Neural network applications in synthetic organic chemistry. I. A hybrid system which performs retrosynthetic analysis

A computational method is proposed for devising synthetic routes which uses retrosynthetic analysis and embodies chemical intuition using a hybrid expert system that incorporates a collection of neural networks and rule-based components. Each neural network will act as an agent. The authors use list structures which are productions of context-sensitive language to represent the target molecule and subsequent molecular fragments arising from each successive set of disconnections. The internal representation thus produced will be parsed to detect certain types of features, such as strategic bonds and functional groups. A feature vector is assembled using the parser output and presented to the set of neural networks as input