Abstract
Mass cytometry technology enables the simultaneous measurement of over 40 proteins on single cells, providing deep insights into white blood cell heterogeneity. However, current statistical methods often collapse this rich single-cell data into summary statistics, discarding valuable multivariate information. This paper introduces a statistical framework that analyzes raw, uncompressed data to improve replicability and expose multivariate patterns with associated uncertainty profiles. By proposing multivariate Poisson log-normal and logistic linear mixed models, the authors demonstrate that generative models are a valid alternative to univariate testing. Using Hamiltonian Monte Carlo for Bayesian uncertainty quantification, the models successfully reproduce findings from a pregnancy study while identifying increased protein-to-protein correlations between the first and third trimesters.
The Challenge of High-Dimensional Single-Cell Data
Mass cytometry (CyTOF) allows researchers to measure dozens of proteins across thousands of cells per second. Despite this resolution, the standard analytical pipeline typically involves "gating" (clustering) cells into known types and then calculating summary statistics, such as median expression levels. This "premature summarization" is problematic; it ignores how proteins covary within individual cells and often fails to account for donor-level variability or measurement uncertainty. By discarding the multivariate structure of the raw data, researchers may lose statistical power or encounter misleading results.
Proposed Methodology: Multivariate Mixed Models
To address these limitations, the authors propose two complementary regression models that operate directly on uncompressed data:
- Poisson Log-Normal Mixed Model (PLMM): This model treats protein counts as response variables predicted by experimental conditions (e.g., pregnancy trimesters). It accounts for correlations between markers through nested random effects at both the cell and donor levels.
- Logistic Linear Mixed Model (LLMM): This model uses protein expression levels as predictors to determine the experimental condition.
Inference is performed using Hamiltonian Monte Carlo (HMC) via the Stan programming language. This Bayesian approach generates full posterior distributions, allowing for rigorous uncertainty quantification beyond simple p-values.
Resolving Contradictions through Causal Inference
A significant contribution of this work is the use of Directed Acyclic Graphs (DAGs) to interpret discrepancies between the PLMM and LLMM. When these models yield conflicting results, it often signals the presence of "collider" or "pipe" confounders.
In a re-analysis of a 2017 pregnancy study involving 18 subjects, the authors encountered such a contradiction. While the PLMM showed increases in signaling proteins pSTAT3 and pSTAT5 during the third trimester, the LLMM initially suggested a decrease. By applying their causal framework, the authors identified pSTAT1 as a "collider." Removing pSTAT1 from the model resolved the contradiction, proving the apparent decrease was a statistical artifact.
Key Findings and Replicability
The models successfully corroborated known biological signals, such as the increase in pSTAT1 during late pregnancy. More importantly, they uncovered a novel multivariate insight: as pregnancy progresses, the "connectivity" or correlation between signaling proteins in the immune system increases. By utilizing raw data, the models proved more sensitive to subtle signals that are typically obscured by median-based summarization.
Limitations and Future Directions
The primary challenge for these multivariate models is computational intensity. Fitting the PLMM to 180,000 cells took approximately six days on an eight-core machine. Future work may explore Variational Inference as a faster alternative for exploratory phases. Additionally, while the current workflow still relies on an initial gating step, the authors aim to integrate clustering and regression into a single unified Bayesian model to propagate uncertainty throughout the entire pipeline. To support the community, the authors have released the R packages CytoGLMM and cytoeffect.
Summary updated