DMG bulk regulatory network

To avoid technical artifacts resulting from batch effects across datasets as well as transcriptional bias between autopsy versus biopsy specimens, we generated a consensus network across 120 DMG gene expression profiles containing autopsy vs. biopsy information, integrating individual ARACNe3 networks reverse engineered from the following sample sets: CBTTC_Autopsy_Network (15 samples), CBTTC_Biopsy_Network (30 samples), PNOC_Biopsy_Network (31 samples), StJude_Autopsy_Network (28 samples), StJude_Biopsy_Network (16 samples). We accounted for the small sample size in each individual network by setting a higher FDR threshold and integrating the results weighted by the number of samples per network so that only the most robust interactions remained. As such, each network was pruned with a mutual information threshold calculated to control the FDR at 0.50 and subsequently by MaxEnt pruning (<i>i.e.,</i> DPI pruning). Regulons from each network were integrated using a consensus scoring approach where Association Weight (<i>i.e.</i> likelihood) and Association Mode (<i>i.e.</i> tfmode) values were averaged for each target over all networks, and each networks’ data was weighted according to the number of gene expression profiles used to reverse engineer the network.The final network comprised regulons with 50 targets for 6,138 regulatory proteins, with 7,002,347 interactions, allowing inference of accurate regulons for proteins whose expression was too low to be detected at the single-cell level.