Bayesian model selection reveals biological origins of zero inflation in single-cell transcriptomics

Kwangbom Choi, Yang Chen, Daniel A. Skelly, Gary A. ChurchillbioRxiv 2020 · DOI

TL;DR

Single-cell RNA sequencing (scRNA-seq) data is notoriously "noisy" with many zero counts, leading to debate about their origin (technical vs. biological) and how to handle them. This paper uses Bayesian model selection to show that zero inflation in scRNA-seq is primarily biological, driven by cell type and sex, not just technical artifacts. Despite this, they recommend a simpler, non-zero-inflated model for general analysis.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) allows us to study gene expression at the individual cell level, revealing incredible biological complexity. However, scRNA-seq data is characterized by a high proportion of zero counts (zeros), meaning a gene appears to have no expression in a cell. This "zero inflation" is a major challenge for data analysis and interpretation. There's a significant debate in the field: are these zeros due to technical issues like failed capture or sequencing errors (technical dropout), or do they reflect genuine biological processes like genes being truly off in certain cells (biological zeros)? Some researchers propose imputing these zeros to fill in "missing" data, while others argue that the zeros are informative and should be modeled directly. This paper aims to resolve this by rigorously investigating the origins of zero inflation in scRNA-seq data.

Methods

The core of this paper's approach is Bayesian model selection using a criterion called the Expected Log Predictive Density (ELPD). They developed a software package called scRATE to implement this. For each gene, they compare four statistical models:

  • Poisson (P): A basic count model.
  • Negative Binomial (NB): Accounts for overdispersion (more variability than Poisson).
  • Zero-Inflated Poisson (ZIP): Adds a component for "extra" zeros beyond what Poisson predicts.
  • Zero-Inflated Negative Binomial (ZINB): Combines NB with a zero-inflation component.

scRATE uses leave-one-out cross-validation to estimate the ELPD, which helps select the model that best predicts unseen data, penalizing both under- and over-fitting. They also incorporate Generalized Linear Models (GLMs) with an offset term to account for differences in sequencing depth (total UMI counts per cell), effectively normalizing the data to rates of expression. Crucially, they also include biological covariates like cell type and sex in their GLMs to see if these explain the observed zero inflation.

Datasets

The primary dataset used is from mouse heart cells (Skelly et al., 2018), profiling over 10,000 cells. They also analyzed two additional datasets for validation:

  • Mouse kidney cells (Park et al., 2018)
  • Human peripheral blood mononuclear cells (PBMCs) (10X Genomics, 2018)

Key Insights and Ideas

  • Zero inflation is real and biological: The paper provides strong evidence that zero inflation in scRNA-seq is not primarily a technical artifact but is largely driven by biological factors like cell type and sex. This challenges the notion that all zeros are simply "dropout" events to be imputed away.
  • Cell type is a major driver: When cell type is included as a covariate in the models, the number of genes classified as zero-inflated drops significantly. This suggests that differences in gene expression patterns between cell types are a key source of apparent zero inflation.
  • Sex also plays a role: For sex-specific genes (like Xist and Ddx3y), their zero inflation status changes when sex is accounted for, highlighting its biological contribution.
  • Estimated zero inflation parameter (π₀) is unreliable: The paper demonstrates that the estimated proportion of zero inflation from a ZINB model can be misleading, especially when the data is not truly zero-inflated or when models are mis-specified. This means simply looking at the π₀ value from a ZINB fit isn't a good indicator of biological zero inflation.
  • Simpler models can be better: Despite finding evidence for zero inflation, the authors ultimately recommend the Negative Binomial (NB) model (without zero inflation) as a suitable reference model for general scRNA-seq analysis. This is because the NB model, when accounting for biological covariates, can capture much of the heterogeneity that a ZINB model might try to explain with extra zeros.

Key Results

  • Using the heart dataset, they identified 1,474 genes as zero-inflated (ZI) at a 0 SE threshold, dropping to 76 genes at a more conservative 2 SE threshold.
  • When accounting for cell type, the number of ZI genes dropped dramatically. For example, out of 76 genes initially classified as ZI (at 2 SE), 72 were no longer classified as ZI after including cell type.
  • For sex-specific genes like Xist and Ddx3y, their ZI classification changed when sex was included as a covariate.
  • Simulations showed that the estimated zero inflation parameter (π₀) from a ZINB model could be highly overestimated when applied to data that was not truly zero-inflated.
  • The NB model provided accurate estimates of mean and variance even when applied to ZINB simulated data, suggesting its robustness.

Limitations and Future Work

  • Under-powered detection: The paper notes that for more stringent thresholds (like 2 SE), detecting ZI genes becomes under-powered with sample sizes up to 10,000 cells, suggesting larger sample sizes might reveal more high-confidence ZI genes.
  • Unexplained ZI genes: While many ZI genes were explained by cell type and sex, a number remained without a clear biological explanation. The authors suggest these could point to other hidden biological factors like cell cycle stage, immune cell activation, or misclassified cell types, and that scRATE can be a useful tool for identifying such genes for further investigation.
  • Model selection vs. classification: While model selection is useful for demonstrating zero inflation, they caution against using it as a gene-specific classifier for downstream analysis due to potential inflated type-I error rates. They suggest model averaging as an alternative, though computationally demanding.
  • Recommendation for NB: Their recommendation for the NB model as a reference, despite finding biological zero inflation, implies a pragmatic approach: the complexity of ZINB might not always be necessary or provide more reliable inferences than a well-specified NB model with covariates. Future work could explore optimal ways to incorporate biological covariates into simpler models.

Summary updated

Abstract

Abstract Single-cell RNA sequencing is a powerful tool for characterizing cellular heterogeneity in gene expression. However, high variability and a large number of zero counts present challenges for analysis and interpretation. There is substantial controversy over the origins and proper treatment of zeros and no consensus on whether zero-inflated count distributions are necessary or even useful. While some studies assume the existence of zero inflation due to technical artifacts and attempt to impute the missing information, other recent studies of argue that there is no zero inflation in scRNA-Seq data. We apply a Bayesian model selection approach to unambiguously demonstrate zero inflation in multiple biologically realistic scRNA-Seq datasets. We show that the primary causes of zero inflation are not technical but rather biological in nature. We also demonstrate that parameter estimates from the zero-inflated negative binomial distribution are an unreliable indicator of zero inflation. Despite the existence of zero inflation of scRNA-Seq counts, we recommend the generalized linear model with negative binomial count distribution (not zero-inflated) as a suitable reference model for scRNA-Seq analysis.
BibTeX
@misc{choi2020bayesian,
  author = {Kwangbom Choi and Yang Chen and Daniel A. Skelly and Gary A. Churchill},
  title = {{Bayesian model selection reveals biological origins of zero inflation in single-cell transcriptomics}},
  year = {2020},
  doi = {10.1101/2020.03.03.974808},
  howpublished = {bioRxiv},
}
Figure