Development of New Molecular Descriptors Based on Flare Software Considering Three-Dimensional Chemical Structures

Computational science and machine learning have attracted considerable attention for accelerating drug discovery. Comparative molecular field analysis (CoMFA) is a widely used method in three-dimensional (3D)-quantitative structure–activity relationships and is highly interpretable owing to the visualized structural information on electrical and steric properties; however, CoMFA requires a common structure in all compounds used in machine learning because the compounds are arranged based on this structure. In this study, we developed Flare descriptors that consider the properties of 3D chemical structures and can be used even when the structures in a data set do not have a common structure. The predictive performance of the proposed Flare descriptors was demonstrated through case studies using four data sets, and its predictive ability was higher than that of the RDKit descriptors. In addition, the interpretability of the Flare descriptors was examined and their effectiveness was confirmed.