Application of nonlinear and local modeling methods for 3D QSAR study of class I antiarrhythmics

Abstract

Locally weighted regression and artificial neural networks were employed to find a quantitative relationship between recovery time and molecular structures for class I antiarrhythmics. Nonlinear and local models have been built between score vectors of columns of Comparative Molecular Field Analysis (CoMFA) as independent variables and recovery time values as dependent variables. Other method by applying cheaply computed descriptors invariant to roto-translation with artificial neural networks were also used. Predictive ability of the methods was tested by a separate set of compounds, and the performance of both procedures proved to be acceptable, and comparable to CoMFA. This study clearly demonstrates the need and ability of nonlinear algorithms in building of 3D QSARs. The methods presented here do not assume any particular functional form for developing quantitative models between molecular descriptors and biological activity.