Data-Driven Molecular Structure Generation for Inverse QSPR/QSAR Problem

In 2010, we (Funatsu and Miyao) proposed a Bayesian approach for the inverse quantitative structure–property and structure–activity relationship (QSPR/QSAR) problem in molecular design [Molecular Informatics (2010), 29:111–125]. While QSPR/QSAR workflow predicts physicochemical properties or biological activities from chemical structures, inverse QSPR/QSAR workflows adopt QSPR/QSAR models to propose novel chemical structures having desired properties or activities. The proposed Bayesian approach efficiently identifies the descriptor regions where corresponding chemical structures have the desirable output of a model within the domain of applicability of the model. Since then, methods have been developed for inverse QSPR/QSAR analyses, including fragment-based exhaustive structure generation and the incorporation of nonlinear QSPR/QSAR modeling methods [J. Chem. Inf. Model (2016), 56:286–299]. In this chapter, we explain the concept of our inverse QSPR/QSAR workflow, particularly in terms of adopting a Bayesian approach, and present a proof-of-concept study of structure generation for thrombin inhibitors. Furthermore, related methods for the inverse QSPR/QSAR problem are briefly reviewed in light of our approach.