MethylMix: an R package for identifying DNA methylation-driven genes.

UNLABELLED: DNA methylation is an important mechanism regulating gene transcription, and its role in carcinogenesis has been extensively studied. Hyper and hypomethylation of genes is an alternative mechanism to deregulate gene expression in a wide range of diseases. At the same time, high-throughput DNA methylation assays have been developed generating vast amounts of genome wide DNA methylation measurements. Yet, few tools exist that can formally identify hypo and hypermethylated genes that are predictive of transcription and thus functionally relevant for a particular disease. To accommodate this lack of tools, we developed MethylMix, an algorithm implemented in R to identify disease specific hyper and hypomethylated genes. MethylMix is based on a beta mixture model to identify methylation states and compares them with the normal DNA methylation state. MethylMix introduces a novel metric, the 'Differential Methylation value' or DM-value defined as the difference of a methylation state with the normal methylation state. Finally, matched gene expression data are used to identify, besides differential, transcriptionally predictive methylation states by focusing on methylation changes that effect gene expression.
AVAILABILITY AND IMPLEMENTATION: MethylMix was implemented as an R package and is available in bioconductor.

1 Introduction

DNA methylation is one of the most studied epigenetic aberrations underlying oncogenesis. Besides genetic mutations, hyper and hypomethylation of genes is an alternative mechanism that is capable of altering the normal state and driving a wide range of diseases. Prior studies have identified hypo or hypermethylation based on heuristic measures for example in breast cancer (Hill ). Additionally, computational methods have been developed to identify differentially methylated regions for specific DNA methylation platforms (Aryee ; Wang et al., 2012; Warden et al., 2013). However, few methods formalize the identification of DNA methylation driven genes using a model-based approach. We identified three key criteria that should be addressed to derive key methylation-driven genes. First, the determination of the degree of methylation cannot hinge on arbitrary thresholds as is commonly done. Second, the assessment of a gene as hyper or hypomethylated must be made in comparison to normal tissue. Finally, the identification of genes that are hyper or hypomethylated should be transcriptionally predictive effect, thereby implying that their methylation is functionally relevant. We designed MethylMix to accommodate these three criteria to identify methylation driven genes in diseases.

2 Algorithm

MethylMix integrates DNA methylation from normal and disease samples and matched disease gene expression data via a three-step algorithm: The final output of MethylMix is genes that are both transcriptionally predictive and differential together with the parameters of their methylation states. Additionally, a matrix of DM-values is part of the output and can be used in subsequent analysis, for example to define methylation driven subgroups using clustering algorithms.

Step i: Genes are filtered by identifying transcriptionally predictive methylation. First, each CpG site is associated with its closest gene. Next, MethylMix requires that the DNA methylation of a CpG site has a significant effect on its corresponding gene expression in order for the gene to be considered a methylation-driven gene. We define such genes as transcriptionally predictive genes.

Step ii: The methylation states of a gene are identified using univariate beta mixture modeling to identify subgroups of patients with similar DNA methylation level for a specific CpG site. We use the Bayesian Information Criterion (BIC) to select the number of methylation states by iteratively adding a new mixture component if the BIC score improves. Each beta mixture component is referred to as a methylation state and represented by its mean methylation level.

Step iii: Hyper and hypomethylated genes are defined relative to normal by comparing the methylation levels of each methylation state to the mean of the DNA methylation levels of normal tissue samples using a Wilcoxon rank sum test. Based on this test, Differential Methylation values or DM-values are created defined as the difference of a methylation state with the normal methylation state. Genes with methylation states different from normal are called differential genes.

3 Functions and examples

MethylMix was implemented in the statistical language R and is provided as an R package in the supplementary data. MethylMix contains two key functionalities; the creation of MethylMix models for a set of genes of any size and the visualization of a MethylMix plot for each gene. MethylMix needs three datasets: normal DNA methylation data, disease DNA methylation data and matched disease gene expression data. The normal DNA methylation data should ideally be from the same tissue or cell type as the disease DNA methylation data. We provided example data for 14 genes from 251 glioblastoma patients from The Cancer Genome Atlas (TCGA) (McLendon ) in the package. The 14 genes were selected based on their documented differential DNA methylation status in glioblastoma in the literature (Etcheverry ; Hegi ). First, a MethylMix model is created for all genes as follows:

> library(MethylMix)

> data(METcancer)

> data(METnormal)

> data(MAcancer)

> MethylMixResults = MethylMix(METcancer,METnormal,MAcancer)

MethylMix will first investigate for each gene if it is transcriptionally predictive by building a linear regression model that estimates the association between DNA methylation and gene expression. MethylMix only selects genes with a significant inverse relationship (P value < 0.01) resulting in nine transcriptionally predictive genes in this example. Then a MethylMix model is created for these nine transcriptionally predictive genes and MethylMix reports how many methylation states each gene has. For large datasets with more genes, MethylMix can be run in parallel mode and take advantage of multiple cores.

Next, a MethylMix model plot can be created for each gene visualizing the beta mixture model and the methylation states that were identified for a particular gene. Additional parameters can be passed to the plot function by adding the normal methylation data and the matched gene expression data. These additional parameters will visualize the 95% confidence interval of the normal DNA methylation data and the relationship with matched gene expression data.

> MethylMix_PlotModel(’MGMT’,METcancer, MethylMixResults,MAcancer,METnormal)

For example, Figure 1 displays the MethylMix model for MGMT showing two methylation states, whereby the low methylation state matches the normal methylation and the high methylation state corresponds to hypermethylation of MGMT, a well-known case of hypermethylation influencing treatment of glioblastoma patients (Hegi ). Next, Figure 2 shows the inverse correlation between DNA methylation and matched gene expression of MGMT.

Fig. 1.

MethylMix model for the MGMT gene based on 251 glioblastoma patients from TCGA

Fig. 2.

Inverse correlation of DNA methylation and gene expression for MGMT in 251 glioblastoma patients from TCGA

4 Conclusion

MethylMix is an R package that identifies hyper and hypomethylated genes using a beta mixture modeling approach. MethylMix also quantifies the effect DNA methylation has on gene expression, thereby identifying transcriptionally predictive DNA methylation events. MethylMix can be used both to study single genes as in the example above or in parallel mode to build MethylMix models genome wide. MethylMix requires a large cohort to identify methylation states and capture DNA methylation heterogeneity present in a particular disease. We used MethylMix and their associated DM-values to identify driver genes (Gevaert and Plevritis, 2013; Gevaert , 2014), on dataset sizes of 100 samples ore more, and on more than 4000 TCGA cases across 12 tissues to identify methylation driven subgroups (Gevaert ). In summary, MethylMix offers a new tool to identify methylation-driven genes providing a complimentary source of information to copy number and mutation spectra to identify disease driver genes.

Funding

This work was partially supported by NIH/NCI R01 CA160251 and NIH/NCI U01 CA176299.

Conflict of Interest: none declared.