Applications for neural networks in chemistry. 2. A general connectivity representation for the prediction of regiochemistry

Abstract

A general method for the prediction of organic reactions by a backpropagation neural network is described. Neural networks trained using modified Dugundji-Ugi BE-matrix representations gave excellent predictions of the regiochemistry for three different types of reactions: Markovnikov addition to alkenes, Diels-Alder and retro-Diels-Alder reactions, and Saytzeff elimination. The networks were able to extract reactivity information from examples of the reactions to develop an internal representation of the reactions without explicitly incorporating rules into the network. Since the neural network was better at interpolating than extrapolating, it is important that the training set span the set of possible reactions. The method of representation used is sufficiently general to handle most classes of organic reactions.