Abstract
Immune checkpoint blockade (ICB) has transformed non-small cell lung cancer (NSCLC) treatment, but durable clinical responses remain limited, underscoring the need for robust predictive biomarkers. We integrated multiomics profiling with machine learning to systematically identify determinants of ICB efficacy. Comparative evaluation of 22 survival algorithms across four NSCLC cohorts (n=156) led to the development of an Accelerated Oblique Random Survival Forest model, which outperformed conventional Cox regression and deep learning methods in predictive accuracy (training C-index=0.864; test C-index=0.748). Single-cell RNA sequencing of an immunotherapy-treated cohort revealed that high-risk tumors harbor malignant epithelial subclusters expressing growth differentiation factor 15 (GDF15), a transforming growth factor-β superfamily member implicated in immune evasion. Single-cell non-negative matrix factorization identified GDF15 as a network hub regulating proliferative dominance. External validation using melanoma cohorts (GSE91061) confirmed the pan-cancer predictive relevance of GDF15 and its associated tumor cluster. Functional studies utilizing GDF15-knockdown Lewis lung carcinoma cells showed no significant effect on intrinsic tumor proliferation or growth under immune stress (both p>0.05). GDF15 deletion significantly potentiated PD-1 inhibitor efficacy in vivo, reducing tumor mass by 94.41±6.53 % (SH1) and 94.54±5.21 % (SH2) compared with 3.39±54.90 % in empty vector controls (p<0.01 for all comparisons). CD8(+) T cell infiltration was also substantially enhanced (81.62±4.79 % [SH1] and 123.50±10.02 % [SH2] vs. 29.63±22.17 % [EV], p<0.05). These findings implicate GDF15 as a regulator of the immunosuppressive tumor microenvironment. Our findings position GDF15 as a first-in-class biomarker for predicting ICB resistance; they establish a translational framework that bridges computational prediction with single-cell mechanistic insights to inform NSCLC immunotherapy.