Genomic determinants of overall survival in early-stage non–small cell lung cancer: a large-scale clinical sequencing cohort analysis with independent cohort comparison

Despite curative-intent surgery, approximately 35% of patients with early-stage non–small cell lung cancer (NSCLC) die within five years. Identifying genomic determinants of survival may refine risk stratification and inform adjuvant therapy decisions. We analyzed 4,097 patients with primary NSCLC from the MSK-IMPACT 50 K clinical sequencing cohort (discovery) and 1,176 patients from the ALCHEMIST adjuvant chemotherapy trial cohort (independent comparison). Tumor mutation burden (TMB), whole genome doubling (WGD), individual gene mutations, and mutational signatures were evaluated for association with overall survival (OS) using multivariable Cox regression adjusted for age, sex, and TMB. Co-mutation patterns were assessed in 3,064 patients sequenced on uniform gene panels. Higher TMB (adjusted hazard ratio [HR] = 1.34 per log10 unit; 95% CI, 1.22 to 1.49; P = 8.2 × 10⁻⁹) and WGD (adjusted HR = 1.35; 95% CI, 1.21 to 1.51; P = 8.9 × 10⁻⁸) independently predicted worse OS, with no significant interaction ( P = 0.25). Gene-level analysis identified SMARCA4 (HR = 1.45; P = 4.7 × 10⁻⁵), TP53 (HR = 1.42; P = 1.7 × 10⁻⁹), KEAP1 (HR = 1.41; P = 2.6 × 10⁻⁶), and STK11 (HR = 1.25; P = 0.002) as adverse prognostic factors, while EGFR mutations were protective (HR = 0.72; P = 9.4 × 10⁻⁷). In the ALCHEMIST comparison cohort, STK11 mutations were substantially underrepresented compared with the real-world MSK-IMPACT cohort (6.9% v 13.7%; enrichment ratio 0.51). In this setting, TMB, WGD, and mutations in TP53, KEAP1, SMARCA4, and STK11 were associated with worse overall survival (and EGFR mutation with better survival), although the observational design and the absence of stage and treatment data preclude causal or treatment-predictive claims. The depletion of adverse-prognosis mutations in the adjuvant chemotherapy trial population warrants attention in biomarker-driven trial design.