Regression Analysis of Digitized Molecular Structures in Organic Reactions

Abstract

Molecular shape and size can be important factors in the control of organic reactions. In this presentation, we will report a statistical method to extract important information of optimal molecular shape and size for a reaction. We employ finely divided molecular structures (0,1 vectors calculated from 3D molecular structures) as steric descriptors. By using regularized regression (LASSO/Elastic Net), we perform regression analysis between reaction outcomes and the steric descriptors. Based on the coefficients of the resulting regression models, we can visualize important partial structures for the reaction.