Tree-structured topic modelling of single-cell gene expression data uncovers hierarchical relationships between immune cell types

Abstract

Immune cells undergo a series of differentiation steps following a lineage-tree structure stemming from hematopoietic stem cells. During differentiation of immune cells in both homeostasis and pathological processes, many gene regulatory mechanisms are shared by fully differentiated immune cell sub-types. In order to characterize these features quantitatively, we proposeLaRCH, a tree-structured embedded topic model. In this model, single-cell gene expression profiles are represented by a mixture of topics consisting of latent features that follow an underlying tree structure, mirroring that of cellular differentiation–nested cluster structures. We present findings of our model trained on simulated single-cell RNA sequencing (scRNA-seq) based on cell-sorted bulk RNA-seq data as well as on a scRNA-seq dataset of over 1.2 million cells from healthy individuals and individuals diagnosed with systemic lupus erythematosus (SLE). The cellular topic profiles estimated by our model markedly improve clustering accuracy over traditional latent variable models and illustrate transcriptomic differences between SLE phenotypes, revealing a pivotal role of multiple immune cell types in disease progression and relapse. Ultimately,LaRCHcaptures the hierarchical context between cellular subtypes by simultaneously identifying shared and distinct latent features amongst subsets of heterogeneous samples of cells.