Deciphering context-dependent epigenetic program by network-based prediction of clustered open regulatory elements from single-cell chromatin accessibility

The paper introduces enCORE, a computational framework that utilizes enhancer-enhancer interaction networks to identify clustered open regulatory elements (COREs) from single-cell ATAC-seq data, thereby enabling the deciphering of context-dependent epigenetic programs and the discovery of disease-relevant regulators in processes like hematopoiesis and colorectal cancer.

The Gist

Imagine your body is a massive, bustling city. Every cell in your body is a unique building in this city—a library, a hospital, a power plant, or a bakery. Even though every building is made of the exact same blueprint (your DNA), they all do very different jobs. How does a liver cell know to filter blood while a skin cell knows to protect you from the sun?

The answer lies in the control switches scattered throughout the blueprint. These switches are called enhancers. They tell the cell's machinery which genes to turn on and which to keep off.

The Problem: The "Pixelated" Map

For a long time, scientists looked at these switches one by one, like trying to understand a city by looking at individual lightbulbs. This is called "element-centric" analysis. But in reality, these switches don't work alone. They form neighborhoods or districts where they all talk to each other, creating a powerful, coordinated signal to run the cell.

Think of a Super-Enhancer as a "Downtown District" where all the major lights are on, creating a massive glow that defines the city's identity.

The problem is that when we look at a single cell (a single building) using current technology, the signal is very faint and "pixelated." It's like trying to see the whole downtown district through a foggy, broken window. We can see a few lights, but we miss the big picture of how the whole neighborhood works together. This makes it hard to understand how cells change during development or how diseases like cancer hijack these neighborhoods.

The Solution: enCORE (The "City Planner's AI")

The authors of this paper created a new tool called enCORE. You can think of enCORE as a smart city planner that uses a network map to reconstruct the whole neighborhood, even when the view is foggy.

Here is how it works, using a simple analogy:

1. The Foggy Window (scATAC-seq): The researchers start with data from single cells. It's messy and incomplete, like a map where some streets are missing and some lights are flickering.
2. Connecting the Dots (The Network): Instead of just looking at one light, enCORE looks at how the lights "blink together." If two switches in different parts of the genome tend to turn on and off at the same time, enCORE assumes they are neighbors in the same district.
3. Building the Neighborhood (COREs): It groups these connected switches into Clustered Open Regulatory Elements (COREs). These are the reconstructed "downtown districts" of the cell.
4. Ranking the Districts: Not all districts are equal. Some are just quiet suburbs (poised/ready for the future), and some are the bustling city centers (active right now). enCORE has two modes:
   • Potential Mode: Looks at all districts, including the ones that are just getting ready to wake up.
   • Active Mode: Filters out the sleepy suburbs to focus only on the districts currently running the show.

What Did They Discover?

Using this new "city planner," the researchers looked at three different scenarios:

1. The Immune System (The Police Force)
They looked at blood cells. Just like a police force has different units (traffic, SWAT, detectives), blood cells have different types (T-cells, B-cells, Monocytes).

• The Finding: enCORE perfectly identified the "headquarters" for each unit. It found the specific neighborhoods of switches that make a cell a T-cell or a B-cell. It also found that these neighborhoods are where the "risk factors" for diseases like Lupus or Asthma are hidden. It's like finding out that the reason a city has traffic jams isn't one broken light, but a specific district of poorly coordinated intersections.

2. Blood Development (The Growth of a City)
They tracked how stem cells (the raw construction materials) turn into mature blood cells.

• The Finding: enCORE could see the "future plans" of the city. Even in early stem cells, it detected the faint outlines of the neighborhoods that would eventually become macrophages (a type of immune cell). It's like seeing the blueprint for a future skyscraper in a pile of dirt before the first brick is laid. This helps us understand how cells "decide" their fate.

3. Colorectal Cancer (The Rogue City)
Cancer is like a city where the zoning laws have been broken. The "downtown districts" are lighting up in the wrong places, causing chaos.

• The Finding: enCORE found the specific rogue neighborhoods driving the cancer. It identified a gene called USP7 as a key culprit.
• The Breakthrough: They didn't just find the problem; they simulated a "fix." Using a computer model, they "turned off" the rogue neighborhood (the CORE) and predicted that the cancer gene (USP7) would calm down. This suggests that targeting this specific neighborhood could be a new way to treat colorectal cancer.

