Methods and algorithms for predicting the properties of chemical compounds by common fragments of molecular graphs

Abstract

Methods and algorithms for predicting the properties of chemical compounds by common fragments of their molecular graphs are described. The prediction algorithms are based on determination of a measure of structural proximity (distance) between molecular graphs, which depends on the size of their common fragment. The prediction procedure involves the following steps: partitioning the property classes of the training sample compounds into subclasses of structurally similar compounds; seeking structurally typical compounds and their fragments in each subclass; classifying control compounds according to their distances from the training sample compounds or fragments of classes; forming a set of essential fragments of samples potentially responsible for the properties exhibited by the compounds. The algorithms were successfully tested in the BACC system for analyzing and classifying biologically active compounds designed at the Institute of Mathematics, Siberian Branch, Russian Academy of Sciences.