A critical diagnostic gap. Lung adenocarcinoma (LUAD) is the most common subtype of non-small cell lung cancer and carries a poor prognosis, particularly at advanced stages where the 5-year survival rate remains very low. A central challenge is the lack of specific molecular biomarkers to support early diagnosis and targeted therapy.
The role of long non-coding RNAs. Long non-coding RNAs (lncRNAs) are RNA molecules longer than 200 nucleotides that do not encode proteins but regulate gene expression at transcriptional and post-transcriptional levels. Dysregulation of lncRNAs has been increasingly recognized as a driver of cancer development, making them promising candidates as diagnostic or prognostic biomarkers.
Prior lncRNA findings in LUAD. Several lncRNAs have already been linked to LUAD biology. LINC00472 suppresses cell proliferation and promotes apoptosis; FENDRR and LINC00312 show diagnostic value; and lOC100132354 promotes tumor angiogenesis through VEGFA/VEGFR2 signaling. However, systematic large-scale screening of lncRNA biomarkers for LUAD remained limited.
A machine learning opportunity. Machine learning methods offer a rigorous and unbiased way to identify the subset of biomarkers from thousands of candidates that have the greatest diagnostic power. Combined with co-expression network analysis, they can also reveal the biological pathways through which lncRNAs influence LUAD development.
Large-scale genomic data from TCGA. The study used lncRNA and mRNA expression profiles from 513 LUAD tumor tissue samples and 59 adjacent normal tissue samples, all downloaded from The Cancer Genome Atlas (TCGA). Clinical data for 522 patients were included, with the patient cohort being predominantly female (53.8%), with a median age of 65, and mostly Stage I (54.2%) disease.
Differential expression analysis. Differentially expressed lncRNAs (DElncRNAs) and mRNAs (DEmRNAs) were identified using DESeq2 with strict thresholds: false discovery rate below 0.05 and absolute log2 fold change greater than 2. This yielded 260 DElncRNAs and 1364 DEmRNAs between LUAD tumor and normal tissue.
Three-stage feature selection. The identification of optimal diagnostic lncRNA biomarkers proceeded in three steps: LASSO regression reduced the 260 DElncRNAs to 35 candidates; random forest analysis ranked these 35 by their mean decrease in accuracy; and forward-wrapper selection with SVM determined that 8 lncRNAs achieved the highest classification accuracy in ten-fold cross-validation before plateauing.
Three classification models tested. The final 8 lncRNA biomarkers were used to build three machine learning classifiers: random forests, decision tree, and support vector machine (SVM). Each model was evaluated by AUC, sensitivity, and specificity using ROC curve analysis via the pROC package in R.
Eight biomarkers identified. The optimal diagnostic lncRNA panel consists of: LANCL1-AS1, MIR3945HG, LINC01270, RP5-1061H20.4, BLACAT1, LINC01703, CTD-2227E11.1, and RP1-244F24.1. Of these, LANCL1-AS1 and MIR3945HG were down-regulated in LUAD, while the remaining six were up-regulated compared to normal tissue.
Exceptional model performance. The random forest model achieved an AUC of 0.999 with sensitivity 99.8% and specificity 98.3%. The SVM model also achieved AUC 0.999, with perfect specificity of 100% and sensitivity of 98.4%. The decision tree model performed somewhat lower, with AUC 0.937, sensitivity 99%, and specificity 93.2%.
Individual biomarker diagnostic value. Each of the eight individual lncRNAs also showed diagnostic power, with individual AUC values all exceeding 0.89. This suggests that each lncRNA carries independent discriminative information for distinguishing LUAD from normal tissue, not just the combined panel.
Novel biomarkers. Six of the eight lncRNAs (LANCL1-AS1, LINC01270, RP5-1061H20.4, LINC01703, CTD-2227E11.1, and RP1-244F24.1) had not previously been reported in LUAD, making this study the first to identify them as LUAD-associated diagnostic markers.
Building the co-expression network. Weighted Gene Co-expression Network Analysis (WGCNA) was applied to identify modules of highly correlated lncRNAs and mRNAs. A Pearson correlation matrix was computed for all lncRNA-mRNA pairs, then converted to an adjacency matrix using a soft thresholding power of 20, which enforced a scale-free network topology.
Nine co-expression modules identified. Hierarchical clustering identified nine distinct modules. Correlation analysis between each module and tumor versus adjacent-normal status identified two modules most relevant to LUAD: the pink module (negatively associated with tumor, r = -0.55, p = 1.50E-44) and the green module (positively associated with tumor, r = 0.56, p = 1.10E-47).
