Refining ensembles of predicted gene regulatory networks based on characteristic interaction sets

Different ensemble voting approaches have been successfully applied for reverse-engineering of gene regulatory networks. They are based on the assumption that a good approximation of true network structure can be derived by considering the frequencies of individual interactions in a large number of predicted networks. Such approximations are typically superior in terms of prediction quality and robustness as compared to considering a single best scoring network only. Nevertheless, ensemble approaches only work well if the predicted gene regulatory networks are sufficiently similar to each other. If the topologies of predicted networks are considerably different, an ensemble of all networks obscures interesting individual characteristics. Instead, networks should be grouped according to local topological similarities and ensemble voting performed for each group separately. We argue that the presence of sets of co-occurring interactions is a suitable indicator for grouping predicted networks. A stepwise bottom-up procedure is proposed, where first mutual dependencies between pairs of interactions are derived from predicted networks. Pairs of co-occurring interactions are subsequently extended to derive characteristic interaction sets that distinguish groups of networks. Finally, ensemble voting is applied separately to the resulting topologically similar groups of networks to create distinct group-ensembles. Ensembles of topologically similar networks constitute distinct hypotheses about the reference network structure. Such group-ensembles are easier to interpret as their characteristic topology becomes clear and dependencies between interactions are known. The availability of distinct hypotheses facilitates the design of further experiments to distinguish between plausible network structures. The proposed procedure is a reasonable refinement step for non-deterministic reverse-engineering applications that produce a large number of candidate predictions for a gene regulatory network, e.g. due to probabilistic optimization or a cross-validation procedure.