DNA methylation plays an essential role in cancer. Differential variability (DV) in cancer was recently observed that contributes to cancer heterogeneity and has been shown to be crucial in detecting epigenetic field defects, DNA methylation alterations happening early in carcinogenesis. As neighboring CpG sites are highly correlated, here, we present a new method to detect differentially methylated regions (DMRs) that uses combined signals from differential methylation and DV between sample groups. We demonstrated in simulation studies the superior performance of the new method than existing methods that use only one type of signals when true DMRs have both. Applications to DNA methylation data of breast invasive carcinoma (BRCA) and kidney renal clear cell carcinoma (KIRC) from The Cancer Genome Atlas (TCGA) and BRCA from Gene Expression Omnibus (GEO) suggest that the new method identified additional cancer-related DMRs that were missed by methods using one type of signals. Replication analyses using two independent BRCA data sets suggest that DMRs detected based on DV are reproducible. Only the new method identified epigenetic field defects when comparing normal tissues adjacent to tumors and normal tissues from age-matched cancer-free women from the GEO BRCA data and confirmed their enrichment in the progression to breast cancer.

BACKGROUND: There is growing evidence that DNA methylation alterations may contribute to carcinogenesis. Recent data also suggest that DNA methylation field defects in normal pre-neoplastic tissue represent infrequent stochastic "outlier" events. This presents a statistical challenge for standard feature selection algorithms, which assume frequent alterations in a disease phenotype. Although differential variability has emerged as a novel feature selection paradigm for the discovery of outliers, a growing concern is that these could result from technical confounders, in principle thus favouring algorithms which are robust to outliers. RESULTS: Here we evaluate five differential variability algorithms in over 700 DNA methylomes, including two of the largest cohorts profiling precursor cancer lesions, and demonstrate that most of the novel proposed algorithms lack the sensitivity to detect epigenetic field defects at genome-wide significance. In contrast, algorithms which recognise heterogeneous outlier DNA methylation patterns are able to identify many sites in pre-neoplastic lesions, which display progression in invasive cancer. Thus, we show that many DNA methylation outliers are not technical artefacts, but define epigenetic field defects which are selected for during cancer progression. CONCLUSIONS: Given that cancer studies aiming to find epigenetic field defects are likely to be limited by sample size, adopting the novel feature selection paradigm advocated here will be critical to increase assay sensitivity.

MOTIVATION: There is a growing number of studies generating matched Illumina Infinium HumanMethylation450 and gene expression data, yet there is a corresponding shortage of statistical tools aimed at their integrative analysis. Such integrative tools are important for the discovery of epigenetically regulated gene modules or molecular pathways, which play key roles in cellular differentiation and disease. RESULTS: Here, we present a novel functional supervised algorithm, called Functional Epigenetic Modules (FEM), for the integrative analysis of Infinium 450k DNA methylation and matched or unmatched gene expression data. The algorithm identifies gene modules of coordinated differential methylation and differential expression in the context of a human interactome. We validate the FEM algorithm on simulated and real data, demonstrating how it successfully retrieves an epigenetically deregulated gene, previously known to drive endometrial cancer development. Importantly, in the same cancer, FEM identified a novel epigenetically deregulated hotspot, directly upstream of the well-known progesterone receptor tumour suppressor pathway. In the context of cellular differentiation, FEM successfully identifies known endothelial cell subtype-specific gene expression markers, as well as a novel gene module whose overexpression in blood endothelial cells is mediated by DNA hypomethylation. The systems-level integrative framework presented here could be used to identify novel key genes or signalling pathways, which drive cellular differentiation or disease through an underlying epigenetic mechanism. AVAILABILITY AND IMPLEMENTATION: FEM is freely available as an R-package from http://sourceforge.net/projects/funepimod.

Recently it has been observed that cancer tissue is characterised by an increased variability in DNA methylation patterns. However, how the correlative patterns in genome-wide DNA methylation change during the carcinogenic progress has not yet been explored. Here we study genome-wide inter-CpG correlations in DNA methylation, in addition to single site variability, during cervical carcinogenesis. We demonstrate how the study of changes in DNA methylation covariation patterns across normal, intra-epithelial neoplasia and invasive cancer allows the identification of CpG sites that indicate the risk of neoplastic transformation in stages prior to neoplasia. Importantly, we show that the covariation in DNA methylation at these risk CpG loci is maximal immediately prior to the onset of cancer, supporting the view that high epigenetic diversity in normal cells increases the risk of cancer. Consistent with this, we observe that invasive cancers exhibit increased covariation in DNA methylation at the risk CpG sites relative to normal tissue, but lower levels relative to pre-cancerous lesions. We further show that the identified risk CpG sites undergo preferential DNA methylation changes in relation to human papilloma virus infection and age. Results are validated in independent data including prospectively collected samples prior to neoplastic transformation. Our data are consistent with a phase transition model of carcinogenesis, in which epigenetic diversity is maximal prior to the onset of cancer. The model and algorithm proposed here may allow, in future, network biomarkers predicting the risk of neoplastic transformation to be identified.

