The clinical problem. Lung adenocarcinoma (LUAD) is the most common subtype of non-small cell lung cancer, accounting for roughly 40% of all lung cancer cases. Despite advances in targeted therapies and immune checkpoint inhibitors, overall survival remains poor due to wide variation in how tumors respond to immunotherapy.
Root cause: immune heterogeneity. The tumor immune microenvironment differs dramatically between patients, and existing classification systems rely on limited immune signatures or single-platform data without comprehensive validation. The molecular drivers behind these differences are poorly understood.
Enter TPX2. Targeting Protein for Xklp2 (TPX2), a regulator of cell division, has been implicated not only in tumor proliferation but also in shaping the immune environment through cell-cycle signaling. Its precise role in LUAD immune heterogeneity had not been systematically studied.
Study objectives. Researchers combined bulk RNA sequencing, single-cell transcriptomics, multi-omics profiling, and machine learning to map the immune landscape of LUAD, identify immune subtypes, build a predictive scoring model, and characterize TPX2 as a key molecular driver.
Large-scale multi-cohort data. Bulk RNA sequencing data came from TCGA-LUAD along with seven independent GEO cohorts (GSE13213, GSE50081, GSE30219, GSE42127, and others). All samples were processed on a TPM scale, log-transformed, and batch-corrected using the ComBat algorithm to eliminate cross-study technical variation.
Single-cell atlas. Rather than reprocessing raw files, the team leveraged a pre-curated integrated NSCLC single-cell RNA sequencing resource aggregating six GEO cohorts covering thousands of cells with standardized annotations. This enabled direct downstream analysis without quality-control duplication.
Immune subtype assignment. Tumor immune subtypes were assigned using the ImmuneSubtypeClassifier R package, which applies an ensemble learning framework to classify samples into pan-cancer immune classes. Only samples with a classification confidence score above 0.6 were retained to ensure reliable subtype calls.
Immune infiltration profiling. The ESTIMATE algorithm calculated stromal and immune scores. Gene Set Variation Analysis with 24 immune and stromal cell-type signatures quantified cell-type composition. T cell functional states including exhaustion, cytotoxicity, and activation were scored using the TCellSI framework.
Feature selection pipeline. Differentially expressed genes between the IFN-gamma Dominant and Inflammatory subtypes were filtered by fold change and adjusted p-value, then screened by univariate Cox regression for survival association. This produced a candidate gene list for model training.
Mime1 framework. The open-source Mime1 machine learning framework evaluated ten algorithms and their combinations, including stepwise Cox regression and elastic-net regularization. The TCGA-LUAD cohort served as the training set while five independent GEO cohorts provided external validation.
Final model selection. The StepCox forward plus elastic-net model with alpha equal to 0.1 produced the most stable and reproducible performance based on cross-cohort Harrell's C-index and time-dependent ROC/AUC at one, three, and five years. The resulting composite score was named the Differential Phenotype Immune Score (DPIS).
Single-cell projection. DPIS was not retrained at the single-cell level. Instead, the 10-gene DPIS panel was used to compute a module activity score across individual cells, enabling spatial mapping of the DPIS program without applying bulk Cox coefficients to single cells.
Three predominant classes. Among 445 robustly classified TCGA-LUAD samples, three immune subtypes covered 98.2% of cases: IFN-gamma Dominant (58.9%), Inflammatory (24.3%), and Wound Healing (16.7%). Lymphocyte Depleted, Immunologically Quiet, and TGF-beta Dominant subtypes were too rare for meaningful comparison.
Clinical differences. The IFN-gamma Dominant subtype was enriched in advanced-stage tumors with higher proportions of T3-T4 lesions and N2-N3 nodal involvement, reflecting more aggressive disease. The Inflammatory subtype was overrepresented in early-stage disease. Survival differed significantly across groups (p = 0.038), with Inflammatory patients faring best.
Tumor microenvironment profiles. The Inflammatory subtype had the highest immune and stromal scores, consistent with enhanced immune cell infiltration. The Wound Healing subtype showed the highest tumor purity and lowest immune infiltration, consistent with an immune-excluded, suppressive microenvironment that explains its poor prognosis.
DPIS as independent predictor. After adjusting for sex, stage, and detailed T/N/M categories in multivariable Cox analysis, DPIS remained an independent risk factor for overall survival with a hazard ratio of 3.65 (95% CI 1.83-7.25, p = 0.000226), confirming it adds prognostic value beyond standard clinical variables.
TIDE prediction results. Within TCGA-LUAD, TIDE-estimated immunotherapy response rates differed markedly by subtype: Inflammatory tumors had the highest predicted responder fraction (approximately 59%), followed by IFN-gamma Dominant (approximately 36%), while Wound Healing tumors had only about 8% predicted responders.
