Knowledge Discovery in Reaction Databases:  Landscaping Organic Reactions by a Self-Organizing Neural Network

A new method based on a Kohonen neural network and physicochemical variables for describing reaction centers is developed, showing how a set of chemical reactions with the same reaction center can automatically be classified, clearly revealing different levels of similarities of the reactions under investigation.

Abstract

Chemists have always derived their knowledge about chemical reactions by inductive learning from observations on a series of individual chemical reactions. Predictions of the products of chemical reactions are made by analogy. With the availability of large reaction databases this process can be automated. In this paper a new method based on a Kohonen neural network and physicochemical variables for describing reaction centers is developed for this purpose. The results with two reaction datasets show how a set of chemical reactions with the same reaction center can automatically be classified, clearly revealing different levels of similarities of the reactions under investigation. The relative positions of reactions and clusters in the two-dimensional Kohonen map offer extra chemical information. A third reaction dataset is used to show how a trained Kohonen network can be used to predict reaction types for organic reactions.