Statistical inference for cell type deconvolution

Dongyue Xie, Lin Gui, Jingshu WangJournal of the Royal Statistical Society Series B (Statistical Methodology) 2026 · DOI

TL;DR

This paper introduces MEAD, a novel statistical framework for cell type deconvolution that accounts for measurement errors and platform-specific scaling differences. MEAD provides more reliable statistical inference for cell type proportions, especially when comparing individuals, by explicitly modeling uncertainty and correlations, a crucial step for robust biological discovery.

What Problem Does This Paper Address?

Estimating the proportions of different cell types within a tissue sample (cell type deconvolution) is vital for understanding biological processes and disease. Typically, this is done by using reference datasets (like single-cell RNA sequencing) to infer the composition of bulk RNA sequencing data. However, this process is fraught with challenges:

  • Platform Differences: Data from different technologies (e.g., bulk RNA-seq vs. single-cell RNA-seq) have inherent scaling effects and measurement noise that can bias results.
  • Uncertainty in Estimates: Current methods often treat the estimated cell type proportions as if they were perfectly known. This "naive" approach can lead to inflated false positives and unreliable conclusions when comparing cell type compositions across different individuals or conditions.
  • Gene-Gene Correlations: Genes within a sample are not independent, and these correlations can complicate statistical inference.

MEAD aims to address these issues by providing a statistically sound framework for both estimating cell type proportions and performing valid inference on these estimates, particularly for cross-individual comparisons.

Methods

MEAD introduces a statistical framework that tackles the challenges of cell type deconvolution through several key innovations:

  • Measurement Error Adjustment: MEAD explicitly models and corrects for measurement errors and platform-specific scaling differences between reference and target data. This is achieved by incorporating these factors into a generalized linear model.
  • Identifiability Conditions: The paper establishes the theoretical conditions under which cell type proportions can be uniquely identified, even with arbitrary gene-specific scaling differences across platforms. This is a crucial theoretical contribution.
  • Valid Statistical Inference: MEAD provides methods for constructing confidence intervals and performing hypothesis tests that account for the uncertainty in the estimated cell type proportions. This is achieved through techniques like sandwich-type estimators and a Softplus transformation to handle proportions near zero.
  • Accounting for Gene-Gene Correlations: The framework incorporates gene-gene correlations into the inference process, leading to more robust results.
  • Weighting Strategy: MEAD proposes a gene-weighting strategy that is inversely proportional to gene noise variance, improving estimation efficiency.

Datasets

The paper uses the following datasets for evaluation:

  • Pancreatic Islet scRNA-seq Data: From Xin et al. (2016) and Segerstolpe et al. (2016), used for both reference data and cross-platform deconvolution benchmarks.
  • Neuron Development scRNA-seq Data: From Jerber et al. (2021), used for benchmarking across multiple individuals.

The paper also generates synthetic data based on parameters estimated from real scRNA-seq data to rigorously test the method's performance under controlled conditions.

Key Insights and Ideas

  • Ignoring Uncertainty is Risky: Treating estimated cell type proportions as ground truth in downstream analyses can lead to misleading conclusions, especially when comparing groups of individuals. MEAD's explicit modeling of uncertainty is critical for reliable inference.
  • Identifiability is Non-Trivial: Simply having more data (more individuals) doesn't automatically guarantee that cell type proportions can be identified. Specific structural properties in the gene expression data are required, particularly when dealing with cross-platform scaling differences.
  • Gene Selection Isn't Always Optimal: The common practice of selecting "marker genes" for deconvolution doesn't necessarily improve estimation accuracy and can sometimes be detrimental. MEAD's approach of weighting genes by their noise variance is more effective.
  • Cross-Platform Differences Matter: MEAD's ability to explicitly model and correct for gene-specific scaling differences across platforms is a significant advantage over methods that ignore these effects.

Key Results

  • Improved Inference: MEAD provides statistically valid confidence intervals for cell type proportions, outperforming methods that ignore estimation uncertainty.
  • Accurate Estimation: In simulations and real-data analyses, MEAD demonstrates competitive to superior estimation accuracy compared to existing methods like CIBERSORT, MuSiC, and OLS.
  • Robustness: MEAD shows improved performance in estimating differences in cell type proportions between groups, particularly in real-world scenarios with complex data characteristics.
  • Practical Utility: The framework is designed to be applicable to various downstream tasks, including comparing cell type proportions across individuals and relating them to covariates.

Limitations and Future Work

  • Non-Informative Scaling Ratios Assumption: MEAD's inference procedure relies on the assumption that gene-specific scaling ratios are non-informative (i.e., independent of true gene expression levels). While this simplifies estimation, it may not always hold in practice, especially if scaling ratios are correlated with mean expressions. Relaxing this assumption is a direction for future research.
  • Missing or Over-Partitioned Cell Types: The paper briefly discusses the effects of missing cell types in reference data or over-partitioning cell types into subtypes, noting that these complexities are beyond the scope of the current work.
  • Bayesian Framework: The authors suggest that a hierarchical Bayesian model could offer an alternative approach, potentially relaxing some of MEAD's assumptions, but acknowledge the challenges in fully specifying such a high-dimensional model.

Summary updated

Abstract

Abstract Integrating heterogeneous datasets across different measurement platforms poses fundamental challenges for statistical inference. An important example is cell type deconvolution, where cell type proportions in bulk RNA-seq data are estimated using reference single-cell data from different sources, leading to platform-specific scaling effects, measurement noise, and biological heterogeneity. Existing methods often treat estimated proportions as observed in downstream analyses, potentially compromising validity when comparing multiple individuals. We introduce measurement error adjusted deconvolution, a statistical framework for estimation and inference in deconvolution with externally approximated design matrices. We establish necessary and sufficient conditions for identifiability under arbitrary gene-specific cross-platform scaling differences and develop valid inferential procedures for both individual-level proportions and comparisons across individuals, accounting for gene–gene correlation and shared estimation uncertainty. Simulations and real-data analyses demonstrate competitive estimation accuracy and reliable statistical inference.
BibTeX
@article{xie2026statistical,
  author = {Dongyue Xie and Lin Gui and Jingshu Wang},
  title = {{Statistical inference for cell type deconvolution}},
  journal = {Journal of the Royal Statistical Society Series B (Statistical Methodology)},
  year = {2026},
  doi = {10.1093/jrsssb/qkag054},
}
Figure