Detecting copy number variations from single-cell chromatin sequencing data by AtaCNV

Xiao-Chen Wang, Zijie Jin, Yang Shi, Ruibin XibioRxiv 2023 · DOI

TL;DR

This paper introduces AtaCNV, a novel computational algorithm designed to accurately detect copy number variations (CNVs) from single-cell ATAC-sequencing (scATAC-seq) data. By addressing the inherent noise and sparsity of scATAC-seq, AtaCNV enables more precise identification of tumor subclones and characterization of cancer cell heterogeneity, which is crucial for understanding cancer evolution and developing targeted therapies.

What Problem Does This Paper Address?

Single-cell ATAC-sequencing (scATAC-seq) is a powerful tool for mapping chromatin accessibility at the single-cell level, offering insights into cellular states and tumor heterogeneity. A key application of scATAC-seq in cancer research is the identification of tumor subclones and distinguishing them from normal cells, often relying on the detection of copy number variations (CNVs). However, detecting CNVs from scATAC-seq data is notoriously challenging. This is primarily due to:

  • High Noise and Sparsity: scATAC-seq data is inherently noisy and sparse, making it difficult to discern true biological signals from technical artifacts.
  • Confounding Factors: Unlike DNA sequencing, scATAC-seq read depth is influenced not only by CNVs but also by variations in chromatin accessibility across the genome. This means a region with low read counts could be due to a copy number loss or simply inaccessible chromatin.
  • Unknown CNV Structure: The number and location of CNVs are generally unknown beforehand, requiring robust methods to infer them from the data.

Existing methods for CNV detection from scATAC-seq either suffer from limited resolution or require complementary DNA sequencing data, leaving a gap for a standalone, high-resolution method.

Methods

The core of this paper is the development of AtaCNV, a computational algorithm that tackles the challenges of CNV detection in scATAC-seq data through a multi-step approach:

  1. Preprocessing and Quality Control: Standard procedures are applied to raw scATAC-seq data, including fragment filtering, doublet removal, and quality control using the ArchR package.
  2. Binning and Read Count Matrix Generation: The genome is divided into 1 Mbp bins, and a bin-by-cell read count matrix is generated. Regions with low mappability or excessive zero counts are removed.
  3. Smoothing: To mitigate noise, a one-order dynamic linear model is used to smooth the read counts within each cell's bin matrix. This helps to denoise the signals.
  4. Normalization: This is a critical step. AtaCNV normalizes the smoothed read counts to account for factors beyond CNVs, such as chromatin accessibility and mappability.
  5. Baseline Selection: The algorithm first aims to identify a "baseline" read depth from normal cells. If normal cells are explicitly provided, their median read depth across bins is used. If not, AtaCNV employs clustering to identify a high-confidence group of normal cells within the tumor sample.
  6. Deconvoluting CNVs: The normalized read depths are then used to estimate copy number ratios, effectively separating the CNV signal from chromatin accessibility effects.
  7. Joint Segmentation: AtaCNV applies a multi-sample BIC-seq algorithm to jointly segment all single cells. This means that breakpoints are identified across multiple cells simultaneously, allowing the algorithm to pool information and improve denoising and signal detection. This joint segmentation is key to achieving high resolution.
  8. CNV Burden Score: A score is calculated for each cell to quantify its CNV burden, which is then used to distinguish malignant from non-malignant cells.

Datasets

AtaCNV was benchmarked and validated using a combination of simulated data and real-world scATAC-seq datasets from various cancer types:

  • Simulated Data: The simATAC package was used to generate synthetic scATAC-seq data with simulated CNV events. This allowed for controlled evaluation of AtaCNV's performance under different conditions (e.g., varying coverage, noise levels, number of subclones, and CNV sizes – arm-level and small-scale 10-20 Mbp).
  • Real scATAC-seq Datasets:
    • Basal Cell Carcinoma (BCC) dataset (paired normal and tumor samples).
  • Pancreatic Ductal Adenocarcinoma (PDAC) datasets.
  • Adult Glioblastoma (aGBM) datasets.
  • Pediatric Glioblastoma (pGBM) datasets.
  • Ovarian Cancer (OC) dataset.
  • Endometrial Cancer (EC) dataset.
  • 10X PBMC dataset (used as a reference for simulation).

The paper also mentions the use of matched whole-exome sequencing (WES) data for BCC samples and paired scRNA-seq data for gynecologic cancers to validate findings.

