SparseMDC: Implementation of SparseMDC Algorithm

Martin Barron, Jun LiOpenalex Unknown 2018 · DOI · OpenAlex

Abstract

Implements the algorithm described in Barron, M., and Li, J. (Not yet published). This algorithm clusters samples from multiple ordered 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 SparseMDC as outlined in the paper. We recommend that users further select their marker genes using the magnitude of the cluster centers.

Introduction

In biological research, particularly within the realm of single-cell transcriptomics, data is frequently collected from ordered populations, such as cells undergoing developmental transitions or responding to time-course treatments. Standard clustering methods often fail to maintain the inherent links between cell types across these conditions or struggle with high-dimensional noise. SparseMDC is an R package developed to address these challenges by clustering high-dimensional data across multiple ordered conditions while simultaneously identifying the marker genes driving these transitions.

Methodology

The core of the SparseMDC package is a penalized clustering framework designed to achieve two distinct types of sparsity:

  1. Marker Gene Selection ($\lambda_1$): This penalty regularizes the number of genes used to define clusters, ensuring that results are driven by a small, interpretable set of features rather than global noise.
  2. Condition-Dependent Expression ($\lambda_2$): This penalty controls the differences in mean expression values across conditions for each cluster. By penalizing unnecessary differences, it effectively "links" clusters across conditions, allowing for divergence only where supported by the data.

The package utilizes an iterative solver involving functions such as mu_solver, update_c, and update_mu to optimize cluster assignments and centers. For larger datasets, a parallelized version (sdc_mpar) is available.

Automated Parameter Estimation

A significant challenge in sparse clustering is the selection of appropriate penalty values. SparseMDC provides automated "calculators" to estimate these parameters using bootstrap repetitions based on the null distribution of the data.

Parameters for the lambda1_calculator function, used to estimate the marker gene selection penalty.

The lambda1_calculator (shown above) and the lambda2_calculator are essential for tuning the algorithm to specific datasets.

The lambda2_calculator function handles the estimation of the condition-dependent expression penalty.

Data and Implementation

The package includes the Biase Data set, which contains FPKM gene expression measurements for 49 mouse embryo cells across three developmental stages (zygote, two-cell, and four-cell). This dataset serves as a benchmark for demonstrating the algorithm's ability to track cell fate inclination.

Key implementation features include:

  • Integrated Pre-processing: The pre_proc_data function handles normalization and centralization, which are critical for accurate marker identification.
  • Gap Statistic Integration: The sparsemdc_gap function helps users estimate the optimal number of clusters ($K$) within the sparse framework.

Conclusion and Limitations

SparseMDC offers a robust framework for discovering marker genes and linking clusters in ordered populations. However, its design specifically targets sequential or hierarchical relationships, and it may be less effective for unordered conditions. Furthermore, while the algorithm provides statistical selection of marker genes, the authors recommend manual refinement based on cluster center magnitudes to ensure biological relevance.

Summary updated

Abstract

Implements the algorithm described in Barron, M., and Li, J. (Not yet published). This algorithm clusters samples from multiple ordered 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 SparseMDC as outlined in the paper. We recommend that users further select their marker genes using the magnitude of the cluster centers.
BibTeX
@inproceedings{barron2018sparsemdc,
  author = {Martin Barron and Jun Li},
  title = {{SparseMDC: Implementation of SparseMDC Algorithm}},
  booktitle = {Openalex Unknown},
  year = {2018},
  doi = {10.32614/cran.package.sparsemdc},
}
Figure