The research, led by Xiujuan Hu, Guoxin Zhang, and Yue Zhang with senior authors Fang Zhou and Guanhua Li at Tianjin Chest Hospital and Tianjin Medical University Cancer Institute and Hospital, began with a deceptively simple question: what happens when the two great engines of cancer—chronic inflammation and metabolic dysregulation—operate together? Scientists have long known that persistent inflammatory signaling can recruit immunosuppressive cell populations and sustain a pro-tumor cytokine network, while metabolic reprogramming in both cancer cells and immune cells starves cytotoxic lymphocytes of the nutrients they need to function. But the combined prognostic and mechanistic relevance of these intertwined processes in lung adenocarcinoma had remained poorly defined.

To interrogate this intersection systematically, the team started with a curated set of 117 genes from the Molecular Signatures Database representing systemic inflammation-metabolism dysregulation in cancer. By combining differential expression analysis between tumors and normal lung tissue with univariate Cox regression for overall survival, they narrowed the field to 16 core genes, including ADM, PFKP, MT2A, PLAUR, CDA, PSMA3, TNFRSF12A, UBE2L6, LGALS3, ATOX1, ATP2A2, HSPA5, CP, IL13RA1, CFB, and F10. Genomic characterization revealed that 71.28 percent of the 289 TCGA-LUAD tumors analyzed carried at least one somatic alteration in this gene set, with PLG, MET, and HNF4A showing the highest mutation frequencies at 10, 7, and 6 percent respectively.

What sets this study apart methodologically is the sheer scale of its machine learning search. Rather than relying on a single algorithm, the researchers systematically evaluated 101 survival-learning algorithm combinations, spanning least absolute shrinkage and selection operator regression, ridge regression, elastic net, stepwise Cox regression, CoxBoost, random survival forest, supervised principal components, survival support-vector machines, gradient boosting machines, and partial least-squares Cox regression. Using TCGA-LUAD as the training cohort and six independent Gene Expression Omnibus cohorts totaling 714 additional samples for external validation, the random survival forest combined with supervised principal components, or RSF + SuperPC, emerged as the top performer with a mean concordance index of 0.648 across validation cohorts.

The resulting risk score separated patients into high- and low-risk groups with significantly divergent survival across all seven cohorts. In the TCGA training set, the hazard ratio was 2.10 with a 95 percent confidence interval of 1.58 to 2.81 and a log-rank P value below 0.0001. External cohort hazard ratios ranged from 1.93 in GSE13213 to 26.82 in GSE31210, though precision varied in smaller datasets. Time-dependent area under the curve values reached 0.814 for one-year survival in GSE13213 and 0.789 for five-year survival in GSE31210, demonstrating that the model’s discriminative power, while variable across platforms and time horizons, consistently replicated in the direction of its survival stratification.

Beyond prognosis, the risk score proved to be a window into the tumor’s immunological character. High-risk tumors showed reduced stromal and immune infiltration and elevated tumor purity, as measured by the ESTIMATE algorithm, with Spearman correlations of −0.17 for StromalScore, −0.14 for ImmuneScore, and 0.14 for inferred tumor purity, all statistically significant after multiple-testing correction. Immunophenoscores from The Cancer Immunome Atlas differed between risk groups in three of four CTLA4 and PD1 checkpoint categories, and immune deconvolution using seven independent algorithms through TIMER2.0 revealed broad differences in lymphoid and myeloid cell populations between high- and low-risk tumors.

Among the 16 model genes, ATP2A2 showed the strongest positive correlation with the risk score at r = 0.72, followed closely by PLAUR and TNFRSF12A at r = 0.70 each. ATP2A2 encodes SERCA2, a sarco/endoplasmic reticulum calcium ATPase that sits at the heart of intracellular calcium signaling networks governing metabolic activity, inflammasome activation, and immune effector function. The gene was consistently overexpressed in tumor tissue relative to normal or adjacent lung in TCGA-LUAD and four of five independent GEO datasets, with immunohistochemistry on patient samples confirming higher protein staining in tumors. Expression climbed with histological grade, pathological stage, and primary tumor category, and high ATP2A2 predicted shorter overall survival in GSE26939 and GSE31210 with hazard ratios of 1.67 and 2.04 respectively.

