Bayesian gamma-negative binomial modeling of single-cell RNA sequencing data

Abstract

Background: Single-cell RNA sequencing (scRNA-seq) is a powerful profiling technique at the single-cell resolution. Appropriate analysis of scRNA-seq data can characterize molecular heterogeneity and shed light into the underlying cellular process to better understand development and disease mechanisms. The unique analytic challenge is to appropriately model highly over-dispersed scRNA-seq count data with prevalent dropouts (zero counts), making zero-inflated dimensionality reduction techniques popular for scRNA-seq data analyses. Employing zero-inflated distributions, however, may place extra emphasis on zero counts, leading to potential bias when identifying the latent structure of the data.

Results: In this paper, we propose a fully generative hierarchical gamma-negative binomial (hGNB) model of scRNA-seq data, obviating the need for explicitly modeling zero inflation. At the same time, hGNB can naturally account for covariate effects at both the gene and cell levels to identify complex latent representations of scRNA-seq data, without the need for commonly adopted pre-processing steps such as normalization. Efficient Bayesian model inference is derived by exploiting conditional conjugacy via novel data augmentation techniques.

Conclusion: Experimental results on both simulated data and several real-world scRNA-seq datasets suggest that hGNB is a powerful tool for cell cluster discovery as well as cell lineage inference.

Keywords: Bayesian; Hierarchical modeling; Single-cell RNA sequencing.

Fig. 1

Graphical representation of the hierarchical gamma-negative binomial (hGNB) model

Fig. 2

Mean-difference (MD) plot for S1/CA1 dataset. The solid red line represents the local regression fit to the data

Fig. 3

Low-dimensional representations of the S1/CA1 dataset. Panels correspond to (a) PCA (on total-count normalized data), (b) ZIFA (on total-count normalized data), (c) ZINB-WaVE, and (d) hGNB

Fig. 4

Average silhouette width in scRNA-seq datasets (a) S1/CA1, (b) mESC, and (c) V1. Silhouette widths were computed in the low-dimensional space, using the groupings provided by the authors of the original publications. PCA and ZIFA were applied with both unnormalized (RAW) data and after total count (TC) normalization

Fig. 5

Lineage inference on the OE dataset. The low dimensional data representation derived by hGNB were used to cluster cells by RSEC. The minimum spanning tree (MST) of the derived clusters constructed by slingshot is also displayed