Figure 2

Network modelling of epigenetic genes – phenotypes in ALL derived from PGnet. The complete tripartite network produced by PGNet is available as Suppl. Fig. 1 (Additional file 14) and Suppl. Table 2 (Additional file 3). Here, panel (a) shows a bipartite subset of 33 statistically significant associations linking 23 distinct epigenetic seed genes (ESGs, yellow triangle) to 10 distinct leukemia phenotypes (yellow squares). Distinct background colors of ESG gene names indicate distinct molecular function or biological process related to or targeted by epigenetic regulation (Suppl. Table 1 (Additional file 2)). Two LPs and one ESG with asterisk were selected to show the details of a biomodule derived by PGnet (Figure 4). Red genes are those for which the expression is up-regulated in the associated ALL phenotype, while blue ones are with down-regulated expression, and grey ones are related to more than one phenotype with alternate up-down regulations (details of the full network in Suppl. Table 2 (Additional file 3)). Panel (b) shows a tripartite network that includes GEMs (grey circles) and focuses on the circled subset of Panel a: the "ALL relapse". Panel (c) is a robust sub-network associated to ALL outcome (relapse vs. continuous complete remission (CCR)). Three ESGs and 53 GEMs were obtained by 100 repetitions of 3-fold cross-validation of PGnet operating on the subset of ALL arrays comprising 87 patients experiencing either "CCR" or "relapse" (n > 32, details in Suppl. Methods (Additional file 1)). Note that ALL subtypes associated PGnet in panel a) and b) were derived from all 132 patients (in this case the biomodules of the sub-network 2b pertain to patients with relapse and not everything else). The robust sub-network in panel c) was conducted for training a predictor of outcome. Only 87 out of the 132 samples contained outcome information related to ALL relapse ("relapse" or "CCR" of ALL). Please note that some patients did have an outcome of secondary AML, a distinct form of leukemia, and were excluded from panel 2c because this disease occurs at a later stage and the authors of the dataset did not disclose the outcome of ALL for these patients. Thus the biomodules of Figure 2c overlap partially with those of Figure 2b.