SparseDC: Implementation of SparseDC Algorithm

Jun Li, Martin BarronOpenalex Unknown 2017 · DOI · OpenAlex

Abstract

Implements the algorithm described in Barron, M., Zhang, S. and Li, J. 2017, "A sparse differential clustering algorithm for tracing cell type changes via single-cell RNA-sequencing data", Nucleic Acids Research, gkx1113, doi:10.1093/nar/gkx1113. This algorithm clusters samples from two different populations, links the clusters across the conditions and identifies marker genes for these changes. The package was designed for scRNA-Seq data but is also applicable to many other data types, just replace cells with samples and genes with variables. The package also contains functions for estimating the parameters for SparseDC as outlined in the paper. We recommend that users further select their marker genes using the magnitude of the cluster centers.

Introduction

In single-cell biology, a recurring challenge is the comparison of two distinct populations—such as healthy versus diseased tissue or different developmental stages. Traditional clustering methods often analyze each condition in isolation, which complicates the process of "tracing" cell types across conditions. It becomes difficult to determine if a cluster in one condition truly corresponds to a cluster in another, and even harder to identify the specific genes driving these biological shifts.

SparseDC addresses this by providing a unified framework for simultaneous clustering and marker gene identification. It moves beyond simple clustering into "differential clustering," allowing researchers to observe how a specific cell type's gene expression profile shifts across conditions.

Methodology

The core of the package is the SparseDC algorithm, which utilizes a penalized clustering approach. Key features include:

Joint Clustering and Linking

Unlike traditional methods, SparseDC clusters samples from both populations simultaneously. This ensures that clusters are linked across conditions, providing a direct mapping of cell types.

Sparsity via Regularization

The algorithm employs L1 regularization (penalty terms) to ensure that only a small, relevant subset of genes is used to define clusters and differences. This focuses the analysis on high-magnitude changes that define biological identity rather than background noise.

Parameter Estimation

The package includes specialized functions, lambda1_calculator and lambda2_calculator, which use bootstrapping to estimate optimal penalty values.

  • Lambda 1 ($\lambda_1$): Controls the number of marker genes selected for each cluster.
  • Lambda 2 ($\lambda_2$): Controls the number of genes showing condition-dependent expression within a cell type.
Parameters for the Lambda 1 calculator, including data inputs and bootstrap settings.
Parameters for the Lambda 2 calculator, mirroring the structure of the Lambda 1 estimation.

Iterative Optimization

The algorithm iteratively updates cluster memberships and cluster centers to minimize a cost function, incorporating a soft-thresholding operator to maintain sparsity.

Datasets and Implementation

The package utilizes both real-world and simulated data for evaluation:

  • Biase Data: A dataset from Biase et al. (2014) studying cell fate in mouse embryos, containing FPKM gene expression for zygote, two-cell, and four-cell embryo stages.
  • Simulated Data: A built-in simulator allows users to generate synthetic datasets with known marker genes and cluster structures.

Key Results

The implementation provides tools to achieve several critical outcomes:

  1. Automated Marker Selection: Identifies marker genes based on the magnitude of cluster centers.
  2. Optimal Cluster Detection: The sparsedc_gap function helps mathematically determine the most likely number of cell types using the gap statistic.
  3. Cross-Condition Mapping: Links clusters even when the number of clusters differs between conditions.

Limitations

While powerful, the algorithm has limitations. The bootstrapping process for parameter estimation and the gap statistic can be computationally intensive for large datasets. Additionally, the algorithm's success is highly dependent on proper normalization and data centralization. Users are encouraged to use specialized scRNA-seq normalization methods before applying SparseDC.

Summary updated

Abstract

Implements the algorithm described in Barron, M., Zhang, S. and Li, J. 2017, "A sparse differential clustering algorithm for tracing cell type changes via single-cell RNA-sequencing data", Nucleic Acids Research, gkx1113, &lt;<a href="https://doi.org/10.1093%2Fnar%2Fgkx1113" target="_top">doi:10.1093/nar/gkx1113</a>&gt;. This algorithm clusters samples from two different populations, links the clusters across the conditions and identifies marker genes for these changes. The package was designed for scRNA-Seq data but is also applicable to many other data types, just replace cells with samples and genes with variables. The package also contains functions for estimating the parameters for SparseDC as outlined in the paper. We recommend that users further select their marker genes using the magnitude of the cluster centers.
BibTeX
@inproceedings{li2017sparsedc,
  author = {Jun Li and Martin Barron},
  title = {{SparseDC: Implementation of SparseDC Algorithm}},
  booktitle = {Openalex Unknown},
  year = {2017},
  doi = {10.32614/cran.package.sparsedc},
}
Figure