How to select an optimal neural model of chemical reactivity?

The paper aims at methodological studies on selection of optimal neural network that performs modeling of chemical reactivity of a given group of chemical compounds. The problem (prediction of biological activity in enzymatic reaction catalyzed by ethylbenzene dehydrogenase) is taken as a case study for assessment of various types of neural networks. The main goal of the study is to select the best type of the network, optimal dimension of the layers, proper parameters of the network as well as the optimal form of data representation for purpose of neural-based modeling of complex empirical data. Various approaches (linear networks, regression and classification multiple layer perceptrons, generalized regression neural networks) are compared and tested.