The ultimate goal of drug discovery is to create molecules with desired biological properties. This is obviously a difficult task because the molecular space is tremendously large and discrete with a wide variety of molecules. For example, there are only 10^8 molecules synthesized, but 10^60 molecules are estimated to be existing. Computer-aided molecular design is attracting great attention as a promising solution for efficient drug discovery. An accurate computational method allows us to find molecules with target properties but it entails huge computational costs for high-throughput virtual screening over known databases. As an alternative strategy, we propose to use deep learning techniques, which promise more accurate and yet fast predictions. We have developed several deep learning models showing substantially better performance in molecular property predictions than conventional methods. Then, we applied them to virtual screening of chemical compounds via drug-target interactions. In addition, a graph-based deep generative model is used for scaffold-based molecular design to mimic what medicinal chemists do. In this talk, we present the performance of those methods for drug discovery and raise some issues that should be addressed to overcome current limitations.