Bayesian Spatial Models for Cancer Imaging Data

Abstract

The tumor microenvironment is complex, consisting of many discrete components which interact in complex ways. Of these components, one that is of particular interest is the immune composition of the tumor microenvironment. This composition refers not only to which types of immune cells are present or absent in a given tumor, but also how these cell types tend to spatially interact (or not) with one another. In this dissertation, we present three novel methods that utilize various types of imaging data to attempt to better understand these relationships and their associations with clinical and biological outcomes of interest.

The first of these is the SPatial Analysis of paRtitioned Tumor-Immune imagiNg (SPARTIN) method. SPARTIN utilizes data derived from high definition images of hematoxylin and eosin (H&E) stained biopsies; specifically, data on cell types and locations. Cells from these biopsies can be classified as either tumor cells or immune cells, and the location of each cell can be ascertained. SPARTIN quantifies interaction between tumor cells and immune cells by partitioning the biopsies into non-overlapping sub-regions and computing a novel metric of interaction on each of the resulting sub-regions. We applied SPARTIN to a set of 335 high definition images of Skin Cutaneous Melanoma (SKCM) biopsies, and found several significant associations with various patient level outcomes of interest, including gene expression, cancer subtype, and overall survival.

The second method, the Dual random effect and main effect selection model for Spatially structured regression (DreameSpase), builds on SPARTIN. While SPARTIN focused on the quantification of heterogeneity of tumor-immune interaction between biopsies, the partitioning of biopsies into non-overlapping sub-regions raises a natural question about the heterogeneity in interaction within biopsies, particularly with respect to biopsy level gene expression. Given a partitioned image of an H&E biopsy, DreameSpase jointly models the heterogeneity of interaction both within and between biopsies with respect to biopsy level gene expression data. This model enabled us to study the association between a targeted subset of genes known to be associated with specific immune cells. We found that several genes that have been previously shown to be associated with patient prognosis in melanoma were associated with both inter- and intra-biopsy heterogeneity, including PLXNC1, S100A8, and S100A9.

The third method, the Spatial Logistic Auxiliary Control method (SLAC), operates on data from multiplex images. Unlike H&E imaging, multiplex imaging data can have theoretically unlimited cellular phenotypes that are measured in a single biopsy. While this technology is incredibly powerful in its ability to identify arbitrarily specific cell types in the tumor microenvironment, it also presents novel challenges as to the best way to jointly model the resulting interactions. To address this problem, we introduce the SLAC method. The SLAC method offers a flexible framework in which to jointly model the interaction between the various cell types present in a multiplex image and perform joint regularization and selection on those interactions. We applied the SLAC method to a data set of 1,268 multiplex images of non-small cell lung cancer (nSCLC), and found that interaction between certain types of PD-L1 expressing cells were more common in patients who responded to immunotherapy than those who did not.