Identifying the binding of molecules and targets is essential in drug discovery. Recently, studies using deep learning have been presented to reduce the cost of this process. Since the amount of experimental data in molecules is small, studies introducing a multi-task learning method have been published. In this study, a form of learning similar targets together based on the chemical similarity between the ligand sets of the target to be predicted was applied in the model for predicting the binding of molecules to the target. We looked into whether this method could improve performance over a single task-learning method. As a result, it was found that this multi-task learning method improves the average performance of the model, especially for targets that the model did not predict well. Furthermore, by applying the knowledge distillation technique, we examined whether this method improves performance in predicting the binding of molecules. As a result, the multi-task learning model, which received knowledge from a single work learning model, showed the highest performance.