.In the early stage of drug development, high-throughput screening is carried out using experimental assays. It is currently estimated that there are over 10⁶³ potential compounds as drug candidates. Consequently, there has been growing interest in virtual drug screening as a cost-effective and time-efficient approach. Virtual screening methods can be divided into two types: an AI-based approach, which leverages machine learning models trained on existing experimental data, and a docking simulation, which is based on three-dimensional structures of target proteins. Whereas various techniques have been proposed in both approaches, a significant gap still exists in the virtual and real-world scenarios, such as imbalanced data and dynamics properties of molecules. In this presentation, we will introduce our efforts in practical evaluation of protein-compound evaluation in both the AI-based and the structure-based approaches. In the former, we have improved the model's generalizability with self-training method to address the lack of experimentally validated negative samples in the public databases. In the latter, we have successfully achieved molecular dynamics-based protein-drug screening by utilizing the supercomputer Fugaku. These achievements represent significant advances in next-generation computer-assisted drug discovery.