Fig. 2: Networks of subtypes produced by six different methods.

The networks are displayed for a METABRIC-A, b METABRIC ER−, c METABRIC ER+. The nodes are colored and labeled according to the particular subtyping method. An edge between nodes indicates that the corresponding subtypes had a significant number of samples in common. Dashed circles indicate the nodes (subtypes) that were clustered together by the MCL network clustering method. a Application of methods A–F resulted in subtype systems with 3, 5, 5, 6, 3, 3 subtypes, respectively. The three subtypes identified by Method A were labeled A1–A3, and similarly for other methods. A network was formed having nodes these 25 subtypes. Markov clustering of this network resulted in five clusters of subtypes (identified with dashed ovals), plus the singleton D3 (n = 31). For each cluster of subtypes, a subset of METABRIC-A (labeled T1–T5) that formed the core samples of the cluster was identified, with remaining samples in T0 (n = 92, 9.2%). b Methods A–F applied to this domain resulted in 4, 4, 4, 3, 3, 3 subtypes, respectively. Markov cluster analysis of the network of 21 subtypes identified 3 clusters of subtypes, with three unclustered subtypes. Core samples assigned to each cluster formed NT1–3 with unclassified samples comprising NT0. c In METABRIC ER+ samples, Methods A–F resulted in 5, 4, 6, 8, 3, 3 clusters, respectively. Markov cluster analysis applied to the network of 29 clusters resulted in four clusters that grouped 9, 5, 8, and 7 nodes, respectively. Core sample assignment to these clusters created subtypes PT1–4 and unclassified samples as PT0.