Detection of differentially abundant cell subpopulations discriminates biological states in scRNA-seq data

Jun Zhao, Ariel Jaffe, Henry Li, Ofir Lindenbaum, Esen Sefik, Ruaidhrí Jackson, Xiuyuan Cheng, Richard A. Flavell, Yuval KlugerbioRxiv 2019 · DOI

TL;DR

This paper introduces DA-seq, a novel algorithm that goes beyond traditional clustering to identify subtle, differentially abundant cell subpopulations in single-cell RNA sequencing (scRNA-seq) data. By uncovering these hidden cell groups, DA-seq provides deeper biological insights into complex diseases and developmental processes that were previously missed.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) generates vast amounts of data, allowing researchers to study cellular heterogeneity. A common task is comparing cells from different biological states (e.g., healthy vs. diseased, different time points). Traditional methods often rely on clustering cells into predefined groups and then comparing the proportions of cells from each state within these clusters. However, this approach can miss important differences if the differentially abundant (DA) subpopulations don't align neatly with existing clusters or if the differences are subtle and spread across multiple clusters. This limitation can lead to overlooking crucial biological insights, especially in complex scenarios like disease progression or development where cellular dynamics are intricate.

Methods

DA-seq tackles this problem with a multi-scale, unbiased approach. It involves four key steps:

  1. Multiscale Score Calculation: For each cell, DA-seq computes a "score vector" that quantifies the relative abundance of cells from the two biological states within its local neighborhood. This is done across multiple neighborhood sizes (scales) to capture differences at various resolutions.
  2. DA Measure Computation: The multiscale scores are then fed into a logistic regression classifier. This classifier learns to predict the biological state of a cell based on its neighborhood composition, producing a single "DA measure" for each cell. Cells with extreme DA measures (highly enriched in one state's neighborhood) are flagged.
  3. Clustering DA Cells: The flagged DA cells are then clustered using a graph-based approach (specifically, a shared nearest neighbor graph) in their original gene expression space. This step identifies contiguous regions of cells that exhibit differential abundance.
  4. Gene Marker Identification: Finally, DA-seq identifies characteristic gene markers for each detected DA subpopulation. This is achieved using a novel embedded feature selection method called stochastic gates (STG), which aims to find a minimal set of genes that best distinguish the DA subpopulation from other cells.

Datasets

The authors applied DA-seq to several publicly available scRNA-seq datasets, including:

  • Melanoma patients: Data from Sade-Feldman et al. [7] comparing immune cells from responders and non-responders to checkpoint immunotherapy.
  • Embryonic mouse dermal cells: Data from Gupta et al. [14] comparing cells from embryonic days E13.5 and E14.5 to study hair follicle development.
  • COVID-19 patients: Data from Chua et al. [5] comparing immune cells from severe and moderate COVID-19 cases.
  • Aging mouse brain: Data from Ximerakis et al. [15] comparing brain cells from young and old mice.
  • Simulated datasets: Two simulated datasets were used for benchmarking, one based on [7] and another a perturbed Gaussian mixture model.

The paper also mentions validation using data from Fan et al. [26] and Liao et al. [6].

Key Insights and Ideas

  • Beyond Clusters: DA-seq's core innovation is its ability to detect DA subpopulations that are not confined to predefined clusters, allowing for the discovery of novel cell states.
  • Multiscale Approach: By considering neighborhoods at different scales, DA-seq can capture DA signals that might be missed by single-scale methods, whether too broad or too narrow.
  • Unbiased Discovery: The algorithm does not require prior knowledge of cell types or markers, making it a powerful tool for exploratory data analysis.
  • Revealing Hidden Biology: The method consistently uncovered new DA subpopulations in existing datasets, leading to novel biological insights that were overlooked by previous analyses.

Key Results

  • DA-seq successfully recovered known DA subpopulations identified in previous studies across various biological systems (melanoma, embryonic development, COVID-19, aging brain).
  • Crucially, DA-seq also identified new DA subpopulations in these datasets that were not reported in the original publications.
  • For example, in the melanoma dataset, DA-seq identified a subpopulation (DA4) that was enriched in responders and highly expressed CTLA4, a gene previously associated with non-responders in other cell types, suggesting a nuanced role in treatment response.
  • In the COVID-19 study, DA-seq identified distinct neutrophil subpopulations (DA1 and DA5) with differential abundance between severe and moderate cases, providing insights into inflammatory mechanisms.
  • The method's gene selection approach (STG) provides a more interpretable and potentially more accurate characterization of DA subpopulations compared to standard differential expression analysis.

Limitations and Future Work

  • Parameter Sensitivity: The choice of parameters, particularly the thresholds for identifying DA cells (τ_L, τ_H) and the range of k-values for neighborhood analysis, can influence the results. While the paper provides guidance, careful selection is necessary.
  • Computational Complexity: Constructing k-nearest neighbor graphs can be computationally intensive for very large datasets, though the authors mention fast approximate approaches.
  • Future Directions: The authors suggest exploring neural network classifiers directly on input features as a potential improvement for DA-seq. They also envision extending DA-seq to compare multiple biological states simultaneously (beyond pairwise comparisons) by adapting the classifier to a multi-class setting. Furthermore, DA-seq could be applied to analyze cell populations based on specific marker expression within a single experiment.

Summary updated

Abstract

Abstract Traditional cell clustering analysis used to compare the transcriptomic landscapes between two biological states in single cell RNA sequencing (scRNA-seq) is largely inadequate to functionally identify distinct and important differentially abundant (DA) subpopulations between groups. This problem is exacerbated further when using unsupervised clustering approaches where differences are not observed in clear cluster structure and therefore many important differences between two biological states go entirely unseen. Here, we develop DA-seq, a powerful unbiased, multi-scale algorithm that uniquely detects and decodes novel DA subpopulations not restricted to well separated clusters or known cell types. We apply DA-seq to several publicly available scRNA-seq datasets on various biological systems to detect differences between distinct phenotype in COVID-19 cases, melanomas subjected to immune checkpoint therapy, embryonic development and aging brain, as well as simulated data. Importantly, we find that DA-seq not only recovers the DA cell types as discovered in the original studies, but also reveals new DA subpopulations that were not described before. Analysis of these novel subpopulations yields new biological insights that would otherwise be neglected.
BibTeX
@misc{zhao2019detection,
  author = {Jun Zhao and Ariel Jaffe and Henry Li and Ofir Lindenbaum and Esen Sefik and Ruaidhrí Jackson and Xiuyuan Cheng and Richard A. Flavell and Yuval Kluger},
  title = {{Detection of differentially abundant cell subpopulations discriminates biological states in scRNA-seq data}},
  year = {2019},
  doi = {10.1101/711929},
  howpublished = {bioRxiv},
}
Figure