Chromatin Accessibility

Accessible regions are fundamentally characterized by nucleosome-depleted regions (NDRs) where canonical histone octamers are evicted or repositioned. This physical clearance exposes bare DNA to transcription factors and the transcriptional machinery. Deep learning models like DeepSEA and Enformer implicitly learn the sequence determinants of nucleosome positioning—including poly(dA:dT) tracts that resist bending and specific dinucleotide periodicities—to predict accessibility directly from raw sequence.

Different experimental assays capture distinct aspects of accessibility:

• ATAC-seq: Uses Tn5 transposase to simultaneously fragment and tag open chromatin. Provides high signal-to-noise ratio with low cell input requirements.
• DNase-seq: Uses DNase I endonuclease to digest nucleosome-depleted DNA. Offers nucleotide-resolution footprinting capability.
• FAIRE-seq: Enriches for nucleosome-depleted DNA via phenol-chloroform extraction. No enzymatic bias but lower resolution.
• MNase-seq: Digests linker DNA between nucleosomes, providing the inverse signal—protected rather than accessible regions.

Raw sequencing reads are transformed into continuous accessibility signals through a standardized computational pipeline:

1. Adapter trimming and quality filtering of raw FASTQ files
2. Alignment to reference genome using splice-aware aligners like Bowtie2 or BWA
3. Filtering of mitochondrial reads and PCR duplicates
4. Tn5 shift correction (ATAC-seq): offsetting + strand reads by +4 bp and - strand reads by -5 bp to account for transposase binding footprint
5. Peak calling with algorithms like MACS2 to identify statistically enriched regions over background
6. Signal normalization to generate bigWig coverage tracks for model input

Chromatin accessibility is highly cell-type-specific and dynamically remodeled during development, differentiation, and disease. A genomic region may be accessible in hepatocytes but occluded in neurons. This specificity is driven by pioneer transcription factors—such as FOXA1 and PU.1—that bind nucleosomal DNA and initiate chromatin opening. Multi-task deep learning architectures exploit this by simultaneously predicting accessibility across hundreds of cell types from the same DNA sequence, learning shared sequence grammar while capturing cell-type-specific regulatory logic through learned embeddings.

Accessible regions are not monolithic; they are classified into distinct regulatory categories based on shape, size, and co-occurring epigenomic marks:

• Promoters: Broad, symmetrical accessibility peaks near transcription start sites, marked by H3K4me3
• Enhancers: Narrower, often asymmetric peaks distal to genes, marked by H3K27ac and H3K4me1
• Insulators: Bound by CTCF, creating sharp accessibility boundaries that demarcate topologically associating domains
• Silencers: Accessible regions that repress transcription, often marked by H3K27me3 Models like BPNet resolve these distinctions at base-pair resolution by predicting the exact binding profile shape.

Deep learning models have revealed that DNA sequence alone is highly predictive of chromatin accessibility. Key sequence features include:

• Transcription factor motif density: Clusters of TF binding motifs create cooperative binding environments
• GC content: Accessible regions often exhibit elevated GC content due to CpG island proximity at promoters
• Short tandem repeats: Specific repeat expansions can alter local nucleosome positioning
• Sequence context beyond motifs: Flanking nucleotides influence TF binding affinity through DNA shape features (minor groove width, roll, propeller twist) The Enformer architecture captures these determinants across 200 kb contexts, modeling long-range enhancer-promoter interactions.