Metabolic Landscape Decoding: A Machine Learning Approach to Predicting Immunotherapy Response Across Cancer Types

Background Immune checkpoint blockade (ICB) has revolutionized cancer therapy, yet its efficacy remains limited due to heterogeneous patient responses across cancer types. Reliable predictive biomarkers are urgently needed to optimize patient stratification. Methods We integrated transcriptomic data from 10 independent ICB‐treated cohorts ( n = 772) across six cancer types and developed a metabolism‐based signature (MBS) using machine learning algorithms. The MBS was constructed from metabolism‐related genes (MRGs) and validated in the external cohort. We further explored the immune landscape associated with MBS using single‐cell RNA sequencing (scRNA‐seq), bulk RNA‐seq, and multiplex immunohistochemistry (mIHC). Functional validation of the top MBS gene, CA12, was performed in vitro and in vivo using lung adenocarcinoma models. Results The MBS demonstrated robust predictive performance for ICB response, with AUCs of 0.77, 0.71, and 0.75 in the training, validation, and testing cohorts, respectively. Low MBS scores were associated with enhanced immune infiltration, including increased CD8 + T cells and favorable survival outcomes. CA12 emerged as a key contributor to the MBS and was significantly upregulated in tumors, correlating with immune exclusion and poor prognosis. CA12 silencing enhanced CD8 + T cell infiltration and synergized with anti‐PD‐1 therapy in a murine lung metastasis model, leading to improved survival and reduced tumor burden. Conclusions The MBS is a clinically relevant, metabolism‐based biomarker that predicts immunotherapy response across cancer types. Targeting CA12 may reverse immune exclusion and enhance ICB efficacy, offering a promising strategy for metabolic immune modulation in cancer therapy.