A novel three-CpG signature for lung adenocarcinoma diagnosis: integrating multi-omics, machine learning, and...

Lung adenocarcinoma (LUAD) is a leading cause of cancer-related mortality worldwide. The identification of non-invasive biomarkers for early detection is critical to improving patient outcomes. Integrative analysis of DNA methylation and gene expression represents a promising approach for identifying cancer-specific diagnostic markers. This study aimed to identify and validate a LUAD-specific DNA methylation signature for the early detection of LUAD.

An integrated analysis of DNA methylation data from the Illumina HumanMethylation450K array, including 455 tumor samples, 32 normal samples, and RNA-sequencing (RNA-seq) data from The Cancer Genome Atlas (TCGA) LUAD cohort was performed. Differential methylation and expression were identified, and LUAD-specific markers were selected by excluding probes altered in 10 other cancer types. All combinations of five candidate cytosine-phosphate-guanine (CpG) sites (excluding single-site combinations) were constructed using four machine learning algorithms, generating 104 models. A stringent two-step filtering strategy required each model to achieve an area under the curve (AUC) >0.8, F1 score >0.8, and coefficient of variation (CV) <0.1 in at least four of the five validation datasets [including TCGA test set and four independent Gene Expression Omnibus (GEO) cohorts] and required each CpG combination to be validated by all four algorithms. Candidate methylation markers were experimentally validated by MethyLight quantitative polymerase chain reaction (qPCR) in 97 paired LUAD and 14 benign lung lesion samples.

We identified 6,840 differentially methylated sites (DMSs) and 3,020 differentially expressed genes (DEGs). Integration of multi-omics data revealed 132 genes with correlated methylation-expression patterns. After filtering for LUAD specificity, 36 methylation markers were retained. A three-CpG eXtreme Gradient Boosting (XGBoost) model (cg00817598 in WNT3, cg13470063 in TIE1, and cg23094674 in PLEKHA6) was selected as the optimal classifier, achieving an AUC of 0.995 in TCGA test set and AUC values ranging from 0.864 to 0.986 across the four GEO datasets. The model demonstrated robust cancer-type specificity (AUC: 0.771 in distinguishing LUAD from 565 other cancer samples) and outperformed established markers RASSF1A and SHOX2 (average AUC: 0.725-0.937). Experimental validation confirmed significant hypermethylation of all three CpG sites in LUAD tissues (all P<0.001). The model achieved an AUC of 0.832 in distinguishing LUAD from benign lesions. Methylation levels of cg23094674 and cg00817598 showed positive correlations with tumor size and progressively increased from benign lesions to early-stage and advanced LUAD.

A three-CpG methylation signature was identified that may serve as a highly specific and sensitive diagnostic biomarker for LUAD, supported by rigorous model selection, multi-dataset validation, and experimental confirmation. The ability of the signature to distinguish LUAD from other cancer types and benign lesions highlights its strong potential for clinical translation as a non-invasive assay for detection.