 This paper proposes a novel approach to drug discovery based on protein sequences. The authors demonstrate its effectiveness by applying it to three different tasks, predicting the binding affinity between two molecules, interpreting the binding knowledge learned from the model, and identifying potential new targets for existing drugs. This work provides a promising alternative to traditional structure based drug design methods, especially for proteins, without high-quality 3D structures. This article was authored by Lifan Chen, Sishun Fan, Jie Chang, and others.