A new MIT framework uses machine learning to simultaneously predict molecular properties and generate new molecules using only a small amount of data for training.