Characterization of the Illumina EPIC array for optimal applications in epigenetic research targeting diverse human populations

Abstract

The Illumina EPIC array is widely used for high-throughput profiling of DNA cytosine modifications in human samples, covering more than 850,000 modification sites across various genomic features. The application of this platform is expected to provide novel insights into the epigenetic contribution to human complex traits and diseases. Considering the diverse inter-population genetic and epigenetic variation, it will benefit the research community with a comprehensive characterization of this platform for its applicability to major global populations. Specifically, we mapped 866,836 CpG probes from the EPIC array to the human genome reference. We detected 91,034 CpG probes that did not align reliably to the human genome reference. In addition, 21,256 CpG probes were found to ambiguously map to multiple loci in the human genome, and 448 probes showing inaccurate genomic information from the original Illumina annotations. We further characterized those uniquely mapped CpG probes in terms of whether they contained common genetic variants, i.e., single nucleotide polymorphisms (SNPs), in major global populations, by utilizing the 1000 Genomes Project data. A list of optimal CpG probes on the EPIC array was generated for major global populations, with the aim of providing a resource to facilitate future studies of diverse human populations. In conclusion, our analysis indicated that studies of diverse human populations using the EPIC array would be benefited by taking into account of the technical features of this platform.

Introduction

DNA methylation (i.e., cytosine modification) has been implicated in various human complex traits and diseases [1]. The Illumina Infinium HumanMethylation assay, such as the Infinium 27K and 450K arrays, offers a cost-efficient, high-throughput platform of genome-wide CpG profiling for investigating complex traits and diseases [2,3,4], including for example detecting epigenetic variation between human populations, dissecting the genetic architecture of cytosine modifications, and detecting epigenetic contributors to human diseases [5]. However, it has become known that a set of the Infinium CpG probes did not perform as designed due to technical biases, such as probes ambiguously mapped to the human genome, probes containing common genetic variants in the form of single nucleotide polymorphisms (SNPs), and probes containing mismatched nucleotides [6,7,8,9,10,11,12].

Of particular interest to us is the potential technical bias caused by inter-population genetic variation, given that probes containing common genetic variants for a particular population may not generate reliable profiling data in that population [10]. Therefore, a thorough analysis of the CpG probes for these technical features will help determine which CpG probes may perform more reliably in a given population. We herein described a comprehensive analysis of the Illumina EPIC array, which is the current and widely used Infinium platform designed to interrogate more than 850,000 modification sites across the human genome, including for example sites within and outside of CpG islands, ENCODE (The Encyclopedia of DNA Elements) open chromatin, DNase hypersensitive sites, and > 90% of CpGs from the 450K array [10, 13]. Specifically, we sought to provide a technical characterization of the EPIC array with the aim of facilitating epigenomic research of complex traits and diseases targeting different populations.

Materials and methods

Retrieval of CpG probe sequences and annotations

Probe sequences and genomic annotations of the EPIC array were obtained at the support website maintained by Illumina, Inc. (http://www.illumina.com/). In total, sequences of 866,836 probes (50 bp/probe) were retrieved. Illumina provides basic information such as chromosomal location, genomic position (hg19), and strand. The original genic assignment was checked for accuracy by aligning the probe sequences to the human genome reference (hg19) as described below. The EPIC array employs two types of probes: Type I (2 probes per CpG locus, representing 1 “unmethylated” and 1 “methylated” query sequence) and Type II (one probe per CpG locus).

Obtaining genetic variant information from global populations

We downloaded variant calls in VCF format for 26 global populations from the 1000 Genomes Project [14,15,16,17]. Given that SNPs represent the majority of common genetic variants, we focused on SNPs in this work [14]. We calculated allele frequency for each bi-allelic SNP in each population. In addition to individual populations, we also obtained common SNPs for each of the five major global groups: African (7 populations), European (5 populations), East Asian (5 populations), South Asian (5 populations), and ad Admixed American (4 populations).

Detection of cross-hybridized probes, i.e., ambiguously mapped to the human genome

We aligned 866,836 probe sequences to hg19 following the method presented in a previous study [18]. Briefly, to represent all possible post-bisulfite conversion sequences, four single-stranded genomes (i.e., forward methylated, forward unmethylated, reverse methylated, and reverse unmethylated) were generated in silico. Both forward and reverse strands of hg19 were bisulfite converted in silico, in which all cytosines (C) of non-CpG dinucleotides were converted to thymines (T). For the unmethylated genome, all Cs of CpG sites were also converted to T’s, whereas in the methylated genome, all Cs of CpG sites remain as Cs (Fig. 1a). Type I probes were extracted from the annotated file provided by the Illumina. For type II probes, according to Illumina, some of them contain R nucleotides, representing either A or G due to the presence of CpG sites within the probe sequence. In order to represent all possible probe sequences, we replaced all R nucleotides in these probes with all possible combinations of A and G. In the end, we generated 3,505,864 probe sequences for all array probes. All these probe sequences were then aligned to the in silico converted reference using BLAT [19] with default parameters.

