Benchmarking clustering algorithms on estimating the number of cell types from single-cell RNA-sequencing data

Abstract

Background A key task in single-cell RNA-seq (scRNA-seq) data analysis is to accurately detect the number of cell types in the sample, which can be critical for downstream analyses such as cell type identification. Various scRNA-seq data clustering algorithms have been specifically designed to automatically estimate the number of cell types through optimising the number of clusters in a dataset. The lack of benchmark studies, however, complicates the choice of the methods. Results We systematically benchmark a range of popular clustering algorithms on estimating the number of cell types in a variety of settings by sampling from the Tabula Muris data to create scRNA-seq datasets with a varying number of cell types, varying number of cells in each cell type, and different cell type proportions. The large number of datasets enables us to assess the performance of the algorithms, covering four broad categories of approaches, from various aspects using a panel of criteria. We further cross-compared the performance on datasets with high cell numbers using Tabula Muris and Tabula Sapiens data. Conclusions We identify the strengths and weaknesses of each method on multiple criteria including the deviation of estimation from the true number of cell types, variability of estimation, clustering concordance of cells to their predefined cell types, and running time and peak memory usage. We then summarise these results into a multi-aspect recommendation to the users. The proposed stability-based approach for estimating the number of cell types is implemented in an R package and is freely available from (https://github.com/PYangLab/scCCESS).

Huntington’s Disease Produces Multiplexed Transcriptional Vulnerabilities of Striatal D1-D2 and Striosome-Matrix Neurons

“…3), which seems a class of striatal interneuron according to our preliminary data (Pineda et al, in preparation). These cell types, identified by ACTIONet [20][21][22][23][24][25][26][27][28][29][30][31][32][33][34] , co-clustered perfectly atop each other between controls and the HD samples (magenta and cyan dots, Fig. 1A-C), affirming the consistency of cell-type annotations across phenotypes.…”

“…We analyzed snRNA-seq data by using ACTIONet [20][21][22][23][24][25][26][27][28][29][30][31][32][33][34] (Extended Data Fig. 1), from striatal samples harvested from human striatum and from R6/2 and zQ175 HD model mice, and originally reported in an initial study without attention to the coordinated compartmental transcriptomics examined here (Fig.…”

“…We used the ACTIONet and scran R packages to normalize, batch correct, and cluster single-nucleus gene counts. Batch-corrected data were used as input to the archetypal analysis for cell type identification (ACTION) algorithm [20][21][22][23][24][25][26][27][28][29][30][31][32][33][34] to identify a set of landmark cells or 'archetypes', each representing a potential underlying cell state. Using ACTION-decompositions with varying numbers of archetypes, we employed the ACTION-based network (ACTIONet) framework [20][21][22][23][24][25][26][27][28][29][30][31][32][33][34] to create a multi-resolution nearest neighbor graph.…”

“…Batch-corrected data were used as input to the archetypal analysis for cell type identification (ACTION) algorithm [20][21][22][23][24][25][26][27][28][29][30][31][32][33][34] to identify a set of landmark cells or 'archetypes', each representing a potential underlying cell state. Using ACTION-decompositions with varying numbers of archetypes, we employed the ACTION-based network (ACTIONet) framework [20][21][22][23][24][25][26][27][28][29][30][31][32][33][34] to create a multi-resolution nearest neighbor graph. A modified version of the stochastic gradient descent-based layout method was used in the uniform manifold approximation and projection (UMAP) algorithm 59 , to visualize the ACTIONet graph.…”

Bioinformatics roadmap for therapy selection in cancer genomics

“…Nevertheless, it is important to remark that these are not the only methodologies available and that the sparsity (meaning the high fraction of zeroes present in single-cell matrices) of the analysed data set or the amount of sequenced cells should guide the algorithm selection. Interesting reviews from Du o et al [187] and Yu et al [188] could be helpful for performing this algorithm selection. The manifold learning algorithms are recommended for further exploratory singlecell data visualisation [189].…”