Comparative analysis of MBD-seq and MeDIP-seq and estimation of gene expression changes in a rodent model of schizophrenia

Abstract

We conducted a comparative study of multiplexed affinity enrichment sequence methodologies (MBD-seq and MeDIP-seq) in a rodent model of schizophrenia, induced by in utero methylazoxymethanol acetate (MAM) exposure. We also examined related gene expression changes using a pooled sample approach. MBD-seq and MeDIP-seq identified 769 and 1,771 differentially methylated regions (DMRs) between F2 offspring of MAM-exposed rats and saline control rats, respectively. The assays showed good concordance, with ~56% of MBD-seq-detected DMRs being identified by or proximal to MeDIP-seq DMRs. There was no significant overlap between DMRs and differentially expressed genes, suggesting that DNA methylation regulatory effects may act upon more distal genes, or are too subtle to detect using our approach. Methylation and gene expression gene ontology enrichment analyses identified biological processes important to schizophrenia pathophysiology, including neuron differentiation, prepulse inhibition, amphetamine response, and glutamatergic synaptic transmission regulation, reinforcing the utility of the MAM rodent model for schizophrenia research.

1. Introduction

To date, DNA methylation has been the most examined and best characterized epigenetic mechanism. In mammals, DNA methylation predominantly involves the addition of methyl groups to cytosine residues of cytosine-guanine dinucleotides (CpG sites) via DNA nucleotide methyltransferase enzymes. Regions within the genome displaying increased CpG density are referred to as CpG islands (CGI), and DNA methylation within these (often located at gene promoters) typically leads to transcriptional repression.¹ Patterns of DNA methylation also spatially correlate with patterns of histone methylation, supporting the hypothesis that CpG methylation is involved with regulating chromatin structure and consequent accessibility to transcriptional machinery.^{2, 3} Alterations in DNA methylation have been linked to a variety of diseases, including schizophrenia,⁴ and evidence indicates a direct link between environmental exposure, particularly during the prenatal period, and DNA methylation changes.⁵

Because CpG methylation is fundamental to survival, the development of robust detection methods has been an active research priority in recent decades. Most genome-wide approaches involve microarray or high-throughput sequencing, but vary by how input DNA is treated – the primary approaches being bisulfite conversion-based, affinity capture-based, or restriction endonuclease-based. The gold-standard approach to assessing DNA methylation is bisulfite conversion sequencing (BS-seq), however, sequencing to sufficient coverage (at least 30x) is prohibitively expensive and a large portion of sequence reads are non-informative of methylation.^{2, 6, 7} Reduced representation bisulfite sequencing limits the scale and associated costs of sequencing by treating DNA with a restriction enzyme to minimize input of DNA with low CpG densities² but still requires special bioinformatics considerations and tools. MeDIP- and MBD-seq offer an economical alternative, based on affinity enrichment protocols. MeDIP-seq (methylated DNA immunoprecipitation sequencing) employs a 5-methylcytosine monoclonal antibody to recover methylated fragments from single-stranded DNA,^{8, 9} whilst MBD-seq captures double-stranded methylated DNA fragments using the methyl-binding domain from MECP2 (Supplementary Figure 1).¹⁰ Literature indicates that both protocols are effective, and their results are generally concordant but non-identical.^11–15 Both approaches risk poor coverage of the medium to low CpG density regions of the genome, and neither allows for single-base CpG resolution.^{11, 12} Both detect 5-methylcytosine exclusively, unlike bisulfite conversion, which does not differentiate between 5-methylcytosine and 5-hydroxymethylcytosine.¹¹ Finally, the costs are similar for the two methodologies and we and others have developed protocols that allow for sample multiplexing.^{11, 16} Previous reports favor MBD-seq as somewhat more sensitive (with sufficient coverage), however MeDIP demonstrates no bias for a specific nucleotide sequence and is constrained only by the methyl-CpG density within DNA fragments so it may provide better genome-wide coverage.^{15, 17, 18} An additional drawback to MBD-seq is the lack of stable binding between MBD domain of MECP2 and hemimethylated DNA.¹⁹ An important consideration with both methods is the coverage being non-inclusive of unmethylated CpGs, for which a lack of methylation can only be inferred by a lack of reads. Confidence to make such an inference requires sufficiently saturated read depth. Further, because CGIs are predominantly unmethylated, coverage at these areas of the genome may be limited. Both MeDIP and MBD protocols enrich primarily for methylated low-CpG genomic regions, plus the small subset of (methylated) CpG-dense CGIs.^{20, 21} Reports conflict regarding which protocol better enriches for high-density CpG regions, with the primary consideration being the concentration of salt elution chosen for MBD-seq protocol.^{20, 22} Other established and continuously developing sequencing-based methods are also available to query cytosine methylation,^{12, 23–25} each with incumbent strengths and weaknesses, but here we focus on a direct comparison between the results of MeDIP-seq and MBD-seq assays.

