Finding Biologically Accurate Clusterings in Hierarchical Decompositions Using the Variation of Information

Hierarchical clustering is a popular method for grouping together similar items based on a distance measure between them. These clusters can be used to infer annotations for uncharacterized items. However, in many cases, annotation information for some elements is known beforehand. We present a novel approach for decomposing a hierarchical clustering into the optimal clusters that match a set of known annotations, as measured by the variation of information metric. Our approach is general, and we apply it to two biological domains: finding protein complexes within protein interaction networks and identifying species within metagenomic DNA samples. For both applications, we test the quality of our clusters by using them to predict complex and species membership. We find that our approach generally outperforms the commonly used heuristic methods.