Effect of molecular representation on deep learning performance for prediction of molecular electronic properties

Abstract

Deep learning (DL) can be a useful approach to molecular applications such as the organic light-emitting diode (OLED) development via high-throughput virtual screening. Various representations have been proposed to incorporate molecular structures in DL methods. However, it is yet to be clear which one would be better for accurate prediction of molecular electronic properties. Here, we carried out a comparative study on the performance of four widely used molecular representations to elucidate an optimal solution for DL applications to OLED materials. We implemented six DL models based on the four representations and assessed their accuracies in the prediction of the electronic properties of thermally activated delayed fluorescence (TADF) molecules. The attention gated graph neural network based on molecular graphs showed the highest accuracy for test sets and TADF candidates. Therefore, the molecular graph can be used as an optimal representation to predict the TADF-related molecular properties.