D. A. Devetyarov, M. I. Kumskov

Artificial neural networks (ANN) are widely used in quantitative structure-activity relationship (QSAR) problem to search for new chemical compounds with preset properties. The paper suggests the methodology of ANN application to fuzzy description of molecular structure in QSAR problem. The interpretation rules of "structure-property" model as artificial neural networks are described. The use of fuzzy description allows us to carry out the automatic search for the optimal description of the molecular structure and to efficiently process molecules with flexible structure. We suggest separating the stage of training ANNs from the stage of their application in order to lower the computational costs. As a result, we suggest constructing the prediction model by the use of other machine learning methods and then interpreting the result (the prediction model) as an ANN. For example, in this paper we construct models using the Group Method of Data Handling (GMDH) applied to preliminarily identified clusters of objects. The paper describes the algorithm of interpretation of the machine learning methods as ANNs. Classification of a new molecule to one of the clusters or outliers can be implemented by the use of perceptrons - radial basis functions. Linear models constructed by GMDH and voting of linear models are suggested to implement using multilayer perceptrons. The approach above was applied to prediction of toxicity of chemical compounds. The computational experiments were carried out for both fuzzy and non-fuzzy descriptors as well as with different sets of parameter values used when constructing the descriptors. The results demonstrated that constructed ANN prediction models allow us to achieve prediction accuracy of 87.9% on leave-one-out cross validation.