Purpose

Treatment of advanced non–small-cell lung cancer with immune checkpoint inhibitors (ICIs) is characterized by durable responses and improved survival in a subset of patients. Clinically available tools to optimize use of ICIs and understand the molecular determinants of response are needed. Targeted next-generation sequencing (NGS) is increasingly routine, but its role in identifying predictors of response to ICIs is not known.

Methods

Detailed clinical annotation and response data were collected for patients with advanced non–small-cell lung cancer treated with anti–programmed death-1 or anti–programmed death-ligand 1 [anti-programmed cell death (PD)-1] therapy and profiled by targeted NGS (MSK-IMPACT; n= 240). Efficacy was assessed by Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1, and durable clinical benefit (DCB) was defined as partial response/stable disease that lasted> 6 months. Tumor mutation burden (TMB), fraction of copy number–altered genome, and gene alterations were compared among patients with DCB and no durable benefit (NDB). Whole-exome sequencing (WES) was performed for 49 patients to compare quantification of TMB by targeted NGS versus WES.

Results

Estimates of TMB by targeted NGS correlated well with WES (ρ= 0.86; P<. 001). TMB was greater in patients with DCB than with NDB (P=. 006). DCB was more common, and progression-free survival was longer in patients at increasing thresholds above versus below the 50th percentile of TMB (38.6% v 25.1%; P<. 001; hazard ratio, 1.38; P=. 024). The fraction of copy number–altered genome was highest in those with NDB. Variants in EGFR and STK11 associated with a lack of benefit. TMB and PD-L1 expression were independent variables, and a composite of TMB plus PD-L1 further enriched for benefit to ICIs.

Conclusion

Targeted NGS accurately estimates TMB and elevated TMB further improved likelihood of benefit to ICIs. TMB did not correlate with PD-L1 expression; both variables had similar predictive capacity. The incorporation of both TMB and PD-L1 expression into multivariable predictive models should result in greater predictive power.