Insights into the tumour immune microenvironment using tissue phenomics to drive cancer immunotherapy

Abstract

Background: The tumor immune microenvironment (TIME) may hold critical information for developing and optimizing immuno-therapeutic approaches, identifying predictive signatures, and selecting the most adequate treatment option for a given patient. Tissue phenomics facilitates the use of the TIME to derive predictive conclusions. The visual information content in histological sections is systematically converted into numerical readouts using artificial intelligence (AI). Resulting quantitative descriptors, phenes, of detected structures are mined to yield local expression profiles; this spatial data aggregation detects categories of local environments, which are correlated to clinical, genomic or other-omics data to identify relevant cohort subpopulations. Methods: Exploration of this technology is illustrated by various examples on different cohorts of NSCLC patients: A categorization of n=45 non-IO-treated patients with respect to local immune profiles learned via AI in a hypothesis-free scenario was examined. A deep learning based PD-L1 scoring was compared to 3 pathologist's scoring on n=40 durvalumab-treated patients using the cutoff 25% of tumor cells staining positive for PD-L1 at any intensity. The predictive value of a digital signature combining cell densities of PD-L1 and CD8+was tested on n=163 durvalumab-treated and n=199 non-IO-treated samples. Results: A categorization into biologically interpretable classes learned by AI illustrates the exploratory benefits of tissue phenomics. The scoring algorithm could reproduce survival prediction when compared to pathologist's visual scoring.The digital signature suggests a predictive value for patient stratification into responders and non-responders for durvalumab, while no prognostic value could be found on the non-IO-treated patients. Kaplan-Meier plots for the 2 latter examples will be presented in the poster. Conclusions: Tissue phenomics facilitates the quantitative assessment of the tumor geography and may lead to improved tools for biomarker analysis and diagnosis. Analysis on larger and prospective datasets are to be conducted in the future to strengthen the findings.