NRF2–COX2–PGE2 axis drives immune cold tumors and predicts resistance to combination immunotherapy in hepatocellular carcinoma

BACKGROUND AND AIMS: Atezolizumab plus bevacizumab (Atez/Bev) is the first-line immunotherapy for advanced hepatocellular carcinoma (HCC), yet many patients show primary or acquired resistance. We aimed to identify tumor-intrinsic mechanisms driving immune cold tumor microenvironments (TMEs) and Atez/Bev resistance. APPROACH AND RESULTS: We used a genetically heterogeneous immunocompetent HCC mouse model generated by hydrodynamic injection of a barcoded oncogene library, integrating single-cell RNA sequencing, spatial transcriptomics, and bulk RNA sequencing of human HCC tissues. Retrospective analyses included a multi-institutional registry of 549 Atez/Bev-treated patients and 199 surgically resected HCC patients. External validation used RNA-seq data from 247 patients registered in the IMbrave150 and GO30140 trials. In mice, Atez/Bev eliminated hot tumors, while pre-existing cold tumors with exhausted effector T cells and Tregs predominated later. Barcode analysis revealed enrichment of NFE2L2 (NRF2) in resistant tumors. NRF2 overexpression suppressed immune infiltration and conferred resistance, which was reversed by NRF2 or COX2 inhibition. In human cohorts, both non-responders and patients with acquired resistance exhibited high NRF2/COX2 expression and immune exclusion, confirmed by spatial profiling. High NRF2 activity predicted shorter progression-free survival (PFS), whereas concomitant COX2 inhibitor use correlated with longer PFS. Plasma prostaglandin E2 (PGE 2 ) independently predicted poor response and survival. In IMbrave150/GO30140, NRF2 and prostaglandin pathway activation correlated with Atez/Bev resistance but not sorafenib response, validating a treatment-specific mechanism. CONCLUSIONS: Tumor-intrinsic activation of the NRF2-COX2-PGE 2 axis drives immune cold TMEs and mediates Atez/Bev resistance in HCC. Targeting this pathway may enhance efficacy, and plasma PGE 2 represents a non-invasive biomarker for stratification.