Motivation: Despite advances in the gene annotation process, the functions of a large portion of gene products remain insufficiently characterized. In addition, the in silico prediction of novel Gene Ontology (GO) annotations for partially characterized gene functions or processes is highly dependent on reverse genetic or functional genomic approaches. To our knowledge, no prediction method has been demonstrated to be highly accurate for sparsely annotated GO terms (those associated to fewer than 10 genes). Results: We propose a novel approach, information theory-based semantic similarity (ITSS), to automatically predict molecular functions of genes based on existing GO annotations. Using a 10-fold cross-validation, we demonstrate that the ITSS algorithm obtains prediction accuracies (precision 97% recall 77%) comparable to other machine learning algorithms when compared in similar conditions over densely annotated portions of the GO datasets. This method is able to generate highly accurate predictions in sparsely annotated portions of GO, where previous algorithms have failed. As a result, our technique generates an order of magnitude more functional predictions than previous methods. A 10-fold cross validation demonstrated a precision of 90% at a recall of 36% for the algorithm over sparsely annotated networks of the recent GO annotations (about 1400 GO terms and 11000 genes in Homo sapiens). To our knowledge, this article presents the first historical rollback validation for the predicted GO annotations, which may represent more realistic conditions than more widely used cross-validation approaches. By manually assessing a random sample of 100 predictions conducted in a historical rollback evaluation, we estimate that a minimum precision of 51% (95% confidence interval: 43-58%) can be achieved for the human GO Annotation file dated 2003.