DUBStepR is a scalable correlation-based feature selection method for accurately clustering single-cell data

Abstract

Feature selection (marker gene selection) is widely believed to improve clustering accuracy, and is thus a key component of single cell clustering pipelines. Existing feature selection methods perform inconsistently across datasets, occasionally even resulting in poorer clustering accuracy than without feature selection. Moreover, existing methods ignore information contained in gene-gene correlations. Here, we introduce DUBStepR (Determining the Underlying Basis using Stepwise Regression), a feature selection algorithm that leverages gene-gene correlations with a novel measure of inhomogeneity in feature space, termed the Density Index (DI). Despite selecting a relatively small number of genes, DUBStepR substantially outperformed existing single-cell feature selection methods across diverse clustering benchmarks. Additionally, DUBStepR was the only method to robustly deconvolve T and NK heterogeneity by identifying disease-associated common and rare cell types and subtypes in PBMCs from rheumatoid arthritis patients. DUBStepR is scalable to over a million cells, and can be straightforwardly applied to other data types such as single-cell ATAC-seq. We propose DUBStepR as a general-purpose feature selection solution for accurately clustering single-cell data.

Single-cell and bulk transcriptome sequencing identifies two epithelial tumor cell states and refines the consensus molecular classification of colorectal cancer

“…1 ). We performed de novo clustering on CRC-SG1 epithelial cells using DUBStepR 10 for feature selection, and then re-clustered the cells using differentially expressed genes (DEGs) between the initial clusters (Supplementary Fig. 1 and Methods ).…”

“…To cluster the 15,920 high-quality epithelial cells described above, we first used DUBStepR, a correlation-based feature selection algorithm that outperforms existing methods across diverse clustering benchmarks, to identify an informative set of genes 10 (Supplementary Fig. 1 ).…”

Feature selection revisited in the single-cell era

“…While filters are the most common options for pre-processing and feature selection from single-cell transcriptomics data, the application of wrapper methods is gaining much attention with a range of approaches built and extends on classic methods with the primary goal of facilitating downstream analyses such as cell type classification. Some examples include the application of classic methods such as greedy-based optimisation of entropy [ 66 ], nature-inspired optimisation such as using GA [ 67 , 68 ], and their hybrid with filters [ 69 – 71 ] or embedded methods [ 72 ]. More advanced methods include active learning-based feature selection using SVM as a wrapper [ 73 ] and optimisation based on data projection [ 74 ].…”

“…Muto et al [ 88 ] performed filter-based differential analysis on both chromatin and gene levels based on Cicero estimated gene activity scores [ 89 ]. Finally, DUBStepR [ 71 ], a hybrid approach that combines a correlation-based filter and a regression-based wrapper for gene selection from scRNA-seq data, can also be applied to scATAC-seq data. Collectively, these methods and tools demonstrate the utility and impact of feature selection on scATAC data for cell-type identification, motif analysis, regulatory element and gene interaction detection among other applications.…”

“…While we found a small number of them employed hybrid approaches (e.g. [ 71 , 72 ]), most are relatively straightforward combinations (such as stepwise application of filter and then wrapper methods) as have been used previously for bulk data analyses. The application of ensemble and deep learning-based feature selection methods is even sparser in the field.…”