Directional Generative Networks; Comparison to Evolutionary Algorithms, Using Measurements for Molecules

Various methods have been proposed to search for molecules with desired properties. However, it is still a challenging task due to the vastness of the search space. In recent years, machine learning methods such as deep learning have achieved remarkable results in various fields. Although machine learning methods depend on data sets, it is sometimes difficult to prepare good-quality data for some tasks, and unsupervised learning methods have also been studied. A method called Directional Generative Networks (DGN) uses Deep Learning models as regression models. DGNs can generate molecules with desirable features by using the values of an evaluation function, even in the absence of a dataset of molecules. Given the lack of applications of this method, this study attempts to extend the evaluation function for molecular search applications using DGN. While previous studies have used three different functions to evaluate molecules, this study adds three more, resulting in six evaluation functions compared to the Evolutionary Algorithm (EA). Furthermore, DGN incorporates random numbers as input, and we discuss the impact of the degree of freedom of the input during the training phase. A large degree of freedom in the input increases the number of variations of molecules generated, and it causes the training models to become more challenging to converge. Conversely, when the degree of freedom of the input is small, the training convergence becomes more manageable, but the variation of the generated molecules decreases.