Fig. 1

a Social network approach to clustering patient samples. First we transform/encode the mutation/voxel data, then compute all patient–patient similarities. At each order of similarities, clustering is based on similarities in that order, resulting in different clustering solutions. Shown here from left to right: features, 1st-order, 2nd-order, 3rd-order, ‘true’ communities. Note that links between the same sample but different orders are not shown (e.g. A always has a strong self link), but are used in the similarity calculations. b Flow diagram of HOCUS analysis: Feature data (such as imaging voxels or mutations are supplied to HOCUS to generate higher-order features from which sample-sample similarities are calculated. The HOCUS order is selected by comparing sample-sample similarity kernels with an external criterion. Clustering is then done followed by survival and downstream analyses