E.I. Prokhorov

The work is devoted to the methods of predicting the properties of chemical compounds («structure - property» problem). As part of the solution of the main problem author proposes formulation of the problem of constructing the recognizing models and the bounds of permissibility for them. Problem of constructing the bounds of permissibility reduces to the problem of classification, called in this paper the problem of the second level. We give the necessary theoretical basis for this approach and show the advantages of using artificial neural networks with one output neuron for the classification of the second level. In the present article the author discusses the adaptation of the description of the training sample for solving the original problem and the problem of the second level. As a method to adapt the description of the training sample we propose the evolutionary selection of descriptors. As a result of adaptation for calculating bounds of permissibility of the ANN we use its special space of descriptors. Among other things we show the necessity of evolutionary selection as a means of reducing the original space of descriptors for solving the problem of overtraining when using neural networks. Moreover, using the adaptive descriptions for calculating bounds of permissibility greatly speeds up the screening of large databases of unexplored compounds. The features of using the artificial neural networks for classification problems in chemistry are described. The necessary recommendations on the choice of architecture of neural networks to solve «structure - property» problem are given. The paper proposes general methodology of research in the prediction of properties of unexplored chemical compounds using artificial neural networks. We describe an approach to the consistent prediction by using sets of recognizing models and its bounds of permissibility. Current paper includes the results of testing of this approach in the analysis of a sample of enzyme inhibitors of cell division.