Machine Learning–Derived Intrinsic Gene Signatures Predict Outcomes to PD-1 Inhibitors in Lung Adenocarcinoma

To better understand why patients with non–small cell lung cancer (NSCLC) experience variable benefit from PD-1 inhibitors, we trained machine learning models on transcriptomic data from 57 treated individuals. When ranking the genes most strongly linked with therapeutic outcome, lung adenocarcinoma (LUAD) displayed a pronounced enrichment of tumor-intrinsic genes (69%), in contrast to both the broader NSCLC cohort (36%) and lung squamous cell carcinoma (LUSC) (33%). A signature constructed from LUAD-specific intrinsic transcripts provided the most reliable classification of treatment response, yielding a mean ROC AUC of 0.957 and an accuracy of 0.9—substantially higher than signatures derived from extrinsic programs or from intrinsic sets in NSCLC or LUSC. LUAD patients with elevated intrinsic-signature activity showed significantly longer overall survival (p = 0.034). Functional annotation of the LUAD intrinsic genes highlighted enrichment of pathways related to cell-cycle control and senescence. This signature also demonstrated positive associations with several immune checkpoint molecules, including CD274, LAG3, and PDCD1LG2 (Spearman’s ρ > 0.25). The marked divergence in intrinsic transcriptional patterns between LUAD and LUSC suggests that intrinsic gene programs may serve as a particularly informative biomarker for predicting PD-1 inhibitor benefit in LUAD.