High-throughput functional annotation of proteins is a fundamental task in functional proteomics. Protein functions are typically organized in the form of a general-
specific hierarchy, such as the Gene Ontology (GO), which describes when one
functional class is a specialization of its parent class. The hierarchical structure
indicates that if a protein belongs to one class then it also belongs to all ancestor classes up to the root. Most previous work on protein function prediction has
constructed independent classifiers for each function, which ignore the hierarchical
information available in the GO. We develop a framework for combining the local independent SVM predictions with graphical models, both Bayesian networks
(BNs) and Conditional Random Fields (CRFs), which are built upon the hierarchical structure in the GO. Our goal is to increase the overall predictive accuracy by
exploiting this hierarchical information. Compared to the baseline technique (i.e.
independent SVM classifiers), our techniques using BN and CRF yield significant
improvement on two large data sets constructed from the Uniprot database.