CytoGLMM: Conditional Differential Analysis for Flow and Mass Cytometry Experiments

Christof Seiler, Anne-Maud Ferreira, Lisa M. Kronstad, Laura J. Simpson, Mathieu Le Gars, Elena Vendrame, Catherine A. Blish, Susan HolmesbioRxiv 2020 · DOI

Abstract

Flow and mass cytometry are important modern immunology tools for measuring expression levels of multiple proteins on single cells. The goal is to better understand the mechanisms of responses on a single cell basis by studying differential expression of proteins. We focus on cell-specific differential analysis and one fixed cell type. In contrast, most current methods learn cell types and perform differential analysis jointly. Our narrower field of application allows us to define a more specific statistical model with easier to control statistical guarantees. Differential analysis of marker expressions can be difficult due to marker correlations and inter-individual heterogeneity, particularly for studies of human immunology. We address these challenges with two multiple regression strategies: A bootstrapped generalized linear model and a generalized linear mixed model. On simulated datasets, we compare the robustness towards marker correlations and heterogeneity of both strategies. For paired experiments, we find that both strategies maintain the target false discovery rate under medium correlations and that mixed models are statistically more powerful under the correct model specification. For unpaired experiments, our results indicate that much larger patient sample sizes are required to detect differences. We illustrate the CytoGLMM R package and workflow for both strategies on a pregnancy dataset. Our approach to find differential proteins in flow and mass cytometry data reduces biases arising from maker correlations and safeguards against false discoveries induced by patient heterogeneity.

Introduction

In flow and mass cytometry, researchers measure dozens of proteins across millions of individual cells to understand immune responses. However, standard statistical workflows face two major hurdles: marker correlation and inter-individual heterogeneity. Proteins within a cell are often part of the same signaling pathways; if one protein activates another, a standard marginal analysis will flag both as differentially expressed, obscuring the true driver. Furthermore, baseline protein expression varies significantly between human donors. Failing to account for this "patient effect" can lead to natural variation being mistaken for experimental results.

Methodology

CytoGLMM addresses these issues using a "conditional" approach via multiple regression. Instead of asking if a marker changes with the condition in isolation, it asks if the marker changes given the levels of all other markers. This allows the model to "rule out" markers that are only correlated with true drivers.

The package introduces two strategies:

  1. Bootstrapped GLM: Designed for unpaired samples, it uses a patient-level bootstrap—resampling entire donors rather than individual cells—to ensure the model accounts for the hierarchical structure of the data.
  2. Generalized Linear Mixed Model (GLMM): Designed for paired samples (e.g., pre- and post-stimulus from the same donor), it includes a random effect term to explicitly model each donor's unique baseline.

Results and Discussion

Simulations evaluated the robustness of these strategies against varying levels of donor heterogeneity ($\sigma_U$) and marker correlation ($\sigma_B$).

Statistical power comparison between GLM and GLMM strategies across varying levels of donor heterogeneity ($\sigma_U$) and marker correlation ($\sigma_B$).

As shown in Figure 3, GLMMs are significantly more efficient for paired experiments. By treating patient variation as structured signal rather than noise, GLMMs reached an 80% power threshold with only 7 donor pairs, compared to 13 pairs required for the bootstrapped GLM. Both methods successfully maintained the False Discovery Rate (FDR) below the 5% target under medium correlations.

Power and False Discovery Rate (FDR) for unpaired experiments under different correlation ($\rho_B$) and heterogeneity ($\sigma_B$) scenarios.

For unpaired experiments (Figure 2), the results highlight the impact of donor heterogeneity. As $\sigma_U$ increases, power drops significantly, indicating that much larger sample sizes are required to detect differences in unpaired human studies compared to paired designs. Despite the drop in power, FDR control remains robust across most scenarios.

Biological Validation

The CytoGLMM workflow was applied to a mass cytometry dataset tracking immune responses in 16 pregnant women. The analysis identified pSTAT1 as a major predictor of the third-trimester response to IFN$\alpha$ in Natural Killer cells. CytoGLMM provided a cleaner statistical backing for this finding by demonstrating pSTAT1's importance even when accounting for other signaling proteins like pSTAT3 and pSTAT5.

Conclusion

CytoGLMM provides a robust framework for identifying differential proteins in cytometry data while reducing biases from marker correlations and donor heterogeneity. While it requires pre-gated populations and is currently optimized for two-group comparisons, it offers a powerful tool for more precise immunological discovery.

Summary updated

Abstract

Abstract Background Flow and mass cytometry are important modern immunology tools for measuring expression levels of multiple proteins on single cells. The goal is to better understand the mechanisms of responses on a single cell basis by studying differential expression of proteins. We focus on cell-specific differential analysis and one fixed cell type. In contrast, most current methods learn cell types and perform differential analysis jointly. Our narrower field of application allows us to define a more specific statistical model with easier to control statistical guarantees. Results Differential analysis of marker expressions can be difficult due to marker correlations and inter-individual heterogeneity, particularly for studies of human immunology. We address these challenges with two multiple regression strategies: A bootstrapped generalized linear model and a generalized linear mixed model. On simulated datasets, we compare the robustness towards marker correlations and heterogeneity of both strategies. For paired experiments, we find that both strategies maintain the target false discovery rate under medium correlations and that mixed models are statistically more powerful under the correct model specification. For unpaired experiments, our results indicate that much larger patient sample sizes are required to detect differences. We illustrate the CytoGLMM R package and workflow for both strategies on a pregnancy dataset. Conclusions Our approach to find differential proteins in flow and mass cytometry data reduces biases arising from maker correlations and safeguards against false discoveries induced by patient heterogeneity.
BibTeX
@misc{seiler2020cytoglmm,
  author = {Christof Seiler and Anne-Maud Ferreira and Lisa M. Kronstad and Laura J. Simpson and Mathieu Le Gars and Elena Vendrame and Catherine A. Blish and Susan Holmes},
  title = {{CytoGLMM: Conditional Differential Analysis for Flow and Mass Cytometry Experiments}},
  year = {2020},
  doi = {10.1101/2020.12.09.417584},
  howpublished = {bioRxiv},
}
Figure