Prediction of gene function from expression profiles is an intriguing problem that has been attempted with both unsupervised clustering and supervised learning methods. By the incorporation of prior knowledge concerning gene function, supervised methods avoid some of the problems with clustering. However, even supervised methods ignore the fact that the functional classes associated with genes are typically organized in an ontology. Hence, we introduce a new supervised method for learning in such an ontology. It is tested on both an artificial data set and a data set containing measurements from human fibroblast cells. We also give an approach for measuring the classification performance in an ontology.