Composition of selected approaches for different GO term sizes. In our ensemble method, three different approaches based on SVM were applied, including bagged single SVMs, SVMs with hierarchical correction, and Bayesian integration of single-dataset SVMs. Each of them exhibits good performance in a different domain, which is indicated in this figure by the percentage of terms where each of the different methods exhibited the best performance for four different size categories. For smaller GO terms, the single SVM and hierarchical correction often achieve superior performance. For larger GO terms, Bayesian integration of diverse datasets performed better than the other two methods. GO, Gene Ontology; SVM, support vector machine.