Placental methylome analysis from a prospective autism study

The placental tissues were obtained as previously described [19] from the MARBLES (Markers of Autism Risk in Babies: Learning Early Signs) study, a prospective study of the environmental, genetic, and epigenetic factors leading to autism spectrum disorder (ASD) (Hertz-Picciotto, in revision). Pregnant women and women planning a pregnancy were enrolled if they or the father already had a child with ASD; they were thus at significantly (13-fold) higher risk of having another child with ASD, as compared with parents not having a previous child with ASD [4]. MARBLES recruited Northern Californian families from lists of children receiving services for ASD funded through the California Department of Developmental Services. Proband ASD status was confirmed, and the mothers were seen at regular intervals during pregnancy with biosample collection starting at enrollment and repeated at each visit. The placenta, cord blood, and other samples were collected and frozen at birth, and the children were followed to age 36 months when a final developmental diagnosis of autism or typical development was made by ADI-R in addition to the other assessments. DNA was isolated for MethylC-seq analysis from MARBLES placentas from births of 24 children who received a final 36-month clinical diagnosis of ASD by September 29, 2014 and 23 typically developing children matched to the children with ASD by gender and birth year (within 1.5 years), with preference given to those with greater availability of other study data. Due to the naturally occurring high proportion of males to females with autism, seen also in the MARBLES study, only two in each of the autism and typical placentas came from females (Additional file 1: Table S1). For non-MARBLES population control samples, full-term human placental samples were obtained from routine Cesarean sections, as described previously [16]. All participants gave written informed consent for data and sample use, and protocols were approved by UC Davis IRB (protocol # 225645-35).

Children’s development was assessed by trained and reliable examiners with final diagnostic assessments at 36 months. All children are assessed for autism symptoms using the gold standard Autism Diagnostic Observation Schedule-Generic (ADOS-G) [20, 21]. A clinical best estimate diagnosis was given through the consensus of two clinicians based on DSM-IV or DSM-5 criteria. Placental samples were categorized based on whether the child met ADOS and DSM criteria for ASD, showed typical development, (TD), or had impairments in some domains, but did not meet the full ASD diagnostic criteria (ODC for Other Developmental Concerns).

The placental samples were frozen immediately after birth. DNA was extracted using Qiagen’s Puregene kit, sonicated to ~300 bp, and methylated Illumina adapters were ligated to the ends using NEB’s NEBNext DNA library prep kit. The library was bisulfite-converted using Zymo’s EZ DNA Methylation-Lightning Kit, amplified for 14 cycles using PfuTurbo Cx, purified with Agencourt AMPure XP beads, and sequenced on an Illumina HiSeq 2000. Reads were mapped to the hg19 version of the human genome using BS Seeker [22]. To eliminate clonal PCR amplification duplicates, only one read out of those with identical genomic positions was kept; Genome and CpG coverage was estimated by multiplying the number and the length of the mapped reads and dividing by the size of the human genome (Additional file 1: Table S1). CpG site methylation data were combined from both DNA strands.

Methylation data from 17 typical placentas from MARBLES were combined to create a single, high-coverage map of methylation across the genome. Visually annotated PMD and HMD portions of this consensus genome were used to train a two-state hidden Markov model (HMM) to differentiate PMDs and HMDs using an HMM called hidden Markov models of methylation (StochHMM) [23], as previously described [16]. The model was then applied to the same high-coverage methylation data to define the boundaries of PMD/HMDs in the typical placentas. Those boundary chromosome coordinates were used for calculating average percent methylation in both typical and autism placental samples in the MARBLES study. For each sample, the average % methylation over all PMDs and all HMDs was calculated. In addition, % methylation for each PMD and HMD for each individual was calculated, and differences between ASD and TD samples were assessed for significance using two-tailed t tests and false discovery rate (FDR) multiple hypothesis correction (0.05).

Since the majority of fetal samples were male that only contains a single active X chromosome, we used DNA methylation values from specific regions of the X chromosome that are specifically methylated in females due to X chromosome inactivation on the second X chromosome. All CpG islands within HMDs on the X chromosome were selected, and the mean percent methylation was determined. Since a sample containing no female cells would theoretically be expected to have lower methylation over these regions, the percentage of female cells in each sample was estimated and compared between samples for potential differences.

