Infino: Bayesian Inference to Extract Distinguishing Traits of Immune Cell Expression Phenotypes and Compute Immune Infiltrate Abundance in Tumor Microenvironment

Abstract

Several new cancer therapies perform remarkably well for a fraction of patients, but are not universal. The uniqueness of those patients continues to puzzle cancer biologists. Robust immune infiltrate quantification may provide useful insights and biomarkers to clarify why those patients in particular are benefited by treatment. The immune cells that surround a tumor are informative, but they are slow and expensive to study. We estimate the relative abundances of the immune cell types around a tumor by deconvolving gene expression in the region in silico. First, we review the existing methods and their performance on synthetic mixtures of known composition. Our analysis suggests that existing techniques confuse similar cell types and fail to accurately characterize the rate of error in the deconvolution results. We then motivate and develop a Bayesian mixture deconvolution method tailored to RNA-seq data. Critically, we use information from the hierarchy of immune cell types to improve our deconvolution results. Our method produces rich probability distributions of immune infiltration in RNA-seq mixtures. Using the Cancer Genome Atlas clinical dataset, we can assess the association of immune infiltration with response to immunotherapy. The results could help remove a key obstacle to the adoption of new cancer therapies.