Distinguishing causal from tagging enhancers using single-cell multiome data

This study demonstrates that pervasive non-causal "tagging" effects, driven by shared transcription factor binding sites and peak co-accessibility, confound enhancer-gene linking in single-cell multiome data, necessitating the use of fine-mapping methods like SuSiE to distinguish true causal regulatory relationships from spurious correlations.

The Gist

Imagine your body is a massive, bustling city. In this city, genes are the factories that build the products your body needs (like red blood cells or immune defenders). But factories don't just run on their own; they need instructions. These instructions come from enhancers, which are like remote control switches located all over the city. Sometimes, a switch is right next to the factory it controls, but often, it's miles away, connected by invisible wires.

The big challenge for scientists is figuring out which switch controls which factory.

The Problem: The "Echo Chamber" Effect

In the past, scientists tried to solve this by looking at a snapshot of the city. They noticed that when a specific switch (an enhancer) was "on," a specific factory (a gene) was also "on." They assumed, "Aha! That switch must control that factory!"

But the authors of this paper discovered a tricky problem: The Echo Chamber.

Imagine a row of houses. If the lights in House A turn on, the lights in House B and House C often turn on too, not because House A controls them, but because they are all plugged into the same circuit breaker. In biology, many switches are "co-accessible"—they turn on and off together because they are part of the same neighborhood or controlled by the same master switch.

When scientists just look at the correlation (the fact that they turn on together), they get fooled. They think Switch A controls Factory X, when in reality, Switch A is just a "tag" or a mimic. It's like seeing a shadow and thinking it's the person, when really, the shadow is just following the person. These are called tagging enhancers. They aren't the cause; they are just riding along.

The Solution: A New Detective Tool

To fix this, the researchers developed a new way to separate the "real bosses" from the "mimics" using a special dataset called multiome data. Think of this as a high-tech surveillance system that watches both the switches (chromatin accessibility) and the factories (gene expression) in thousands of individual cells at the same time.

They created two "scores" for every switch:

1. The Neighborhood Score (Co-accessibility): How often does this switch turn on with its neighbors?
2. The Factory Score (Co-activity): How often does this switch turn on with a specific factory?

They found that these two scores were almost identical. If a switch was popular with its neighbors, it was also popular with factories. This confirmed that most of the connections scientists had found were just "echoes" (tagging), not real cause-and-effect relationships.

How They Found the Real Bosses

So, how do you find the real switch? The researchers looked for specific clues that only the true controllers have:

• Location: The real switches are often the ones closest to the factory door (the gene's start site).
• The "Green Light" Mark: Real switches often have a specific chemical sticker on them (called H3K27ac) that says, "I am active!"
• The Master Keys: They found that the "echoes" were mostly caused by Pioneer Transcription Factors. Think of these as construction workers who break down walls to open up new areas. When these workers arrive, they flip many switches at once, creating a massive wave of activity that looks like a single switch controlling everything, but is actually just a group effort.

The Proof: The "Fine-Tuning" Test

To prove their method worked, they used a statistical tool called SuSiE (think of it as a super-precise magnifying glass). Instead of just saying "Switch A and Factory X are linked," SuSiE looks at the whole neighborhood and says, "Okay, Switch A, B, and C are all linked, but only Switch B is the actual cause."

When they tested this against real-world experiments (where they physically turned switches off using CRISPR technology), their "fine-mapped" predictions were incredibly accurate. They were much better at guessing the truth than the old methods.

Why This Matters

This is a huge deal for understanding diseases. Many diseases (like blood disorders) are linked to specific genetic switches found in large studies. But if we can't tell which switch is the real cause and which is just a "tag" (a mimic), we might waste years trying to fix the wrong switch.

In short: This paper teaches us that just because two things happen at the same time, it doesn't mean one caused the other. By using smarter math and looking at the "neighborhood" of switches, we can finally stop chasing shadows and start fixing the actual levers that control our health.

Technical Summary

1. Problem Statement

Single-cell multiome technologies (simultaneous scRNA-seq and scATAC-seq) have become a primary tool for linking enhancers (ATAC-seq peaks) to their target genes by correlating chromatin accessibility with gene expression across individual cells. However, a critical limitation exists: correlations among ATAC-seq peaks can induce non-causal "tagging" associations.

Similar to linkage disequilibrium in GWAS, if two peaks are co-accessible (open simultaneously in the same cells), a peak that is merely "tagging" a causal enhancer will appear correlated with a target gene, even if it has no direct regulatory function. The paper addresses the pervasive nature of these tagging effects and the difficulty in distinguishing true causal enhancer-gene links from these statistical artifacts.

