Creation and Selection of Linear Fragment Descriptors in Chemical Graph Mining

Abstract

Linear fragments of molecular structures have been employed as descriptors in the structure activity relationship studies by the cascade model. A variety of new fragment expressions are developed in order to give better descriptions of characteristic substructures. They include hydrogen-bonded fragments, the omission of element symbols in the middle of a fragment, the lowercase letter notation for atoms in an aromatic ring, and the unified atom treatment for the carbonyl (CO) group. We have to select descriptors out of numerous fragments extracted. Correlation coefficients among categorical attributes are defined. We omit a descriptor when it has a high correlation with other descriptors employed. The resulting descriptor set is efficient and effective in the mining of structure activity relationships.