A coherent approach for analysis of the Illumina HumanMethylation450 BeadChip improves data quality and performance in epigenome-wide association studies.

Lehne B, Drong AW, Loh M, Zhang W, Scott WR, Tan ST, Afzal U, Scott J, Jarvelin MR, Elliott P, McCarthy MI, Kooner JS, Chambers JC. · 2015

Read A coherent approach for analysis of the Illumina HumanMethylation450 BeadChip improves data quality and performance in epigenome-wide association studies. by Lehne B, Drong AW, Loh M, Zhang W, Scott WR, Tan ST, Afzal U, Scott J, Jarvelin MR, Elliott P, McCarthy MI, Kooner JS, Chambers JC. (2015) 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-pmc4365767.txt.

Introduction

Benjamin Lehne and a team of co-authors built a full workflow for a widely used DNA methylation chip, in Genome Biology in 2015. Methylation, a chemical mark on DNA, helps control the genome, and the Illumina 450K chip lets labs measure it genome-wide. Yet labs lacked one agreed way to clean and analyze the data. The team built their pipeline on 2,687 people, with 36 samples measured twice, and tested it against existing approaches.

The workflow adds new steps for quality checks, data scaling, and batch fixes that lean on the chip's own control probes. It also sets a fair null model for genome-wide scans using shuffled tests. The pipeline beat existing methods on data quality and on finding sites where gene variants move methylation levels. Cleaner data of this kind sharpens studies that link methylation to disease.

Catalog record

CREATORLehne B, Drong AW, Loh M, Zhang W, Scott WR, Tan ST, Afzal U, Scott J, Jarvelin MR, Elliott P, McCarthy MI, Kooner JS, Chambers JC.
YEAR2015
COLLECTIONPAPERS
FORMATTXT
FILE SIZE49 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC4365767

