Comparison of State-of-the-art Algorithms for de novo Drug Design

Abstract

During the last decade, the application of deep learning in de novo drug design has increased. By employing generative models, in combination with suitable optimization algorithms, the chemical space can be explored to generate new and useful molecular compounds. Multiple models have been published for this purpose. While they all report promising results on various optimization tasks, there is a lack of continuity regarding what tasks on which the models are demonstrated. This makes a comparison of the models unfeasible. In this thesis, we provide a comprehensive evaluation and comparison of five state-of-the-art algorithms for de novo drug design. To this end, the selected models have been submitted to two experiments that evaluate their ability to generate and optimize molecules on a variety of different optimization tasks. The reported results show that a generative language model displays the best performance, both in terms of exploring and exploiting the chemical space. An application of particle swarm optimization on a continuous representation of the chemical space also shows promise, and we suggest further research on more elaborate usages of this method.