CSsingle: A Unified Tool for Robust Decomposition of Bulk and Spatial Transcriptomic Data Across Diverse Single-Cell References

Wenjun Shen, Cheng Liu, Yunfei Hu, Yuanfang Lei, Hau−San Wong, Si Wu, Xin ZhoubioRxiv 2024 · DOI

TL;DR

CSsingle is a new computational tool that accurately deconvolves cell types from bulk and spatial transcriptomic data, even when using diverse single-cell references. It achieves this by incorporating cell size correction and robustly handling technical and biological variations, leading to improved accuracy and novel biological insights, particularly in cancer research.

What Problem Does This Paper Address?

Understanding the cellular makeup of tissues is crucial for deciphering biological processes and disease mechanisms. While single-cell RNA sequencing (scRNA-seq) offers high resolution, it's costly and not always feasible for large-scale studies. Bulk RNA sequencing (bulk RNA-seq) and spatial transcriptomics (ST) are more accessible but provide averaged or spatially localized but not single-cell resolution. Deconvoluting these data types to infer cell type proportions is challenging due to:

  • Cellular Heterogeneity: Different cell types have varying RNA content due to differences in cell size, which can bias deconvolution if not accounted for.
  • Technical and Biological Variations: Differences in experimental protocols, sequencing depth, and biological states between reference scRNA-seq data and bulk/ST data can lead to inaccurate estimations.
  • Noise and Unrelated Cell Types: ST data, in particular, can be noisy, and reference datasets might contain cell types not present in the mixture.

Existing methods often struggle to address these issues simultaneously, leading to suboptimal accuracy and robustness.

Methods

CSsingle introduces a novel, unified approach for deconvolution with several key innovations:

  • Cell Size Correction: For bulk RNA-seq, CSsingle explicitly accounts for cell size differences by using ERCC spike-in controls to estimate cell size coefficients. This is a significant advancement, as previous methods often assumed equal RNA content across cell types.
  • Iteratively Reweighted Least Squares (IRLS): CSsingle employs an IRLS algorithm to minimize a weighted squared error. This robustly handles technical and biological variations between the mixture and reference data.
  • "Sectional Linear Relationship" for Initialization: The method leverages a property of marker genes to derive robust initial estimates for cell type proportions, improving the efficiency and accuracy of the IRLS algorithm.
  • Fine-Scale Spatial Analysis: For ST data, CSsingle enables the detection of enriched cell types at the single-spot resolution, rather than relying on pre-defined regions, offering a more granular view of tissue architecture.

Datasets

The study utilized a wide array of datasets for training and validation, including:

  • Cell lines: HEK and Jurkat cells (used to demonstrate cell size effects).
  • Human tissues: Pancreatic islets, peripheral blood mononuclear cells (PBMCs), normal squamous-columnar junction (N-SCJ), Barrett's esophagus (BE), esophageal adenocarcinoma (EAC), normal esophagus (NE), normal gastric cardia (NGC), normal duodenum (ND), mouse brain cortex, and human pancreas.
  • Publicly available datasets: Numerous datasets from GEO, TCGA, and other repositories were used, covering various tissues and experimental platforms (scRNA-seq, bulk RNA-seq, microarrays, and ST).
  • Simulated datasets: Artificial bulk and ST datasets were generated to benchmark performance against known ground truths.

Key Insights and Ideas

  • Cell size is a critical, often overlooked, factor in bulk deconvolution. Ignoring it can lead to systematic overestimation of larger cell types and underestimation of smaller ones.
  • A unified approach for bulk and ST data is highly valuable. CSsingle's ability to handle both data types with a single framework simplifies complex analyses.
  • Mosaic Columnar Cells (MCCs) are key players in esophageal adenocarcinoma (EAC). Their significant enrichment in BE and EAC, and their dynamic changes during immunotherapy, suggest a crucial role in disease progression and treatment response.
  • Spatial resolution matters for understanding tissue dynamics. CSsingle's ability to identify cell types at the single-spot level in ST data opens new avenues for studying cellular interactions and tissue microenvironments.

Key Results

  • Improved Accuracy: CSsingle consistently outperformed 10 conventional deconvolution methods in accuracy and robustness across various datasets, including those with significant technical and biological variations. For example, in the HEK/Jurkat cell mixture experiment, CSsingle provided accurate estimates, while other methods showed systematic biases.
  • Clinical Relevance in EAC: The study identified MCCs as a predominant cell type in BE and EAC, and their signatures were found to be predictive and prognostic markers for immunotherapy response in EAC patients.
  • Spatial Deconvolution Performance: CSsingle achieved high accuracy in estimating cell type proportions in mouse brain tissue ST data, outperforming other state-of-the-art methods.
  • Robustness Across Platforms: CSsingle demonstrated versatility by accurately deconvoluting data from diverse scRNA-seq protocols (e.g., Smart-seq2, 10x Chromium, CEL-Seq2).

Limitations and Future Work

  • Cell Size Correction in ST: While CSsingle's cell size correction is crucial for bulk RNA-seq, its impact on ST data was found to be less significant. The authors attribute this to factors like incomplete cell capture in 2D slices and spatial complexity, which can obscure cell size effects. Future work could explore more advanced methods for accounting for these ST-specific challenges.
  • Signature Matrix Construction: The paper explores different strategies for constructing signature matrices, highlighting CSsingle's flexibility. Further research could investigate optimal strategies for specific tissue types or experimental conditions.
  • Biomarker Discovery: The identification of MCCs as potential biomarkers for EAC immunotherapy response is a significant finding. Future work could involve prospective validation of these biomarkers in clinical trials.

Summary updated

Abstract

Abstract We introduce CSsingle, a novel method that enhances the decomposition of bulk and spatial transcriptomic (ST) data by addressing key challenges in cellular heterogeneity. CSsingle applies cell size correction using ERCC spike-in controls, enabling it to account for variations in RNA content between cell types and achieve accurate bulk data deconvolution. In addition, it enables fine-scale analysis for ST data, advancing our understanding of tissue architecture and cellular interactions, particularly in complex microenvironments. We provide a unified tool for integrating bulk and ST with scRNA-seq data, advancing the study of complex biological systems and disease processes. The benchmark results demonstrate that CSsingle outperforms existing methods in accuracy and robustness. Validation using more than 700 normal and diseased samples from gastroesophageal tissue reveals the predominant presence of mosaic columnar cells (MCCs), which exhibit a gastric and intestinal mosaic phenotype in Barrett’s esophagus and esophageal adenocarcinoma (EAC), in contrast to their very low detectable levels in esophageal squamous cell carcinoma and normal gastroesophageal tissue. We revealed a dynamic relationship between MCCs and squamous cells during immune checkpoint inhibitors (ICI)-based treatment in EAC patients, suggesting MCC expression signatures as predictive and prognostic markers of immunochemotherapy outcomes. Our findings reveal the critical role of MCC in the treatment of EAC and its potential as a biomarker to predict outcomes of immunochemotherapy, providing insight into tumor epithelial plasticity to guide personalized immunotherapeutic strategies.
BibTeX
@misc{shen2024cssingle,
  author = {Wenjun Shen and Cheng Liu and Yunfei Hu and Yuanfang Lei and Hau−San Wong and Si Wu and Xin Zhou},
  title = {{CSsingle: A Unified Tool for Robust Decomposition of Bulk and Spatial Transcriptomic Data Across Diverse Single-Cell References}},
  year = {2024},
  doi = {10.1101/2024.04.07.588458},
  howpublished = {bioRxiv},
}
Figure