New molecular surface-based 3D-QSAR method using Kohonen neural network and 3-way PLS

Abstract

Comparative molecular field analysis (CoMFA) has been widely used as a standard three dimensional quantitative structure–activity relationship (3D-QSAR) method. Although CoMFA is a useful technique, it does not always reflect real ligand–receptor interaction. Molecular interactions between the ligand and receptor are mainly occurred near the van der Waals surface of ligand. All grid points surrounding whole molecule in CoMFA are not important as molecular descriptors. If each molecule is represented by physico-chemical parameters on molecular surface, more precise and realistic 3D-QSAR is possible. We developed a new surface-based 3D-QSAR method using Kohonen neural network (KNN) and three-way partial least squares (3-way PLS). This method was applied to 25 dopamine 2 (D2) receptor antagonists for validation. First, the 3D coordinates of all sampling points on the van der Waals surface were projected into the 2D map by KNN. Each node in the map was coded by the associated molecular electrostatic potential (MEP) value of the original sampling point. Then, the correlation between the MEP values of all 2D maps and D2 receptor antagonist activities was analyzed by 3-way PLS. The statistics of the 3-way PLS model was excellent and the coefficients back-projected on the van der Waals surface had reasonable 3D distribution. Lastly, all data was divided into the calibration and validation sets by D-optimal designs and the activities of validation set were predicted. The external validation suggested that 3-way PLS is better than standard (2-way) PLS for prediction.