Machine learning predicts cancer subtypes and progression from blood immune signatures.

Abstract

Clinical adoption of immune checkpoint inhibitors in cancer management has highlighted the interconnection between carcinogenesis and the immune system. Immune cells are integral to the tumour microenvironment and can influence the outcome of therapies. Better understanding of an individual's immune landscape may play an important role in treatment personalisation. Peripheral blood is a readily accessible source of information to study an individual's immune landscape compared to more complex and invasive tumour bioipsies, and may hold immense diagnostic and prognostic potential. Identifying the critical components of these immune signatures in peripheral blood presents an attractive alternative to tumour biopsy-based immune phenotyping strategies. We used two syngeneic solid tumour models, a 4T1 breast cancer model and a CT26 colorectal cancer model, in a longitudinal study of the peripheral blood immune landscape. Our strategy combined two highly accessible approaches, blood leukocyte immune phenotyping and plasma soluble immune factor characterisation, to identify distinguishing immune signatures of the CT26 and 4T1 tumour models using machine learning. Myeloid cells, specifically neutrophils and PD-L1-expressing myeloid cells, were found to correlate with tumour size in both the models. Elevated levels of G-CSF, IL-6 and CXCL13, and B cell counts were associated with 4T1 growth, whereas CCL17, CXCL10, total myeloid cells, CCL2, IL-10, CXCL1, and Ly6Cintermediate monocytes were associated with CT26 tumour development. Peripheral blood appears to be an accessible means to interrogate tumour-dependent changes to the host immune landscape, and to identify blood immune phenotypes for future treatment stratification.