Abstract
Lung adenocarcinoma (LUAD) is a frequently diagnosed cancer type, and many patients have already reached an advanced stage when diagnosed. Thus, it is crucial to develop a novel and efficient approach to diagnose and classify lung adenocarcinoma at an early stage. In our study, we combined in silico analysis and machine learning to develop a new five-gene-based diagnosis strategy, which was further verified in independent cohorts and in vitro experiments. Considering the heterogeneity in cancer, we used the MATH (mutant-allele tumor heterogeneity) algorithm to divide patients with early-stage LUAD into two groups (C1 and C2). Specifically, patients in C2 had lower intratumor heterogeneity and higher abundance of immune cells (including B cell, CD4 T cell, CD8 T cell, macrophage, dendritic cell, and neutrophil). In addition, patients in C2 had a higher likelihood of immunotherapy response and overall survival advantage than patients in C1. Combined drug sensitivity analysis (CTRP/PRISM/CMap/GDSC) revealed that BI-2536 might serve as a new therapeutic compound for patients in C1. In order to realize the application value of our study, we constructed the classifier (to classify early-stage LUAD patients into C1 or C2 groups) with multiple machine learning and bioinformatic analyses. The 21-gene-based classification model showed high accuracy and strong generalization ability, and it was verified in four independent validation cohorts. In summary, our research provided a new strategy for clinicians to make a quick preliminary assisting diagnosis of early-stage LUAD and make patient classification at the intratumor heterogeneity level. All data, codes, and study processes have been deposited to Github and are available online.