scDALI: Modelling allelic heterogeneity of DNA accessibility in single-cells reveals context-specific genetic regulation

Abstract

While the functional impact of genetic variation can vary across cell types and states, capturing this diversity remains challenging. Current studies, using bulk sequencing, ignore much of this heterogeneity, reducing discovery and explanatory power. Single-cell approaches combined with F1 genetic designs provide a new opportunity to address this problem, however suitable computational methods to model these complex relationships are lacking.Here, we developed scDALI, an analysis framework that integrates single-cell chromatin accessibility for unbiased cell state identification with allelic quantifications to assay genetic effects. scDALI builds on Gaussian process regression and can differentiate between homogeneous (pervasive) allelic imbalances and cell state-specific regulation. As a proof-of-principle, we applied scDALI to whole Drosophila embryos from F1 crosses, profiling sciATAC-seq at three embryonic stages. Even in these very complex samples, scDALI discovered hundreds of peaks with heterogeneous allelic imbalance, having effects in specific lineages and/or developmental stages. Our study provides a general strategy to identify the cellular context of allelic imbalance, a crucial step in linking genetic traits to cellular phenotypes.

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

“…Thus we also assessed simulations using φ = 3 to evaluate method robustness when the data was substantially overdispersed relative to a Binomial model (the model used by the GFL in airpart). airpart was assessed via simulation across various settings summarized in Supplementary Table S2, and compared to another statistical method for detecting heterogeneity of allelic ratio in scRNAseq, scDALI (Heinen et al, 2021), which models allele-specific chromatin accessibility using Gaussian Process regression.…”

“…In previous work, researchers often used a Binomial model (Castel et al, 2020) or a Beta-Binomial model for the allelic counts (Skelly et al, 2011;Castel et al, 2015;Edsgärd et al, 2016;Santoni et al, 2017;Heinen et al, 2021;Choi et al, 2019;Zitovsky and Love, 2020), whereas BSCET uses a linear regression for the CTS AI test (Fan et al, 2021). For the datasets examined in the Results, either SMART-seq2 single-cell datasets, or spatially-or time-resolved bulk RNA-seq, we found that a Binomial assumption was sufficient for grouping cell types or conditions by AI, as many genes had minimal over-dispersion relative to a Binomial model.…”

“…We evaluated airpart across a variety of simulation datasets, and in comparison to a newly developed method for detecting heterogeneous allelic imbalance, scDALI (Heinen et al, 2021). On the simulated dataset with 10 cell types of different allelic ratio, as described in Section 2.5, the Generalized Fused Lasso (GFL) with Binomial likelihood tended to have higher ARI than the GFL with Gaussian likelihood or our nonparametric method when the number of genes within a gene cluster was less than 20 (Figure 2A).…”

“…When the AI only exists in one or more specific cell type(s) or the AI varies among cell types, we refer to this phenomenon as differential allelic imbalance (DAI). As single-cell studies providing sufficient coverage and read length for allelic quantification are only now emerging, we are aware of only one related statistical method for detecting DAI, scDALI (Heinen et al, 2021), which models allele-specific chromatin accessibility using Gaussian Process regression.…”