Figure 5
From: A neural network based model effectively predicts enhancers from clinical ATAC-seq samples
Enhancer predictions in individuals’ islets. (a) Schema of our framework for predicting enhancers from islet ATAC-seq profiles of 19 individuals. (b) Distribution of Receiver operating characteristic (ROC) areas under the curve (AUC) values, precision recall (PRC) AUC values, and accuracies (enhancer probability > 0.5) for PEAS enhancer predictions in 19 individuals. Note that these models are consistently predictive across 19 individuals. (c) PEAS enhancer probability distributions for OCRs containing rs11100782 (left panel) and rs7320023 (right panel) stratified based on individuals’ genotypes, where genotypes are ordered with respect to the allelic impact on chromatin accessibility. Note that PEAS enhancer probabilities correlate with genotypes for these two loci. (d) Left panel: Genome browser session for the islet caQTL OCR that contain rs11100782 variant. This OCR is a ChromHMM enhancer and predictions at the individual level using PEAS are depicted under chromatin accessibility profile for each individual. If a peak is not called for an individual at this locus, PEAS do not provide predictions (hence no bars and probabilities). Individual samples are ordered with respect to genotypes, starting from ‘AA’ genotype that is associated with open chromatin. Right panel: Genome browser session for an islet caQTL OCR that contain rs7320023 variant to summarize PEAS predictions at the individual level. Note that this region has not been annotated as an enhancer using ChromHMM. Islet samples are sorted based on genotypes starting from ‘GG’ genotype associated with open chromatin. (e) Distribution of point biserial correlations between PEAS probabilities and genotypes (e.g., in Fig. 5c) for all studied caQTLs and random OCRs that harbor variants. P-value was calculated using the Mann-whitney U test. Note that association between genotype and PEAS probabilities are specific to islet caQTLs. (f) Distribution of ChromHMM annotations and PEAS annotations for T2D-disease state associated OCRs. Note that PEAS predictions are at the individual level, therefore a genomic region is considered multiple times across individuals. Using enhancer predictions at the individual level improved the annotations for these disease-associated loci.