Key Insights and Ideas

  • Bridging the Gap: AtaCNV successfully bridges the gap in detecting CNVs from scATAC-seq data, a task previously hindered by noise and confounding factors.
  • Importance of Normalization: The paper highlights that proper normalization, specifically accounting for chromatin accessibility, is crucial for accurate CNV detection from scATAC-seq. Simple normalization is insufficient.
  • Joint Segmentation Power: The use of multi-sample joint segmentation significantly improves the accuracy and resolution of CNV detection by leveraging information across multiple cells.
  • High-Resolution CNV Detection: AtaCNV's ability to detect small-scale CNVs (10-20 Mbp) is critical, as these can represent important genetic differences between tumor subclones that might be missed by lower-resolution methods.
  • Interplay of Genetics and Epigenetics: The identification of subclones with distinct CNV profiles and cellular states suggests a complex interplay between genetic (CNVs) and epigenetic (chromatin accessibility) plasticity in driving tumor heterogeneity and evolution.

Key Results

  • Superior Performance: AtaCNV demonstrated superior performance in both simulated and real-world scATAC-seq datasets compared to existing methods like Copy-scAT and a simple baseline method, showing higher precision and recall.
  • Accurate Tumor/Normal Classification: AtaCNV effectively distinguished malignant from non-malignant cells across multiple cancer types, often with higher accuracy than Copy-scAT.
  • Subclone Identification: The algorithm successfully identified distinct tumor subclones with unique CNV profiles in glioblastoma, endometrial, and ovarian cancer samples. These subclones often exhibited different transcriptomic features, highlighting their functional significance.
  • Robustness: AtaCNV's performance remained largely consistent across varying sequencing coverages, noise levels, and numbers of tumor subclones in simulations.
  • Detecting Small CNVs: In some cases, AtaCNV identified focal CNVs that were missed by Copy-scAT, underscoring its advantage in high-resolution detection. For example, in EC samples, AtaCNV detected amplifications at 10q22 and 19q12, while Copy-scAT only detected the amplification at 10q22.

Limitations and Future Work

  • Dependence on Normal Cells: While AtaCNV has strategies for selecting normal cells when they are not explicitly provided, its performance is generally better when high-confidence normal cells are available for normalization.
  • Computational Cost: While not explicitly detailed, complex algorithms like joint segmentation can be computationally intensive, which might be a consideration for very large datasets.
  • Integration with Multi-omics: The authors suggest that integrating scATAC-seq data with other single-cell modalities, such as scRNA-seq, could further improve CNV detection and provide a more comprehensive understanding of tumor heterogeneity. They showed that CNV profiles inferred from scATAC-seq and scRNA-seq were largely consistent, suggesting the feasibility of multi-omics CNV analysis.
  • Further Validation: While validated across multiple cancer types, continued application and validation in diverse cancer contexts and with larger cohorts would further solidify AtaCNV's utility.

Summary updated

Abstract

Abstract Singe-cell assay of transposase accessible chromatin sequencing (scATAC-seq) can unbiasedly profile genome-wide chromatin accessibility in single cells. In single-cell tumor studies, identification of normal cells or tumor clonal structures often rely on copy number variations (CNVs). However, CNV detection from scATAC-seq is difficult due to the high noise, sparsity, and confounding factors. Here, we describe AtaCNV, a computational algorithm that accurately detects high resolution CNVs from scATAC-seq data. We benchmark AtaCNV using simulation and real data and find AtaCNV’s superior performance. Analyses of 10 scATAC-seq datasets shows that AtaCNV could effectively distinguish malignant from non-malignant cells. In glioblastoma, endometrial and ovarian cancer samples, AtaCNV identifies subclones at distinct cellular states, suggesting important interplay between genetic and epigenetic plasticity. Some tumor subclones only differ in small-scale CNVs, demonstrating the importance of high-resolution CNV detection. These data show that AtaCNV can aid the integrative analysis for understanding the complex heterogeneity in cancer.
BibTeX
@misc{wang2023detecting,
  author = {Xiao-Chen Wang and Zijie Jin and Yang Shi and Ruibin Xi},
  title = {{Detecting copy number variations from single-cell chromatin sequencing data by AtaCNV}},
  year = {2023},
  doi = {10.1101/2023.10.15.562383},
  howpublished = {bioRxiv},
}
Figure