Reasons for ensemble averaging and its drawback.

<p>A) The hypotheses space <i>H</i> (black shape) contains all networks that can be represented by the applied mathematical framework. There might be no single optimum, as several different network structures might score equally well and thus are equally valid (blue area). Additionally, optimization procedures starting from different initial parameterization might get stuck at local optima and create suboptimal predictions (red dots). If the applied framework is adequate, the reference structure is included in <i>H</i> (green square) and could be predicted by the optimization procedure. Otherwise, predicted high scoring networks should be at least similar to the reference. Here, all high scoring predicted networks are very similar to each other and to the reference. In such a case, the frequency of an interaction in all networks is a reliable indicator for the confidence of an effector-target gene relation, thus applying ensemble voting is advisable. B) Depending on the reference structure, the applied mathematical framework, and the available experimental data, several groups of topologically different high scoring networks might be predicted by a probabilistic reverse-engineering algorithm (blue areas <i>I</i>, <i>II</i> and <i>III</i>). Combining all of these structurally strongly different networks by ensemble voting would obscure characteristics of individual groups of networks. We suggest the presence of sets of co-occurring interactions as a reasonable criterion for identification and delimitation of these groups.