Figure 8

Effect of probe set classes and similarity on MCL clustering (m8 and m27). A – Changes in the pruning score relative to the probe set class (classes are colour-coded as indicated in panel D), and to the network size used for the clustering process. The abscissa indicates the network size reduction (in percentage) obtained by merging the alternative probe sets with a similarity ranging from 100% (6%) to 0% (35%). Size reduction 0% corresponds to the full network. B – Same data as in A, but the abscissa indicates the real size of the sub-networks used for clustering. For each probe set class, the different subsets corresponding to the different used networks are indicated by small circles (from the full network, on the right, to the network with 35% reduction, on the left). C – Inter-chip reproducibility of MCL clustering in m8 and m27. The reproducibility of one cluster corresponds to the number of common probe sets divided by the geometric mean of the cluster sizes in m8 and m27. The reproducibility of clustering is the weighted mean of the ten first clusters (the weights are the inverse of the geometric mean of the cluster sizes). D – Intra-chip reproducibility of MCL clustering in m27 in networks that were reduced by probe set merging relative to the full network.