Technical supplement to “ Consistent probabilistic outputs for protein function prediction ”

Protein function prediction, in the context of the Gene Ontology, is a task that consists of answering, for a fixed protein X, a large number of binary questions of the form: " Does protein X belong to GO term Y ? " Those binary classification problems are strongly related because the ontology consists of nested classes. Two natural requirements for this prediction problem are • that the set of predictions be consistent, i.e., that if a protein is assigned a GO term, then it is all also assigned all the ancester GO terms, and • that high-confidence predictions can be produced with a quantified confidence level. Methods of structured classification proposed in machine learning Taskar et al. [2003] could in theory be used to tackle this problem. However, two practical difficulties that need to be surmounted are the large amount of missing data and the large scale of the 1