Fig. 5: Predicting gene dysregulations in disease using a DNN.
From: Deep neural network prediction of genome-wide transcriptome signatures – beyond the Black-box
a We tested if the DNN with two hidden layers of 250 hidden nodes each could be used to predict causative changes in disease states. We did this by analyzing gene expression changes from known diseases, as available in the Expression Atlas repository. By applying the disease changes to the transcription factor input layer, we could observe how these changes projected down to the target genes. Next, we removed the disease-fold changes of each TF independently and observed the changes in correlation between predicted and observed dysregulation of the target genes. Thus, we could rank the TFs on predicted causative disease changes on the target genes. b TF rankings significantly overlapped with GWAS in 10 of 22 diseases. (−log10 P values shown as red bars). The test was repeated with significantly differentially expressed TFs removed from the set, leaving 7 TF rankings to overlap with GWAS. (−log10 P values shown as teal bars). The corresponding area under the curves for the TF rankings (all TFs) are shown to the right, with 20 of 22 diseases having an AUROC greater than the expected, as generated under the null hypothesis.
