replying to G.L. Meeks et al. Communications Biology https://doi.org/10.1038/s42003-023-05658-5 (2023).

In the paper by Li et al.1, we reported DNA methylation (DNAm) of 946 CpG sites influenced by genetic variants in the conventional model (LA-naïve), and 135 CpGs with genetic effects that significantly differed by local ancestry (LA-specific). Meeks et al. raised a concern that 37.5% of the 946 CpG sites were inadequately controlled as they contained nearby commons variants (Probe-SNPs). They argued that different allele frequencies of probe-SNP between populations of African ancestry and European ancestry led to spurious findings in the paper. In response to Meeks et al.’s concern, using a Methyl-Seq approach in a subset of the samples (N = 211), we were able to show that a large proportion (47.1%-71.6%) of the significant SNP-CpG associations originally reported by Li et al. using Illumina HumanMethylation 450K (HM450K)1 were replicated. We did not observe significant difference in the replication rates between CpGs with and without probe-SNPs, indicating that there is no substantial evidence to suggest that CpGs with probe-SNPs biased our published findings.

Associations of the reported SNP-CpG pairs are replicated using a bisulfite sequencing method

It is well known that the methylation probe containing SNP within 10 base pair (bp) biases the call of CpG methylation in an array-based assay2,3. In practice, CpG sites with SNP within 10bp are filtered out in analyses4 although a few studies removed such CpG sites beyond 10bp window5. Other studies examined the effects of probe-SNP on significant CpG sites in post-hoc analyses6. In our paper1, we removed the polymorphic CpG sites (the ones that overlay with SNPs) and CpG sites with probe-SNP within 10bp based on the annotation file provided by Illumina Infinium. Additionally, following the previous report7, we removed CpG sites with detection p-value > 1e−12, a more stringent threshold than recommended by Illumina (p = 0.01). The use of a stringent detection p-value could more effectively filter out low quality CpG sites and enhance the quality of DNAm array data8.

Methyl-seq serves as a gold standard to validate array-based methylation detection9,10. To confirm the quality of DNA methylation in our previous study, we re-profiled DNA methylation of 211 samples (Supplementary Table 1) that were included in the previous paper1 using Agilent SureSelectXT Methyl-Seq (Supplementary Data 1). No significant differences in demographic variables were observed between the 211 samples with Methyl-seq data and the original discovery group except for smoking (p = 0.04) (Supplementary Table 1). A total of 547 out of the 946 CpG sites in the conventional model, and 77 out of 135 CpGs in the LA-specific model were measured by both platforms. The 547 CpGs in the conventional model had 728 significant SNP-CpG associations, while the 77 CpGs in the LA-specific model had 87 significant associations. To replicate the original findings identified using HM450K data, we first re-conducted the association analyses for the significant SNP-CpG pairs using Methyl-seq data, then we investigated whether the replication rates would differ between CpGs with and without probe-SNPs. The overall replication rates were 71.6% for the 728 significant SNP-CpG pairs in the conventional model (Supplementary Data 2), and 47.1% for the 87 significant pairs in the LA-specific model (Supplementary Data 2). A similar trend of replication rates was observed in the original results of Li et al.1: the replication rate in the LA-specific model was consistently lower than that in the conventional model. Importantly, we found no significant difference in replication rates between CpGs with and without probe-SNPs (p = 0.15 for CpGs in the LA-naïve model; p = 1.00 for CpGs in the LA-specific model) (Fig. 1a).

Fig. 1: Comparing the replication rate between CpGs with probe-SNPs and CpGs without probe-SNPs.
figure 1

The Methyl-seq data (N = 211) were used to replicate the meQTLs identified by Li et al. CpG sites with common probe-SNP within 50bp were defined using the SNPs in a the 1000 Genomes and b the study sample population, respectively. The p-values were derived from χ2 test.