Paradox of the IFN-gamma subtype. Although IFN-gamma Dominant and Inflammatory subtypes both showed sensitivity to immune checkpoint blockade in SubMap analysis, their baseline survival differed substantially. This paradox motivated development of DPIS to capture the molecular differences between these two apparently immunotherapy-sensitive groups.
External cohort validation. Across four independent immunotherapy cohorts (GSE207422, GSE126044, GSE135222, and IMvigor210), Inflammatory and IFN-gamma Dominant subtypes consistently showed better survival and higher response rates. In IMvigor210, survival curves for immune-subtype-mapped patients showed pronounced separation (p = 0.004).
Wound Healing subtype resistance. The Wound Healing subtype remained poorly responsive across all cohorts, consistent with its immune-excluded microenvironment enriched for TGF-beta, WNT, and extracellular matrix remodeling pathways that create stromal barriers blocking immune cell access to tumors.
Robust prognostic stratification. DPIS stratified patients into high-risk and low-risk groups with significantly different overall survival across all six cohorts (p = 0.003 in TCGA; p below 0.001 in four GEO cohorts). Time-dependent ROC analyses showed AUC values ranging from 0.58 to 0.86, with several cohorts achieving one-year AUC values above 0.8.
Ten-gene signature. The final DPIS model incorporated ten genes: TPX2, UBE2C, CDC20, BIRC5, SCGB3A1, SFTPB, CACNA2D2, CYP4B1, MYBL2, and SUSD2. These represent a combination of cell proliferation-associated genes and lung lineage transcriptional markers, together capturing the immune-cold, high-proliferation state.
Single-cell localization. When DPIS was projected onto the NSCLC single-cell atlas, DPIS-positive cells were highly concentrated in the malignant compartment (approximately 44% of DPIS-positive cells versus 15% in DPIS-negative cells). Among malignant cells, Proliferating Cancer cells exhibited the highest DPIS scores, with strong spatial overlap between DPIS-high and proliferating cell populations.
Distinct from simple proliferation. DPIS-high malignant cells did not colocalize with Ki-67 activity on UMAP embeddings, indicating that DPIS captures a distinct malignant state beyond a generic cell-cycle proliferation marker. Regulatory network analysis identified top transcription factors governing this state, including ZNF443, ZNF429, ZNF92, NFE2, and E2F7.
Pan-cancer prognostic role. Across 32 TCGA tumor types, high TPX2 expression was broadly associated with adverse outcomes. In LUAD and multiple other cancers, higher TPX2 levels correlated with reduced overall survival (HR above 1, p below 0.05), supporting a conserved oncogenic function across cancer types.
Functional correlation profile. In LUAD, TPX2 expression tightly aligned with cell-cycle hallmarks including E2F targets, G2M checkpoint, mitotic spindle activity, and MYC/mTORC1 signaling. Conversely, TPX2 negatively correlated with most immune cell infiltration signatures including CD8+ T cells, NK cells, and dendritic cells, defining an immune-cold, proliferation-dominated tumor state.
Cell line experiments. Western blotting confirmed markedly elevated TPX2 protein in four lung cancer cell lines (A549, H1650, H1299, 95D) compared to minimal expression in non-malignant bronchial epithelial cells. Immunohistochemistry data from the Human Protein Atlas showed higher TPX2 protein in lung tumor tissues versus normal lung, with nuclear and cytoplasmic localization in malignant cells.
Functional assays after knockdown. siRNA-mediated TPX2 silencing in H1299 and 95D cells significantly suppressed proliferation, reduced migratory capacity in wound healing assays, and dramatically increased apoptosis from roughly 9% to over 40% of cells. In T-cell co-culture experiments, TPX2 depletion enhanced T-cell-mediated killing and upregulated CXCL10 and CXCL11 chemokines that recruit immune effector cells.
A proliferation-driven immune suppression model. This study proposes that immune heterogeneity in LUAD is partially explained by spatially constrained proliferative programs where TPX2-overexpressing tumor cells locally sculpt immunosuppressive niches, impair antigen presentation, and reduce chemokine signaling that would otherwise recruit immune effectors.
Practical use of DPIS. High-DPIS tumors show TPX2 overexpression, limited immune infiltration, and poor clinical outcomes, while low-DPIS tumors display immune-activated profiles consistent with higher checkpoint inhibitor sensitivity. DPIS therefore provides a clinically applicable framework for stratifying patients before immunotherapy decisions.
Therapeutic targeting strategy. Pharmacologic inhibition of the TPX2-AURKA-E2F axis could potentially reprogram immune-cold, proliferative tumor states and enhance the efficacy of checkpoint blockade. This rationale supports combining cell-cycle-targeted agents with immunotherapy in precision oncology trials.
Study limitations and future directions. The DPIS model was built from retrospective public datasets and requires prospective validation in immunotherapy trial cohorts. The mechanistic link between TPX2 and antigen presentation machinery warrants further study, and clinical-grade assays for DPIS scoring would need development before bedside application.