Evaluation of two algorithms measuring homologous recombination deficiency status in prognostic assessment for treatment-naïve non-small cell lung cancer.

OBJECTIVE

Patients with homologous recombination deficiency (HRD) demonstrate distinct clinicopathological and prognostic features. However, standardised and clinically validated HRD detection methodologies specifically tailored for non-small cell lung cancer (NSCLC) have yet to be established. Further research is needed to clarify the precise role and clinical implications of HRD in NSCLC.

METHODS

A cohort of 580 treatment-naïve NSCLC patients was retrospectively enrolled. Comprehensive genomic profiling (CGP) was performed for all patients, and HRD status was evaluated using two genomic scar score (GSS)-based algorithms: a machine learning-based GSS (ML-GSS) and a continuous linear regression-based GSS (CLR-GSS). To assess the diagnostic performance (sensitivity and specificity) of the ML-GSS and CLR-GSS algorithms for HRD detection, immunohistochemical (IHC) staining was conducted for two HRD-related biomarkers: Schlafen 11 (SLFN11) and RAD51. Survival analysis, including progression-free survival (PFS), along with multivariable Cox proportional hazards models, was performed to compare the prognostic value of the two HRD algorithms.

RESULTS

Among all patients, 146 (25.2%) and 46 (7.9%) were classified as HRD-positive (HRD+) by ML-GSS and CLR-GSS, respectively. Using SLFN11 IHC expression as the reference standard, comparative analysis demonstrated that ML-GSS exhibited significantly higher sensitivity but lower specificity than CLR-GSS. This trend was consistently observed in RAD51 staining analysis. Compared to HRD-negative (HRD-) patients, ML-GSS-defined HRD+ cases displayed distinct clinicopathological and genomic features, including a higher prevalence of homologous recombination (HR)-related genes mutations, mutations, mutations, elevated tumor mutation burden (TMB), and increased copy number variations (CNVs). In contrast, CLR-GSS-defined HRD+ patients were only enriched for mutations, mutations, and elevated TMB. Furthermore, ML-GSS-defined HRD+ status was associated with significantly worse prognosis following first-line therapy compared to HRD- patients. Univariate and multivariable Cox analyses identified ML-GSS-defined HRD+ and mutations as significant predictors and independent risk factors, respectively. No such associations were observed in the CLR-GSS-defined HRD+ cohort.

CONCLUSIONS

ML-GSS demonstrated superior performance to CLR-GSS in assessing chromosomal instability (CIN) and showed greater clinical utility. We recommend the ML-GSS algorithm as a robust and clinically validated tool for HRD/CIN evaluation in NSCLC. Furthermore, ML-GSS-defined HRD+ status was identified as both a significant predictor and an independent risk factor.