Comparing bulk and single-cell methodologies and models to profile gene expression, chromatin accessibility and regulatory links in endothelial cells treated with TNFα

This study demonstrates that while bulk and single-cell RNA/ATAC sequencing profiles of TNFα-treated endothelial cells reveal similar biological pathways, the choice between bulk and single-cell methodologies significantly impacts the predictions of enhancer-to-gene regulatory models, leading to different prioritizations of causal genes for complex traits like coronary artery disease.

The Gist

Imagine your body is a massive, bustling city. Inside every cell of this city, there is a library (the DNA) containing the blueprints for how to build and run everything. However, most of the instructions in the library aren't for building the main structures; they are for the switches (regulatory elements) that turn specific lights on or off.

Scientists have spent years finding broken switches in the DNA of people with heart disease (like Coronary Artery Disease) or high blood pressure. But here's the problem: Finding a broken switch is easy; figuring out which light it controls is incredibly hard. A switch might be located in a hallway far away from the room it controls, and the blueprint doesn't always say which room it belongs to.

This paper is a comparison of two different ways to solve this "switch-to-light" mystery in the cells that line your blood vessels (endothelial cells).

The Two Detectives: The Crowd vs. The Individual

The researchers set up a controlled experiment. They took human blood vessel cells and "stressed" them out with a chemical called TNFα (simulating inflammation, like when you have an infection or heart disease). Then, they asked two different teams of detectives to map out which switches control which lights.

1. The "Crowd" Detective (Bulk Method)
Imagine taking a bucket of 10,000 cells, mashing them all into a blender, and measuring the average noise level.

• How it works: This is the traditional method. It gives you a very clear, loud signal of what the average cell is doing. It's great at hearing the big, obvious changes.
• The Analogy: It's like listening to a choir. You hear the beautiful harmony of the whole group perfectly, but you can't tell if the person in the back row is singing a slightly different note.

2. The "Individual" Detective (Single-Cell Method)
Imagine walking up to each of the 10,000 cells one by one and asking, "What are you doing right now?"

• How it works: This is the new, fancy method. It looks at every single cell individually. It's great at spotting unique behaviors and differences between cells.
• The Analogy: It's like interviewing every single choir member. You might miss the quietest singers (because they are hard to hear individually), but you discover that the person in the back row is actually singing a different tune than the rest!

The Big Discovery: They Don't Always Agree

The researchers expected both detectives to come back with the same map. And for the most part, they did! Both methods agreed on the general "vibe" of the city: the inflammation was real, and the same major pathways were active.

However, when it came to the specific "broken switches" linked to heart disease, the detectives disagreed.

Here is where the paper gets interesting:

• The "Crowd" Detective looked at a specific broken switch and said, "This controls the BCAR1 light."
• The "Individual" Detective looked at the exact same broken switch and said, "No, that one actually controls the CFDP1 light."

In another case, the "Individual" detective found a link to a gene called IL6R (which is crucial for inflammation) that the "Crowd" detective completely missed.

Why Does This Matter?

Think of it like a detective trying to catch a criminal.

• If you use the Crowd method, you might arrest the wrong suspect (BCAR1) because the evidence looked good on average.
• If you use the Individual method, you might catch the right suspect (CFDP1) or find a new suspect (IL6R) that the crowd method ignored.

The paper shows that the tool you choose changes the answer you get.

The Takeaway for Everyone

If you are a scientist trying to cure heart disease, you can't just pick one method and stick with it.

• If you only use the "Crowd" method, you might waste years trying to fix the wrong gene.
• If you only use the "Individual" method, you might miss some important clues that are only visible when you look at the big picture.

The Lesson: To truly understand how our genes cause disease, we need to use both detectives. We need to listen to the choir and interview the individuals. Only by combining these two perspectives can we be sure we are fixing the right "light switch" to treat heart disease and high blood pressure.

In short: The map you draw depends entirely on the lens you look through.

Technical Summary

1. Problem Statement

Genome-wide association studies (GWAS) have identified thousands of non-coding genetic variants associated with complex diseases, but pinpointing the specific causal genes regulated by these variants remains a major challenge.

• The Gap: Statistical models used to predict "enhancer-to-gene" (regulatory) links rely on gene expression and chromatin accessibility data. Historically, these models (e.g., ABC) were trained on bulk data (averages of cell populations). Recently, single-cell (sc) multiome technologies (simultaneous scRNA-seq and scATAC-seq) have enabled the development of new models (e.g., scE2G) that capture cell-to-cell variability.
• The Question: It is currently unknown whether predictions from bulk-based models and single-cell-based models are consistent when applied to the same biological system. If they differ, it could significantly impact the prioritization of causal genes for functional validation in disease contexts.

2. Methodology

The authors utilized a well-controlled in vitro experimental system to directly compare bulk and single-cell modalities.

