Meffil: efficient normalization and analysis of very large DNA methylation datasets.
Min JL, Hemani G, Davey Smith G, Relton C, Suderman M. · 2018
Read Meffil: efficient normalization and analysis of very large DNA methylation datasets. by Min JL, Hemani G, Davey Smith G, Relton C, Suderman M. (2018) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc6247925.txt.
Introduction
Min and co-authors present a software package for very large DNA methylation data sets. Their paper appeared in Bioinformatics. The package is called meffil and runs in R. Methylation data sets keep growing in size and genome coverage. Standard tools struggle with the memory load. Meffil was built to check quality, clean, and analyse data from Illumina chip arrays.
A rebuilt cleaning method cuts memory use without slowing the work. The method can also absorb group effects and batch effects in one step. Settings that once needed hand tuning are now estimated by the program. Data split across sites can be cleaned without sharing private records. That feature helps joint studies across many centres. The code is free online for other teams to use.
Catalog record
| CREATOR | Min JL, Hemani G, Davey Smith G, Relton C, Suderman M. |
|---|---|
| YEAR | 2018 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Meffil: efficient normalization and analysis of very large DNA methylation datasets By J L Min, G Hemani, G Davey Smith, C Relton, M Suderman, John Hancock Published in: Bioinformatics DOI: 10.1093/bioinformatics/bty476 ABSTRACT Abstract Motivation DNA methylation datasets are growing ever larger both in sample size and genome coverage. Novel computational solutions are required to efficiently handle these data. Results We have developed meffil, an R package designed for efficient quality control, normalization and epigenome-wide association studies of large samples of Illumina Methylation BeadChip microarrays. A complete re-implementation of functional normalization minimizes computational memory without increasing running time. Incorporating fixed and random effects within functional normalization, and automated estimation of functional normalization parameters reduces technical variation in DNA methylation levels, thus reducing false positive rates and improving power. Support for normalization of datasets distributed across physically different locations without needing to share biologically-based individual-level data means that meffil can be used to reduce heterogeneity in meta-analyses of epigenome-wide association studies. Availability and implementation https://github.com/perishky/meffil/ Supplementary information Supplementary data are available at Bioinformatics online. 1 INTRODUCTION DNA methylation is the addition of methyl groups to cytosine bases in the DNA sequence, most often in the context of a CpG dinucleotide, a cytosine followed by a guanine. The addition or loss of methyl groups is often associated with changes in gene expression, and through epigenome wide associations studies (EWAS) it has been shown to associate with a wide range of complex traits. A number of technologies have been developed for interrogating DNA methylation including microarrays and sequencing-based methods. The Illumina Infinium HumanMethylation450 BeadChip (450k array) can be used to measure DNA methylation of 485k CpG sites, comprising just under 2% of the total genomic CpG content mainly clustered around the transcription start sites (Michels et al., 2013). The new Illumina Infinium MethylationEPIC BeadChip (EPIC array) expands this coverage to ∼850k sites to include enhancer regions identified by ENCODE (Hong et al., 2016) and FANTOM5 (Andersson et al., 2014). Batch effects present a well-known challenge to microarray analysis (Teschendorff et al., 2011), particularly in datasets composed of thousands of samples since they cannot all possibly be processed at the same times and by the same technical personnel (Leek et al., 2010). This unwanted variation can increase both false negative and false positive rates if correlated with the outcome of interest, and controlling for this is not trivial, especially as sample sizes continue to grow. Following the popularity of quantile normalization for analyzing gene expression microarrays (Bolstad et al., 2003), many variations based on quantile normalization have been developed for DNA methylation microarrays (Lehne et al., 2015; Teschendorff et al., 2013; Touleimat and Tost, 2012), however, all assume that global methylation does not vary between samples (Hicks and Irizarry, 2015). When this does not hold, most notably between tumor and normal samples, between different tissue types, or when there are batch differences between cases and controls, quantile normalization can remove biological variation along with technical variation [e.g. (Fortin et al., 2014b; Heiss