用于分析空间单细胞蛋白成像数据的多细胞类型模型及其在卵巢癌中的应用

Background: Understanding the tumor immune microenvironment (TIME) is essential for advancing cancer research and improving treatment strategies. Multiplex immunofluorescence (mIF) is a spatial proteomics imaging technique enabling simultaneous analysis of multiple markers in preserved tissues. However, mIF-derived cell abundance data pose statistical challenges, such as zero-inflation, over-dispersion, hierarchical cell relationships, and repeated measures, that must be addressed to extract meaningful insights and enhance translational impact. Methods: We developed a novel Bayesian multi-cell type analysis model that simultaneously models the relationship of immune cell abundances with clinical and epidemiological factors, while incorporating the biological relationships between immune cell populations. We applied this model to three large studies assessing the TIME of high-grade serous ovarian cancer: Nurses' Health Study I/II (NHSI/II) (N=321), African American Cancer Epidemiology Study (AACES) (N=92), and University of Colorado Ovarian Cancer Study (UCOCS) (N=103). The mIF staining for these studies was performed using the AKOYA Biosciences OPAL TM 7-Color Automation IHC Kit with the Vectra ® 3 Automated Quantitative Pathology Imaging System (0.499µm/pixel) utilized for image collection. InForm and HALO were utilized for spectral unmixing and cell phenotyping, respectively. Our analysis examined associations between immune cell infiltration (T-cells, B-cells, macrophages) and clinical variables (cancer stage, age at diagnosis, debulking status) with comparisons to the single-cell type model. Results: In the NHSI/II analysis, our multi-cell type model detected a positive association between age at diagnosis and abundance levels of 6 of the 7 cell types in the analysis while the single-cell type model only detected 2 of the 7. This indicates improved association detection, with our multi-cell type model. We also observed that our multi-cell type model had narrower credible intervals (CIs) for all 7 cell types demonstrating higher accuracy in the association estimation. With cancer stage as the predictor in the NHSI/II analysis, neither model detected an association although our multi-cell type model had narrower CIs for 3 of the 7 cell types compared to the single-cell type model. Despite not capturing any associations between the predictors (age, stage, debulking status) and immune cell populations in the AACES or UCOCS, our Bayesian multi-cell type model had narrower CIs for every cell type in both studies for each predictor. Discussion: Our Bayesian multi-cell type model offers a flexible framework for incorporating immune cell relationships and is well-suited for cancer studies of the TIME utilizing TMAs, regions of interest, or whole-slide imaging data.