Airpart: interpretable statistical models for analyzing allelic imbalance in single-cell datasets

Abstract

Motivation: Allelic expression analysis aids in detection of cis-regulatory mechanisms of genetic variation, which produce allelic imbalance (AI) in heterozygotes. Measuring AI in bulk data lacking time or spatial resolution has the limitation that cell-type-specific (CTS), spatial- or time-dependent AI signals may be dampened or not detected.

Results: We introduce a statistical method airpart for identifying differential CTS AI from single-cell RNA-sequencing data, or dynamics AI from other spatially or time-resolved datasets. airpart outputs discrete partitions of data, pointing to groups of genes and cells under common mechanisms of cis-genetic regulation. In order to account for low counts in single-cell data, our method uses a Generalized Fused Lasso with Binomial likelihood for partitioning groups of cells by AI signal, and a hierarchical Bayesian model for AI statistical inference. In simulation, airpart accurately detected partitions of cell types by their AI and had lower Root Mean Square Error (RMSE) of allelic ratio estimates than existing methods. In real data, airpart identified differential allelic imbalance patterns across cell states and could be used to define trends of AI signal over spatial or time axes.

Availability and implementation: The airpart package is available as an R/Bioconductor package at https://bioconductor.org/packages/airpart.

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

Overview of airpart framework. airpart takes as input allele-specific read counts, quantified upstream of our method. Known cell annotation or cell clusters derived from total counts are also part of the input to airpart. Following QC steps, clustering is performed on genes based on their allelic signal over cells. Then during the modeling step, a partition of the cell groups is generated as shown in heatmap, either by a GFL or a non-parametric method. Estimated coefficients of this gene cluster using GFL inform the prior of hierarchical Bayesian model. Finally, airpart outputs estimates of allelic ratio for each gene and cell group, as well as s-value or adjusted P-value for AI and DAI test, respectively. Multiple visualizations of input data, gene clustering and fitted parameters are available as functions within airpart software

Fig. 2.

Performance comparison of airpart variants and scDALI on simulation datasets. (A) Boxplot of partition accuracy among three variants of airpart. y-axis is ARI among 200 iterations. cnt, the higher mean count; n, number of genes within a gene cluster. (B) Boxplot of RMSE per gene for estimation of the allelic ratio for n = 40 cells among 400 iterations. Each gene has an underlying U-shape pattern described in the Section 2.6. (C) Boxplot demonstrating airpart without cell-type grouping step and scDALI performance on each cell type at DAI = 0.2. The highlighted dots inside the boxes represent the simulated allelic ratios

Fig. 3.

Evaluation of airpart on two scRNA-seq experiments. (A) Violin plot of estimated allelic ratio on Larsson’s dataset with n indicating the number of cells. Color represents different partition groups. (B) Forest plot for Larsson’s dataset, showing top 40 genes with smallest s-value. Dotted line denotes allelic ratio = 0.5 (C) Step plot and heatmap of results for Deng’s dataset. This gene cluster partitioned cell types into five groups denoted by highlighted dots in the step plot. (D–G) Selected genes displaying airpart fitted model on Gutierrez-Arcelus’s data: (D) decreasing trend, (E) increasing trend, (F) up-down pattern and (G) down-up pattern