A new approach for 3D-QSAR studies based on the volume learning artificial neural network

QSAR/3D-QSAR methods typically-generate large descriptor datasets, thus hampering attempts to use artificial neural networks (ANNs) alone to build non-linear QSAR/3D-QSAR models due to their slow rate of convergence in these cases. To address this problem, we use the Volume Learning Algorithm (VLA) that combines a Self-Organizing Map (SOM) of Kohonen to compress the input data into a discrete set of clusters that become new inputs for an ANN algorithm to construct 3D-QSAR models. The VLA retains the spatial information of the input data set, automatically captures information-rich regions, and incorporates other special features to minimize the influence of "noisy" data on predictions. VLA's speed and inherent non-linearity improves its ability to build robust models and to predict activities of new compounds. The superior performance of the VLA over Partial Least Squares (PLS) regression for use in QSAR/3D-QSAR applications is demonstxated for several biologically relevant data sets.