Figure 2
From: A neural network based model effectively predicts enhancers from clinical ATAC-seq samples
ATAC-seq profiles and enhancer predictions in different human cell types. (a) Pairwise spearman’s correlations of genome-wide ATAC-seq read distributions for studied cell types: GM12878, CD14+, PBMC, CD4+ T, EndoC-βH1, naïve CD8+ T, K562, MCF7, and islets (n = 19). Samples first clustered based on their lineage, then the cell type, then the individuals. (b) Distribution of ChromHMM annotations for ATAC-seq peaks called (OCRs) in CD4+ T, GM12878, CD14+ monocytes, PBMCs, EndoC-βH1, naïve CD8+ T, K562, MCF7, and one representative islet sample. For each analysis, ChromHMM annotations in the cognate cell type were used. Note that 19–50% of ATAC-seq peaks mapped to enhancers. (c) The log2 ratio of normalized features for group means: enhancer/promoter and enhancer/other in nine different cell types. A representative islet sample is shown here. Note that enhancers have different data characteristic than promoters and “other” regulatory elements and these characteristics were conserved across cell types. (d) Receiver operating characteristic (ROC) area under the curve (AUC) values based on five-fold cross-validation using different algorithms: neural network, random forest, support vector machines (SVM), k-nearest neighbor (KNN), quadratic discriminate analysis (QDA), and naïve Bayes. Neural network models are utilized in the PEAS framework based on their overall performance.