and Brenner, 2015)]. A feature of 450k and EPIC arrays is the inclusion of control probes–probes that do not assay biological variation and only vary due to technical effects. Functional normalization (FN; Fortin et al., 2014b) exploits control probes to separate biological variation from technical variation, and its performance compares favorably to other approaches (Fortin et al., 2014b; Heiss and Brenner, 2015; Lehne et al., 2015; Liu and Siegmund, 2016; Maksimovic et al., 2015). Many DNA methylation datasets using 450k and EPIC arrays have now been generated independently and are being used in EWAS to discover associations between CpG sites and a variety of exposures, complex traits and disease risks. Much like genome-wide association studies (GWAS), the widespread use of the Illumina platform has made possible large-scale meta-analyses organized by international consortium-based efforts that are combining ever larger numbers of subjects to reach statistical power to detect robust associations. However, unlike GWAS, it comes with a number of crucial challenges that have not been fully resolved. First, most existing software tools for quality control and normalizing DNA methylation levels were not designed to handle datasets comprising thousands of samples, and recently developed tools in R (Gorrie-Stone, 2018) and Java (Almeida et al., 2017) may not provide desired functionality. Second, sharing of individual-level data is prohibited due to ethical considerations, so meta-analyses are liable to encounter heterogeneity introduced when datasets are normalized independently. Third, there is currently no universally accepted standard approach that addresses all aspects of dataset quality control and normalization of large datasets. Fourth, there is no standardized approach for selecting and comparing statistical models for EWAS, including selection and generation of covariates. Fifth, there is currently no standard for sharing quality control information and EWAS findings. We have developed meffil (Efficient algorithms for analyzing DNA methylation data) to provide solutions in a user-friendly and open source R package (https://github.com/perishky/meffil). Figure 1 shows the meffil work-flow from raw data to quality control to normalized data to EWAS. Meffil includes functionality for identifying low quality methylation measurements, discovering and rectifying sample mismatches, merging datasets containing both 450k and EPIC arrays, removing confounding effects of cell type heterogeneity and assessing the quality of observed associations. In this paper, we describe its implementation and evaluate the computational and statistical advantages that it achieves, while demonstrating where limitations might still exist. [Figure: The workflow of meffil] 2 MATERIALS AND METHODS 2.1 Data Full details of the Accessible Resource for Integrated Epigenomic Studies (ARIES; Relton et al., 2015) and Genetics of Overweight Young Adults (GOYA; Paternoster et al., 2011) datasets are provided in the Supplementary Material. DNA methylation was quantified in bisulfite-converted genomic DNA using 450k arrays for all samples. Some samples were removed due to genotype and gender mismatches and methylation quality (low detection scores, low number of beads, methylated/unmethylated ratio, strong dye-bias, post-normalization checks). Samples were normalized using FN using meffil. After normalization we checked for batch effects including bisulfite-conversion plate (‘plate’) and beadchip (‘slide’). 2.2 Implementation of meffil Meffil is designed around a re-implementation of FN as implemented in the minfi R package (Fortin et al., 2014b). Output using default settings and without enhancements is therefore identical to minfi (Aryee et al., 2014). Meffil uses the illuminaio R package (Smith et al., 2013) to parse Illumina IDAT files into QC objects (Fig. 1) which contain raw control probe summaries, quantile distributions of raw probe intensities, poor quality probes based on detection P-values and number of beads, predicted sex (Aryee et al., 2014), predicted cellular composition (Houseman et al., 2012) when a cell type reference is specified and batch variable values. As in minfi, probe intensities are dye-bias and background corrected using the ‘noob’ method (Triche et al., 2013). Control probes are summarized as 42 different control types in a control matrix with one row for each control type and one column for each sample. This summary object is all that is needed to perform quality control, sample and CpG site filtering, identification of batch effects and the normalization of sample quantiles, the first normalization step of FN. In this step, probe intensity quantiles are normalized between samples by fitting linear models with these quantiles to the top principal components of the control matrix. The