SCeQTL: an R package for identifying eQTL from single-cell parallel sequencing data

Yue Hu, Xi Xi, Qian Yang, Xuegong ZhangBMC Bioinformatics 2020 · DOI

TL;DR

This paper introduces SCeQTL, an R package designed to identify expression quantitative trait loci (eQTLs) from single-cell genomic and transcriptomic data. It addresses the unique challenges of single-cell data, such as excess zero values, by employing a zero-inflated negative binomial regression model, offering a more sensitive approach than traditional bulk eQTL methods.

What Problem Does This Paper Address?

Understanding how genetic variations influence gene expression is crucial for deciphering complex biological processes. Traditional eQTL analysis, typically performed on bulk tissue samples, has been well-established. However, the advent of single-cell sequencing technologies, which allow for parallel sequencing of both the genome and transcriptome of individual cells, opens up new avenues for eQTL studies. These single-cell datasets possess distinct characteristics, most notably a high proportion of "drop-out" events (unobserved gene expression due to technical limitations) and inherent biological stochasticity, leading to an excess of zero values. Existing eQTL methods, often assuming normal or Poisson distributions, struggle with this excess zero inflation, potentially missing important associations or yielding unreliable results. This paper aims to bridge this gap by developing a method and a corresponding R package, SCeQTL, specifically tailored for eQTL analysis in single-cell data.

Methods

The core of SCeQTL lies in its use of a zero-inflated negative binomial (ZINB) regression model. This model accounts for the two main sources of zero values in single-cell RNA-seq data:

  1. "Drop-out" events: Where a gene is expressed but not detected due to technical limitations.
  2. True zero expression: Where a gene is genuinely not transcribed in a cell.

The ZINB model treats the data as arising from two processes:

  • A probability ($p$) that the gene is "off" (resulting in a zero observation).
  • A probability ($1-p$) that the gene is expressed, with its expression levels following a negative binomial (NB) distribution.

SCeQTL models two potential types of differences in gene expression associated with genotype variations:

  • "Status difference": Differences in the "on-off" transcription status (modeled by the zero ratio, $p$).
  • "Expression level difference": Differences in the abundance of expressed genes (modeled by the mean of the negative binomial distribution, $\mu$).

The association between genotype ($s$) and gene expression is tested by examining the correlation of $p$ and/or $\mu$ with $s$ using the following linear relationships: $\ln\left(\frac{p}{1-p}\right) = \alpha_1 + \beta_1s$ and/or $\ln(\mu) = \alpha_2 + \beta_2s$

Parameters are estimated using maximum likelihood. Hypothesis testing is performed by comparing the log-likelihoods of a full model (including genotype as a predictor) and a reduced model (excluding genotype). The deviance statistic, derived from the log-likelihood ratio, follows a chi-square distribution and is used to test for significant associations ($\beta_1 \neq 0$ or $\beta_2 \neq 0$). The package also incorporates options for covariate correction and multiple test correction (Benjamini-Hochberg and q-value methods).

Datasets

The paper utilizes two types of datasets:

  • Simulated Data: The authors generated simulated genotype and gene expression data for 1500 cells, involving three single nucleotide variations (SNVs) and 20 genes. This simulation was designed to mimic four experimental scenarios with varying effect sizes, focusing on differences in zero ratios and/or negative binomial means across genotypes. This simulated dataset serves as a benchmark for evaluating SCeQTL's performance against Matrix eQTL.
  • Real Data: A scRNA-seq dataset of human preimplantation embryo cells from different embryo days (E5, E6, E7) was used. This dataset, while lacking cell-level genotype information, provided cell attribute groups (embryonic days) to mimic genotype groups. This allowed for the evaluation of SCeQTL's ability to detect associations with other grouping factors beyond genetic variants and to assess its performance on real-world noisy single-cell data. The authors also mention using a dataset with cell lineage information (EPI, TE, PE) to further validate SCeQTL's findings against known biological literature.

Key Insights and Ideas

  • Single-cell data demands specialized eQTL methods: The excess of zero values in scRNA-seq data fundamentally alters the statistical landscape, making traditional bulk eQTL methods less effective.
  • Two types of genetic effects in single cells: Genetic variants can influence not only the level of gene expression but also the probability of a gene being expressed at all (transcription status). SCeQTL's ability to detect both is a key advantage.
  • ZINB is a powerful model for scRNA-seq: The zero-inflated negative binomial distribution accurately captures the bimodal nature of single-cell expression data, with a distinct peak at zero and a distribution for non-zero counts.
  • SCeQTL outperforms Matrix eQTL in complex scenarios: While performance is similar when only one aspect (zero ratio or expression level) differs, SCeQTL shows superior power in detecting associations when both transcription status and expression levels are affected by genotype.
  • Biological validation of findings: The real-data experiments demonstrated that SCeQTL can identify biologically relevant associations, such as lineage-specific genes, which align with existing literature.

Key Results

  • Simulation experiments: SCeQTL generally achieved higher or comparable Area Under the Curve (AUC) values in Receiver Operating Characteristic (ROC) curves compared to Matrix eQTL, particularly in scenarios where both transcription status and expression levels differed between genotypes. In some cases, SCeQTL's power rose earlier and more dramatically, indicating higher sensitivity.
  • Real data experiments:
    • SCeQTL identified significant eQTLs that were missed by Matrix eQTL, especially in cases where differences in zero ratios were prominent, even if non-zero expression levels were similar.
  • When analyzing cell lineage as a grouping factor, SCeQTL identified numerous genes with significant associations, many of which are known lineage-specific genes, with some showing obvious transcription status differences.

Limitations and Future Work

  • Computational Cost: The authors acknowledge that SCeQTL's computation cost can be relatively high for whole-genome eQTL analysis due to iterative parameter estimation. However, they suggest this is less of an issue for single-cell studies where the number of relevant SNVs might be limited, or when analyzing associations with other cell-type grouping factors.
  • Future Directions: The paper highlights the ongoing development of single-cell genomics technologies and expresses hope that SCeQTL will serve as a valuable tool for exploring gene regulation at the single-cell level as these technologies mature and become more widely available for large-scale studies.

Summary updated

BibTeX
@article{hu2020sceqtl,
  author = {Yue Hu and Xi Xi and Qian Yang and Xuegong Zhang},
  title = {{SCeQTL: an R package for identifying eQTL from single-cell parallel sequencing data}},
  journal = {BMC Bioinformatics},
  year = {2020},
  doi = {10.1186/s12859-020-3534-6},
}
Figure