Leveraging locus-specific epigenetic heterogeneity to improve the performance of blood-based DNA methylation biomarkers

Abstract

Background: Variation in intercellular methylation patterns can complicate the use of methylation biomarkers for clinical diagnostic applications such as blood-based cancer testing. Here, we describe development and validation of a methylation density binary classification method called EpiClass (available for download at https://github.com/Elnitskilab/EpiClass ) that can be used to predict and optimize the performance of methylation biomarkers, particularly in challenging, heterogeneous samples such as liquid biopsies. This approach is based upon leveraging statistical differences in single-molecule sample methylation density distributions to identify ideal thresholds for sample classification.

Results: We developed and tested the classifier using reduced representation bisulfite sequencing (RRBS) data derived from ovarian carcinoma tissue DNA and controls. We used these data to perform in silico simulations using methylation density profiles from individual epiallelic copies of ZNF154, a genomic locus known to be recurrently methylated in numerous cancer types. From these profiles, we predicted the performance of the classifier in liquid biopsies for the detection of epithelial ovarian carcinomas (EOC). In silico analysis indicated that EpiClass could be leveraged to better identify cancer-positive liquid biopsy samples by implementing precise thresholds with respect to methylation density profiles derived from circulating cell-free DNA (cfDNA) analysis. These predictions were confirmed experimentally using DREAMing to perform digital methylation density analysis on a cohort of low volume (1-ml) plasma samples obtained from 26 EOC-positive and 41 cancer-free women. EpiClass performance was then validated in an independent cohort of 24 plasma specimens, derived from a longitudinal study of 8 EOC-positive women, and 12 plasma specimens derived from 12 healthy women, respectively, attaining a sensitivity/specificity of 91.7%/100.0%. Direct comparison of CA-125 measurements with EpiClass demonstrated that EpiClass was able to better identify EOC-positive women than standard CA-125 assessment. Finally, we used independent whole genome bisulfite sequencing (WGBS) datasets to demonstrate that EpiClass can also identify other cancer types as well or better than alternative methylation-based classifiers.

Conclusions: Our results indicate that assessment of intramolecular methylation density distributions calculated from cfDNA facilitates the use of methylation biomarkers for diagnostic applications. Furthermore, we demonstrated that EpiClass analysis of ZNF154 methylation was able to outperform CA-125 in the detection of etiologically diverse ovarian carcinomas, indicating broad utility of ZNF154 for use as a biomarker of ovarian cancer.

Keywords: Cancer diagnostics; Cell-free DNA; DNA methylation; Intermolecular variation; Ovarian cancer.

Fig. 1

Schematic of the methylation density binary classifier. a Hypothetical dataset of methylation data from a single case and control cfDNA sample containing heterogeneously-methylated DNA b Table characterizing the use of a methylation density classifier cutoff. c Schematic illustrating the EpiClass procedure for determining optimal methylation density and epiallelic fraction cutoffs. The optimal methylation density cutoff is 75%, and the epiallelic fraction cutoff is either 2 or 3. Thus, a sample in this example is called positive if it has 2 or more epialleles that exhibit methylation densities of 75% or higher

Fig. 2

Simulated performance of EpiClass as a function of admixture ratios of epithelial ovarian carcinoma (EOC) to WBC RRBS reads. The performance of the methylation density binary classifier (EpiClass, red) and mean locus methylation classifier (blue) at increasing dilutions of EOC RRBS reads in a background of WBC RRBS reads acquired from Widschwendter et al.[8]. Simulated performance of improved classification over mean methylation at different read depths demonstrated in Additional file 1: Figure S4 and S5. AUC area under the curve, EpiClass methylation density classifier

Fig. 3

Performance of EpiClass in EOC patient and control plasma samples. a The pooled epiallelic fractions of cfDNA methylated epialleles with varying methylation densities in EOC (red, n = 26) and healthy (blue, n = 41) patient plasma samples. Purple-shaded regions indicate overlap between the two plasma sets. b Performance of EpiClass at each methylation density cutoff for the EOC and healthy control plasma samples. Dotted line shows the optimal methylation density cutoff derived from EpiClass. The red dot indicates the ROC curve AUC for the mean methylation cutoff. Measurement metric refers to either 1—FPR, TPR, AUC, or TPR–FPR. c ROC curves showing the classification performance of using the optimal methylation density cutoff determined by EpiClass (red), MSP (orange), or mean methylation cutoff (blue) to identify the EOC and healthy control plasma samples. d Boxplots showing the performance of the epiallelic fraction cutoffs for either the optimal 60% methylation density cutoff determined by EpiClass, MSP, or mean methylation to classify plasma samples from EOC patients (red, n = 26) or healthy controls (blue, n = 41) Y-axes adjusted to ignore healthy control outliers. EOC epithelial ovarian carcinoma, EpiClass methylation density classifier, MDC methylation density cutoff, AUC area under the curve; * indicates p < 0.05, two-sided Wilcoxon rank-sum test; ns = not significant; Mean* methylation was inferred from the fraction of all methylated epialleles. MSP* performance estimated using an MDC of 95%. Supplementary Figure S9 demonstrates no statistical difference between sample cohorts with respect to mean methylation

Fig. 4

Validation of EpiClass cutoff values and corresponding performance for identifying EOC from patient plasma. a The pooled epiallelic fractions of cfDNA methylated epialleles with varying methylation densities in the second EOC (red, n = 24) and healthy (blue, n = 12) patient plasma sample cohort. Purple shaded regions indicate overlap between the two plasma sets. b Performance of EpiClass at each methylation density cutoff for the EOC and healthy control plasma samples. Dotted line shows the optimal methylation density cutoff derived from EpiClass. The red dot indicates the ROC curve AUC for the mean methylation cutoff. Measurement metric refers to either 1—FPR, TPR, AUC, or TPR–FPR. c Receiver operating characteristic curve for the optimal 60% methylation density cutoff on the second plasma cohort. d Boxplots indicating the distribution of sample normalized read counts with intramolecular methylation densities greater than or equal to 60% (EpiClass), 95% (MSP*), or greater than 0% (mean* methylation). EOC epithelial ovarian carcinoma; EpiClass methylation density classifier, MDC methylation density cutoff, AUC area under the curve; *** indicates p < 0.001, two-sided Wilcoxon rank-sum test; ns = not significant; Mean* methylation was inferred from the fraction of all methylated epialleles. MSP* performance estimated using an MDC of 95%

Fig. 5

Counts of endometrioid (n = 9) or serous (n = 14) EOC subtype patient plasma samples from the validation cohort above the CA-125 cutoff (35 U/ml) or EpiClass cutoffs derived from the training cohort. Numbers above each bar indicate the number of samples above each given classification cutoff. Mean* methylation was inferred from the fraction of all methylated epialleles. MSP* performance estimated using an MDC of 95%. No available specificity measurements for CA-125 as the control samples (n = 12) did not have measured CA-125 U/mL concentrations. MDC methylation density cutoff

Fig. 6

Comparison of EpiClass and CancerDetector. a, b Receiver operating characteristic curves (ROCs) based on classification of 10 test sample sets using either the optimal methylation density read counts at the ZNF154 marker region for EpiClass (a, blue), or the estimated sample tumor fraction derived by CancerDetector (b, red) using reads at the ZNF154 marker region. Light shaded lines indicate individual ROC curves for each test sample set. Dark line indicates the mean ROC curve. Light shaded region indicates 1 standard deviation from the mean. c, d Same as A–B except for the chosen top 3 liver cancer markers. e, f Same as C–D except for the top 3 liver cancer marker regions and ZNF154 marker region. AUC area under the curve, std. dev. standard deviation