Training and validation of a gene expression signature for microsatellite instability

Background: Clinical response to cancer immunotherapy can be predicted from two biologically distinct variables: tumor foreignness, manifested in some tumors as deficient DNA mismatch repair (dMMR)/microsatellite instability (MSI), typically measured by qPCR or IHC, and anti-tumor immunity, typically measured by gene expression signatures or IHC. Clinical benefit should be more accurately predicted by the combination of the two variables than by either alone, but measuring both variables requires multiple clinical assays. Here, we investigate the ability of gene expression alone to provide a surrogate measure of both tumor foreignness and immunogenicity, empowering a single assay to measure both axes of predictive biology. In addition, we explore the relationships between tumor foreignness and immunogenicity in both dMMR and MMR-proficient (pMMR) tumors. Methods: Using TCGA datasets from colon, esophageal, stomach and uterine cancers, we trained two algorithms predicting hypermutation: the first detecting loss of expression of mismatch repair genes (MLH1, MSH2, MSH6 and PMS2) and the second identifying expression patterns shared by hypermutated tumors across these four cancer types. A final algorithm synthesizes the two above algorithms into a single score. For independent validation, we evaluated 60 colorectal cancer (CRC) FFPE samples, 30 dMMR and 30 pMMR as previously identified by IHC, and a cohort of 10 MSI and 5 MSS endometrial and neuroendocrine tumors with the NanoString nCounter® platform. Results: We show that our algorithms successfully predicted hypermutation phenotypes and MMR status in TCGA training data and in the independent CRC and endometrial datasets analyzed with NanoString; with an AUROC of 0.94. Additionally, we demonstrate that higher mutational burden is linked to a heightened tumor immune environment as shown by adaptive immune gene signatures. Conclusions: Gene expression proves to be a powerful predictor of microsatellite instability and hypermutation in cancers where dMMR subtypes are known to exist. This discovery raises the possibility that a gene expression algorithm measuring both hypermutation and immune activity may be advantaged in predicting response to checkpoint inhibition in these cancer types. Legal entity responsible for the study: NanoString Technologies. Funding: NanoString Technologies. Disclosure: S.E. Warren, S. Ong, N. Elliott, A. Cesano: Employee and stockholder: NanoString Technologies. P. Danaher: Stockholder: NanoString Technologies.