# Bayesian multi-cell type models for the analysis of complex immune cell populations with application to ovarian cancer

## TL;DR

This paper introduces a Bayesian statistical model to better analyze immune cell populations in ovarian cancer, improving understanding of how these cells relate to cancer stage and treatment.

## Contribution

A novel Bayesian hierarchical model using beta-binomial distribution to analyze multiple immune cell types together in cancer studies.

## Key findings

- The multi-cell type model detected more associations with narrower credible intervals compared to individual cell type analyses.
- The model was applied to ovarian cancer data from three studies, revealing links between immune cell abundance and cancer stage, age, and debulking status.
- The R package BTIME was developed to implement the model and includes a tutorial for broader use.

## Abstract

To understand how the tumor immune microenvironment (TIME) impacts clinical outcomes and treatment response, researchers have been leveraging single-cell protein multiplex imaging techniques. These technologies measure multiple protein markers simultaneously within a tissue sample, providing a more complete assessment of the TIME. However, statistical challenges arise from the over-dispersed and zero-inflated nature of the data and from relationships among different immune cell populations. To address these challenges, we developed a Bayesian hierarchical method using a beta-binomial (BB) distribution to model the abundance of multiple immune cell types simultaneously while incorporating relationships and immune cell differentiation paths. We applied the model to data from three large studies of high-grade serous ovarian tumors (Nurses’ Health Study I/II: N = 321, African American Cancer Epidemiology Study: N = 92, University of Colorado Ovarian Cancer Study: N = 103). We examined associations between cancer stage, age at diagnosis, and debulking status and the abundance of immune cell populations. We compared the multi-cell type model to individual cell type analyses using a Bayesian BB model. The multi-cell type model detected more associations, when present, with narrower credible intervals. To support broader application, we developed an R package, BTIME, with a detailed tutorial. In conclusion, the Bayesian multi-cell type model is flexible in how relationships between cell types are incorporated and can be used for cancer studies that interrogate the TIME.