2. Methodology

The authors developed a quantitative framework to disentangle causal signals from tagging noise using four distinct multiome datasets (comprising 86,000 cells across 6 immune/blood cell types).

• Definition of Scoring Metrics:
  • Co-accessibility Score ($S_{acc}$): For each ATAC-seq peak, this is calculated as the sum of squared correlations with all nearby peaks. It quantifies how strongly a peak is linked to the local chromatin landscape.
  • Co-activity Score ($S_{act}$): For each peak, this is the sum of squared correlations with nearby genes. It represents the standard "peak-gene" linking strength.
• Stratified Co-accessibility Regression (S-CASC): To identify functional categories enriched for causal links, the authors regressed co-activity scores against stratified co-accessibility scores. This allowed them to isolate enrichment signals within specific peak subsets (e.g., peaks near TSS, H3K27ac-marked peaks) while controlling for general co-accessibility.
• Mechanistic Analysis: The study investigated the drivers of co-accessibility by analyzing the relationship between peak-peak correlations and the presence of Transcription Factor Binding Sites (TFBS), specifically looking for shared TFs between peak pairs.
• Validation:
  • CRISPRi Data: Used to validate causal links by comparing non-causal correlations against the "tagging correlation" with a known causal peak.
  • Fine-mapping: Applied the SuSiE (Sum of Single Effects) algorithm to fine-map peak-gene associations, distinguishing causal variants from correlated non-causal ones.
  • External Benchmarks: Evaluated performance using independent CRISPRi and eQTL datasets.

3. Key Contributions

• Quantification of Tagging Effects: The study provides robust evidence that tagging effects induced by peak co-accessibility are pervasive in single-cell multiome data, often mimicking causal regulatory relationships.
• Novel Scoring Framework: Introduction of the Co-accessibility and Co-activity scores as metrics to quantify and compare the strength of local chromatin correlation versus gene association.
• Mechanistic Insight: Identification that co-accessibility is largely driven by the density of Transcription Factor Binding Sites (TFBS) and, crucially, by pioneer transcription factors that activate repressed chromatin regions.
• Improved Fine-mapping: Demonstration that fine-mapped peak-gene links (using SuSiE) significantly outperform marginal (simple correlation) links in identifying true regulatory relationships.

4. Key Results

• Strong Correlation between Scores: Across all datasets, the co-accessibility score and co-activity score were strongly correlated ($r = 0.57–0.73$). This correlation persisted after controlling for read depth, cell subtypes, and measurement noise, confirming it is driven by tagging rather than technical artifacts.
• Validation with CRISPRi: Non-causal peak-gene correlations were found to be strongly correlated ($r = 0.92$) with a peak's tagging correlation to a known causal peak in CRISPRi data, mathematically confirming the tagging hypothesis.
• Enrichment of Causal Signals: Using S-CASC, the authors identified specific functional categories where causal signals are concentrated:
  • TSS Proximity: Peaks closest to a gene's Transcription Start Site showed a 2.91x enrichment (s.e. 0.67).
  • H3K27ac Marks: Peaks overlapping active histone marks showed a 1.41x enrichment (s.e. 0.11).
• TFBS and Pioneer Factors: Co-accessibility scores were substantially driven by the number of TFBS within a peak. Peak-peak correlations were driven by pairs of peaks sharing a TF, with these effects concentrated in a small number of pioneer TFs.
• Performance of Fine-mapping: Links fine-mapped using SuSiE significantly outperformed marginal links when evaluated against CRISPRi and eQTL ground truth data.
• GWAS Relevance: The study provided specific examples where tagging effects obscured the true causal variants in GWAS of blood cell traits, demonstrating the practical impact of these findings on disease genetics.

5. Significance

This paper fundamentally shifts the understanding of how enhancers are linked to genes in single-cell multiome data. It argues that simple correlation-based approaches are insufficient due to pervasive tagging effects caused by co-accessibility.

The significance lies in:

1. Methodological Rigor: Providing a statistical framework (S-CASC, SuSiE fine-mapping) to filter out noise and identify true causal regulatory elements.
2. Biological Insight: Highlighting the role of pioneer transcription factors in creating the chromatin co-accessibility structures that lead to tagging.
3. Translational Impact: Improving the accuracy of linking non-coding GWAS variants to their target genes, which is critical for understanding the genetic architecture of complex diseases, particularly in immune and blood cell contexts.

In conclusion, the authors underscore that accounting for tagging effects is not merely a statistical refinement but a necessity for accurately reconstructing gene regulatory networks from single-cell multiome data.