Enhancing the Prognostic Precision of Unsupervised Lung Cancer Subtyping Using a Network-Assisted Data Mining of Early Mutated Driver Genes

Lung cancer is often asymptomatic and typically diagnosed at an advanced stage, leading to limited progress in improving survival rates. While it is evident that a better prognosis often results from early detection, such molecular markers holding different prognostic values are not fully characterized. Here, we proposed a set of 47 early mutated driver genes (EMDGs), which are somatically mutated cancer driver genes likely to have arisen early, to serve as early prognostic biomarkers for identifying a clinical subtype of lung adenocarcinoma (LUAD). By leveraging an existing framework of network-based patient stratification, clustering performance based on 47 EMDGs outperforms those relying on benchmark cancer gene panels, yielding three prognosis-related LUAD subtypes. Approximately mutated at the same time window (early event), EMDGs likely interact epistatically, in turn, accounting for diverse tumor growth patterns. We thus performed an association rule mining coupled with network analysis to capture the co-occurring (CO) pattern of a significant gene pair likely co-mutated, resulting in an early epistatic gene set (EEGS = [KRAS, RYR2, TP53, KEAP1, STK11, EGFR]). Distinct epistatic patterns among different subgroups were correlated with survival outcomes. For example, patients carrying STK11/KEAP1 co-mutation were enriched in subgroup 3 (worst survival). For practical purposes, we developed a supervised model trained based on unsupervised mutational profiles (transformed heat diffusion scores). Our model exhibited a high predictive capability of survival prognosis (average accuracy 90.55%±3.21), thus eliminating the need for reclustering whenever new patients of unknown subgroups emerge. We anticipate that a future screening assay such as cell-free tumor DNA will benefit from our discovery towards early lung cancer detection, consequently improving overall survival.