A common and deadly cancer subtype. Lung adenocarcinoma (LUAD) is the most prevalent form of lung cancer, accounting for roughly 40% of all lung cancer diagnoses worldwide. Despite advances in treatment, the five-year survival rate for patients with advanced or metastatic LUAD remains below 20%, underscoring the critical need for better ways to predict outcomes and guide therapy.
Drug resistance and immunotherapy challenges. Even when patients initially respond to targeted therapies or immunotherapy, resistance frequently develops over time. Low overall response rates and increasing drug resistance rates highlight why identifying reliable biomarkers - measurable signals that predict how a patient's tumor will behave - is so urgently needed in clinical practice.
The promise of multi-omics integration. Modern cancer research generates vast amounts of molecular data: gene expression (RNA-seq), epigenetic marks (ATAC-seq), single-cell sequencing, and protein expression from tumor tissue. Integrating these data layers with machine learning offers a way to find patterns that no single data type reveals on its own.
The goal of this study. Researchers aimed to build a machine learning-based prognostic scoring system for LUAD using publicly available genomic data, identify key genes driving poor outcomes, and then experimentally validate one particularly promising target gene - PTGES3 - as a potential new therapeutic vulnerability.
Large-scale genomic datasets. RNA sequencing data from two major public sources - TCGA (The Cancer Genome Atlas) and GEO (Gene Expression Omnibus, dataset GSE42127) - were combined and corrected for batch effects to create a merged cohort. Single-cell RNA sequencing data from over 111,000 cells were also incorporated to examine gene behavior at the level of individual cells within tumors.
Ten machine learning algorithms tested. The team systematically evaluated ten different machine learning approaches - including LASSO regression, random forest, support vector machine, gradient boosting, and naive Bayes - and their combinations. Each model's performance was measured by the concordance index (C-index), a standard metric for how well a model predicts survival outcomes, averaged across three independent datasets.
Random Survival Forest wins. The Random Survival Forest (RSF) algorithm outperformed all alternatives. Starting from thousands of candidate genes, Cox regression analysis across both datasets narrowed the list to 28 key prognostic genes, all classified as risk genes (higher expression associated with worse outcomes). These 28 genes formed the basis of the final prognostic score, called the LS score.
SHAP analysis for interpretability. To understand which genes contributed most to the model's predictions, researchers applied SHAP (SHapley Additive exPlanations) - a game-theory-based method that quantifies each feature's contribution. The five most influential genes were KRT6A, PERP, SEC61G, PTTG1, and PTGES3, with KRT6A having the largest individual impact on predicted poor prognosis.
Strong predictive performance validated. In the training dataset (TCGA-LUAD), the model achieved an AUC exceeding 0.9 on ROC analysis - indicating excellent discriminative ability - and was confirmed in two independent validation cohorts. Patients with higher LS scores had significantly worse overall survival across all three datasets.
High-score patients show aggressive tumor biology. High LS scores correlated strongly with later clinical stage (T, N, M staging), greater tumor burden, and more frequent mutations in TP53, TTN, and MUC16. Enrichment analysis revealed that high-score tumors were enriched for pathways linked to angiogenesis, glycolysis, hypoxia, cell cycle dysregulation, and epithelial-to-mesenchymal transition - all hallmarks of aggressive cancer.
Stemness and immune evasion in high-risk tumors. High LS score patients showed elevated tumor stemness signatures (meaning their cancer cells retained more stem-like, treatment-resistant properties) and higher tumor mutational burden. Immune analysis revealed these tumors had more exhausted T cells and fewer CD4+ helper T cells and NK cells, suggesting an immunosuppressive environment that may explain poor responses to immunotherapy.
Drug sensitivity predictions. Using the oncoPredict algorithm, the researchers predicted that high-score patients might be particularly sensitive to docetaxel and lapatinib - insights that, if validated clinically, could help direct treatment choices for the highest-risk patients.
PTGES3 stands out among the key genes. PTGES3 (Prostaglandin E Synthase 3) encodes an enzyme involved in prostaglandin biosynthesis - molecules that regulate inflammation and immune responses. While lower PTGES3 expression initially appears somewhat protective, higher expression becomes strongly associated with poor prognosis as tumors advance through clinical stages.
PTGES3 is broadly expressed in lung cancer cells. Single-cell analysis across over 111,000 tumor cells confirmed that PTGES3 is expressed at high levels across multiple lung cancer cell types and in the surrounding tumor microenvironment - including both malignant cells and supportive stromal cells.
Complex immune interactions. PTGES3 showed positive correlations with exhausted and effector T cells, and negative correlations with CD4+ T cells and natural killer cells. This suggests PTGES3 may help tumors remodel the immune environment in a way that suppresses effective anti-tumor immunity. Single-cell analysis also revealed significantly altered cell communication patterns - particularly with macrophages and fibroblasts - in patients with high PTGES3 expression.
Enhanced signaling in high PTGES3 tumors. Cell-to-cell communication analysis using the CellChat method revealed stronger VEGFA-receptor signaling (promoting tumor blood vessel formation) and SPP1-CD44 interactions (promoting cell adhesion and metastasis) in patients with high PTGES3 expression, linking this gene to multiple mechanisms of aggressive tumor behavior.
Knocking down PTGES3 slows cancer cell growth. Using lentiviral delivery of short hairpin RNA (shRNA) sequences, the researchers silenced PTGES3 expression in two common lung cancer cell lines (H1299 and A549). Multiple assays confirmed that reducing PTGES3 significantly impaired cell proliferation: CCK-8 viability assays showed slower growth over five days, and colony formation experiments showed fewer and smaller colonies.
