非小細胞肺癌における免疫療法奏功予測のための循環腫瘍DNAベースの機械学習モデルの開発と検証

Current biomarkers, such as programmed death-ligand 1 (PD-L1) and tumor mutational burden (TMB), have limited predictive value for immune checkpoint inhibitor (ICI) response in non-small cell lung cancer (NSCLC). Machine learning analysis of circulating tumor DNA (ctDNA) can enhance patient stratification via liquid biopsy genomic signatures. We aimed to build a support vector machine (SVM) model using baseline ctDNA to predict ICI benefit in advanced NSCLC.

We trained an SVM model on pretreatment ctDNA whole-exome sequencing (WES) data from the OAK trial cohort (n=303) to predict durable clinical benefit (DCB): response or stable disease (SD) ≥6 months. We used least absolute shrinkage and selection operator (LASSO) regression to select 41 predictive somatic mutations. Model performance was validated in the independent POPLAR trial cohort (n=97) and a regional multicenter cohort (n=41). Receiver operating characteristic (ROC) curve analysis and Kaplan-Meier methods were used to evaluate predictive accuracy and survival outcomes.

The ctDNA-based SVM model achieved high accuracy across cohorts: area under the ROC curve (AUC) =0.87 in OAK, 0.92 in POPLAR, and 0.83 in the local cohort. Patients classified as SVM-low (score >0.55) had significantly longer median progression-free survival (mPFS) [12.80 vs. 3.00 months; hazard ratio (HR) =0.191; P=0.002] and median overall survival (mOS) (17.30 vs. 6.40 months; HR =0.159; P=0.002) than SVM-high patients. In particular, SVM-low patients with initial SD exhibited better outcomes than SVM-high. This noninvasive model may reduce the need for repeat tissue biopsies and associated costs. Functional enrichment of the 41-gene signature highlighted VEGF, JAK-STAT, and Ras signalling pathways in ICI resistance.

Our ctDNA-based SVM model accurately predicts DCB and survival outcomes in NSCLC patients receiving ICIs. By using a single baseline liquid biopsy, this model could streamline immunotherapy decision-making without requiring longitudinal monitoring.