Disease prediction by cell-free DNA methylation

Abstract

Disease diagnosis using cell-free DNA (cfDNA) has been an active research field recently. Most existing approaches perform diagnosis based on the detection of sequence variants on cfDNA; thus, their applications are limited to diseases associated with high mutation rate such as cancer. Recent developments start to exploit the epigenetic information on cfDNA, which could have substantially wider applications. In this work, we provide thorough reviews and discussions on the statistical method developments and data analysis strategies for using cfDNA epigenetic profiles, in particular DNA methylation, to construct disease diagnostic models. We focus on two important aspects: marker selection and prediction model construction, under different scenarios. We perform simulations and real data analysis to compare different approaches, and provide recommendations for data analysis.

Keywords: DNA methylation; cell-free DNA; epigenetics; liquid biopsy; marker selection; predictive modeling.

Figure 1

Schematic overview of plasma cfDNA methylation mixing procedure and deconvolution methods for disease detection and monitoring. One straightforward approach is to use marker directly for disease detection. Besides using biomarkers directly, signal deconvolution methods can be categorized into either the ‘reference-free approach’ when the external tissue-specific methylation reference in unavailable, or the ‘reference-based approach’ when the tissue-specific profile is known.

Figure 2

Boxplot of classification accuracies for multiple methods in simulations. Marker: Marker directly predict approach. QP: using tissue proportions solved from QP procedure for prediction. NMF approach. True Prop: using simulated true proportion in classification. A total number of 20 simulations are conducted. (A) Low noise level; (B) medium noise level; (C) high noise level.

Figure 3

Scatterplots of NMF estimated reference methylation levels versus true reference methylation levels in four tissues. (A) Small intestines; (B) adipose tissues; (C) adrenal glands; (D) lungs. Relatively strong correlations are observed. Spearman’s correlation is shown in each panel.

Figure 4

Scatterplots of NMF estimated tissue proportions versus true tissue proportions in four tissues. (A) Small intestines; (B) adipose tissues; (C) adrenal glands; (D) lungs. Relatively strong correlations are observed. Spearman’s correlation is shown in each panel.

Figure 5

Barplot for the estimated 14 tissue proportions from real data for HCC patients, healthy controls and pregnant subjects, using QP with external reference available. HCC patients showed an increased proportion of cfDNA originating from liver, while pregnant controls showed an increased proportion of cfDNA originating from placenta.

Figure 6

Boxplot of real data solved tissue proportions for liver and placenta, respectively, among three groups. (A) Tissue proportions for liver among three groups; (B) tissue proportions for placenta among three groups.