Cell cycle arrest after PTGES3 knockdown. Flow cytometry revealed that cells lacking PTGES3 accumulated in specific phases of the cell cycle and failed to progress normally through division. Western blot analysis confirmed reduced levels of key cell cycle proteins Cyclin D1 and CDK4 - regulators that normally drive cells forward through the division cycle.
Increased apoptosis (programmed cell death). Silencing PTGES3 also triggered apoptosis: the ratio of the pro-death protein Bax to the survival protein Bcl-2 increased, and flow cytometry showed a higher proportion of apoptotic cells in PTGES3-knockdown lines compared to controls.
Confirmed in living animals. When PTGES3-knockdown cancer cells were injected under the skin of immunodeficient mice (xenograft model), resulting tumors grew significantly more slowly than tumors from control cells. This in vivo confirmation strengthens the case that PTGES3 is a functionally important driver of lung adenocarcinoma growth.
Finding what controls PTGES3. Understanding which genes regulate PTGES3 expression is important for designing therapeutic strategies. Using SCENIC - a computational method that reconstructs gene regulatory networks from single-cell data - combined with ATAC-seq data (which identifies open chromatin regions where transcription factors bind), the team identified upstream regulators of PTGES3.
ZBTB7A emerges as a key regulator. Cross-referencing SCENIC results with ATAC-seq chromatin accessibility data pointed to ZBTB7A as a key transcriptional regulator. Multiple independent datasets consistently showed a negative correlation between ZBTB7A and PTGES3 - meaning when ZBTB7A is active, PTGES3 tends to be suppressed. A dual-luciferase reporter assay experimentally confirmed that ZBTB7A directly represses PTGES3 transcription.
ZBTB7A as a tumor suppressor in LUAD. ZBTB7A is a known transcriptional repressor that normally suppresses glycolysis and tumor growth. In lung adenocarcinoma, ZBTB7A expression is relatively low, and approximately 6% of LUAD cases carry ZBTB7A loss-of-function mutations. Loss of ZBTB7A repression allows oncogenic pathways - including those controlled by PTGES3 - to become overactive.
A regulatory network connecting multiple genes. Upstream mutations in CSMD3 and KEAP1 were found to influence PTGES3 expression levels and pathway activity, while TP53 mutations affected ZBTB7A expression. The interaction between LGALS9, P4HB, and CD44 signals may further amplify PTGES3-driven immune remodeling in macrophages, creating a complex network sustaining tumor growth and immune evasion.
KRT6A as the top prognostic contributor. The most influential gene in the LS model, KRT6A (Keratin 6A), is a structural protein associated with stem cell-like properties in cancer. High KRT6A expression correlates with lymph node metastasis, TNM staging, and smoking history in non-small cell lung cancer, and prior research suggests it promotes cancer aggressiveness through MYC and other signaling pathways.
PTGES3's dual role - initially protective, then harmful. Unlike purely oncogenic genes, PTGES3 shows a nuanced pattern: at low expression levels it may modestly protect prognosis, but as expression rises - which tends to occur as tumors progress to higher stages - it becomes associated with worse outcomes. This stage-dependent behavior mirrors the known biology of prostaglandins as context-dependent immune modulators.
Broader implications for PTGES3 as a target. PTGES3 has been implicated as a prognostic marker in breast cancer, ovarian cancer, glioblastoma, and liver cancer. Its role across multiple tumor types, combined with its experimental validation in LUAD, suggests it may represent a broadly relevant therapeutic target. Drugs targeting the prostaglandin synthesis pathway are already in clinical use for other conditions, providing a potential starting point for cancer-specific applications.
Limitations to address. The study relied on retrospective public genomic datasets, which may not fully represent the clinical diversity of LUAD. The LS score has not been validated in prospective randomized controlled trials. Protein-level validation of most of the 28 genes (beyond PTGES3) still needs to be performed. Large-scale, multicenter clinical trials will be needed before this scoring system could be applied to guide real-world treatment decisions.
A validated prognostic tool with strong performance. The LS score, built from 28 differentially expressed genes using Random Survival Forest and validated in multiple independent cohorts, achieves excellent predictive accuracy for one-, three-, and five-year survival in lung adenocarcinoma. Its ability to classify patients into high- and low-risk groups could help tailor treatment intensity and selection in clinical practice.
PTGES3 as a validated therapeutic target. Both computational analyses and laboratory experiments confirm that PTGES3 is functionally important for lung adenocarcinoma growth. Its knockdown impairs proliferation, induces cell cycle arrest, and triggers apoptosis in vitro and in vivo. The identification of ZBTB7A as its upstream regulator provides a mechanistic framework for understanding how PTGES3 becomes overactivated and how it might be therapeutically suppressed.
Implications for immunotherapy selection. The immune landscape analysis suggests that high LS score patients have more immunosuppressive tumor microenvironments - potentially explaining why some LUAD patients respond poorly to checkpoint immunotherapy. Understanding these immune signatures could help predict immunotherapy benefit or identify patients who need additional interventions to restore anti-tumor immunity.
The path to clinical translation. Future work should prioritize protein-level validation of the full gene panel, testing the LS score in prospective multicenter settings, and exploring PTGES3 inhibition strategies - potentially in combination with existing targeted therapies or immunotherapies - across different LUAD molecular subtypes to fully realize the clinical potential of these findings.