MOTIVATION: The standard paradigm in omic disciplines has been to identify biologically relevant biomarkers using statistics that reflect differences in mean levels of a molecular quantity such as mRNA expression or DNA methylation. Recently, however, it has been proposed that differential epigenetic variability may mark genes that contribute to the risk of complex genetic diseases like cancer and that identification of risk and early detection markers may therefore benefit from statistics based on differential variability. RESULTS: Using four genome-wide DNA methylation datasets totalling 311 epithelial samples and encompassing all stages of cervical carcinogenesis, we here formally demonstrate that differential variability, as a criterion for selecting DNA methylation features, can identify cancer risk markers more reliably than statistics based on differences in mean methylation. We show that differential variability selects features with heterogeneous outlier methylation profiles and that these play a key role in the early stages of carcinogenesis. Moreover, differentially variable features identified in precursor non-invasive lesions exhibit significantly increased enrichment for developmental genes compared with differentially methylated sites. Conversely, differential variability does not add predictive value in cancer studies profiling invasive tumours or whole-blood tissue. Finally, we incorporate the differential variability feature selection step into a novel adaptive index prediction algorithm called EVORA (epigenetic variable outliers for risk prediction analysis), and demonstrate that EVORA compares favourably to powerful prediction algorithms based on differential methylation statistics. CONCLUSIONS: Statistics based on differential variability improve the detection of cancer risk markers in the context of DNA methylation studies profiling epithelial preinvasive neoplasias. We present a novel algorithm (EVORA) which could be used for prediction and diagnosis of precursor epithelial cancer lesions. AVAILABILITY: R-scripts implementing EVORA are available from CRAN (www.r-project.org).

BACKGROUND: The 27k Illumina Infinium Methylation Beadchip is a popular high-throughput technology that allows the methylation state of over 27,000 CpGs to be assayed. While feature selection and classification methods have been comprehensively explored in the context of gene expression data, relatively little is known as to how best to perform feature selection or classification in the context of Illumina Infinium methylation data. Given the rising importance of epigenomics in cancer and other complex genetic diseases, and in view of the upcoming epigenome wide association studies, it is critical to identify the statistical methods that offer improved inference in this novel context. RESULTS: Using a total of 7 large Illumina Infinium 27k Methylation data sets, encompassing over 1,000 samples from a wide range of tissues, we here provide an evaluation of popular feature selection, dimensional reduction and classification methods on DNA methylation data. Specifically, we evaluate the effects of variance filtering, supervised principal components (SPCA) and the choice of DNA methylation quantification measure on downstream statistical inference. We show that for relatively large sample sizes feature selection using test statistics is similar for M and β-values, but that in the limit of small sample sizes, M-values allow more reliable identification of true positives. We also show that the effect of variance filtering on feature selection is study-specific and dependent on the phenotype of interest and tissue type profiled. Specifically, we find that variance filtering improves the detection of true positives in studies with large effect sizes, but that it may lead to worse performance in studies with smaller yet significant effect sizes. In contrast, supervised principal components improves the statistical power, especially in studies with small effect sizes. We also demonstrate that classification using the Elastic Net and Support Vector Machine (SVM) clearly outperforms competing methods like LASSO and SPCA. Finally, in unsupervised modelling of cancer diagnosis, we find that non-negative matrix factorisation (NMF) clearly outperforms principal components analysis. CONCLUSIONS: Our results highlight the importance of tailoring the feature selection and classification methodology to the sample size and biological context of the DNA methylation study. The Elastic Net emerges as a powerful classification algorithm for large-scale DNA methylation studies, while NMF does well in the unsupervised context. The insights presented here will be useful to any study embarking on large-scale DNA methylation profiling using Illumina Infinium beadarrays.

MOTIVATION: A common difficulty in large-scale microarray studies is the presence of confounding factors, which may significantly skew estimates of statistical significance, cause unreliable feature selection and high false negative rates. To deal with these difficulties, an algorithmic framework known as Surrogate Variable Analysis (SVA) was recently proposed. RESULTS: Based on the notion that data can be viewed as an interference pattern, reflecting the superposition of independent effects and random noise, we present a modified SVA, called Independent Surrogate Variable Analysis (ISVA), to identify features correlating with a phenotype of interest in the presence of potential confounding factors. Using simulated data, we show that ISVA performs well in identifying confounders as well as outperforming methods which do not adjust for confounding. Using four large-scale Illumina Infinium DNA methylation datasets subject to low signal to noise ratios and substantial confounding by beadchip effects and variable bisulfite conversion efficiency, we show that ISVA improves the identifiability of confounders and that this enables a framework for feature selection that is more robust to model misspecification and heterogeneous phenotypes. Finally, we demonstrate similar improvements of ISVA across four mRNA expression datasets. Thus, ISVA should be useful as a feature selection tool in studies that are subject to confounding. AVAILABILITY: An R-package isva is available from www.cran.r-project.org.