Samples were taken from six different locations in ten full-term control placentas (non-MARBLES samples and primers described previously [16]. Genomic DNA was isolated and bisulfite-converted as above, PCR-amplified for 45 cycles using HotStarTaq polymerase (PyroMark PCR Kit, Qiagen), and run on a PyroMark Q24. Primer sequences for the DLL1 locus were from hg19 chr6:170,534,692-170,534,733 spanning three CpG sites:

Information on demographic factors were collected through telephone-assisted interviews. The following variables were analyzed for associations with PMD total average methylation, HMD total average methylation, and the percent of the 20 kb windows with methylation below 60%: sequencing run, order, and coverage; child race (white non-Hispanic [reference], Asian, multi-racial, white Hispanic, non-white Hispanic); and child sex (male [reference]/female).

We performed univariate linear regression using the SAS software version 9.4 for each variable in relation to PMD methylation, PMD methylation adjusted for HMD methylation, HMD methylation, HMD methylation adjusted for PMD methylation, and the percent of the 20 kb windows with methylation below 60% (using a 5% change as the unit) (Additional file 1: Table S6). To facilitate interpretation of regression coefficients, PMD and HMD methylation was expressed in 5 percentage point and 2 percentage point units, respectively, so that coefficients (and effect size estimates based on them) correspond to an approximately 2 SD change in the independent variable. To account for the multiplicity of hypotheses being assessed, we controlled the FDR at 5% [24]. In ad hoc analyses, we examined demographic factors in relation to DLL1 methylation (unit was 5% change). We used the Akaike information criterion (AIC) to select the most parsimonious among the candidate models [25, 26]. To produce a final model, we used sandwich estimators to produce homoscadisticity-robust estimates of the 95% confidence interval given that the conditional variance of outcomes appeared to vary with the regression-based predicted values.

MethylC-seq analysis of 24 ASD and 23 TD placental samples showed that global genome-wide methylation patterns were very similar between individuals, with methylation variability between samples being larger within the PMDs than HMDs (Fig. 1a, b, Additional file 1: Table S1). In particular, there seemed to be a little deviation in the locations of PMD/HMDs and their boundaries. When graphing average methylation in 20 kb windows across the genome, PMD/HMDs show up as a bimodal distribution in the human placenta [16]. Similar bimodal curves were seen in all typical and autism MARBLES placentas (Fig. 1c, d, Additional file 2: Figure S1). Greater inter-individual variability was observed in PMDs as a higher standard deviation over the peaks with lower methylation (Fig. 1d).

Bild vergrößern

Since there is some difference in methylation between individuals, particularly in PMDs, we next asked if higher versus lower average methylation for a given individual (relative to the group) is consistent across chromosomes or different between chromosomal loci. In Fig. 1a, b, individual samples appeared follow a similar pattern across each chromosome, and that pattern was also observed for each sample graphed by average methylation for each chromosome for each individual (Fig. 1e). Individuals with higher or lower methylation in HMDs on one chromosome also tended to have similar higher or lower methylation in HMDs on the other chromosomes. Genome-wide methylation levels in HMDs varied by as much as 4.76% between individuals. For PMDs, the same trend of consistently higher or lower methylation across chromosomes was still observed, but there was a more random mixing of relative methylation levels in PMDs compared to HMDs. Genome-wide methylation levels in PMDs varied by as much as 11.17% between individuals. These genomic individual differences in methylation levels between individuals were observed in both autism and typical placental samples, with negligible differences between diagnostic groups.

Note that in Fig. 1e, a typical placenta was given the same color in both the PMD and HMD graphs (likewise for the autism placentas). Since individuals with relatively high or low methylation in HMDs did not necessarily have the same relatively high or low methylation in PMDs, we next asked how well genome-wide average methylation levels in PMDs and HMDs correlated. For both typical placentas and autism placentas, there are significant positive correlations between methylation levels in PMDs and HMDs, with a combined correlation of 0.64 (Fig. 1f).

Although PMD/HMD patterns were consistent between individual placentas, we next asked how much of the variability in the relative level of methylation between individuals could be due to tissue sampling heterogeneity. More specifically, are methylation levels or patterns different on the maternal versus fetal sides or between the different regions within the placenta? To answer this question, we sampled the outer fetal, inner fetal, inner maternal, and middle maternal portions of a single normal placenta that was not part of the MARBLES study. Figure 2a–d shows that the variability in relative methylation levels between regions within a single placenta are similar to the inter-individual variability seen in typical MARBLES placentas. Specifically, the ~10% range of methylation values for PMD intra-tissue variability shown in Fig. 2d is no greater than the PMD inter-individual variability shown in Fig. 1e.

Bild vergrößern

To examine whether placental sampling location was predictive of methylation level, pyrosequencing analysis was performed over four genomic loci, representing 2 HMD and 2 PMD loci described previously [16]. Samples were from ten normal placentas that were not part of the MARBLES study, and each placenta was sampled from one to five of the six placental regions (Fig. 2e). Samples from the same placental region but different placentas did not co-cluster but appeared fairly randomly distributed relative to the other samples. At all four loci, two-way ANOVA tests showed neither statistically significant effects of individual placenta nor placental location on percent methylation over these loci.

