A Classification Scheme for Inverse Design of Molecules: from Targeted Electronic Properties to Atomicity

Abstract

In machine learning and molecular design, there exist two approaches: discriminative and generative. In the discriminative approach dubbed forward design, the goal is to map a set of features/molecules to their respective electronics properties. In the generative approach dubbed inverse design, a set of electronics properties is given and the goal is to find the features/molecules that have these properties. These tasks are very challenging because the chemical compound space is very large. In this study, we explore a new scheme for the inverse design of molecules based on a classification paradigm that takes as input the targeted electronic properties and output the atomic composition of the molecules (i.e. atomicity or atom counts of each type in a molecule). To test this new hypothesis, we analyzed the quantum mechanics QM7b dataset consisting of 7211 small organic molecules and 14 electronic properties. Results obtained using twenty three different classification approaches including a regularized Bayesian neural network show that it is possible to achieve detection/prediction accuracy > 90%.