(458c) Computer-Aided Drug Virtual Screening Framework Based on the Binding Site Selectivity

Computer-aided drug design (CADD) [1] is a topic of increasing interest in today's novel drug discovery. Compared to traditional experiment-based exploration of active molecules, CADD methods have shown their advantages in reducing the experimental cost and shortening the development cycle by quickly suggesting promising drug molecules that can be verified through focused experiments. As one of the most popular drug discovery methods in CADD, structure-based virtual screening [2] allows rapid evaluation of drug-like molecules and approved drugs at the early stage of a disease outbreak. Among these pharmaceutical evaluations, the binding of drug molecules to the target protein is an important aspect, which is usually used to analyze the expected efficacy of drugs. However, the current virtual screening workflow focuses more on screening speed and larger molecule libraries. The issue of selectivity of drug molecules for different binding sites (actual binding domain of drugs) on proteins was either not considered or not addressed in sufficient detail. Because of the complexity of biological macromolecular structures, there is usually more than one binding pocket or concave region in a protein, but only the specific binding site is valuable for structure-based drug discovery.

Therefore, a computer-aided drug virtual screening framework based on the binding site selectivity has been proposed in our previous work [3], which can be used to streamline the evaluation of binding of drug molecules and the specific binding site. By comparing three drug-protein binding-related metrics, including binding potential, binding affinity, and binding tendency, the selectivity of drug molecules to different binding sites in the target protein is evaluated in this framework. Traditinoal molecular simulation methods, such as molecular docking and molecular dynamics simulation [4], are integrated with machine learning-based models in the framework. The architecture of this framework is shown in Fig. 1. At first, a convolution neural network (CNN)-based deep learning model is developed to accommodate the demands for high-throughput pre-screening in modern drug discovery processes. It is used to reduce the search space by predicting the binding potential of small molecules to specific binding sites of the target protein. Next, two different scoring functions of binding affinity, obtained from molecular docking are further used to cross-verify the binding strength of small molecules. Then, an artificial neural network (ANN)-based machine learning model [5] is used to predict the binding tendency of small molecules to different binding sites according to the geometry and physical/chemical properties of the binding site structures and the interaction information between them, which provides an additional complement to the reliability of the binding potential prediction model and affinity scores. Those molecules considered to have higher pharmaceutical potential after the above screening steps are concentrated for molecular dynamics simulations to verify the dynamic binding stability. Finally, the properties of absorption, distribution, metabolism, excretion, and toxicity (ADMET) of candidate drug molecules are evaluated. The most suitable molecules will be used for manual inspection and subsequent experimental tests.

In this work, the associated molecular simulation methods, data, and machine learning models in the above screening framework will be highlighted through a more detailed case study involving small molecule inhibitors targeting the binding interface of angiotensin converting enzyme 2 (ACE2) and the receptor binding domain (RBD) of the spike protein of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) [6] to show the feasibility and practicality of the proposed framework. In this case, five candidate inhibitors from the DrugBank database which can bind readily to the ACE2-RBD binding interface and have little effect on the catalytic function of ACE2 are selected as ideal inhibitors for further experimental assays. Compared to the known inhibitors, these candidate inhibitors show higher binding site selectivity to the ACE2-RBD binding interface, which aims to decrease potential side effect risks of drug molecules.