A flexible model for correlated count data, with application to multi-condition differential expression analyses of single-cell RNA sequencing data

Yusha Liu, Peter Carbonetto, Michihiro Takahama, Adam Gruenbaum, Dongyue Xie, Nicolas Chevrier, Matthew StephensarXiv 2022 · DOI · arXiv

TL;DR

Poisson mash is a new statistical framework designed to analyze single-cell RNA sequencing (scRNA-seq) data across dozens of experimental conditions simultaneously. By directly modeling raw molecular counts and "borrowing strength" across related conditions using a flexible empirical Bayes prior, it significantly improves the power to detect differentially expressed genes and the accuracy of estimated fold changes compared to traditional methods.

What Problem Does This Paper Address?

In modern genomics, researchers are moving beyond simple "treatment vs. control" experiments. They are now profiling thousands of single cells across dozens of different conditions—for example, exposing immune cells to 45 different signaling molecules (cytokines).

This presents two major challenges for standard "Differential Expression" (DE) analysis:

  1. Complex Sharing Patterns: A gene might respond to only one specific cytokine, a subset of related cytokines, or all of them. We don't know these patterns in advance, but exploiting them could make our analysis much more powerful.
  2. The "Log-Transform" Bias: scRNA-seq data consists of discrete molecular counts, often with many zeros. The standard practice of log-transforming these counts to use Gaussian-based models introduces severe biases and underestimates the true magnitude of expression changes, especially for lowly expressed genes.

Existing tools either handle counts well but treat conditions independently (losing power), or they model sharing well but rely on biased log-transformed data.

Methods

The authors introduce Poisson multivariate adaptive shrinkage (Poisson mash). The method consists of three core components:

  • Poisson Measurement Model: It models the raw UMI counts directly, avoiding the biases of log-transformation.
  • The "mash" Prior: It uses a flexible Empirical Bayes prior—a mixture of multivariate normal distributions. This allows the model to learn from the data itself which conditions tend to show similar effects (e.g., learning that IL-1$\alpha$ and IL-1$\beta$ treatments produce similar gene expression signatures).
  • Variational Inference: Because combining Poisson likelihoods with multivariate normal priors makes the math "intractable" (you can't solve it exactly), the authors use a Gaussian variational approximation to provide fast, scalable posterior computations.

The model also includes an optional RUV (Removing Unwanted Variation) component, which uses latent factor modeling to account for batch effects or other technical noise that might confound the biological signal.

Datasets

The authors evaluated Poisson mash using:

  • Simulated scRNA-seq data: Created via "binomial thinning" of real data. This technique preserves the realistic sparsity and sequencing depth variation of actual experiments while allowing the authors to "ground truth" the biological effects.
  • Cytokine Stimulation Dataset: A massive real-world dataset containing ~142,000 single cells from mice, spanning 8 cell types and 45 different treatment conditions (44 cytokines plus one control).
  • Bulk RNA-seq simulations: To demonstrate that the model is generalizable to other types of multivariate count data.

Key Insights and Ideas

  • Borrowing Strength: The biggest "aha moment" is that modeling all 45 conditions together is better than doing 45 separate analyses. If the model learns that a group of cytokines acts similarly, it can use the data from all of them to more accurately estimate the effect of any single one.
  • Pseudobulk is Sufficient: The authors show that "pseudobulk" analysis—summing counts across cells in the same condition—works remarkably well when combined with a Poisson model and RUV, simplifying the computation without losing the benefits of single-cell resolution.
  • Regularization without Over-smoothing: The Empirical Bayes approach provides "adaptive shrinkage." It shrinks noisy, small effects toward zero (denoising) while preserving the magnitude of strong, clear biological signals.

Key Results

  • Increased Power: In simulations with 25 conditions, Poisson mash RUV achieved significantly higher power to detect differentially expressed genes at the same False Discovery Rate (FDR) compared to established tools like edgeR, MAST, and limma.
  • Superior Accuracy: The Root Mean Squared Error (RMSE) for estimated log-fold changes was drastically lower for Poisson mash, particularly for genes with low baseline expression where other methods typically struggle.
  • Biological Discovery: Applying the tool to the cytokine data revealed clear "modules" of gene response. For example, they identified a core interferon-response signature shared across all cell types, as well as highly specific responses (like a G-CSF response module unique to neutrophils).

Limitations and Future Work

  • The "Small R" Problem: The model relies on having enough conditions to learn the sharing patterns. When the number of conditions is very small (e.g., $R < 10$), it can struggle to distinguish true biological sharing from random experimental noise.
  • Replicate Handling: The current implementation assumes one aggregated "observation" per condition. While the authors provide a theoretical extension for multiple biological replicates per condition in the appendix, it is not yet fully implemented in the software.
  • Computational Intensity: While variational inference is fast, the memory requirements for very large multi-condition datasets can be high (the cytokine analysis required ~100GB of RAM). Future optimizations could focus on reducing this footprint.

Summary updated

Abstract

Detecting differences in gene expression is an important part of single-cell RNA sequencing experiments, and many statistical methods have been developed for this aim. Most differential expression analyses focus on comparing expression between two groups (e.g., treatment vs. control). But there is increasing interest in multi-condition differential expression analyses in which expression is measured in many conditions, and the aim is to accurately detect and estimate expression differences in all conditions. We show that directly modeling single-cell RNA-seq counts in all conditions simultaneously, while also inferring how expression differences are shared across conditions, leads to greatly improved performance for detecting and estimating expression differences compared to existing methods. We illustrate the potential of this new approach by analyzing data from a single-cell experiment studying the effects of cytokine stimulation on gene expression. We call our new method "Poisson multivariate adaptive shrinkage", and it is implemented in an R package available online at https://github.com/stephenslab/poisson.mash.alpha.
BibTeX
@misc{liu2022flexible,
  author = {Yusha Liu and Peter Carbonetto and Michihiro Takahama and Adam Gruenbaum and Dongyue Xie and Nicolas Chevrier and Matthew Stephens},
  title = {{A flexible model for correlated count data, with application to multi-condition differential expression analyses of single-cell RNA sequencing data}},
  year = {2022},
  eprint = {2210.00697},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2210.00697},
}
Figure