Full size image

Meeks et al. define whether a CpG has probe SNPs within 50bp using the 1000 genomes data. Here, we re-examined if our reported CpGs harbored nearby SNPs using the genotype data from the study cohort (4.7 million SNPs with minor allele frequency >0.01)1: among the 946 CpGs identified in the conventional model (LA-naïve), 28 (3.0%) had probe SNPs within 10bp and 157 (16.6%) had probe SNPs within 50bp. Among the 135 CpGs with genetic effects significantly differed by ancestry (LA-specific), only 3 of them (2.2%) had probe SNPs within 10bp and 21 of them (15.6%) had probe-SNPs within 50bp. When evaluating whether there is a significant difference in replication rates between CpGs with and without probe SNPs, we also presented parallel results using the studied population SNP-list as reference and found no significant difference (Fig. 1b). Together, these data suggest that there is no clear evidence for the probe-SNP bias in the results of Li et al.1 as the replication rate in CpGs with probe-SNPs was not significantly different from that in CpGs without probe-SNPs.

The 1000 Genomes versus study population based genomes

Meeks et al. mapped the nearby SNPs for our reported CpG sites using the 1000 Genomes SNP-list and found a large proportion (37–61%) of them contained SNPs within 50bp. However, using the genotype data from our study cohort, we noted that the proportion of CpGs with probe SNPs was lower (2.2–3.0% containing SNPs within 10bp, 15.5–16.5% containing SNPs within 50bp) (Supplementary Fig 1). One reason for the discrepancy was the difference in size between the studied population SNP-list and the 1000 Genomes SNP-list, and we agreed with Meeks et al. that our strict genotype quality control steps1 led to the smaller SNP-list (4.7 million SNPs with minor allele frequency >0.01 in our studied cohort). This strict procedure is appropriate to keep high quality SNPs for meQTL identification, but to avoid probe-SNP bias, a more comprehensive SNP-list, such as the 1000 Genomes, is also a helpful reference to filter CpGs with probe-SNPs. Therefore, to assess whether our previous results were affected by potential probe-SNPs defined by 1000 Genomes, we performed parallel replication analyses using both the studied population SNP-list and the 1000 Genomes SNP-list as references, and the results were consistent: there was no significant difference in replication rates between CpGs with and without probe-SNPs (Fig. 1).

In summary, the concerns highlighted by Meeks et al. underscored the importance of filtering CpGs with probe-SNPs in methylation association and meQTL studies in a mixed ancestry population. Using Methyl-Seq data, we confirmed that replication rates of the significant SNP-CpG associations did not differ significantly between CpGs with and without probe-SNPs. Applying bisulfite methylation sequencing not only prevents the biased methylation detection influenced by nearby SNPs in array-based assay, but can also enable the evaluation of additional allelic/haplotypic genetic-epigenetic effects that array-based methods are blind to11,12. Furthermore, polymorphic repeats between human populations may require specialized long-read techniques13. Altogether, future studies may consider applying Methyl-seq based methods instead of relying on array-based assay.

Methods

Quality control (QC) on the Methyl-seq data was conducted following standard procedure14. Quality of sequence data was examined by using FastQC (ver. 0.11.8). We used Bismark pipelines (ver. v0.22.1_dev)15 to align the reads to the bisulfite human genome (hg19) with default parameters. Quality-trimmed paired-end reads were transformed into a bisulfite converted forward strand version (C → T conversion) or into a bisulfite-treated reverse strand (G → A conversion of the forward strand). Duplicated reads were removed from the Bismark mapping output by deduplicate_bismark. All CpG sites were grouped by sequencing coverage, also known as read depth. Only the CpG sites with coverage > 10x depth were kept to ensure the data quality. The study was approved by the committee of the Human Research Subject Protection at Yale University and the Institutional Research Board Committee of the Connecticut Veteran Healthcare System. Informed consent was obtained from all human participants. All analyses were carried out in accordance with all relevant ethical regulations.