A comparison of feature selection and classification methods in DNA methylation studies using the Illumina Infinium platform

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 beta-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.