ABSTRACT
O. Ivanciuc Department of Organic Chemistry, Faculty of Chemical Technology
University “Politehnica” of Bucharest, Oficiul 12 CP 243, 78100 Bucharest, Romania
Artificial neural networks are general nonlinear models that proved to be very efficient in computing physical, chemical, and biological properties of various classes of chemical compounds. The success of neural networks in structureproperty models depends mainly on the numerical representation of the structure of the compounds in network calibration and prediction. The atomic and molecular graph descriptors used as input data for neural networks are presented together with examples of their computation. Three new neural networks were defined in order to encode into their topology the chemical structure of each compound presented to the network: the Baskin-Palyulin-Zefirov neural device, ChemNet defined by Kireev, and MolNet introduced by Ivanciuc. All three neural models use information from the molecular graph to generate the neural network. The rules that define the three neural models are presented together with examples of network generation from the molecular graph.
