LMD: Cluster-Independent Multiscale Marker Identification in Single-cell RNA-seq Data

Ruiqi Li, Rihao Qu, Fabio Parisi, Francesco Strino, Hon‐Ming Lam, Jay S. Stanley, Xiuyuan Cheng, Peggy Myung, Yuval KlugerbioRxiv 2023 · DOI

TL;DR

This paper introduces LMD, a novel method for identifying "localized genes" – markers specifically expressed in distinct cell groups – from single-cell RNA sequencing data. LMD outperforms existing methods in accurately identifying known markers and revealing novel cell states, even across different samples without the need for complex batch correction.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) generates vast amounts of data, allowing us to understand cellular diversity and function at an unprecedented resolution. A key challenge in analyzing this data is identifying reliable cell markers – genes that are uniquely expressed in specific cell types or states. Traditional methods often rely on pre-clustering cells, which can be sensitive to noise and may miss subtle cell subtypes. Existing cluster-independent methods, while better, can be limited by their inability to capture markers across different scales or their reliance on specific reference points. This paper aims to address these limitations by providing a robust, cluster-independent method that can identify localized markers at multiple resolutions and facilitate cross-sample comparisons.

Methods

LMD employs a three-step workflow:

  1. Cell-cell affinity graph construction: LMD builds a graph where each cell is a node, and edges connect cells that are transcriptionally similar. This similarity is determined by k-nearest neighbors (kNN) based on Euclidean distance after dimensionality reduction (e.g., PCA).
  2. Multi-scale diffusion process: Gene expression data is treated as an initial state on this graph. A diffusion process is then applied, simulating how gene expression spreads across the cell graph over time. The rate at which this diffusion reaches equilibrium is indicative of how "localized" the gene's expression is.
  3. LMD-score calculation: The time-dependent diffusion dynamics are quantified using the Kullback-Leibler (KL) divergence between the initial and diffused states. The area under the normalized KL-divergence curve (the LMD-score) summarizes this localization. A lower LMD-score signifies a more localized gene expression pattern.

Candidate markers identified by LMD can be further grouped into "gene modules" using hierarchical clustering and dynamic tree cutting, which represent functional gene sets associated with specific cell states or biological processes.

Datasets

The authors evaluated LMD on several datasets:

  • Tabula Muris: This includes FACS-based and droplet-based scRNA-seq data from mouse bone marrow, lung, and pancreas.
  • Azimuth demo datasets: Publicly available human and mouse scRNA-seq datasets from various tissues (motor cortex, pancreas, kidney, bone marrow, lung).
  • Mouse embryonic skin data: A novel dataset generated by the authors to study dermal condensate genesis in wild-type and mutant mouse embryos.

LMD was benchmarked against six other marker identification algorithms on nine scRNA-seq datasets.

Key Insights and Ideas

  • Localization is key: The core idea is that genes specifically expressed in a particular cell group will diffuse more slowly across the cell graph compared to genes expressed more broadly. This "localization" is the fundamental principle LMD leverages.
  • Multi-scale analysis: LMD's diffusion process inherently captures gene expression patterns at various scales, from broad cell types to fine-grained subtypes, without explicit parameter tuning for scale.
  • Cluster-independent power: By avoiding pre-clustering, LMD can identify markers for cell populations that might be missed or poorly defined by traditional clustering approaches.
  • Cross-sample comparison without batch correction: LMD's ability to identify conserved and sample-specific gene modules allows for direct comparison of biological conditions without relying on potentially problematic batch effect correction methods.
  • Functional gene modules: Grouping localized markers into modules reveals coherent biological functions and pathways, providing deeper insights into cellular states and processes.

Key Results

  • Superior marker identification: LMD outperformed six other methods in identifying known cell type markers across nine scRNA-seq datasets, ranking first in 5 out of 9 datasets for both tested marker sets.
  • Enhanced cell separation: Markers identified by LMD led to better separation of cell groups compared to other methods, as measured by the density index.
  • Identification of novel cell states: LMD successfully identified novel cell states in the mouse hair follicle dermal condensate system, including a mutant-specific quiescent cell state and intermediate stages of differentiation.
  • Efficient runtime: LMD demonstrated reasonable computational efficiency, processing large datasets (e.g., 30,000 cells) in under an hour.
  • Robustness: The method showed robustness to variations in kNN parameters and input data dimensionality.

Limitations and Future Work

  • Cell graph bias: The paper acknowledges that the cell graph construction itself can introduce bias, potentially favoring genes expressed in well-connected neighborhoods. Future work aims to balance gene distribution with graph geometry.
  • Potential bias from using expression data for graph construction and ranking: Using the same gene expression data for both graph construction and marker ranking might introduce a bias towards highly variable genes.
  • Adaptability to other single-cell techniques: The authors suggest that LMD could be adapted for other single-cell modalities like scATAC-seq and spatial transcriptomics.
  • Further exploration of gene modules: The paper highlights the potential of LMD-derived gene modules for understanding dynamic biological processes and discovering novel cell subtypes.

Summary updated

Abstract

Abstract Identifying accurate cell markers in single-cell RNA-seq data is crucial for understanding cellular diversity and function. Localized Marker Detector (LMD) is a novel tool to identify “localized genes” - genes exclusively expressed in groups of highly similar cells - thereby characterizing cellular diversity in a multi-resolution and fine-grained manner. LMD constructs a cell-cell affinity graph, diffuses the gene expression value across the cell graph, and assigns a score to each gene based on its diffusion dynamics. LMD’s candidate markers can be grouped into functional gene modules, which accurately reflect cell types, subtypes, and other sources of variation such as cell cycle status. We apply LMD to mouse bone marrow and hair follicle dermal condensate datasets, where LMD facilitates cross-sample comparisons, identifying shared and sample-specific gene signatures and novel cell populations without requiring batch effect correction or integration methods. Furthermore, we assessed the performance of LMD across nine single-cell RNA sequencing datasets, compared it with six other methods aimed at achieving similar objectives, and found that LMD outperforms the other methods evaluated.
BibTeX
@misc{li2023lmd,
  author = {Ruiqi Li and Rihao Qu and Fabio Parisi and Francesco Strino and Hon‐Ming Lam and Jay S. Stanley and Xiuyuan Cheng and Peggy Myung and Yuval Kluger},
  title = {{LMD: Cluster-Independent Multiscale Marker Identification in Single-cell RNA-seq Data}},
  year = {2023},
  doi = {10.1101/2023.11.12.566780},
  howpublished = {bioRxiv},
}
Figure