TL;DR
Single-cell ATAC-seq (scATAC-seq) is powerful but lacks a standardized reference for analyzing chromatin accessibility, hindering cell atlas development. This paper introduces "consensus peaks" (cPeaks), a comprehensive reference derived from over 600 bulk ATAC-seq datasets, which improves cell type annotation and rare cell identification in scATAC-seq data.
What Problem Does This Paper Address?
Single-cell ATAC-seq (scATAC-seq) allows us to study chromatin accessibility at the resolution of individual cells, providing insights into gene regulation and cell states. However, unlike single-cell RNA-seq which benefits from well-established reference transcriptomes, scATAC-seq lacks a standardized "feature reference" – a consistent set of genomic regions to define as accessible. This absence makes it difficult to compare results across different studies, build comprehensive cell atlases, and accurately identify rare cell types. Current methods often involve calling peaks specific to each dataset or sample, which can lead to data loss and inconsistencies when trying to integrate information.
Methods
The core of this work is the creation and validation of a "consensus peak" (cPeak) reference. The authors first aggregated over 600 high-quality bulk ATAC-seq datasets to define an initial set of approximately 1.4 million "observed" cPeaks. They then developed a deep convolutional neural network (CNN) model, inspired by existing work, to predict new cPeaks based on DNA sequence patterns. This CNN was trained on observed cPeaks and then used to scan the genome, adding about 0.28 million "predicted" cPeaks to the reference. The consistency of these cPeaks across different tissues, sequencing technologies, and peak-calling methods was rigorously tested. Finally, the authors evaluated the performance of their cPeak reference in downstream scATAC-seq analyses, such as cell type annotation and rare cell identification, comparing it against existing methods.
Datasets
- Training/Reference Construction: 624 high-quality bulk ATAC-seq datasets from ENCODE and ATACdb.
- Validation:
- 231 independent scATAC-seq datasets for evaluating cPeak coverage.
- 789 independent bulk ATAC-seq datasets for validating predicted cPeaks.
- A hematopoietic differentiation scATAC-seq dataset (2,034 cells) for cell type annotation benchmarking.
- A peripheral blood mononuclear cell (PBMC) scATAC-seq dataset (8,728 cells) for rare cell type discovery.
- A human gynecologic malignancy scATAC-seq dataset (74,621 cells) for tumor analysis.
- A human fetal retinal scATAC-seq dataset for developmental analysis.
- Pan-Cancer TCGA dataset for tumor progression analysis.
- CATLAS dataset for rare cell subtype analysis.
Key Insights and Ideas
- Inherent Genomic Features: The consistent "shapes" of cPeaks across diverse experimental conditions (tissue types, sequencing technologies, peak-calling methods) strongly suggest they represent fundamental, inherent genomic features of chromatin accessibility.
- Deep Learning for Expansion: Using a deep CNN model to predict new cPeaks based on DNA sequence is a powerful way to expand the reference and capture regions that might be missed in experimental datasets, especially for cell types or tissues not well-represented in the initial bulk data.
- "Housekeeping" Regions: A subset of cPeaks, termed "housekeeping" cPeaks (accessible in >90% of datasets), are enriched in promoter regions and linked to essential biological pathways, suggesting a conserved role in basic cellular functions.
- Shape Matters for Regulation: cPeaks exhibit distinct "shapes" (unimodal/multimodal, well-positioned/weakly-positioned edges) that correlate with regulatory roles, particularly in development and disease. Well-positioned cPeaks appear to be hallmarks of cellular proliferation and differentiation.
Key Results
- Comprehensive Reference: The authors constructed a reference set of approximately 1.4 million observed cPeaks and 0.28 million predicted cPeaks.
- High Coverage: cPeaks demonstrated high coverage (>98% on average) across a wide range of scATAC-seq datasets.
- Superior Performance: In cell type annotation tasks, cPeaks consistently outperformed existing methods (genomic bins, pseudo-bulk peaks, united peaks, cDHS, CATLAS cCREs), especially when using feature selection.
- Rare Cell Identification: cPeaks significantly improved the identification of rare cell types and subtypes, detecting substantially more differential accessible regions compared to pseudo-bulk peaks.
- Development and Cancer Insights: cPeaks facilitated the analysis of chromatin dynamics during fetal development and tumor progression, revealing patterns of well-positioned cPeaks associated with proliferation and tumorigenesis.
Limitations and Future Work
- Initial Version: The authors acknowledge that their cPeak set is an initial version based on currently available data and that there is room for refinement in defining and annotating cPeaks.
- Cross-Species and Evolution: They suggest extending the cPeak framework to other species and studying conserved chromatin accessibility signatures to understand the evolution of regulatory elements.
- Foundation Models: cPeaks provide a groundwork for developing foundation models for chromatin accessibility data and multi-omics integration.
- Unseen Regions: While the CNN model expands coverage, there may still be unique accessible regions not captured by the current cPeak reference. Future work could focus on incorporating these into the reference.
Summary updated