To investigate this, we made use of a rat schizophrenia model. Since schizophrenia is a neurodevelopmental disorder, we employed a developmental disruption model of the disease. Specifically, the administration of methylazoxymethanol acetate (MAM) on gestational day 17 has been demonstrated to produce robust and reproducible histological, neurophysiological and behavioral deficits consistent with those observed in schizophrenia patients (reviewed in²⁶).^{27, 28} More recently, we showed that a subset of F2 and F3 offspring also display a schizophrenia-like phenotype, and that widespread hippocampal DNA methylation changes were associated with in utero administration in both F1 and F2 progeny.¹⁶ An increasing body of evidence suggests that the ventral hippocampus may represent a key site of pathology in schizophrenia. Indeed, hippocampal pathology is routinely observed in schizophrenia patients and is correlated with both psychosis²⁹ and cognitive dysfunction.³⁰ This has been further examined in preclinical models where the hippocampus has been advanced as a novel site of intervention for schizophrenia.^{31, 32} Here, we query genomic methylation in ventral hippocampal tissue of F2 offspring of methylazoxymethanol acetate (MAM) exposed rats,^{28, 33, 34} via MeDIP-seq and MBD-seq, to extend analyses of our previous study and compare these two methodologies, as well as examine potential effects on gene expression.

2. Materials and Methods

2.1 Animals

The experiment was performed in accordance with the guidelines outlined in the USPH Guide for the Care and Use of Laboratory Animals and was approved by the Institutional Animal Care and Use Committee of the University of Texas Health Science Center at San Antonio. Methyl DNA immunoprecipitation, methyl binding domain capture, and RNA-sequencing experiments used DNA or RNA extracted from tissue from the ventral hippocampus, collected from F2 generation Sprague-Dawley rats. These rats were second generation offspring of animals previously treated with methylazoxymethanol acetate in utero, to induce a schizophrenia-like phenotype in the offspring. As previously described,^{33, 34} MAM exposure was initiated on timed pregnant female Sprague–Dawley rats (Harlan Laboratories), housed individually in plastic tubs. MAM (diluted in saline, 25 mg/kg, intraperitoneal) was administered on gestational day 17, whereas control rats received injections of saline (1 ml/kg, intraperitoneal). Male pups were weaned on postnatal day 21 and were housed with littermates in groups of two to three until adulthood (>PD 60). F2 generation rats were obtained by crossing saline (♂) × saline (♀) and MAM (♂) × MAM (♀). Adult male F2 rats that were MAM-exposed or saline-controls (n=4 per group) were anesthetized with sodium pentobarbital (60 mg/kg, intraperitoneal) and rapidly decapitated. The hippocampus is a discrete brain structure that is readily separated from the adjacent tissue. Here we used a brain matrix to block the brain at the anterior and posterior limits of the hippocampus. The entire hippocampus was manually dissected and divided into dorsal and ventral regions, then snap frozen and stored at −80°C prior to DNA and RNA extraction.

2.2 Sample Preparations

2.2.1 DNA Extraction and Shearing

DNA from each of the eight ventral hippocampal tissue samples (~25mg each) was extracted using the DNeasy Blood and Tissue Kit (Qiagen, cat# 69504) according to the manufacturer’s protocol with modifications (addition of 60 μl proteinase K and incubation for 4 h; elution with 2 × 75 μl AE buffer). For each sample, 1.6 μg of DNA, normalized to 20 ng/μl, was subjected to shearing by the Covaris S220 Ultra Sonicator, using the following conditions: duty cycle of 10%, an intensity of 5200 cycles per burst, 120 s, frequency sweeping mode, 23 W power, and temperature 6–8 °C. The resulting sheared DNA was divided and used for both methylation protocols.

