Computational analyses for biomedical knowledge discovery greatly benefit from the availability of the description of gene and protein functional features expressed through controlled terminologies and ontologies, i.e. of their controlled annotations. In the last years, several databases of such annotations have become available; yet, these annotations are incomplete and only some of them represent highly reliable human curated information. To predict and discover unknown or missing annotations existing approaches use unsupervised learning algorithms. We propose a new learning method that allows applying supervised algorithms to unsupervised problems, achieving much better annotation predictions. This method, which we also extend from our preceding work with data weighting techniques, is based on the generation of artificial labeled training sets through random perturbations of original data. We tested it on nine Gene Ontology annotation datasets; obtained results demonstrate that our approach achieves good effectiveness in novel annotation prediction, outperforming state of the art unsupervised methods.