Supplementary Figure 2: Pairwise similarity of module predictions from different methods.

(a) Pairwise similarity of module predictions from different methods in Sub-challenge 1, averaged over all networks. Similarity was computed based on whether the same genes were clustered together by the two methods. Specifically, a prediction vector P_mk was defined for every method m and network k, specifying for every pair of genes whether they were co-clustered in the same module (Methods). The prediction vectors P_mk of method m for the six networks (k = 1,2,...,6) were then concatenated, forming a single vector P_m representing the module predictions of that method for all six networks. A corresponding distance matrix between the 42 methods was computed as described in Methods (Equation 1) and hierarchically clustered using Ward’s method. The annotation row and column show the method type. The top five methods (1-5) and the consensus (C) are highlighted. The top methods did not converge to similar module predictions (they are not all grouped together in the hierarchical clustering). Related to Fig. 3, which shows similarity of module predictions from individual networks. (b) Comparison of trait-associated modules identified by all challenge methods. Pie-charts show the percentage of trait modules that show overlap with at least one trait module from a different method in the same network (top) and in different networks (bottom). We distinguish between strong overlap, sub-modules, weak but statistically significant overlap, and insignificant overlap (Methods).