The Big Picture

Before this, scientists were trying to understand a complex city by counting individual lightbulbs. enCORE allows them to see the neighborhoods, the traffic patterns, and the districts that define who the cell is and what it's doing.

It's a powerful new lens that helps us see the "big picture" of our biology, one cell at a time, offering new hope for understanding diseases and finding better treatments.

Technical Summary

1. Problem Statement

Large-scale cis-regulatory domains, such as super-enhancers (SEs), are critical for orchestrating robust, cell-state-specific transcriptional programs that define cellular identity and drive disease. However, current single-cell analysis methods face significant limitations:

• Element-Centric Bias: Standard single-cell ATAC-sequencing (scATAC-seq) workflows focus on discrete peaks (individual enhancers), failing to capture the cooperative, domain-level interactions essential for robust gene regulation.
• Loss of Context: Bulk-level assays (e.g., ChIP-seq for H3K27ac) average signals across heterogeneous cell populations, masking cell-state-specific regulatory logic and obscuring the dynamic nature of these domains during development and disease.
• Data Sparsity: Single-cell data is inherently sparse, making it difficult to infer higher-order chromatin structures and enhancer clusters directly from raw accessibility data without a specialized computational framework.

There is a critical need for a method to infer Clustered Open Regulatory Elements (COREs) directly from scATAC-seq data to resolve domain-level epigenetic regulation in a cell-state-aware manner.

---

2. Methodology: The enCORE Framework

The authors developed enCORE, a computational framework that predicts clustered open regulatory elements by integrating three key characteristics of highly interactive enhancer clusters: chromatin accessibility, co-accessibility (inferred interactions), and transcription factor (TF) motif information.

Algorithmic Workflow:

1. Enhancer Candidate Extraction:
   • Identifies potential enhancers based on co-accessibility (correlation coefficient $\ge$ 0.2) and gABC scoring (a modified Activity-By-Contact model) to link enhancers to target genes.
   • Filters candidates located >2kb from Transcription Start Sites (TSS).
2. Initial Clustering & Correction:
   • Constructs an initial co-accessibility network and decomposes it into disconnected components.
   • Link Loss Correction: Merges disconnected components within 500kb if they show high cosine similarity in summed gABC scores, compensating for scATAC-seq sparsity.
   • Distance Threshold Optimization: Automatically calculates a genomic distance threshold to partition merged clusters, maximizing the separation between intra-cluster and inter-cluster distances.
   • Sparsity Correction: Re-includes high-accessibility peaks that were initially filtered out if they show strong co-accessibility with existing cluster members.
3. Network Construction:
   • Builds a pseudo-directed graph where nodes are enhancer candidates and edges represent interaction strength.
   • Edge Weights: Calculated based on co-accessibility, inverse genomic distance, and product of accessibility values.
   • Node Weights: Derived from MinMax-scaled accessibility and TF motif enrichment scores (z-scores).
   • Directionality: Inspired by electronegativity, edge weights are redistributed asymmetrically based on the total accessibility of the connected enhancer clusters.
4. Ranking and Selection:
   • Applies Personalized PageRank to rank enhancer clusters based on network topology and node weights.
   • Identifies the "elbow point" in the ranked distribution to define the final COREs.
5. Two Operational Modes:
   • Potential Option: Captures the original COREs, including poised/primed states (future-responsive).
   • Active Option: Filters COREs by promoter accessibility to exclude poised states, focusing strictly on currently active regulatory programs.

---

3. Key Contributions

• First scATAC-seq Domain Inference: enCORE is the first framework to predict domain-scale enhancer clusters (COREs) solely from scATAC-seq data, moving beyond element-centric analysis.
• Network-Based Topology: Unlike SEs defined by signal intensity (e.g., H3K27ac ChIP-seq), enCORE defines clusters based on network topology (clustering coefficients) and cooperative accessibility, capturing a distinct layer of regulatory architecture.
• Dual-Mode Flexibility: The "Potential" vs. "Active" options allow researchers to study either current cell states or future developmental potentials (poised states).
• Integration of Multi-omics: The framework successfully integrates accessibility, co-accessibility, and TF motifs to reconstruct complex biological hierarchies.

