Development of a Structure Generator to Efficiently Generate New Structure in Target Areas on Chemical Spaces

Abstract

Computational drug design is a growing field in recent drug discovery research. For example, lead compounds, which will have high activity, can be computationally discovered in efficient ways on the first stage of development of new drugs. In this case, various lead compounds are needed since some of them may not satisfy the conditions as medical supplies such as ADMET. Thus we focus on chemical spaces around which highly active compounds exist. New compounds close to those spaces will show the same degree of activity as the highly active compounds. Therefore we develop a new system of structure generation for searching a target area in chemical spaces, which are defined by structural descriptors. First, highly active compounds are manually selected as initial seeds. Then, those seeds are entered to our generator and structures slightly different from the seeds are generated and pooled. Next seeds are selected from the new structure pool with the scores calculated from two indices for each pooled structure. One index is distance from the center of the compounds surrounding a target area, and the other is predicted activity. We used GVK data of ligand binding affinity and showed that the proposed system performed appropriate structure generation.