Functional experiments added weight to the clinical associations. When the researchers overexpressed ATP2A2 in A549 lung adenocarcinoma cells and performed Transwell migration and invasion assays, they observed more crystal-violet-positive cells penetrating both uncoated and basement-membrane-coated membranes compared with empty-vector controls, consistent with increased cellular motility. This in vitro finding provides a plausible mechanistic link to the clinical observation that ATP2A2-high tumors tend to be more advanced and less differentiated.

Perhaps the most striking results emerged from the immunotherapy cohorts. In the OAK trial cohort of 257 patients treated with atezolizumab, high ATP2A2 expression was associated with dramatically shorter progression-free survival, with a hazard ratio of 2.15 and a log-rank P value of 0.00015, and shorter overall survival with a hazard ratio of 2.38 and P below 0.0001. In the smaller POPLAR cohort of 59 patients, overall survival was also shorter in the high-expression group with a hazard ratio of 2.08. Across both OAK and TCGA-LUAD, ATP2A2 expression inversely correlated with stromal, immune, and composite ESTIMATE scores and positively with inferred tumor purity, while seven deconvolution algorithms consistently showed lower abundance of B-cell, T-cell, and dendritic-cell subsets in ATP2A2-high tumors.

Deeper mechanistic probing through the IOBR framework revealed that ATP2A2 was negatively correlated with immune-recruiting chemokines and receptors including CXCL13, CCL5, and CCL19, as well as multiple immunostimulatory molecules, while a smaller set of inflammatory mediators such as CXCL1 and CXCL8 showed positive correlations. Analysis of the seven-step cancer-immunity cycle, quantified by single-sample gene-set enrichment, showed that ATP2A2 was negatively associated with antigen presentation, T-cell priming and activation, recruitment of CD8 T cells, NK cells, and dendritic cells, recognition of malignant cells, and tumor-cell killing, with several associations replicating in both OAK and TCGA-LUAD at adjusted P values below 0.001. Single-cell RNA sequencing of the HRA001130 dataset, which identified 13 distinct cell populations, showed that ATP2A2 expression was highest in ciliated and alveolar type II epithelial cells and low in immune populations, with ATP2A2-high samples showing a lower proportion of T cells and a higher proportion of alveolar epithelial cells. Multiplex immunofluorescence on human lung adenocarcinoma sections provided the final spatial confirmation, revealing fewer CD8-positive cytotoxic T cells within ATP2A2-high, pan-cytokeratin-positive tumor regions compared with ATP2A2-low areas.

The authors are careful to acknowledge the limits of their work. The causal chain linking ATP2A2, calcium homeostasis, and immune microenvironment remodeling is currently supported mainly by expression correlations and mechanistic inference from prior literature, and direct evidence from in vivo gain-of-function or loss-of-function experiments remains lacking. The retrospective nature of public database cohorts also introduces inherent biases and sample heterogeneity that cannot be fully corrected through algorithmic optimization. Nevertheless, by anchoring a machine learning-derived signature to concrete biological mechanisms validated across bulk transcriptomics, immune deconvolution, single-cell resolution, and spatial protein analysis, the study positions ATP2A2 as both a promising prognostic biomarker and a potential therapeutic target for overcoming immunotherapy resistance in lung adenocarcinoma. If future functional studies confirm that modulating SERCA2 activity can reverse the immune-desert phenotype, the calcium pump that keeps muscle cells contracting may find an unexpected second act as a lever for unlocking the immune system’s attack on lung cancer.

Subject of Research: An inflammation-metabolism gene signature and the role of ATP2A2 in immune evasion and immunotherapy resistance in lung adenocarcinoma

Article Title: An AI-derived inflammation-metabolism ecosystem signature reveals ATP2A2 as a microenvironment-remodeling driver of immune evasion in lung adenocarcinoma

Article References: Hu, X., Zhang, G., Zhang, Y., Zhao, Y., Zhou, F., & Li, G. (2026). An AI-derived inflammation-metabolism ecosystem signature reveals ATP2A2 as a microenvironment-remodeling driver of immune evasion in lung adenocarcinoma. Frontiers in Cell and Developmental Biology, 14, Article 1971779. https://doi.org/10.3389/fcell.2026.1971779

Image Credits: AI Generated

DOI: 10.3389/fcell.2026.1971779

Keywords: lung adenocarcinoma, ATP2A2, SERCA2, tumor microenvironment, immune evasion, immunotherapy resistance, machine learning, inflammation, metabolism, calcium signaling, prognostic signature, single-cell RNA sequencing