Hierarchical cost-sensitive algorithms for genome-wide gene function prediction

In this work we propose new ensemble methods for the hierarchical classification of gene functions. Our methods exploit the hierarchical relationships between the classes in different ways: each ensemble node is trained “locally”, according to its position in the hierarchy; moreover, in the evaluation phase the set of predicted annotations is built so to minimize a global loss function defined over the hierarchy. We also address the problem of sparsity of annotations by introducing a cost-sensitive parameter that allows to control the precision-recall trade-off. Experiments with the model organism S. cerevisiae, using the FunCat taxonomy and seven biomolecular data sets, reveal a significant advantage of our techniques over “flat” and cost-insensitive hierarchical ensembles.