• Experimental System: Immortalized human aortic endothelial cells (teloHAEC) under three conditions:
  1. Non-treated (NT).
  2. Treated with TNFα for 4 hours.
  3. Treated with TNFα for 24 hours.
  • Rationale: TNFα induces a robust inflammatory response relevant to coronary artery disease (CAD) and hypertension.
• Data Generation:
  • Bulk Data: Re-analyzed existing bulk RNA-seq, ATAC-seq, H3K27ac ChIP-seq, and HiC data (GSE126200).
  • Single-Cell Data: Generated new 10X Genomics Multiome data (scRNA-seq + scATAC-seq) for the same cell types and conditions.
• Analytical Pipeline:
  1. Concordance Check: Compared differential gene expression (DEG) and differential open chromatin peaks (DOP) between bulk and sc methods.
  2. Regulatory Link Prediction:
     • Bulk Model: Applied the ABC (Activity-by-Contact) model using bulk ATAC-seq, H3K27ac, and HiC data.
     • Single-Cell Model: Applied the scE2G model using sc-multiome data.
  3. Validation via GWAS:
     • Used LD score regression to test heritability enrichment for Coronary Artery Disease (CAD) and Diastolic Blood Pressure (DBP).
     • Fine-mapped GWAS summary statistics using RSparsePro to identify 95% credible sets.
     • Calculated "Causal Signal Density" (sum of Posterior Inclusion Probabilities normalized by genomic size) to assess how well predicted links capture causal variants.
  4. Gene Prioritization: Compared the predicted links against Open Targets association scores for CAD to see which model better connected variants to known causal genes.

3. Key Contributions

• Direct Comparison: First study to systematically compare bulk vs. sc regulatory link predictions (ABC vs. scE2G) within the exact same experimental system (teloHAEC + TNFα).
• Model Discrepancy Identification: Demonstrated that while bulk and sc data profiles are biologically concordant, the predictive models built on them often yield divergent regulatory links.
• Impact on GWAS Interpretation: Showed that the choice of methodology (bulk vs. sc) can change the prioritized causal gene for a specific GWAS locus, potentially leading to different functional hypotheses.

4. Key Results

A. Data Concordance (Bulk vs. sc)

• Expression & Accessibility: Bulk and sc RNA-seq/ATAC-seq results were highly concordant in terms of fold-change correlations and enriched biological pathways.
• Sensitivity Differences: Bulk methods detected significantly more differentially expressed genes (643 vs. 177) and differentially open peaks, likely due to higher statistical power to detect small fold-changes and lower noise compared to sc methods.

B. Regulatory Link Predictions

• Overlap: Only ~15–32% of predicted regulatory links were shared between the ABC (bulk) and scE2G (sc) models.
• Structural Differences:
  • scE2G heavily favored promoter-proximal links (short distances).
  • ABC (bulk) predicted more distal links, as it explicitly excludes promoters from the "distal" regulatory definition.
• Divergent Gene Assignments: The models frequently assigned the same fine-mapped GWAS variant to different genes.
  • Example 1 (IL6R): Only the scE2G model linked CAD variants to the IL6R gene; the ABC model did not.
  • Example 2 (BCAR1/CFDP1): At a specific locus, ABC linked variants to BCAR1, while scE2G linked the same variants to the neighboring gene CFDP1.

C. GWAS Heritability and Fine-Mapping

• Heritability Enrichment: All subsets of regulatory links (ABC-only, scE2G-only, and overlapping) captured significant heritability for CAD and DBP.
• Causal Signal Density: After normalizing for the size of the regulatory regions, all models captured a similar density of causal signals. This suggests that while the specific genes linked differ, the models are generally capturing the correct genomic regions associated with disease.
• Gene Prioritization:
  • Links predicted by both models showed the highest enrichment for Open Targets CAD genes.
  • However, 44 unique CAD candidate genes were linked only by the ABC model, and 59 unique genes were linked only by the scE2G model.
  • The Open Targets association scores for genes linked by scE2G-only were generally higher than those linked by ABC-only, suggesting sc models might be better at identifying high-confidence causal genes in some contexts.

5. Significance and Implications

• Methodological Choice Matters: The study concludes that the choice between bulk and single-cell methodologies is not merely a technical preference but fundamentally alters the list of prioritized causal genes for GWAS loci.
• Experimental Planning: Researchers planning functional follow-up experiments (e.g., CRISPR editing) to validate GWAS hits must consider that a "negative" result in one model might be a "positive" result in the other.
• Future Directions:
  • The authors suggest that integrating both approaches (combining the high sensitivity of bulk data with the cell-type resolution of sc data) may provide the most robust framework for predicting regulatory links.
  • The findings highlight the need to validate predictions with orthogonal methods (e.g., HiC, eQTLs) before committing to expensive functional assays.

In summary, this paper provides critical evidence that bulk and single-cell enhancer-to-gene models are not interchangeable. While they agree on broad biological signals, they diverge significantly at the specific gene-variant level, necessitating a multi-faceted approach to interpreting GWAS data in complex diseases like CAD.