|
Abstract:
|
Single-cell RNA sequencing is a powerful tool to study heterogeneity and dynamic changes in cell populations. Although a wealth of clustering algorithms are available, it remains challenging to describe the multi-scale characteristics and features present in the data. To resolve this challenge, tree architectures are advocated as a natural representation of the internal structure of cell populations. Our post-clustering processing tool, called Dynamic Multi-resolution Reconciled Tree (MR-tree), builds a hierarchical tree structure based on multi-resolution partitions, which is highly flexible and can be coupled with most scRNA-seq clustering algorithms. MR-tree out-performs bottom-up hierarchical clustering approaches because it inherits the advantage of a flat clustering approach, while maintaining the hierarchical structure of cells determined by the similarity measure embedded in the corresponding clustering metrics. Moreover, MR-tree has a built-in mechanism for self-correcting the bias in clustering membership with pooled information at different scale. Application of MR-tree to various scRNA-seq data sets brought insights on sub-cell type and maker gene identification.
|
