Abstract: The development cycle of drugs is long and the cost is huge. The method of computerized virtual drug screening can effectively improve the efficiency of the pilot compounds. This paper proposes a new feature fusion scheme based on attention mechanism, called multi-feature fusion scheme. Combined with the existing graph convolution network based on edge attention, the biological activity prediction task is carried out by using this method for different kinds of bioactive data sets selected from PubChem, the public chemical database. The instability and unreliability caused by manual calculation can be avoided by learning the molecular graph features directly, and multi-feature fusion scheme based on attention makes the model adaptive to fuse multiple edge attribute features. The results show that the method can predict the biological activity of molecules more accurately than other machine learning methods.