Fig. 1

Detection of CpG probes ambiguously mapped to the human genome. a Four single-stranded genomes (i.e., forward methylated, forward unmethylated, reverse methylated, and reverse unmethylated) were generated in silico bisulfite conversion (blue represents Cs of non-CpG; red represents Cs of CpG sites). b Matching criteria for detection of ambiguous probes

The matching criteria were shown in Fig. 1b. Briefly, matches with ≤ 2 mismatches were retained. Only matches with a match at the 50^th nucleotide of probe sequences were retained. Duplicate matches of the same probe that map to the same chromosomal location were also removed. Matches with gaps were also removed. Probes that satisfied the above criteria, while having multiple matches, were defined as ambiguous probes. The chromosomal locations of unambiguous probes were compared with Illumina’s information.

Enrichment analysis with cis-regulatory elements

We used DNase I hypersensitivity sites (DNase), transcription factor binding sites (TFBS), and annotations of histone modification peaks pooled across cell lines, downloaded at the ENCODE Analysis Hub at the European Bioinformatics Institute. For each regulatory element, we calculated the number of overlapping regions with the ambiguous probes (observed). To generate the null (expected) distribution of the number of overlaps between the regulatory elements and array probes, a random set of probes (same number as the ambiguous probes) from the array was randomly selected 10,000 times and overlapped with each regulatory element. The expected mean number of overlaps was derived from the distribution of 10,000 random sets. We then calculated the ratio of observed to mean expected as the enrichment fold and obtained an empirical p-value from the distribution of expected.

Detection of SNP-containing CpG probes by population

We examined the 754,546 unambiguous EPIC array probes for SNPs in each of the 26 populations from the 1000 Genomes Project [15]. We searched for common SNPs, i.e., MAF (minor allele frequency) >0.05 within 20 bp of the interrogated CpG sites in each population. The common SNP-containing probes were summarized by the location of the SNP within a probe. To provide an overview of major global populations, we also examined common SNP-containing probes in the five major global groups.

Results

Evaluation of the cross-hybridized probes in the EPIC array

Out of the total number of 866,836 probes on the EPIC array, in total 91,034 probes did not meet the criteria for ≤ 4 ambiguous bases (i.e., N) and ≤ 2 mismatches. We detected 21,256 probes that were not aligned to unique loci in the human genome reference. Figure 2 shows the summary of the ambiguous EPIC probes in terms of genomic distribution and enrichment of the ENCODE cis-regulatory elements according to Illumina’s annotated chromosomal locations. The ambiguous probes generally followed the distribution of all EPIC probes, while we observed a trend of enrichment with H3K9me3 (empirical p < 0.001). We further compared the mapping results with the original genomic annotations provided by Illumina. In total, 448 probes were mapped to a different gene from the original Illumina annotation file (Supplemental Table 1).

Fig. 2

Summary of the EPIC probes with Illumina annotations. a Ambiguous probes generally follow the distribution of all EPIC probes in terms of relative position to the gene. The black line represents all EPIC probes. The grey bar represents ambiguous probes. b Enrichment analysis indicates that ambiguous probes are more likely to co-localize with H3K9me3. DNase, DNase I hypersensitivity sites; TFBS, transcription factor binding sites; H2A.Z, histone H2A variant; H3, histone H3; H4, histone H4; K, lysine; me1, monomethylation; me2, demethylation; me3, trimethylation; ac, acetylation

Common SNP-containing probes by population

Table 1 shows the summary of common SNPs (MAF > 0.05) detected by population for those unambiguous probes on the EPIC array. Notably, for the East Asian populations there existed the least common SNP-containing probes, 54,810 at MAF >0.05, followed by European (68,155), South Asian (70,068), Admixed American (70,576), and African populations (96,674). SNP-containing probes by population are provided in Supplemental Table 2.

Table 1 Summary of common SNPs within +/− 20 bp of CpG probes by population

Conclusion

Epigenetic research of complex traits and diseases including epigenome-wide association studies (EWAS) rely on the robustness of profiling measurements from high through-put platforms. The EPIC array provides a powerful tool that covers almost 1 million modification sites across the human genome, as well as cost efficiency for population-based studies. However, there have been many instances of technical artifacts that can lead to erroneous results [10, 13, 18, 20, 21]. Cross-hybridization due to imperfect matches (e.g., mismatches/INDELs) and genetic variation (e.g., SNPs) is one of the most significant technical artifacts, where probes can map to multiple places in the genome and consequently assess a mixture of genuine and spurious signals [22]. Multiple studies have identified cross-hybridized probes in both the 450k and EPIC arrays, which affects a significant number of probes (6 to 11% of all probes). These findings contributed to the establishment of quality control practices for problematic probes on the EPIC array, resulting in the exclusion of a number of probes [10, 13, 18, 21]. Here, our technical characterization of the EPIC array showed that special care would be necessary when using this platform in epigenetic studies targeting diverse populations, because a substantial proportion of the interrogated CpG probes on the array contained common SNPs for major global populations. We therefore would like to provide those CpG probes on the array with potential technical biases (Supplemental Tables 1 and 2) as a useful resource for population-specific studies. For example, we can envision that investigators who identify any differential CpGs using the EPIC array will be benefited by referring this resource to make sure their findings are less biased to potential issues caused by cross-hybridization or common SNPs in a particular population, which will be critical for enhancing our knowledge of epigenetic contribution to complex traits and diseases.