Motivation: Misannotation in sequence databases is an important obstacle for automated tools for gene function annotation, which rely extensively on comparison with sequences with known function. To improve current annotations and prevent future propagation of errors, sequence-independent tools are, therefore, needed to assist in the identification of misannotated gene products. In the case of enzymatic functions, each functional assignment implies the existence of a reaction within the organism's metabolic network; a first approximation to a genome-scale metabolic model can be obtained directly from an automated genome annotation. Any obvious problems in the network, such as dead end or disconnected reactions, can, therefore, be strong indications of misannotation.
Results: We demonstrate that a machine-learning approach using only network topological features can successfully predict the validity of enzyme annotations. The predictions are tested at three different levels. A random forest using topological features of the metabolic network and trained on curated sets of correct and incorrect enzyme assignments was found to have an accuracy of up to 86% in 5-fold cross-validation experiments. Further cross-validation against unseen enzyme superfamilies indicates that this classifier can successfully extrapolate beyond the classes of enzyme present in the training data. The random forest model was applied to several automated genome annotations, achieving an accuracy of ~60% in most cases when validated against recent genome-scale metabolic models. We also observe that when applied to draft metabolic networks for multiple species, a clear negative correlation is observed between predicted annotation quality and phylogenetic distance to the major model organism for biochemistry (Escherichia coli for prokaryotes and Homo sapiens for eukaryotes).
Supplementary information: Supplementary data are available at Bioinformatics online.