A benchmark study of machine learning methods for molecular electronic transition: Tree-based ensemble learning versus graph neural network

Abstract

Fluorophores play crucial roles in chemical and biological imaging. An efficient computational model that evaluates the electronic properties of molecules accurately would be a useful tool for discovering novel fluorophores. Tree-based ensemble and graph neural network (GNN) methods have been regarded as attractive models for predicting molecular properties. Here, we present a benchmark test using three tree-based ensemble methods (Random Forest, LightGBM, and XGBoost) and three GNNs (directed message passing neural network [D-MPNN], attention message passing neural network [AMPNN], and DimeNet++) for predicting electronic transition properties such as excitation energies and oscillator strengths. From our benchmark, DimeNet++ was identified as the most accurate model to predict electronic transition properties. The average root mean square error of DimeNet++ for predicting the HOMO-LUMO gap was 0.11 eV whereas those of the other methods exceeded 0.3 eV. D-MPNN predicted fastest without sacrificing accuracy. Our results show that DimeNet++ and D-MPNN may serve as helpful evaluators for novel fluorophore design when combined with molecular generation methods.