scATAC-seq

The count matrix is overwhelmingly sparse, typically with >95% zeros. This arises because a diploid cell has only two copies of any given locus, so chromatin accessibility at a single-cell resolution is fundamentally a binary or near-binary event. Unlike scRNA-seq, which models transcript counts, scATAC-seq is often binarized to indicate open or closed chromatin. This sparsity demands specialized imputation and dimensionality reduction methods like Latent Semantic Indexing (LSI) rather than those designed for count distributions.

The raw data consists of Tn5 transposase cut sites rather than direct molecular counts. Each accessible region generates paired-end sequencing reads that define a DNA fragment. Key data features include:

• Fragment size distribution: Nucleosome-free regions produce short fragments (<147 bp), while nucleosome-bound regions yield longer, periodic fragment lengths.
• Cut site aggregation: Counts are aggregated into peaks or genome-wide windows for analysis.
• Insertion bias: Tn5 has a sequence preference that must be computationally corrected.

Unlike gene-centric scRNA-seq, scATAC-seq directly profiles cis-regulatory elements such as promoters, enhancers, and insulators. The fundamental unit of analysis is the peak—a genomic region of enriched Tn5 insertions. A single gene may be regulated by multiple distal enhancers, making the feature space vast. A typical experiment identifies 50,000–300,000 peaks across the genome, creating a high-dimensional feature space where each peak represents a candidate regulatory element.

Chromatin accessibility profiles are highly cell-type discriminant. Master transcription factors drive cell identity by binding to enhancers and creating accessible chromatin domains. This results in:

• Clustered accessibility patterns: Cells of the same type share peaks at lineage-defining gene loci.
• TF footprinting: Deep sequencing coverage within peaks reveals protected DNA motifs where transcription factors are bound, enabling TF activity inference.
• Dynamic elements: Transient regulatory elements active during differentiation are captured in trajectory analyses.

scATAC-seq assays the entire genome in an unbiased manner, but the signal at any single locus in an individual cell is sparse. This creates a trade-off between breadth and depth:

• Breadth: The assay captures open chromatin across all chromosomes.
• Depth: Per-locus coverage is low, often 1–2 fragments per peak per cell.
• Aggregation requirement: Cells must be grouped into clusters or metacells to achieve sufficient coverage for peak calling, motif enrichment, or TF footprinting. This is a fundamental distinction from bulk ATAC-seq.

Multiple technical factors introduce structured noise that must be computationally modeled:

• Transposition efficiency: Variable Tn5 activity across cells creates differences in library complexity.
• GC content bias: Regions with extreme GC content show biased amplification.
• Mapping ambiguity: Repetitive genomic regions cause multi-mapping reads that are typically discarded.
• Batch effects: Differences in cell lysis, transposition time, and sequencing depth across samples require integration methods like Harmony or scVI adapted for binary data.

The computational fusion of scATAC-seq with scRNA-seq and other modalities into a unified latent representation. Methods like Seurat WNN learn cell-specific modality weights to capture holistic cellular states, linking distal regulatory elements identified by chromatin accessibility to the gene expression programs they control. This resolves the central challenge of connecting enhancer usage to transcriptional output.

The reconstruction of transcription factor–target gene interactions from single-cell data. scATAC-seq provides the critical cis-regulatory evidence—open chromatin at promoters and enhancers—that complements co-expression networks. Tools like SCENIC+ integrate accessibility and expression data to map the regulatory logic controlling cell identity, identifying master regulators that govern lineage commitment.

A supervised approach that projects cell-type annotations from a well-characterized reference atlas onto a new query dataset. For scATAC-seq, this involves mapping cells into a shared latent space with scRNA-seq references using methods like canonical correlation analysis or anchor-based integration, enabling annotation of chromatin-defined populations without requiring prior knowledge of marker genes.

The alignment of multiple scATAC-seq datasets from different conditions, donors, or technologies into a shared embedding. Algorithms like Harmony apply mixture model-based corrections to harmonize peak matrices while preserving biological variation. This is essential for large-scale cell atlas projects that must correct for technical batch effects without erasing genuine epigenetic heterogeneity.

The computational ordering of cells along continuous developmental paths based on chromatin accessibility dynamics. By tracking the opening and closing of regulatory elements across pseudotime, researchers identify the sequence of transcription factor binding events that drive differentiation. This reconstructs epigenetic remodeling during processes like hematopoiesis from static snapshot data.

Large-scale pretrained transformer models like Geneformer and scGPT that learn universal cell representations from massive single-cell corpora. When fine-tuned on scATAC-seq data, these models leverage contextual attention mechanisms to predict chromatin accessibility patterns, identify regulatory elements, and perform zero-shot cell-type annotation, dramatically reducing the need for task-specific training.