8041 Background: The molecular characterization of non-small cell lung cancer (NSCLC) and targeting therapies are improving outcomes for many patients. However, about 40% remain without identifiable molecular drivers. Understanding the co-mutational landscape of NSCLC may contribute to close that gap but has not yet been sufficiently characterized beyond e.g. TP53 co-mutations that associated with adverse outcome (Galina et al. J. Thor. Onc . 2024). Previously, the hierarchical Dirichlet mixture model (HDMM) provided important insight into acute leukemia biology (Papaemmanuil et al. NEJM 2016, Turki et al. EHA plenary abstract 2025), reason why we adopted this method for the first time to decompose the NSCLC genetics landscape. Methods: Here, we leveraged the HDMM with multinomial distribution for unsupervised learning of NSCLC genetics on the large AACR GENIE BPC NSCLC cohort (n = 1,846 patients, Choudhury et al. Clin. Can. Res . 2023) including 300 features for 60 genes. Preprocessing excluded columns with < 5 mutations, features duplicating information or with high rates of missingness, leaving the most relevant 85 genetic features for HDMM analysis including single mutations, gene amplifications and deletions. Within UICC stages I, II-III and IV, overall survival (OS) was analyzed with the Kaplan Meier method comparing the new components with the logrank test. Results: Among 1926 genetically distinct tumors of 1785 NSCLC patients with detectable genetic alterations, the HDMM identified four components assembling patients with similar landscape. Their driver patterns involved either co-mutated TP53/KRAS , TP53/EGFR or PDGFR/KDR and other mutations including established risk groups, i.e. with amplifications in ALK , BRAFV600E, cMET exon 14 or single EGFR or KRAS mutations, ROS1 or NTRK translocations. As expected, UICC stages I-IV stratified this cohort significantly for OS. Next, we tested whether genetic components refined outcomes within the same UICC stage. Indeed, for stage I NSCLC the genetic components associated with significant OS differences (p = 0.0065). Patients with PDGFR/KDR had the highest OS, while patients of the TP53/KRAS component had the lowest median OS with a Kaplan Meier estimate of 82.4 months as compared to 116.7 in the TP53/EGFR cluster. Also, among patients with stage II-III NSCLC, the three components associated with significant differences in OS (p = 0.023). Not taking into account other driver mutations, the components also associated with distinct OS in patients with stage IV NSCLC (p = 0.0001). Conclusions: This analysis of the co-mutational landscape is complementing existing genetic classes in NSCLC. We identified a new PDGFR / KDR component and support the relevance of previously described co-occurring TP53 / EGFR and TP53 / KRAS . The significantly distinct OS in resectable NSCLC may open new risk adapted treatment strategies.
Ugwu et al. (Thu,) studied this question.