2.2.2 Methyl-Binding Domain Sequencing (MBD-seq)

The MBD assay was performed by mixing 500ng Covaris sheared DNA (25 μl) with 110 μl Buffer B and subjecting 119 μl to the MethylCap Kit (Diagenode, Denville, NJ, USA), according to the manufacturer’s instruction (remaining volume used as input control). Methylated DNA was captured using methyl-binding domains from MeCP2 fused with GST, and anti-GST magnetic beads. The beads were washed with Buffer B according to the protocol, and eluted in a single elution using 50 μl high elution buffer. The captured product was then purified using the QIAquick PCR Purification Kit (Qiagen, Germantown, MD, USA) according to protocol instructions, with elution in 30 μl buffer EB. The NEBNext Ultra^TM DNA Library Prep Kit for Illumina (New England Biolabs (NEB), Ipswich, MA, USA) was used to perform end repair, adapter ligation, and PCR amplification, according to the manufacturers’ protocol. Sample cleanup was performed with size selection prior to PCR amplification using Agencourt AMPure XP beads (Beckman Coulter, Indianapolis, IN, USA). Samples underwent PCR amplification (12 cycles) using unmethylated adapters (diluted 1:10) obtained from the NEBNext Multiplex Oligos for Illumina Kit to allow for multiplexing of samples. Amplified products were verified for size (~250–500 bp, peak at 350 bp) using the Agilent High Sensitivity DNA Kit (Agilent Technologies, Santa Clara, CA, USA) and samples were pooled in equimolar quantities for sequencing analysis. Samples underwent cluster generation and 200bp paired-end sequencing using TruSeq v3 chemistry on the Illumina HiSeq2500 instrument (Illumina, San Diego, CA, USA).

2.2.3 Methylated DNA Immunoprecipitation Sequencing (MeDIP-seq)

MeDIP-seq was performed as previously described, using a novel multiplex approach.¹⁶ The NEBNext Ultra^TM DNA Library Prep Kit for Illumina was used to perform end repair, adapter ligation, and sample cleanup without size selection on 1 μg Covaris sheared DNA, according to the manufacturers’ protocol (starting volume 50 μl sheared DNA and 5.5 μl buffer EB). Unmethylated adapters (undiluted) obtained from the NEBNext Multiplex Oligos for Illumina Kit allowed for multiplexing of samples. Adapter-ligated samples underwent immunoprecipitation with a 5-methylcytosine antibody using the MagMeDIP Kit (Diagenode), according to manufacturer instructions with the modification that the starting volume of DNA for the IP incubation mix was 23 μl (volume of water adjusted). Owing to the high pH of the elution buffer used, we performed a pH adjustment by adding 8 μl of 100 mM Tris-HCl to the 100 μl volume of sample, and then performed sample cleanup using Agencourt AMPure XP beads, according to the standard protocols (starting with 1.8 × the sample volume for the resuspended bead mix, including two 70% ethanol washes, and ending with resuspension in 26 μl 10 mM Tris-HCl). PCR amplification (12 cycles) and cleanup were carried out using the NEBNext UltraTM DNA Library Prep Kit for Illumina, incorporating the primer provided in the NEBNext Multiplex Oligos for Illumina Kit, according to manufacturers’ protocol. Amplified products were verified for size (~300–1000 bp, peak at 500 bp) using the Agilent High Sensitivity DNA Kit and samples were pooled in equimolar quantities for sequencing analysis. Samples underwent 200 bp paired-end sequencing runs on the Illumina HiSeq2500 instrument.

2.2.4 RNA Extraction and Sequencing (RNA-seq)

RNA was extracted from each of the eight rat ventral hippocampal tissue samples (~30mg each) using the miRNeasy Mini Kit (Qiagen). Briefly, tissues were homogenized in 300 μl Qiazol, and underwent total RNA purification according to the manufacturer’s protocol, samples were eluted in a total of 60 μl RNase-free water (2 × 30 μl elutions). RNA was assessed for quality (all samples had RIN >7.5), and saline control samples and MAM-treated samples were pooled separately in equimolar concentrations. Due to costs and the preliminary nature of this pilot work, RNA-sequencing was performed on saline control and F2 MAM-treated pooled RNA rather than individual samples. For each pool, 500ng of RNA underwent sample preparation using the Illumina TruSeq Stranded Total RNA Kit w/Ribo-Zero^TM Globin (Illumina), according to manufacturer instructions. Amplified products were validated for size (~220–400 bp, peak at 280 bp) using the Agilent High Sensitivity DNA Kit and samples underwent 200 bp paired-end sequencing runs on the Illumina HiSeq2500 instrument, according to the manufacturer’s instructions, using reagents outlined above.

