Effector T-cell cytolytic activity modules derived from CD3+ single cells from human primary triple-negative breast cancer (TNBC) in multiple solid tumors to predict response to immune checkpoint blockade therapy (ICB).

Abstract

1073 Background: The prognostic and predictive value of tumor infiltrating lymphocytes for ICB has been recognized in a variety of tumor types, including TNBC. Nonetheless, our understanding of the mechanistic aspects of T cell activation remains incomplete. We hypothesize that a specific effector phenotype of T cell cytolytic activity (ECA) is a consistent feature of epithelial tumors, possibly varying by tumor types with a range of inflammatory features. Methods: We evaluated 6,311 purified CD3+ single cells from human primary TNBC and computed sample set enrichment scores of a set of previously published immune metagenes. Following unsupervised clustering of the enrichment scores of the entire single cell population, two subgroups of cells with highest and lowest average enrichment score of T cell cytolytic activity formed a basis for detecting functional gene expression modules. Spectral decomposition and Jackstraw analysis estimated eight modules with overlapping sets of genes. Each gene expression module was then used to train a Random Forest classifier of ECA phenotype. Results: We discovered that our module-derived classifiers were prognostic not only in TNBC samples obtained from both TCGA (N = 150) and METABRIC (N = 320) datasets but also in 14 other tumor types encompassing 6,000 samples. For example, patient samples from TCGA dataset predicted to be in group ECA ‘High’ have better progression-free survival (p-value: 0.0098l; HR: 0.30) and better overall survival (p-value: 0.0066; HR: 0.17). In both breast datasets, gene within the classifier are relatively under-expressed in ER+ tumors as opposed to HER2+ and TNBC (p-value < 2.2e-16). In a dataset of normal, pure DCIS and mixed DCIS (GSE26304;N = 114), the same genes were relatively under-expressed in DCIS samples relative to invasive tumors (p-value < 2.2e-16). Additionally, in a pre-therapy tumor dataset of fifty-one advanced melanoma patients treated with Nivolumab, who previously either progressed on ipilimumab or were ipilimumab-naïve, our module-derived classifier was able to classify responders and non-responders with 77% accuracy (p-value = 0.02) and was associated with progression-free survival (p-value = 0.03; HR: 0.28). Conclusions: Our study highlights one important application of single-cell genomics in our understanding of immune microenvironment and potentially identify new immunotherapy targets.