It takes a long time for a drug to go from research and development to clinical application, and the investment cost during the period can reach one billion yuan. The combination of medicine and artificial and the development of big data of biochemistry contribute to sharply increasing drug activity data, and traditional experimental methods for drug activity prediction and discovery are hard to meet the needs of drug research and development. Algorithms are used to assist drug development and solve various problems during the process to significantly accelerate drug development. Traditional machine learning methods, especially random forests, support vector machines, and artificial neural networks, can improve drug activity prediction accuracy. Due to the multi-layer neural networks of deep learning, the model can process high-dimensional input variables and there is no need to limit the amount of input data characteristics manually. Deep learning can build a more complex function, and its application in drug research and development can further improve the efficiency of each step of drug research. Widely used deep learning models in drug activity are mainly DNN (deep neural networks), RNN(recurrent neural networks), and AE (auto encoder). GAN (generative adversarial networks) is often used in combination with other models for data enhancement due to its ability to generate data. Researches and applications of deep learning in drug molecule activity prediction in recent years showed that the accuracy and efficiency of deep learning models were higher than traditional experimental methods and traditional machine learning methods. Therefore, deep learning is expected to become the most critical auxiliary calculation model in drug research and development in the next decade.