2.3 Data Analyses

2.3.1 Methylation Analysis

Raw sequencing data was processed by the Illumina base-calling pipeline and de-multiplexed. Sequence quality was ascertained using FastQC (http://www.bioinformatics.bbsrc.ac.uk/projects/fastqc), and based on that output, the reads were trimmed 15 bp at the 5’ end and filtered for quality (>30% N calls or >10% of the sequence with Phred quality <20 were omitted) using CutAdapt.³⁵ The R environment was used for all subsequent downstream analysis. Trimmed filtered reads were aligned to the rat reference genome obtained from UCSC, version rn6 (BSgenome.Rnorvegicus.UCSC.rn6) using QuasR³⁶ with the Rbowtie aligner set to default parameters except four allowed mismatches (-N = 4, based on the quality control metrics from FastQC).³⁷

The Bioconductor package MEDIPS³⁸ was used to identify differentially methylated regions (DMRs) derived from both affinity capture methodologies. Commands were executed with a genome window size of 100 nucleotides (based on sequence length) and the extend parameter set at 250 nucleotides (based on the average gap distance between the 100 bp paired-reads). We had previously tested windows of 50bp (MEDIPS default), however expanding to 100bp merged many DMRs, thereby increasing read counts within each window. To begin, each sample was verified to have sufficient read saturation for reproducibility, and then analyzed for CpG content and CpG enrichment. For both assays, DMRs between saline and MAM-treated rats were identified using the edgeR functionality in MEDIPS³⁹ with a minimum of five reads required per 100bp window in order to test for differential methylation, with quantile normalization enabled, and a FDR threshold of 0.05. Adjacent locations were merged using the mr.edgeR.s.m command, followed by annotation of the results. Annotation data was obtained from the Rattus norvegicus genes (Rnor_6.0) dataset from biomaRt (http://www.biomart.org) and from the rn6 refFlat file from UCSC Genome Browser.⁴⁰ Annotation was performed within Galaxy.org^41–43 using joining tools within the “Operate on Genomic Intervals” toolkit and files from the UCSC Main Table Browser.^{44, 45} We created genome tracks for an example DMR within UCSC Genome Browser, on the Rat Jul. 2014 (RGSC 6.0/rn6) assembly. To further compare our MeDIP-seq data analysis to our previous analysis,¹⁶, we converted the Rn5 DMR coordinates identified in our first study to Rn6 coordinates, using the UCSC genome database liftover functionality, and then looked for concordance between datasets.

2.3.2 RNA Analysis

Raw sequencing data was processed by the Illumina base-calling pipeline and de-multiplexed. Sequence quality was ascertained using FastQC (http://www.bioinformatics.bbsrc.ac.uk/projects/fastqc), and based on the output, reads were trimmed 15 bp at the 5’ end and filtered for quality (>30% N calls or >10% of the sequence with Phred quality<20 were omitted) using CutAdapt.³⁵ Trimmed sequences from the two pools (MAM-treated and saline) were aligned to the rat genome (Ensembl release 81, Rnor_6.0 ftp://ftp.ensembl.org/pub/release-81/fasta/rattus_norvegicus/dna/) using Tophat2⁴⁶ on an in-house Linux cluster with the following command: tophat2 -N 4 -r 150 -p 8 --read-edit-dist 4 rn6 trimmed_R1.fastq.gz trimmed_R2.fastq.gz --transcriptome-index rn6.gtf. The resulting transcriptome .bam files were assembled and quantified using the Cufflinks program, and differential expression between the control pool and the treated pool was ascertained using the Cuffdiff program.⁴⁷ Because we were comparing two pooled groups only, with no replicates, results were filtered by: 1) sum of normalized sequence counts across both pools 10; and 2) log2(fold_change) of greater than 1.5 or less than -1.5, including transcripts that were exclusive to either treatment or control (log2(fold)change = +/− ∞).

2.3.3 Gene Ontology (GO) Enrichment

We performed GO enrichment analysis of biological processes using the PANTHER (Protein ANalysis THrough Evolutionary Relationships) Classification System. Since the regulatory effects of each DMR are not necessarily known, we generated a list of genes that either intersected, or were upstream or downstream to each DMR, provided the genes were identified by both the MeDIP-seq and MBD-seq methods. In addition, we generated a list of genes that were differentially expressed between MAM-exposed F2 progeny and control rats based on a log2(fold_change) >±1.5 and the sum of normalized sequence counts across both pools being 10. Resulting gene lists were analyzed for GO enrichment using the PANTHER Overrepresentation Test, at the Gene Ontology Consortium (http://www.geneontology.org/page/go-enrichment-analysis).⁴⁸ This tool connects to the PANTHER Classification System, a maintained database for GO annotations.^{49, 50} Analysis was based on the GO Ontology database version released 2017-02-28, and utilized the “GO biological process complete” data set for R. norvegicus with Bonferroni correction for multiple testing. Results tables list significant shared GO terms (or parents of GO terms) used to describe the set of genes that the user supplied, the expected and observed gene frequency, fold enrichment, and corrected p-value.

