Methodology for hyperparameter tuning of deep neural networks for efficient and accurate molecular property prediction

This paper presents a methodology of hyperparameter optimization (HPO) of deep neural networks for molecular property prediction (MPP). Most prior applications of deep learning to MPP have paid only limited attention to HPO, thus resulting in suboptimal values of predicted properties. To improve the efficiency and accuracy of deep learning models for MPP, we must optimize as many hyperparameters as possible and choose a software platform to enable the parallel execution of HPO. We compare the random search, Bayesian optimization, and hyperband algorithms, together with the Bayesian-hyperband combination within the software packages of KerasTuner and Optuna for HPO. We conclude that the hyperband algorithm, which has not been used in previous MPP studies, is most computationally efficient; it gives MPP results that are optimal or nearly optimal in terms of prediction accuracy. Based on our case studies, we recommend the use of the Python library KerasTuner for HPO.