A Measure of Domain of Applicability for QSAR Modelling Based on Intelligent K‐Means Clustering
Robert Stanforth, Evgueni Kolossov, Boris Mirkin · QSAR & Combinatorial Science · 2007
Abstract The prediction of a biological activity using a Quantitative Structure–Activity Relationship (QSAR) model is valid only if the compound in question is inside the model's domain of applicability. The existing methods for determining the domain of applicability in descriptor space suffer from problems including poor handling of nonconvex training sets and computational inefficiency. In this paper, we propose a cluster‐based approach to modelling the domain of applicability, which may overcome some of the shortcomings of the existing approaches described. We investigate applying an intelligent version of the K‐means clustering algorithm to this problem, modelling the training set as a collection of clusters in the descriptor space. A test compound is assigned ‘fuzzy membership’ of each individual cluster, from which an overall distance may be calculated. Finally, we experimentally assess how this cluster‐based approach compares with the existing methods.