DNA methylation signatures of cervical pre-invasive and invasive disease: An epigenome-wide association study

Abstract

Epigenetic alterations are essential in the development of cancers, while epigenome-wide exploration in cervical cancer has been limited. In this epigenome-wide association study (EWAS) we explore differential DNA methylation signatures associated with CIN (cervical intraepithelial neoplasia) grade 3 and cervical cancer to better understand potential drivers and biomarkers of cervical carcinogenesis. 247 women were recruited between 2014 and 2020 (N = 119 benign, N = 74 CIN3/CGIN [cervical glandular intraepithelial neoplasia] and N = 54 cancer). Methylation signatures were obtained from exfoliated cervical cells and sequenced using the Illumina 850 k array. Logistic regression and conditional analyses were used to test for independent associations between Cytosine-phosphate-Guanine (CpG) sites and case-control status, with adjustment for batch, chip, age, and human papillomavirus (HPV) status. 409 CpG sites were strongly associated with CIN3/cancer (p-value <5 × 10^-8). Following conditional analysis, two CpG sites located in PAX1 (cg16767801) and NREP-AS1 genes (cg23642047) were independently associated with case status, yielding an area under the curve (AUC) of 0.92 (AUC = 0.97 for invasive disease). In a validation dataset (CIN3 only) PAX1/NREP-AS1 yielded a combined AUC of 0.77. Methylation markers offer promise for use in cervical screening, particularly as triage tests and self-sampling. We have identified a novel combined methylation marker that offers a high accuracy for the detection of CIN3 or worse.

Keywords: EWAS; PAX; cervical cancer; cervical intraepithelial neoplasia; epigenome; human papillomavirus; methylation.

FIGURE 1

Associations between methylation levels and CIN3 (N = 73) or ICC (N = 54) status (vs. N = 114 controls) in the study population. (A) Manhattan plot showing the p‐values, measuring the strength of association, derived from logistic models adjusted for HPV status and represented on the‐log10 scale (Y‐axis). CpG sites (N = 843,611) are ordered by their position on the genome (X‐axis). The epigenome‐wide significance level is set to the Bonferroni‐corrected threshold (horizontal red line). (B) Heatmap of absolute correlations between the CpG sites (N = 727) significantly associated with CIN3 or ICC status in the model adjusted on HPV status, or further adjusted on cg16767801. To avoid spurious correlations, these are adjusted on the case/control status. CpG sites are ordered by hierarchical clustering with complete linkage using the absolute correlation as a similarity metric. (C) Circos plot showing the p‐values from conditional analyses on‐log10 scale. Models are adjusted on HPV status (Model 0), and further adjusted on cg16767801 (Model 1), and cg23642047 (Model 2). CpG sites (N = 727) are ordered based on the same hierarchical clustering. Lead CpG sites from each of the models are indicated with vertical dotted lines. Points are coloured by chromosome as in panel (A).

FIGURE 2

Receiver operator curves (ROC) for logistic models including cg16767801 (in blue), or cg23642047 (in green) or both (in red) as predictors and all cases (A), cancer (B) or CIN3 (C, D) as outcome. Models were fitted on 1000 training sets with 80% of the cases and 80% of the controls and performances were evaluated on a test set with the remaining 20% of participants. The pointwise average, 5th and 95th percentiles of the True and False Positive Rates and Area Under the Curve (AUC) are reported. Results presented in panels A, B and C were obtained using our study population (N = 114 controls, N = 73 CIN3 cases and N = 54 cancer cases). An independent validation set (accession number GSE14375s, with N = 54 controls and N = 42 CIN3 cases) was used in panel D.