Local and global intrinsic dimensionality estimation for better chemical space representation

Abstract

In this paper, local and global intrinsic dimensionality estimation methods are reviewed. The aim of this paper is to illustrate the capacity of these methods in generating a lower dimensional chemical space with minimum information error. We experimented with five estimation techniques, comprising both local and global estimation methods. Extensive experiments reveal that it is possible to represent chemical compound datasets in three dimensional space. Further, we verified this result by selecting representative molecules and projecting them to 3D space using principal component analysis. Our results demonstrate that the resultant 3D projection preserves spatial relationships among the molecules. The methodology has potential implications for chemoinformatics issues such as diversity, coverage, lead compound selection, etc.