Detecting differential alternative splicing events in scRNA-seq with or without Unique Molecular Identifiers

doi:10.1371/journal.pcbi.1007925

Detecting differential alternative splicing events in scRNA-seq with or without Unique Molecular Identifiers

Fig 2

Exon-inclusion level estimation, false positive rate comparison and power comparison of simulation study.

Informative read counts of 1,000 exon groups for 600 cells (300 vs 300) were simulated using a generative model shown in S1 Fig with cell-specific technical parameters estimated from CA1 pyramidal cells in the scRNA-seq data generated by Zeisel et al. [16]. To simulate data that resemble real scRNA-seq data, one CA1 pyramidal cell was randomly selected and read coverage for each gene in that cell was used to obtain true gene concentration at the population level (). Based on the generative model accounting for technical noise, we generated 6,000 reads on average for each cell. (A) Scatter plots of 1,842 true exon-inclusion levels, estimated by percent spliced in (PSI) for a given exon, against estimates from SCATS and a naïve method that ignores technical noise. SCATS models technical noise to quantify usage of an alternatively spliced exon while the naïve method simply estimates the inclusion level by , where Y₊ and Y₋ are the informative read counts of the “+” and “-” exon groups across cells. SCATS estimates are closer to the ground truth. (B) Quantile-quantile plots of the p-values from SCATS, Census and DEXSeq under the null hypothesis (Δ = 0). X-axis represents uniform theoretical quantiles between 0 and 1 in –log₁₀ scale. Y-axis represents observed p-value quantile in –log₁₀ scale. Uniformly distributed data should follow the red dashed line. P-values of SCATS are more uniformly distributed while those from Census are right-skewed and those from DEXSeq are left-skewed. (C) Type I error comparison of SCATS, Census and DEXSeq with different significance levels (α = 0.05, 0.01, 0.005). Similar to (B), SCATS has better type I error control than Census and DEXSeq. (D) Power comparison of SCATS, Census and DEXSeq. Barplots show the estimated power under different effect sizes (Δ = 0.1,0.2,0.3,0.4), where significance was evaluated at 0.05, 0.01, and 0.005 levels, respectively. Colors indicate different methods. SCATS outperformed Census and DEXSeq across all effect sizes, especially when Δ = 0.1. DEXSeq is conservative in detecting DAS events.

doi: https://doi.org/10.1371/journal.pcbi.1007925.g002