CoRE-ATAC

CoRE-ATAC classifies cis-regulatory elements (cis-REs) identified by ATAC-seq into functional categories (promoters, enhancers, insulators) to predict their regulatory roles from chromatin accessibility data.

Key Features:

• Deep learning framework: Uses a deep learning model to predict functional classes of cis-REs from ATAC-seq data.
• Multi-modal data encoders: Integrates DNA sequence (reference or personal genotypes) with ATAC-seq cut sites and read pileups via novel encoders.
• Training regimen: Trained on datasets from four cell types with six samples per replicate.
• Cross-cell-type performance: Evaluated on seven additional cell types not included in training, achieving mean average precision 0.80 and mean F1 score 0.70.
• Enhancer activity prediction: Predicts enhancer activity in human islet samples and identifies cis-REs with gain or loss of function associated with genetic alterations, validated by MPRAs.
• Single-nucleus ATAC applicability: Infers cis-RE functions from aggregate single nucleus ATAC-seq (snATAC) data from human blood-derived immune cells.
• Increased functional resolution: Enhances the functional interpretation of ATAC-seq maps to resolve individual- and cell-specific variations in cis-RE activity.

Scientific Applications:

• Enhancer activity mapping: Predicts and validates enhancer activity changes in human islet samples using MPRA validation.
• Immune cell regulatory annotation: Infers cis-RE functions from aggregate snATAC of blood-derived immune cells to annotate regulatory elements in rare populations.
• Cross-cell-type annotation: Enables functional annotation of accessible chromatin across multiple cell types, including cell types not used in training.
• Disease-related regulatory variation: Supports investigation of individual- and cell-specific regulatory disruptions that may underlie disease.

Methodology:

Implements a deep learning framework with novel encoders that combine DNA sequence (reference or personal genotypes), ATAC-seq cut sites, and read pileups; trained on four cell types with six samples per replicate and evaluated on seven additional cell types (reported mean average precision 0.80 and mean F1 0.70).