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A coherent approach for analysis of the Illumina HumanMethylation450 BeadChip improves data quality and performance in epigenome-wide association studies By Benjamin Lehne, Alexander W Drong, Marie Loh, Weihua Zhang, William R Scott, Sian-Tsung Tan, Uzma Afzal, James Scott, Marjo-Riitta Jarvelin, Paul Elliott, Mark I McCarthy, Jaspal S Kooner Published in: Genome Biology DOI: 10.1186/s13059-015-0600-x ABSTRACT Abstract DNA methylation plays a fundamental role in the regulation of the genome, but the optimal strategy for analysis of genome-wide DNA methylation data remains to be determined. We developed a comprehensive analysis pipeline for epigenome-wide association studies (EWAS) using the Illumina Infinium HumanMethylation450 BeadChip, based on 2,687 individuals, with 36 samples measured in duplicate. We propose new approaches to quality control, data normalisation and batch correction through control-probe adjustment and establish a null hypothesis for EWAS using permutation testing. Our analysis pipeline outperforms existing approaches, enabling accurate identification of methylation quantitative trait loci for hypothesis driven follow-up experiments. Electronic supplementary material The online version of this article (doi:10.1186/s13059-015-0600-x) contains supplementary material, which is available to authorized users. BACKGROUND DNA methylation is involved in the regulation of numerous biological processes, including gene expression [1], cell differentiation [2] and X-chromosome inactivation [3]. Altered DNA methylation has been linked to complex human diseases including cancer [4], schizophrenia [5], multiple sclerosis [6] and type 2 diabetes [7-9]. Recent technological developments, in particular the release of the Illumina Infinium HumanMethylation450 BeadChip (450 K methylation array), make it possible to measure DNA methylation on a genome-wide scale [10]. However, the 450 K methylation array includes multiple different probe types, each using different chemistry. Furthermore the methylation assay involves bisulphite conversion of DNA and other steps that introduce assay variability and batch effects. Multiple methods have been proposed for analysis of the complex data generated by the 450 K methylation array [11-17]; however, there is currently no consensus on the optimal analysis pipeline. We propose a comprehensive approach to the analysis of 450 K methylation array data. Our method was developed using data from over 2,600 samples from the London Life Sciences Prospective Population (LOLIPOP) study, including 36 samples measured in duplicate and identifies differential methylation on a single-marker level. Our pipeline, termed CPACOR (incorporating Control Probe Adjustment and reduction of global CORrelation), performs superiorly to published methods, and provides a blueprint for the analysis of large-scale Epigenome-Wide Association Studies (EWAS). RESULTS AND DISCUSSION Initial quantification and quality control We analysed two DNA methylation datasets: a population study of type 2 diabetes comprising 2,687 samples; and a technical replication dataset comprising 36 samples measured in duplicate (Materials and Methods). To maximise the impact of technical factors in the replication dataset, the initial and repeat sample analyses were carried out in separate batches. We performed an initial top-level quality control following analysis recommendations given by Illumina. We excluded 22 samples (sample call rate <98% or incorrect gender). The distributions for methylation values differ between autosomal and gender chromosome markers (Additional file 1: Figure S1); we therefore analyse these separately. Markers that are predicted to cross-hybridise [18], with a SNP in the probe-sequence, or that measure methylation at non-CpG sites were retained but flagged. Evaluating the detection P value threshold We initially used a detection P value of P <0.05 for marker calling based on Illumina recommendations. We noted though that calculated detection P values reported by minfi [15] range from 1 to 2.2 × 10−16, with values lower than 2.2 × 10−16 reported as zero (Additional file 1: Figure S2). To investigate the impact of detection P value threshold, we first evaluated call rates on the Y-chromosome among females in the population study; these are expected to be zero for all 416 markers. In contrast, we found that >50% of Y-chromosome markers had non-zero call rates in females (Figure 1), suggesting that the default detection P value (P <0.05) is not sufficient to prevent spurious results. When the detection P value threshold is lowered to P <10−16 the proportion of Y-chromosome markers with non-zero call rate in females is reduced from 55% to 10%. The majority of these remaining markers represent previously unidentified cross-hybridising probes (Additional file 1: Table S1). A more stringent detection threshold does not impact materially on Y-chromosome calling in males (Figure 1 and Additional file 1: Figure S3). [Figure: Marker call rates on the Y-chromosome. Distribution of call rates for 416 Y-chromosome markers in males (red points and red line) and females (green bars). Y-chromosome markers in females are represented in light green if their respective probes sequences are predicted to cross-hybridise with multiple genomic regions. Values greater than 80 are represented by numbers. (A) For a detection threshold P <0.05 more than 50% of Y-chromosome markers show non-zero call rates (call rate >0.05%) in females, even though females do not possess a Y chromosome. (B) For a detection threshold P <10−16 only 10% of Y-chromosome markers show non-zero call rates in females. Marker call rates in males (shown in red) are not materially affected by the more stringent detection P value threshold.] To extend these findings to autosomal markers, we quantified the proportion of extreme values (outliers) at each marker in the population study as a metric for quality of marker calling (Methods). Adoption of a more stringent detection P value threshold (P <10−16) reduces the proportion of outlying values, especially at markers with lower call rates, consistent with improved calling (Additional file 1: Figure S4). As a final test, we compared results for the 36 samples that were measured in duplicate. We observe a higher correlation (P = 2.91 × 10−11) between duplicate pairs when a detection P value threshold of P <10−16 is applied compared to a threshold P <0.05 (Additional file 1: Figure S5), providing further evidence for improved quantification of methylation with a more stringent detection P value threshold. This approach provides a roadmap for researchers to determine the detection P value threshold that is optimal for their dataset. Based on our results, we chose P <10−16 as detection P value threshold, providing a high accuracy at minimal loss of data. We recalculated sample call rates and excluded one further sample from the population study dataset with a call rate below 98% leaving 2,664 samples for further analysis. Data normalisation Data normalisation is frequently applied in the analysis of microarray data to reduce technical biases across measurements. To establish a consensus approach for normalisation of the 450 K methylation array we assessed the performance of 10 different normalisation methods [11,14,18-21] using the relationship between beta values for the 36 samples measured in duplicate (Additional file 1: Figure S6). The highest correlations between the paired measurements of methylation were achieved after quantile normalisation of intensity values for markers, subdivided by probe type, probe sub-type and colour channel (Additional file 1: Figure S7 and S8, Table S2). Functional normalisation (FN) [22], subset within array normalisation (SWAN) [20] and quantile normalisation of beta values performed significantly worse, while within-array approaches showed little or no improvement compared to non-normalised data. While correlation between technical replicates assesses Type-I statistical error, it may not assess over-normalisation. To quantify the ability to detect true signal after each normalisation method, we performed spike-in simulations based on the population study. Case–control status

Read the complete text