Development of Predictors for PD-1/PD-L1-directed Therapy of Non-Small Cell Lung Cancer (NSCLC) by Gene Expression Profiling of Small Diagnostic Biopsies (DBX).

e15121 Background: Immune checkpoint inhibitors (ICI) targeting PD-1/PD-L1 are standard second-line treatments of stage IV NSCLC and preferred first-line therapy in selected patients (pts). However, only a minority of pts respond. These can be only partially enriched by predictive biomarkers applicable to small DBX available in most pts with stage IV NSCLC such as PD-L1 expression and tumor mutational burden. Methods: Expression analysis of 770 immune-related genes in small surplus DBX using the NanoString nCounter platform was established following serial dilution experiments and adaptation. Clinical endpoints were best response (PD vs. SD/PR; PD vs. SD vs. PR), time-to-treatment-failure (TTF), and overall survival (OS) following ICI therapy. Predictive feature sets were selected by ensemble-based penalized linear or logistic regression. Prediction models for all endpoints were trained using support vector machine classification or regression. Scores of prediction quality were derived using 20-fold cross validation. The study was approved by the responsible IRB. Results: 127 pts with advanced NSCLC were enrolled at a single center. Extraction of qualitatively and quantitatively sufficient RNA was possible in 84 pts, of which 55 (33 adeno-, 21 squamous- and 1 adenosquamous carcinoma) generated admissible nCounter reads. Mean age at diagnosis was 62.6 years (37.1-84.7) with 22 females and 33 males. PD-L1 tumor proportion score was ≥1% in 57.1% and ≥50% in 16.3%. Median follow-up after start of ICI therapy was 11.9 months, and median OS since ICI therapy was 10.8 months. The best achieved cross-validated accuracy was 80% (based on the expression of 3 genes) for PD vs. SD/PR and 78% (5 genes) for PD vs. SD vs. PR. Best score (integrated area under the time-specific ROC curve) achieved for OS was 0.67 (5 genes) and 0.79 for TTF (3 genes). Validation in an independent cohort is ongoing. Conclusions: Comprehensive gene expression profiling is feasible in a proportion of routine DBX from NSCLC pts. Validation of more restricted gene panels is expected to expand the assessable pt population.