Hypomethylation on cfDNA within transposable elements can predict ovarian malignancy in women with adnexal masses

ObjectiveTumor transformation is accompanied by widespread methylation changes, in the form of cytosine modifications, in transposable elements (TEs). The purpose of this study was to investigate whether changes in methylation and hydroxymethylation in the form of 5mC and 5hmC within transposable elements in circulating cell-free DNA (cfDNA) can serve as predictive markers of ovarian malignancy. MethodsHealthy women as well as women undergoing surgery for various benign, borderline, or malignant adnexal masses and gynecologic conditions were selected for this study. 5hmC-Seal or LABS-seq were performed on cfDNA isolated from prospectively collected serum and plasma samples. We built two models using either differentially hydroxymethylated (5hmC) or methylated (5mC) features within TEs to establish models predictive of ovarian malignancy in women with adnexal masses. ResultsWe isolated cfDNA and analyzed a total of 522 samples (plasma, n = 363; serum, n = 159) from women receiving care at the University of Chicago Medical Center. Multivariate modelling using age, cell deconvolution and the top 38 differentially hydroxymethylated (5hmC) transposable elements was able to accurately predict malignancy in women with an area under the curve (AUC) of 0.854 and a 95% confidence interval: 0.746- 0.962 (95% CI) in plasma and an AUC of 0.893 (95% CI: 0.806- 0.98%) in serum. 5mC-based modelling using the top 25 positive and negatively correlated features, based on differentially methylated transposable element promoters, was able to accurately predict malignancy with an AUC of 0.970 (accuracy of 0.936). ConclusionsWe developed two multivariate models based on cfDNA-derived 5hmC or 5mC modified transposable elements to accurately predict ovarian malignancy in women. Furthermore, we show that our 5hmC-based model applies to plasma (EDTA) and serum samples, which are commonly collected in clinical practice. The results of our study warrant further investigation of the predictive performance of our models in large-scale cohorts in the future.