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.

A deep learning framework for predictions of excited state properties of light emissive molecules

“…To the best of our knowledge, we have achieved a notably high prediction quality for the oscillator strength with the outof-sample R 2 approaching 0.92. The level is in significant contrast with the previous works, where R 2 is reported in the limited range of 0.7-0.8, 14,19 and magnified predictive uncertainties can be encountered. 13,20 The current research reveals that in addition to a sophisticated ML model, an extended torsional-angle profile between the donor and acceptor can be equally important for the accurate prediction of the transition dipoles.…”

Machine-Learned Electronically Excited States with the MolOrbImage Generated from the Molecular Ground State

“…20 More importantly, the literature report for the machine-learned electronic transition intensity across chemical compound space is rather limited. 15,21,26 The promising prediction of μ and f mainly comes from system-specific ML models. 18,29−31 As such, there is still a pressing need to develop a ML model that maps computationally cheap descriptors to electronic transition energies and intensities with similar accuracy as the high-level method.…”

“…A number of ML strategies (Figure a) have been developed to probe the molecular excited state. − The ultimate goal is to reach an accuracy comparable to those of high-level computational methods such as the second-order approximate coupled cluster (CC2) and multireference configuration interaction (MRCI). On the basis of the TDDFT-computed bands and Coulomb matrix descriptor, the Δ TDDFT CC2 -ML method has reproduced CC2 excitation energies for S 1 and S 2 states with mean absolute errors (MAEs) of 0.09 and 0.16 eV, respectively, for 21786 small organic molecules .…”

“…Recently, SchNarc, which combines SchNet with the SHARC (surface hopping including arbitrary couplings) approach, has been developed to compute the excited state dynamics of CH 2 NH 2 + and isoelectronic molecules . More importantly, the literature report for the machine-learned electronic transition intensity across chemical compound space is rather limited. ,, The promising prediction of μ and f mainly comes from system-specific ML models. ,− As such, there is still a pressing need to develop a ML model that maps computationally cheap descriptors to electronic transition energies and intensities with similar accuracy as the high-level method.…”

De novo creation of fluorescent molecules via adversarial generative modeling

“…The advent of machine learning brings about a variety of opportunities for the exploration of unknown chemical space and predictions of molecular properties. In particular, numerous efforts have been made to investigate the excited state properties of organic semiconductors, [15][16][17][18][19][20][21] and relevant de novo design frameworks 14,22 have also been proposed to develop new molecules with desired functionalities. For uorescent materials, machine learning has given successful predictions for the experimental quantum yields and emission energies, 21,23 by using empirically designed chemical descriptors.…”