Discriminating malignancy of pulmonary nodules using a peripheral blood leukocyte DNA methylation-based SEMO network model

Background Accurately discriminating malignant pulmonary nodules from benign lesions remains challenging, particularly in subcentimeter lesions. This study developed and internally validated a peripheral blood leukocyte (PBL) DNA methylation-derived Selective Epigenomic Module Optimization (SEMO) network model to improve risk stratification. Methods A prospective specimen collection and retrospective analysis were performed. Preoperative PBL samples from 149 patients (102 were subcentimeter lesions) with surgically resected and pathologically confirmed pulmonary nodules were analyzed via targeted methylation sequencing of 500 aging/immune-related gene promoters. Protein-protein interaction (PPI) sets and functional pathway-based gene sets were integrated to derive SEMO features by calculating differential β-value distributions. Three models including clinical-radiological Baseline, SEMO-only, and Combined classifiers were established using elastic-net regression. Model performance was internally validated using 1,000 stratified bootstrap replicates with optimism correction and patient-level out-of-bag (OOB) prediction, no external validation was performed. Clinical utility was evaluated via decision curve analysis (DCA). Results The SEMO-only and Combined models demonstrated comparable discrimination, with an apparent AUC of 0.996 (95% CI 0.988-1.000) and 0.993 (95% CI 0.978-1.000). Bootstrap internal validation yielded optimism-corrected AUCs of 0.966 and 0.966, with the more conservative patient-level OOB AUCs were 0.905 (95% CI 0.837–0.956) and 0.921 (95% CI 0.859–0.966). The calibration intercept, calibration slope, and Brier score were 0.019, 0.954, and 0.065 for SEMO-only model and -0.454, 1.390, and 0.076 for Combined model. In an exploratory analysis, nodules were stratified by predicted probability into low-risk (<0.37), intermediate-risk (0.37–0.87), and high-risk (≥0.87) categories. The subcentimeter evaluation as a prespecified exploratory endpoint, SEMO-only and Combined models maintained AUCs of 0.920 and 0.931. Conclusion The PBL-derived SEMO model may serve as a non-invasive preoperative risk-stratification tool for patients with CT-detected nodules highly suspicious for malignancy who are being considered for surgery. External validation is required before clinical implementation.