This chapter critically reviews some of the important methods being used for building quantitative structure-activity relationship (QSAR) models using the artificial neural networks (ANNs). It attends predominantly to the use of multilayer ANNs in the regression analysis of structure-activity data. The highlighted topics cover the approximating ability of ANNs, the interpretability of the resulting models, the issues of generalization and memorization, the problems of overfitting and overtraining, the learning dynamics, regularization, and the use of neural network ensembles. The next part of the chapter focuses attention on the use of descriptors. It reviews different descriptor selection and preprocessing techniques; considers the use of the substituent, substructural, and superstructural descriptors in building common QSAR models; the use of molecular field descriptors in three-dimensional QSAR studies; along with the prospects of "direct" graph-based QSAR analysis. The chapter starts with a short historical survey of the main milestones in this area. © 2006 Springer-Verlag.
CITATION STYLE
Baskin, I. I., Palyulin, V. A., & Zefirov, N. S. (2008). Neural networks in building QSAR models. Methods in Molecular Biology, 458, 137–158. https://doi.org/10.1007/978-1-60327-101-1_8
Mendeley helps you to discover research relevant for your work.