RgCop-A regularized copula based method for gene selection in single-cell RNA-seq data

Snehalika Lall, Sumanta Ray, Sanghamitra BandyopadhyayPLOS Computational Biology 2021 · DOI

TL;DR

This paper introduces RgCop, a novel method for selecting informative genes from single-cell RNA sequencing (scRNA-seq) data. By leveraging a robust copula-based correlation measure and regularization, RgCop effectively identifies genes that improve downstream analysis like cell clustering and classification, even in the presence of technical noise.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) generates vast amounts of data, but extracting meaningful biological insights requires careful preprocessing. A critical step is gene selection, where the goal is to identify a subset of genes that best represent cell types and states. Traditional methods often rely on identifying highly variable or highly expressed genes. However, scRNA-seq data is notoriously noisy due to technical variations, which can lead to unstable and unreliable gene sets with these conventional approaches. This noise can hinder accurate cell clustering and classification, making it difficult to understand cell development and disease. The paper aims to address this by developing a gene selection method that is robust to technical noise and can identify a more stable and predictive set of genes.

Methods

The core of RgCop lies in its use of copula correlation (Ccor), a measure that captures multivariate dependencies between genes. Unlike traditional correlation measures, Ccor is scale-invariant and more robust to noise, making it well-suited for scRNA-seq data.

The method employs a regularized objective function that balances two criteria:

  • Max Relevancy: Identifying genes that have a high Ccor with cell type labels.
  • Min Redundancy: Ensuring that selected genes are not overly dependent on each other, thus avoiding redundant information.

This objective function is optimized using a forward selection wrapper approach. An l₁ regularization term is incorporated to penalize large coefficients, further refining the selected gene set and controlling for potential overfitting. The regularization parameter ($\gamma$) is tuned using simulated data.

Datasets

The study evaluated RgCop using several publicly available scRNA-seq datasets:

  • Yan: Human preimplantation embryo and embryonic stem cell data.
  • Pollen: Single-cell RNA sequencing data from a developing cerebral cortex.
  • Muraro: Human pancreas cell tissue data.
  • PBMC68k: Peripheral blood mononuclear cell data.

Additionally, the authors utilized simulated data generated by Splatter for tuning the regularization parameter and for performance evaluation under various noise conditions. They also used datasets from the Batch-effect-removal-benchmarking repository to assess robustness across different batches. A large-scale dataset of Adult Mouse Brain (690k cells) was also used for a case study.

Key Insights and Ideas

  • Copula's Robustness: The paper highlights that copula-based measures, like Ccor, are inherently more robust to technical noise in scRNA-seq data due to their scale-invariant properties. This is a significant advantage over methods relying solely on variation or expression levels.
  • Balancing Relevancy and Redundancy: The objective function's dual focus on maximizing gene-to-class relevancy while minimizing gene-to-gene redundancy is key to obtaining a parsimonious and informative gene set.
  • Improved Downstream Performance: RgCop's selected genes consistently lead to better performance in downstream tasks like clustering and classification, demonstrating the biological relevance of the selected features.
  • Annotation of Unknown Cells: The selected genes are not only useful for clustering but also enable accurate annotation of previously unknown cell types, a crucial step in scRNA-seq analysis.

Key Results

  • Clustering Performance: RgCop achieved higher Adjusted Rand Index (ARI) scores compared to several state-of-the-art unsupervised feature selection methods on real datasets (Yan, Muraro, Pollen).
  • Classification Accuracy: RgCop outperformed other supervised feature selection methods in classification accuracy across multiple datasets and classifiers (GBM, NNET, SVM). For instance, on the Yan dataset with GBM, RgCop achieved an accuracy of 0.98 ± 0.01, compared to MRMR's 0.97 ± 0.01.
  • Stability: The method demonstrated good stability by maintaining high similarity scores when tested with noisy data, outperforming competing methods.
  • Rare Cell Identification: RgCop successfully identified rare cell populations (e.g., CD14+ monocytes, dendritic subtypes) in complex datasets.
  • Large-Scale Data: The method showed efficacy on a very large dataset (690k cells), achieving a good ARI score of 0.78.
  • Batch Effect Robustness: RgCop demonstrated robustness when applied to datasets with different batches, achieving high ARI scores.

Limitations and Future Work

  • Computational Cost: The paper acknowledges that RgCop, being a wrapper-based method, can be more computationally intensive than filter-based methods, especially for very large datasets.
  • Parameter Tuning: While the regularization parameter ($\gamma$) was tuned using simulated data, further exploration of optimal tuning strategies for diverse real-world datasets might be beneficial.
  • Future Directions: The authors suggest that RgCop can be a valuable tool for computational biologists in the initial stages of scRNA-seq downstream analysis, implying its potential for broader application and integration into existing pipelines. The ability to annotate unknown cells also opens avenues for developing more sophisticated supervised annotation tools.

Summary updated

BibTeX
@article{lall2021rgcop,
  author = {Snehalika Lall and Sumanta Ray and Sanghamitra Bandyopadhyay},
  title = {{RgCop-A regularized copula based method for gene selection in single-cell RNA-seq data}},
  journal = {PLOS Computational Biology},
  year = {2021},
  doi = {10.1371/journal.pcbi.1009464},
}
Figure