---

4. Key Results

A. Validation in Peripheral Blood (PBMC)

• Cell-Type Specificity: CORE-associated genes showed significantly higher cell-type specificity compared to non-CORE active enhancers and typical enhancers (TEs), mirroring the behavior of Super-Enhancers (SEs).
• Master Regulators: enCORE successfully recovered lineage-defining master regulators (e.g., GATA3 in NK/T cells, EBF1 in B cells, MAFB in myeloid cells) specific to their respective cell types.
• Comparison with SEs: While COREs overlapped with SEs, they captured a unique subset of genes (e.g., KLF4, RBPJ) missed by SEs, suggesting COREs identify distinct regulatory layers.
• Superiority over scCUT&Tag: enCORE outperformed SEs derived from single-cell H3K27ac CUT&Tag data in identifying master regulators, demonstrating robustness against the sparsity of single-cell histone modification data.

B. Chromatin Interactions and Disease Variants

• Interaction Enrichment: CORE constituents exhibited significantly higher chromatin interaction frequencies (validated by Hi-C and HiChIP data) compared to non-CORE regions and SEs, particularly in poised enhancer states.
• GWAS Variant Enrichment: COREs were highly enriched for fine-mapped GWAS variants associated with immune diseases (e.g., SLE, CLL, Asthma, T1D).
  • Example: CLL risk variants were enriched in B-cell COREs; Asthma variants in T-cell COREs.
  • Novel Insight: COREs captured RXRA upregulation in CLL, a gene missed by SE analysis, linking it to disease pathogenesis.
  • CAD Risk: Coronary Artery Disease (CAD) risk loci were specifically enriched in monocyte-derived COREs, implicating LIPA regulation.

C. Hematopoietic Lineage Reconstruction (Bone Marrow)

• Lineage Trajectory: Applying enCORE to Bone Marrow Mononuclear Cells (BMMC) successfully reconstructed the hematopoietic hierarchy, showing continuity from Hematopoietic Stem Cells (HSC) to mature lineages.
• Developmental Continuity: CORE profiles revealed that Bone Marrow monocytes contain early epigenetic signatures of macrophage differentiation (e.g., DCSTAMP, MEF2A, BHLHE41) that are attenuated in SE-based summaries, highlighting enCORE's ability to detect "anticipatory" epigenetic imprints.

D. Colorectal Cancer (CRC) Reprogramming

• Disease Stratification: enCORE effectively distinguished CRC epithelial cells from normal cells using H3K27ac profiles, outperforming SE-based classification.
• Therapeutic Targets:
  • Identified exclusive enrichment of the FOXM1 motif in COREs.
  • Discovered USP7 as a potential therapeutic candidate exclusively linked to COREs.
  • In Silico Validation: Computational perturbation (AlphaGenome) of the USP7-proximal CORE predicted a reduction in USP7 expression, supported by HiChIP loop data linking the CORE to the USP7 promoter.
• Metastasis Signatures: COREs showed higher clustering coefficients and enriched metastasis-related genes compared to SEs, suggesting a role in invasion-associated transcriptional programs.

---

5. Significance and Impact

• Deciphering Epigenetic Logic: enCORE provides a powerful tool to decode the "regulatory grammar" of the genome at the single-cell level, revealing how coordinated enhancer clusters drive cell identity and disease.
• Bridging Sparsity and Complexity: By leveraging network topology, enCORE overcomes the sparsity limitations of scATAC-seq to infer higher-order structures that are invisible to traditional peak-calling methods.
• Clinical Relevance: The framework successfully links non-coding disease risk variants (GWAS) to specific cell states and identifies novel therapeutic targets (e.g., USP7 in CRC), offering a pathway from epigenetic discovery to actionable hypotheses.
• Future Directions: The authors propose that domain-level representations will become increasingly valuable as single-cell atlases expand, enabling the study of state-resolved epigenetic programs in development, immunity, and cancer.

In summary, enCORE represents a paradigm shift from element-centric to domain-centric single-cell epigenomics, offering a robust, network-based approach to understanding context-dependent gene regulation.