A neural network based model effectively predicts enhancers from clinical ATAC-seq samples

Abstract

Enhancers are cis-acting sequences that regulate transcription rates of their target genes in a cell-specific manner and harbor disease-associated sequence variants in cognate cell types. Many complex diseases are associated with enhancer malfunction, necessitating the discovery and study of enhancers from clinical samples. Assay for Transposase Accessible Chromatin (ATAC-seq) technology can interrogate chromatin accessibility from small cell numbers and facilitate studying enhancers in pathologies. However, on average, ~35% of open chromatin regions (OCRs) from ATAC-seq samples map to enhancers. We developed a neural network-based model, Predicting Enhancers from ATAC-Seq data (PEAS), to effectively infer enhancers from clinical ATAC-seq samples by extracting ATAC-seq data features and integrating these with sequence-related features (e.g., GC ratio). PEAS recapitulated ChromHMM-defined enhancers in CD14+ monocytes, CD4+ T cells, GM12878, peripheral blood mononuclear cells, and pancreatic islets. PEAS models trained on these 5 cell types effectively predicted enhancers in four cell types that are not used in model training (EndoC-βH1, naïve CD8+ T, MCF7, and K562 cells). Finally, PEAS inferred individual-specific enhancers from 19 islet ATAC-seq samples and revealed variability in enhancer activity across individuals, including those driven by genetic differences. PEAS is an easy-to-use tool developed to study enhancers in pathologies by taking advantage of the increasing number of clinical epigenomes.

ATAC-seq signal processing and recurrent neural networks can identify RNA polymerase activity

“…experiments. Our results with histone marks are consistent with recent efforts to utilize ATAC-seq to identify regions labeled by ChromHMM as enhancers (Thibodeau et al, 2018). However, here we infer the ChIP-seq histone modification data directly from nucleotide-resolution ATAC-seq peaks.…”

“…It has been utilized in a 1 wide range of applications, from classifying types of chronic lymphocytic leukemia cells (Rendeiro et al, 2016), to TF motif discovery (Setty and Leslie, 2015), or discriminating among brain cell types (Fullard et al, 2018). Moreover, a recent study (Thibodeau et al, 2018) attempted to identify gene enhancer regions using ATAC-seq peaks.…”

A pitfall for machine learning methods aiming to predict across cell types

“…This approach involves training a model in one or more cell types and then evaluating it in one or more other cell types (Figure 1b). Researchers have used this approach to identify cis-regulatory elements [13][14][15][16][17][18], impute epigenomics assays that have not yet been experimentally peformed [19,20], and predict CpG methylation [21]. The cross-cell type strategy is typically adopted when the goal is to yield predictions in cell types for which experimental data is not yet available.…”

Combining signal and sequence to detect RNA polymerase initiation in ATAC-seq data

“…Machine learning is a natural tool to classify data derived from genomics assays, particularly ATAC-seq. A wide range of machine learning applications for ATAC-seq datasets have been developed, from classifying types of chronic lymphocytic leukemia cells [10], to TF motif discovery [11], discriminating among brain cell types [12], and identifying gene enhancer regions using ATAC-seq peaks [13]. Given regions of polymerase initiation are dense with transcription factor binding motifs and have a characteristics sequence bias [6], we reasoned that any predictor would benefit from leveraging sequence information.…”