Estimating allele-specific expression of SNVs from 10x Genomics Single-Cell RNA-Sequencing Data

Abstract

With the recent advances in single-cell RNA-sequencing (scRNA-seq) technologies, estimation of allele expression from single cells is becoming increasingly reliable. Allele expression is both quantitative and dynamic and is an essential component of the genomic interactome. Here, we systematically estimate allele expression from heterozygous single nucleotide variant (SNV) loci using scRNA-seq data generated on the 10x Genomics platform. We include in the analysis 26,640 human adipose-derived mesenchymal stem cells (from three healthy donors), with an average sequencing reads over 120K/cell (more than 4 billion scRNA-seq reads total). High quality SNV calls assessed in our study contained approximately 15% exonic and >50% intronic loci. To analyze the allele expression, we estimate the expressed Variant Allele Fraction (VAFRNA) from SNV-aware alignments and analyze its variance and distribution (mono-and bi-allelic) at different cutoffs for required minimal number of sequencing reads. Our analysis shows that when assessing SNV loci covered by a minimum of 3 unique sequencing reads, over 50% of the heterozygous SNVs show biallelic expression, while at minimum of 10 reads, nearly 90% of the SNVs are bi-allelic. Consistent with single cell studies on RNA velocity and models of transcriptional burst kinetics, we observe a substantially higher rate of monoallelic expression among intronic SNVs, signifying the usefulness of scVAFRNA to assess dynamic cellular processes. Our analysis demonstrates the feasibility of scVAFRNA estimation from current scRNA-seq datasets and shows that the 3'-based library generation protocol of 10x Genomics scRNA-seq data can be highly informative in SNV-based analyses.

scReQTL: an approach to correlate SNVs to gene expression from individual scRNA-seq datasets

“…Additional attenuation of this constraint is possible through reducing the value of minR used in the analysis. Indeed, while in this study we apply minR > 10, which retained between 308 and 721 input SNVs per sample, in our prior research we show that at minR > 5 the number of SNVs is higher by an order of magnitude (26). Second, scReQTL appears to have relatively low power to detect cis-acting (on the same gene) SNVs (See Supplementary Figure 3).…”

“…In the present study, we have included SNV loci with minR > 10. Our previous research shows that for current 10× Genomics scRNA-seq datasets, minR > 5 provides a reasonable balance between VAFRNA confidence and SNV inclusivity (26). At lower cutoffs (i.e.…”

“…Adipose cells and erythrocytes were found in all three samples, whereas naïve-B cells were seen in N5 and N7 and neutrophils -in N8 (Figure 3). VAFRNA is assessed from the individual cell alignments at sites with heterozygous SNV calls using ReadCounts (22), which estimates the number of sequencing reads harboring the variant and the reference nucleotide (nvar and nref, respectively), and calculates VAFRNA (VAFRNA = nvar / (nvar + nref)) at each heterozygous SNV site of interest (26). To address stochasticity of sampling, estimations of VAFRNA require a threshold of minimal number of unique sequencing reads (minR).…”

“…To address stochasticity of sampling, estimations of VAFRNA require a threshold of minimal number of unique sequencing reads (minR). Our previous research shows that current scRNA-seq datasets can contain hundreds of SNV sites covered by minimum of 10 sequencing reads (minR > 10) and thousands of SNV sites with minR > 5 (26). In the herein presented analysis, we used VAFRNA estimated at sites with minR > 10; from here on, we refer to these loci as informative.…”

“…At lower cutoffs (i.e. minR = 3) stochasticity of sampling can affect the VAFRNA estimation (26). In addition, low cutoffs are expected to include SNVs in genes expressed at low levels, where additional technical noise can affect the accuracy of the estimations.…”