Abstract
Lung Cancer remains the leading cause of cancer deaths in the USA and worldwide. Non-small cell lung cancer (NSCLC) harbors high transcriptomic intratumor heterogeneity (RNA-ITH) that limits the reproducibility of expression-based prognostic models. In this study, we used multiregional RNA-seq data (880 tumor samples from 350 individuals) from both public (TRACERx) and internal (MDAMPLC) cohorts to investigate the effect of RNA-ITH on prognosis in localized NSCLC at the gene, signature, and tumor microenvironment levels. At the gene level, the maximal expression of hazardous genes (expression negatively associated with survival) but the minimal expression of protective genes (expression positively associated with survival) across different regions within a tumor were more prognostic than the average expression. Following that, we examined whether multiregional expression profiling can improve the performance of prognostic signatures. We investigated 11 gene signatures collected from previous publications and one signature developed in this study. For all of them, the prognostic prediction accuracy can be significantly improved by converting the regional expression of signature genes into sample-specific expression with a simple function-taking the maximal expression of hazardous genes and the minimal expression of protective genes. In the tumor microenvironment, we found a similar rule also seems applicable to immune ITH. We calculated the infiltration levels of major immune cell types in each region of a sample based on expression deconvolution. Prognostic analysis indicated that the region with the lowest infiltration level of protective or highest infiltration level of hazardous immune cells determined the prognosis of NSCLC patients. Our study highlighted the impact of RNA-ITH on the prognostication of NSCLC, which should be taken into consideration to optimize the design and application of expression-based prognostic biomarkers and models. Multiregional assays have the great potential to significantly improve their applications to prognostic stratification.