Development of gene expression signatures characterizing the tumor-immune interaction.

Abstract

205 Background: The efficacy of anti-tumor immunity depends on diverse factors, including not just abundance of immune cell populations but also activities of those populations and of tumor cells. Many of these processes are onerous to assay, but all are reflected in a tumor’s gene expression profile. Using a novel method, we develop gene expression signatures measuring a variety of biological processes underlying the tumor-immune interaction. These signatures fall into categories including antigen availability, structural barriers to immune infiltration, inhibitory signaling by both immune and tumor cells, inhibitory metabolism, pro-immune signaling, killing of tumor cells, tumor receptiveness to immune signaling, and tumor proliferation and death. Methods: We develop a method to train signatures of biological processed by synthesizing biological knowledge and large gene expression datasets. For a given process, we use literature searches and expert knowledge to derive lists of candidate genes. We then evaluate the co-expression of these candidate genes in data from The Cancer Genome Atlas (TCGA), discarding genes whose co-expression patterns are incompatible with their measuring their putative biological process. This approach safeguards the interpretability of our signatures: we only report signatures whose genes show evidence for measuring the desired biology. Finally, we further exploit co-expression patterns to obtain optimal weights for each signature gene. Results: We attempted to train signatures of over 30 biological processes involved in immune oncology. Of these, 17 candidate gene sets displayed sufficient evidence for measuring their putative biology. We show these signatures provide granular but intelligible descriptions of both immunotherapy datasets and single samples. We find they improve power in differential expression analyses and in training of predictors of drug response. Conclusions: The signatures we derive convert gene expression data into measurements of biological processes central to immune oncology, and they improve statistical power and interpretation of results in immunotherapy studies. Our training procedure ensures these signatures measure their intended biology.