Module composition. The pink module contained 44 down-regulated mRNAs and one down-regulated lncRNA (RP11-389C8.2). The green module contained 241 up-regulated mRNAs and two up-regulated lncRNAs: CTD-2510F5.4 and TMPO-AS1. Network visualization showed RP11-389C8.2 co-expressed with 44 mRNAs, while CTD-2510F5.4 and TMPO-AS1 were each co-expressed with over 240 mRNAs.
Hub lncRNAs in the network. The extensive co-expression of CTD-2510F5.4 and TMPO-AS1 with hundreds of mRNAs in the green module suggests these lncRNAs may serve as regulatory hubs that coordinate broad transcriptional programs relevant to LUAD tumor biology.
Green module pathways. Functional annotation of the 241 mRNAs in the tumor-associated green module revealed highly significant enrichment in cell cycle (p = 1.82E-35), DNA replication (p = 1.02E-19), and p53 signaling pathway (p = 1.71E-05) as the top three KEGG pathways. These are core oncogenic processes governing tumor proliferation and genomic instability.
Pink module pathways. The pink module, which is negatively associated with LUAD, was enriched for blood vessel morphogenesis, angiogenesis, cell adhesion, and cell migration by GO analysis, with leukocyte transendothelial migration and cell adhesion molecules as enriched KEGG pathways. This suggests that loss of vascular and adhesion regulation may accompany LUAD development.
Mechanistic implications. The enrichment of cell cycle and p53 pathway genes co-expressed with TMPO-AS1 and CTD-2510F5.4 suggests these lncRNAs may contribute to LUAD by dysregulating cell cycle checkpoints and compromising p53-mediated tumor suppression. These findings provide a mechanistic hypothesis for future experimental validation.
TMPO-AS1 as a prognostic hub. Prior studies have shown TMPO-AS1 is a prognostic biomarker for LUAD and influences prognosis through regulation of cell cycle and cell adhesion. The WGCNA results independently confirm its upregulation and central network position in LUAD, strengthening evidence for its clinical relevance.
qRT-PCR confirmation. Eight biomarker lncRNAs were tested by quantitative RT-PCR in tissue samples from five LUAD patients and five normal adjacent controls. Six of the eight showed expression patterns consistent with TCGA analysis (two down-regulated, four up-regulated). Only RP5-1061H20.4 and LINC01703 showed discordant results, which may reflect sample size limitations.
GEO dataset validation. Independent validation used two Gene Expression Omnibus (GEO) datasets: GSE32863 (60 LUAD and 60 normal controls) and GSE104854 (9 LUAD and 9 normal controls). The expression patterns of selected DEmRNAs (ETV4, TUBB3, EPAS1, PECAM1) and DElncRNAs (BLACAT1, TMPO-AS1) were examined in these datasets.
Mostly concordant results. ETV4, TUBB3, and BLACAT1 were up-regulated while EPAS1 and PECAM1 were down-regulated in LUAD in the GEO datasets, consistent with TCGA findings. The one exception was TMPO-AS1, which appeared down-regulated in the GEO data despite being up-regulated in TCGA, a discrepancy the authors acknowledge.
Known biomarker BLACAT1 confirmed. BLACAT1, originally identified in bladder cancer and subsequently shown to act as an oncogenic lncRNA in multiple cancers including non-small cell lung cancer, was the top-ranked DElncRNA in this LUAD dataset. Its consistent up-regulation across TCGA and GEO data confirms the reliability of the integrated analysis pipeline.
A validated diagnostic lncRNA panel. This study identified eight lncRNAs as optimal diagnostic biomarkers for LUAD using a systematic machine learning pipeline applied to a large TCGA dataset. The high AUC values (above 0.999 for the best models) and validation across independent cohorts support the potential clinical utility of this panel for LUAD diagnosis.
Novel candidates for functional investigation. Six of the eight biomarkers were not previously characterized in LUAD, representing new candidates for mechanistic investigation. Understanding how these lncRNAs regulate tumor biology could reveal new therapeutic targets beyond their diagnostic value.
MIR3945HG as a cross-cancer biomarker. MIR3945HG, previously identified as a candidate diagnostic marker for tuberculosis and a high-value biomarker in lung squamous cell carcinoma, was found to be down-regulated in LUAD as well. This pattern suggests MIR3945HG may act as a tumor suppressor lncRNA with roles across multiple lung diseases.
Limitations and next steps. The qRT-PCR validation cohort was small (n=5 per group), and the biological functions of most identified lncRNAs in LUAD remain unknown. Future work should include larger clinical validation studies and in vitro or in vivo experiments to determine how these lncRNAs mechanistically contribute to LUAD development and whether they can serve as therapeutic targets.