonists of human 5-hydroxytryptamine receptor subtype 1A (5-HT1A ). Subsequently, intrinsic functions of these ligands (agonists/antagonists) were modeled by support vector machine (SVM), using five 2D molecular fingerprints and the 3D Topomer distance. Five kernel functions, including linear, polynomial, RBF, Tanimoto and a novel Topomer kernel based on Topomer 3D similarity were used to develop linear and nonlinear classifiers. These classifiers were validated through cross-validation, yielding a classification accuracy ranging from 80.4 % to 92.3 %. The performance of different kernels and fingerprints was analyzed and discussed. Linear and nonlinear models were further interpreted through the illustration of underlying classification mechanism. The computation protocol has been automated and demonstrated through our online service. This study expands the scope and applicability of similarity-based methods in cheminformatics, which are typically used for the identification of active molecules against a target protein. Our findings provide a good starting point for further systematic classifications of other GPCR ligands and for the data mining of large chemical libraries.