We next asked whether the inter-individual variability in methylation observed in PMDs could be attributed to differences in levels of maternal blood contamination. Because global methylation levels in the blood are typically high, similar to other adult somatic tissues, the amount of blood in the placental sample would proportionally raise the amount of observed methylation in PMDs. To address this in the MARBLES placentas, we utilized the epigenetic feature that females have higher methylation in CpG island promoters on the inactive X chromosome [27]. We reasoned that if maternal cells were contributing to placental methylation levels, higher levels of maternal cells in the placentas of male offspring would correspond to increased methylation in X chromosome CpG island promoters. Figure 2f, however, shows that there is no correlation between autosomal PMD methylation and X chromosome CpG island promoter methylation. However, four of the MARBLES samples were from female offspring (marked with a purple box), and increased methylation is observed in their X chromosome CpG island promoters, as would be expected from a female genome.

In order to determine if DNA methylation in placenta could be used as a biomarker for early autism detection, we first performed linear regression tests on both global average HMD methylation and global average PMD methylation, looking at the effects not only of child diagnosis (autism or typical) but also sex, the sequencing run number, the order of sequencing (since Illumina sequencing chemistries changed during the course of the study), and the average sequencing coverage. None of the factors, including diagnosis, had statistically significant effects on either average HMD or PMD methylation (data in Additional file 1: Table S2). In addition, no significant difference in percent methylation between ASD and TD samples was observed over any chromosome (data in Additional file 1: Table S3). Average methylation in individual promoters, CpG islands, gene bodies, and non-overlapping 20 kb windows were also tested as above, but no statistically significant differences between ASD and TD samples were found after FDR correction.

To determine if there were methylation differences within chromatin states, hidden Markov models of chromatin states (ChromHMM) mappings for the human placenta were investigated [28]. Figure 3 shows the average methylation, across autism/typical placental samples, for each chromatin state. This shows, for example, not only the average methylation for heterochromatin but also the range of methylation found for individual heterochromatin elements. Within the human placenta, both active and poised transcription start sites (TSSs) show low methylation; however, active TSSs show the lowest methylation levels. Transcribed regions show the highest methylation and heterochromatin, Polycomb-regulated, quiescent, and many enhancer regions show partial methylation. In order to determine if ASD and typical placentas differed in their global levels of methylation in any of the chromatin states, average genome-wide methylation for each chromatin state was calculated for each MARBLES placental sample. Two-tailed t tests showed no statistically significant differences between ASD and TD samples for any of the ChromHMM states after FDR correction (Additional file 1: Table S4).

Bild vergrößern

We next tested average methylation in individual PMDs and HMDs using two-tailed t tests and FDR multiple hypothesis correction (0.05). No PMDs showed statistically significant differences between autism and typical placental sample; however, one HMD showed significance with a p value of 0.0267 (Additional file 2: Figure S2). It is located on chromosome 6, 45 kb downstream from the DLL1 gene and 16 kb downstream of the LOC15444 non-coding RNA. This differentially methylated HMD includes a region with the enhancer marks H3K4me1 and H3K27ac in the fetal brain (Fig. 4a). Evidence for three-dimensional interaction between this putative enhancer and the DLL1 promoter is provided by visualization of this locus in published HiC data (Fig. 4b) [29]. A pyrosequencing analysis of bisulfite-converted DNA using primers designed to the enhancer region validated significantly increased DNA methylation in ASD compared to TD samples (Fig. 4c). In contrast, samples from the “other developmental concerns” (ODC) group in MARBLES with behavioral features that did not fall into either TD or ASD categories showed intermediate methylation levels at this DLL1 locus (Fig. 4d, Additional file 1: Table S5). Similar to the other specific loci shown in Fig. 2e, methylation analysis of the DLL1 locus by pyrosequencing did not reveal detectable differences based on location of the placental sample (Fig. 4e).

Bild vergrößern

Sequencing run, order, and coverage and child race/ethnicity and sex were evaluated as predictors of average PMD and HMD values, as well as the percentage of 20 kb windows less than 60% methylation, in univariate linear regression analysis. None of these factors was a significant predictor of methylation though child race/ethnicity showed a trend toward an association with PMD methylation prior to FDR correction (Additional file 1: Table S6). Sequencing and demographic variables were not independently significantly associated with DLL1 locus methylation, and the most parsimonious model selected by AIC included only ASD status, which was associated with higher average DLL1 methylation (estimate = 0.24, 95% CI −0.04, 0.51, p = 0.09).