Parallel clustering of single cell transcriptomic data with split-merge sampling on Dirichlet process mixtures

Abstract

Source code is publicly available on https://github.com/tiehangd/Para_DPMM/tree/master/Para_DPMM_package.

VPAC: Variational projection for accurate clustering of single-cell transcriptomic data

“…improvement on NMI compared to Para_DPMM, even the clustering on these datasets was claimed as a challenging task [18,19]. In addition, our method consistently achieves stable performance, which demonstrates its higher robustness than the baseline methods.…”

“…The dataset measures the expression of 32,738 genes in 32,695 cells, which were enriched from fresh peripheral blood mononuclear cells (PBMCs) and sequenced by Illumina NextSeq 500 instrument. Clustering on this dataset was claimed as a challenging task by Z.Sun et al [18] and T.Duan et al [19].…”

“…including Para_DPMM, the state-of-the-art method for clustering discrete UMI counts [19], and pcaReduce, another method combining dimension reduction with clustering [20].…”

“…In order to fairly evaluate the performance of our model, we compared the performance of VPAC with other two baseline methods on the three T cell datasets of different scales provided by T.Duan et al [19]. The small scale dataset (S-Set) consists of 1200 cells with the 1000 top variable genes, the medium scale dataset (M-Set)…”

“…A key advantage of scRNA-seq is that it can be used to identify putative cell types using unsupervised clustering, which is essential to fully understand the cell identity [18]. Para-DPMM further improves the clustering quality by introducing a Dirichlet process prior to automatically infers the number of clusters from the dataset, and a split-merge mechanism is adopted to improve convergence and optimality of the result [19]. Given the high dimensionality of single-cell data, clustering directly on the original dimension may affect the performance due to the intrinsic noise of single-cell data, and usually demands tedious preprocessing such feature selection, which may restrict the robustness of clustering.…”