Characterization of the tumor microenvironment landscape and deep learning-guided prediction of prognosis in lung adenocarcinoma with bulk RNA sequencing data.

Abstract

e21025 Background: The tumor microenvironment (TME) plays an important role in tumor progression and treatment response, therefore profoundly affecting patient outcomes. Efforts to characterize the TME in lung adenocarcinoma are emerging but have been limited by the sample size and lack of treatment timepoints. Methods: We characterized changes in lung TME using the xCell algorithm to distinguish 64 immune and stroma cell types from bulk RNA sequencing data. The correlation between subtype cell population in lung TME and various clinical and biological characteristics was analyzed from over 500 lung adenocarcinoma (LUAD) samples from The Cancer Genome Atlas and an independent cohort of 48 advanced LUAD patients with treatment annotations (treatment-naive, residual disease, and progressive disease). In addition, we used key features in lung TME to predict prognosis using deep learning algorithms. Results: We found significant changes in both immune and stroma cell populations according to various clinical parameters such as smoking history, cancer stage, and treatment status. Specific sub-populations within lung TME correlate with survival outcomes based on Kaplan-Meier survival analyses. CD4- and CD8-positive T-cells are enriched in early stage disease and depleted in late stage disease, suggesting evolution of the TME during cancer progression. Consistent with previous reports, scores of immune cell populations associated with worse survival, such as T helper type 2 cells, are increased in late stage disease. Smoking history also reshapes the lung TME as populations correlated with better survival are decreased in smokers. We also found variations in sub-populations according to the driver oncogenes, with a less abundant lymphoid compartment in EGFR mutant samples compared to KRAS driven samples. Interestingly, we found higher scores of macrophage populations in residual disease following targeted-therapy treatment compared to pre-treatment. Finally, using machine and deep learning methods we identified a panel of 12 key features within the lung TME which could be used to predict prognosis. Conclusions: We comprehensively characterized immune and stroma cell type changes in the lung TME utilizing bulk RNA-seq data, and evaluated the association of sub-type cell populations with different clinical and biological features. Key features in lung TME could be used to predict prognosis.