Robust enhancer-gene regulation identified by single-cell transcriptomes and epigenomes

Fangming Xie, Ethan J. Armand, Zizhen Yao, Hanqing Liu, Anna Bartlett, M. Margarita Behrens, Yang Eric Li, Jacinta Lucero, Chongyuan Luo, Joseph R. Nery, António Pinto‐Duarte, Olivier Poirion, Sebastian Preißl, Angeline Rivkin, Bosiljka Tasic, Hongkui Zeng, Bing Ren, Joseph R. Ecker, Eran A. MukamelbioRxiv 2021 · DOI

TL;DR

This paper introduces a novel computational framework to accurately link gene regulatory elements (enhancers) to the genes they control, overcoming the challenge of gene co-expression that plagues existing methods. By integrating multiple single-cell molecular data types and employing a stringent statistical approach, they identified over 10,000 robust enhancer-gene pairs in the mouse brain, providing a more reliable map of gene regulation.

What Problem Does This Paper Address?

Understanding how genes are regulated is fundamental to biology. Cis-regulatory elements (CREs), such as enhancers, play a crucial role by controlling gene expression. Identifying which enhancer regulates which gene is challenging, especially in complex tissues like the brain, which contains hundreds of cell types. Traditional methods often rely on correlating gene expression with enhancer activity. However, genes with similar functions tend to be expressed together (co-expression), leading to spurious correlations that falsely link enhancers to the wrong genes. This paper aims to develop a method that can reliably distinguish genuine enhancer-gene interactions from these co-expression artifacts.

Methods

The authors developed a multi-step computational approach to identify robust enhancer-gene links:

  1. Data Integration: They integrated single-cell transcriptomic (scRNA-seq) and epigenomic (snATAC-seq for chromatin accessibility and snmC-seq for DNA methylation) data from the mouse primary motor cortex.
  2. Metacell Generation: To handle the complexity of single-cell data and integrate modalities, they generated "metacells." These are aggregated profiles from groups of similar cells, allowing for a more stable and adjustable resolution for analysis.
  3. Correlation Analysis: They correlated enhancer epigenetic signals with gene expression across these metacells.
  4. Stringent Statistical Framework: This is the core innovation. They developed a permutation-based procedure to control for gene co-expression. This involved comparing observed correlations against two null distributions:
  5. Shuffling Metacells: This decouples epigenetic and transcriptomic signals, effectively removing both enhancer-gene correlation and gene co-expression.
  6. Shuffling Regions: This retains the gene co-expression structure but randomly assigns enhancers, serving as a more stringent control for genuine enhancer-gene links.
  7. Pairs passing the "shuffling regions" null distribution were termed "linked," while those passing the less stringent "shuffling metacells" null were termed "correlated."
  8. Cross-Validation: Predicted links were further validated using independent 3D chromatin conformation data (snm3C-seq).

Datasets

The study utilized a large dataset of over 200,000 single cells from the mouse primary motor cortex, comprising:

  • scRNA-seq (single cell)
  • snmC-seq (single nucleus; DNA methylation)
  • snATAC-seq (single nucleus; chromatin accessibility)

These datasets were publicly available from a previous study.

Key Insights and Ideas

  • Gene co-expression is a major confounder: The paper strongly demonstrates that simple correlation-based methods are highly susceptible to false positives due to co-expressed genes, making it difficult to identify true regulatory relationships.
  • A stringent null model is crucial: The "shuffling regions" null distribution proved to be a more effective control for identifying genuine enhancer-gene links compared to shuffling metacells, as it better accounts for gene co-expression.
  • Multi-modal data integration is powerful: Combining transcriptomic and epigenomic data, especially at the single-cell level, provides a richer picture of gene regulation than any single modality alone.
  • 3D chromatin conformation data validates physical interactions: The correlation between predicted enhancer-gene links and chromatin contact frequency supports the idea that these links represent actual physical interactions in the genome.
  • Cell type granularity matters: The choice of cell type resolution (number of metacells) can influence the number of identified links, highlighting the importance of careful parameter selection.

Key Results

  • The authors identified over 10,000 robust enhancer-gene pairs in the mouse brain.
  • Their stringent method ("linked" pairs) identified a subset of "correlated" pairs that showed stronger associations, suggesting they are more likely to be genuine enhancer-gene interactions.
  • For example, the gene Stim2 was linked to a nearby enhancer with high confidence, and this link was further supported by 3D chromatin contact data.
  • Comparisons with existing methods like CICERO and the ABC model showed that their "linked" pairs had higher scores, indicating stronger associations.

Limitations and Future Work

  • Weak interactions are hard to detect: The authors acknowledge that their correlation-based method has limitations in detecting weak enhancer-gene interactions that might not rise above the background noise of gene co-expression.
  • Causality requires experimental validation: While their method identifies strong candidates for causal interactions, true causality cannot be inferred from correlational analysis alone.
  • Future work: The authors suggest that future experimental validation, including large-scale assays, will be needed to confirm the predicted links. They also note that larger datasets with more diverse tissues or cell types could help address the limitation of not identifying constitutive enhancer-gene links present in all cell types.

Summary updated

Abstract

Abstract Integrating single-cell transcriptomes and epigenomes across diverse cell types can link genes with the cis -regulatory elements (CREs) that control expression. Gene co-expression across cell types confounds simple correlation-based analysis and results in high false prediction rates. We developed a procedure that controls for co-expression between genes and integrates multiple molecular modalities, and used it to identify >10,000 gene-CRE pairs that contribute to gene expression programs in different cell types in the mouse brain.
BibTeX
@misc{xie2021robust,
  author = {Fangming Xie and Ethan J. Armand and Zizhen Yao and Hanqing Liu and Anna Bartlett and M. Margarita Behrens and Yang Eric Li and Jacinta Lucero and Chongyuan Luo and Joseph R. Nery and António Pinto‐Duarte and Olivier Poirion and Sebastian Preißl and Angeline Rivkin and Bosiljka Tasic and Hongkui Zeng and Bing Ren and Joseph R. Ecker and Eran A. Mukamel},
  title = {{Robust enhancer-gene regulation identified by single-cell transcriptomes and epigenomes}},
  year = {2021},
  doi = {10.1101/2021.10.25.465795},
  howpublished = {bioRxiv},
}
Figure