Abstract CT095: Mutation-based subtypes identified in the 1L NSCLC study BFAST using unsupervised machine learning clustering based on consensus non-negative matrix factorization

Abstract Background: Patients with the same disease stage and other clinical parameters often have markedly different prognoses and display variable responses to standard treatments, even within oncogenic subsets. This clinical variability may be driven by molecular heterogeneity and mutational complexity that may be captured by mutational signatures with the potential for impact on patient management. Methods: BFAST (NCT03178552) is a Phase II/III global, multicenter, open-label, multi-cohort screening and interventional umbrella trial designed to evaluate the safety and efficacy of targeted therapies in patients with unresectable, advanced or metastatic NSCLC selected based on the presence of oncogenic somatic mutations or a positive bTMB score. Patients were screened using blood-based NGS for enrollment into an appropriate treatment cohort based on their mutation or a non-interventional natural history cohort (Cohort Z). This exploratory analysis applied an unsupervised clustering using machine learning based on Consensus Non-Negative Matrix Factorization (cNMF) to identify genomic-based molecular subtypes in 1L NSCLC on liquid-based NGS samples from the BFAST trial using 25% of the available n=8583 as a discovery set. The clusters identified were evaluated in Cohort Z, which included patients treated with local SoC, for their association with OS. Results: Unsupervised cNMF on screening samples resulted in three clusters, cNMF1 n=1434(36%), cNMF2 n=879(22%), and cNMF3 n=1664(41%). These three clusters stratified Cohort Z patients (n=226) for survival and appeared to capture distinct biologies. Patients in cluster cNMF2 had the shortest median OS (8.1 months, 95% CI 6.05-10.9) and were enriched in KRAS mutations as well as mutations in CDKN2A, FGFR2, TERT, MYCN and PTEN. Patients in cluster cNMF1 had a longer median OS (11.4 months, 95% CI 9.40-21.9) and were enriched in NF1 as well as some of the less prevalent driver mutations such as MET, ALK, and BRAF, reflecting a relatively upstream node in the RAS/MAPK pathway. Patients in cluster cNMF3 had the longest median OS (17.9 months, 95% CI 13.1-NA) and were enriched in EGFR mutations as well as PIK3CA, FGFR1, and CTNNB1. These mutated genes could be found in all clusters but were generally concentrated in one. The OS HR of cNMF1 vs cNMF2 was 0.56 (95% CI 0.39-0.81), and OS HR of cNMF3 vs cNMF2 was 0.41 (95% CI 0.27-0.63). Importantly, although one cluster was enriched for a driver mutation, survival was impacted by the genomic contexture captured in the cNMF cluster. For example, patients with mutant KRAS in cNMF1 (n=16) had longer OS than those in cNMF2 (n=52): 9.4 months (95% CI 5.5-NA) versus 6.6 months (95% CI 4.8-9.5; p<0.05). EGFR mutant patients in cNMF3 (n=36) had longer OS than those in cNMF1 (n=21): 25.7 (95% CI 16.3-NA) versus 9.9 (95% CI 7.6-NA) months; p<0.05). Conclusions: This unsupervised ML approach based on mutational data from an NGS panel has defined three molecular subtypes that appear to capture broader biology beyond one single driver alteration. Additional research is needed to validate these observations in independent NGS data sets, and to explain their prognostic vs. predictive nature with respect to treatment received. Study sponsored by F. Hoffmann-La Roche Ltd/Genentech, Inc. Citation Format: Habib Hamidi, Zoe June Assaf, Solange Peters, Shirish M. Gadgeel, Clelia Cahuzac, Michael S. Mathisen, Tony Mok, Rafal Dziadziuszko, Sarah M. Shagan. Mutation-based subtypes identified in the 1L NSCLC study BFAST using unsupervised machine learning clustering based on consensus non-negative matrix factorization [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 2 (Late-Breaking, Clinical Trial, and Invited Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_2):Abstract nr CT095.