A prediction model based on DNA methylation biomarkers and radiological characteristics for identifying malignant from benign pulmonary nodules

Abstract

Background: Lung cancer remains the leading cause of cancer deaths across the world. Early detection of lung cancer by low-dose computed tomography (LDCT) can reduce the mortality rate. However, making a definitive preoperative diagnosis of malignant pulmonary nodules (PNs) found by LDCT is a clinical challenge. This study aimed to develop a prediction model based on DNA methylation biomarkers and radiological characteristics for identifying malignant pulmonary nodules from benign PNs.

Methods: We assessed three DNA methylation biomarkers (PTGER4, RASSF1A, and SHOX2) and clinically-relevant variables in a training cohort of 110 individuals with PNs. Four machine-learning-based prediction models were established and compared, including the K-nearest neighbors (KNN), random forest (RF), support vector machine (SVM), and logistic regression (LR) algorithms. Variables of the best-performing algorithm (LR) were selected through stepwise use of Akaike's information criterion (AIC). The constructed prediction model was compared with the methylation biomarkers and the Mayo Clinic model using the non-parametric approach of DeLong et al. with the area under a receiver operator characteristic curve (AUC) analysis.

Results: A prediction model was finally constructed based on three DNA methylation biomarkers and one radiological characteristic for identifying malignant from benign PNs. The developed prediction model achieved an AUC value of 0.951 in malignant PNs diagnosis, significantly higher than the three DNA methylation biomarkers (0.912, 95% CI:0.843-0.958, p = 0.013) or Mayo Clinic model (0.823, 95% CI:0.739-0.890, p = 0.001). Validation of the prediction model in the testing cohort of 100 subjects with PNs confirmed the diagnostic value.

Conclusion: We have shown that integrating DNA methylation biomarkers and radiological characteristics could more accurately identify lung cancer in subjects with CT-found PNs. The prediction model developed in our study may provide clinical utility in combination with LDCT to improve the over-all diagnosis of lung cancer.

Keywords: Biomarkers; CT; DNA methylation; Lung cancer; Pulmonary nodules.

Fig. 1

Comparison of the studied DNA methylation expressions in patients with benign PNs, and patients with malignant PNs in a training cohort. Scatter plots show the distribution of relative normalized methylation values for each of the 3 genes determined by q-PCR. The paired t-test was performed

Fig. 2

Receiver-operator characteristic (ROC) curve analysis of the three models in a training cohort. The area under the ROC curve (AUC) for each model conveys its accuracy for diagnosing malignant PNs. The prediction model produced a higher AUC value for identifying malignant PNs comparing with the panel of the three DNA methylation biomarkers and the Mayo Clinic model

Fig. 3

Comparison of the studied DNA methylation expressions in patients with benign PNs, and patients with malignant PNs in an independent cohort. Scatter plots show the distribution of relative normalized methylation values for each of the 3 genes determined by q-PCR. The paired t-test was performed

Fig. 4

Comparison of ROC curves generated using the prediction model, panel of the three DNA methylation biomarkers, and Mayo Clinic model in an independent cohort. The prediction model produced the highest AUC value of the three models