Immune Profiling among Colorectal Cancer Subtypes using Dependent Mixture Models

Abstract Comparison of transcriptomic data across different conditions is of interest in many biomedical studies. In this paper, we consider comparative immune cell profiling for early-onset (EO) versus late-onset (LO) colorectal cancer (CRC). EOCRC, diagnosed between ages 18-45, is a rising public health concern that needs to be urgently addressed. However, its etiology remains to be poorly understood. We work towards filling this gap by identifying homogeneous T cell subpopulations that show significantly distinct characteristics across the two tumor types, and to identify others that are shared between EOCRC and LOCRC. Such inference may reveal underlying determinants of clinically observed differences in the two disease subpopulations. We develop dependent finite mixture models where immune subtypes enriched under a specific condition are characterized by terms in the mixture model with common atoms but distinct weights across conditions, whereas common subtypes are characterized by sharing both atoms and relative weights. The proposed model defines a variation of mixtures of finite mixture models, facilitating the desired comparison by introducing highly structured multi-layer Dirichlet priors. The model allows us to explicitly compare features across conditions. We illustrate inference with simulation studies and data examples. Results identify EO-enriched and LO-enriched T cells subtypes whose biomarkers are found to be linked to mechanisms of tumor progression. The findings reveal distinct characteristics of the immune profiles in EOCRC and LOCRC, and potentially motivate insights into treatment and management of CRC.