Lung cancer prevalence. Lung cancer is the third most common cancer in the United States, with an overall five-year survival rate of approximately 17%. Non-small cell lung cancer (NSCLC) accounts for 80-85% of all lung malignancies, and adenocarcinoma is its most prevalent histological subtype, carrying a mortality rate of about 80%.
Early detection imperative. Improving patient outcomes depends heavily on early detection and accurate diagnosis. Identifying reliable biomarkers for lung adenocarcinoma could enable earlier intervention and more targeted treatment approaches, ultimately reducing cancer-related mortality.
Deep learning for biomarker discovery. While previous studies have used machine learning techniques such as one-class logistic regression to identify cancer stem cell biomarkers, few have applied deep learning specifically to biomarker identification in lung adenocarcinoma. This study aims to bridge that gap by combining deep learning analysis with experimental RT-PCR validation.
RNA-Seq dataset. The study used RNA-Seq data from the GDAC (Global Data Assembly Centers) dataset, comprising 522 samples and 20,531 genes. This included 506 lung adenocarcinoma patients and 16 healthy controls, along with clinical parameters such as gender, tumor stage, TNM classification, and survival outcomes.
Preprocessing pipeline. Duplicate genes and samples were removed using filtering code in R. The remaining 20,531 genes were normalized using the Limma package. Differentially expressed genes (DEGs) were identified based on a log fold change of 2 or greater and a P-value below 0.05, yielding 3,513 DEGs for downstream analysis.
Gene symbol standardization. All gene symbols were standardized according to HGNC nomenclature. Where Ensembl identifiers were initially retrieved, approved gene symbols were used consistently, with original IDs retained for reference and reproducibility.
Neural network design. The researchers constructed a feed-forward neural network with three fully connected hidden layers containing 128, 64, and 32 neurons respectively. Each hidden layer used the ReLU activation function, while the output layer employed a sigmoid function for binary class probability prediction.
Training and regularization. The model was trained using the Adam optimizer with a learning rate of 0.001 and binary cross-entropy loss. Dropout regularization was applied to hidden layers, and early stopping based on validation loss prevented overfitting. A 70/30 train-test split was used with no overlap between sets.
Evaluation metrics. Performance was assessed using AUC, accuracy, F1 score, R-squared, confusion matrix, and the Brier score - which measures the mean squared error between predicted probabilities and true class labels. All metrics were calculated on the independent held-out test set.
Exceptional predictive accuracy. The deep learning model achieved a classification accuracy of 98.44% on the test set, with a Brier score of just 0.0013 and an AUC of 1.0, demonstrating near-perfect discrimination between lung adenocarcinoma and healthy samples.
Key gene identification. CYP3A7 was found to have the highest effect size among all analyzed genes. Combined ROC analysis using a generalized linear model identified A2M, CYP2C9, and SIRPD as the top diagnostic gene combination, achieving a sensitivity of 0.96.
Functional enrichment. Gene Ontology and KEGG pathway enrichment analyses revealed significant involvement of identified DEGs in critical biological pathways, reinforcing the clinical relevance of the discovered biomarkers.
Clinical parameter analysis. The team examined 55 DEGs for correlations with demographic and clinical factors including age, tumor size, lymph node involvement, distant metastasis, and cancer stage using Spearman correlation techniques.
Survival analysis. Kaplan-Meier survival curves and Cox regression analyses identified prognostic genes among the top-selected DEGs. Genes meeting the criteria of hazard ratio between -1 and 1 with P-value below 0.05 were classified as prognostic markers.
External validation. The expression levels of candidate biomarker genes were validated using independent datasets from GDAC and multiple GEO datasets (GSE19804, GSE75037, GSE131907, GSE37745, and GSE166720), confirming the consistency of the bioinformatics findings.
Study cohort. Blood samples were collected from 30 lung adenocarcinoma patients (stages 1-3, no distant metastasis, treatment-naive) and 30 age- and sex-matched healthy controls. All participants provided informed consent.
RNA extraction and cDNA synthesis. Five milliliters of peripheral blood were collected from each participant, with RNA extraction performed using the RNeasy Midi Kit. cDNA synthesis was carried out using the Viva 2-step RT-PCR Kit with primers designed using AlleleID6 software.
Real-time PCR protocol. Quantitative PCR was performed using HotTaq EvaGreen qPCR mix. The fluorescent dye EvaGreen binds to double-stranded DNA for detection, and results were interpreted based on amplification and melt curves.
Expression validation. Real-time PCR confirmed that CYP2C9, KRT14, and PECAM1 were significantly upregulated in lung adenocarcinoma patients compared to healthy controls (P less than 0.001). Conversely, A2M was significantly downregulated in patients.
Bridging computation and experiment. The concordance between the deep learning predictions and the RT-PCR results demonstrates the effectiveness of AI-driven approaches in identifying clinically relevant biomarkers. This experimental verification strengthens the case for these genes as potential diagnostic tools.
Blood-based detection potential. The fact that these biomarker expression changes are detectable in peripheral blood samples suggests the possibility of developing non-invasive blood-based diagnostic tests for lung adenocarcinoma.
Integrated approach. This study demonstrates the power of combining deep learning bioinformatics with experimental validation for biomarker discovery. The identified genes - A2M, CYP2C9, KRT14, PECAM1, and SIRPD - represent promising candidates for early lung adenocarcinoma detection and disease management.
Clinical translation potential. The bioinformatics-identified genes could serve as a foundation for developing diagnostic panels that enable earlier cancer detection. The RT-PCR validation confirms that AI can effectively identify biomarkers that hold up under experimental scrutiny.
Future directions. Larger validation cohorts and prospective clinical trials will be needed to fully establish the diagnostic utility of these biomarkers. The methodology itself - deep learning discovery followed by experimental confirmation - provides a replicable framework for biomarker identification in other cancer types.