In this work, we investigate the impact of incorporating emotion classes on the task of predicting emojis from Twitter texts. More specifically, we first show that there is a correlation between the emotion expressed in the text and the emoji choice of Twitter users. Based on this insight we propose a few simple methods to incorporate emotion information in traditional classifiers. Through automatic metrics, human evaluation, and error analysis, we show that the improvement obtained by incorporating emotion is significant and correlate better with human preferences compared to the baseline models. Through the human ratings that we obtained, we also argue for preference metric to better evaluate the usefulness of an emoji prediction system.