3. Results

3.1 Data Output and Alignment Statistics

DNA extracted from the ventral hippocampus of male rats (four F2 offspring of MAM-treated rats and four F2 offspring of saline-control rats) was subjected to MeDIP-seq and MBD-seq protocols. The eight samples (Table 1) produced similar numbers of high quality paired-end reads across the two assays, the mean number of reads for MeDIP-seq was 18,560,572 (standard deviation (s)=2,730,060) for MAM-treated and 17,275,587 (s=988,655) for saline-control rats, and the mean number of reads for MBD-seq was 19,362,935 (s=2,190,907) for MAM-treated and 19,163,209 (s=2,175,500) for saline-control rats. The mean gap between paired reads among all samples was 263 (s=6.6). After read-trimming and quality filtering, the paired end reads were aligned to the most current release of the R. norvegicus genome (UCSC rn6), with similar results across the two assays (Table 1). Overall, the MBD-seq experiment somewhat outperformed MeDIP-seq in terms of data output and alignment, including a notable ~10% improved concordant pairing.

Table 1.

Alignment statistics for MBD-seq and MeDIP-seq data aligned with QuasR/Rbowtie

Assay	Total Reads	Mean	Left Aligned Reads	Right Aligned Reads	Aligned Pairs	% Alignment	% Concordant Pairs	% Multiple Alignments
MBD-seq_MAM	19326475	19362935 (s=2190907)	17752215	17182900	16532136	90.4	77.2	17.3
	17183342		11232684	13684791	10716433	83.2	65.1	16.9
	18574542		12078897	14763677	11533871	82.9	64.1	18.9
	22367381		12517428	15279227	11971937	82.9	65.1	17.4
MBD-seq_Saline	21574301	19163209 (s=2175500)	19856902	19101503	18374213	90.3	76.6	17.9
	19463667		17918551	17310277	16668391	90.5	77.1	17.8
	19327057		17776414	17179842	16528874	90.4	77.1	17.5
	16287811		12130976	14835641	11579746	82.8	64.1	19.1
MeDIP-seq_MAM	16117023	18560572 (s=2730060)	11773923	14349366	10906813	81.1	58.7	18.1
	17183342		12533063	17183342	11589324	80.9	58.9	16.7
	18574542		13522594	16470305	12494126	80.7	58.2	17.9
	22367381		20017729	19023865	17846973	87.3	69.9	16.9
MeDIP-seq_Saline	17597846	17275587 (s=988655)	12804209	15605425	11845928	80.7	58.6	17.4
	16034400		11686974	14260000	10809308	80.9	58.4	17.8
	17077133		12361734	15033414	11364634	80.2	57.2	16.9
	18392967		16476887	15600987	14597428	87.2	69.1	17.2

3.2 Saturation Analysis and CpG Enrichment

To determine whether the input sequence data coverage was sufficiently deep for reproducibility, saturation analysis was performed using the MEDIPS R package.⁵¹ Genome-wide coverage saturation analysis produced two values for each sample: the saturation correlation (maxTruCor) which is based on sequential calculations using half of the reads, and the estimated correlation (maxEstCor) which is based on artificially doubling that half. Actual correlation and reproducibility for the full set of reads will theoretically fall between the results of these true and estimated values. Among the eight MeDIP-seq samples, mean maxTruCor=0.964 (s=0.002) and mean maxEstCor=0.982 (s=0.001). Similarly, among the eight MBD-seq assays, mean maxTruCor=0.976 (s=0.003) and mean maxEstCor=0.988 (s=0.001). In all samples for both assays, our sets of mapped reads were sufficiently large to generate reproducible genome-wide coverage profiles (Supplementary Table 1).

To ascertain whether the pull-down assays enriched the samples for CpG sites as expected, two enrichment scores are reported by MEDIPS: RelH (based on the relative frequency of CpGs within the binned regions) and GoGe (based on the observed/expected ratio of CpGs within the binned regions). For the eight MeDIP-seq assays, the average enrichment.score.relH was 1.727 and the average enrichment.score.GoGe was 1.400 indicating that the experiment successfully enriched the samples for methylated cytosine beyond a genome baseline value of 1. For the eight MBD-seq assays, mean enrichment scores were considerably higher - indicative of more successful CpG enrichment within the samples (enrichment.score.relH was 2.447 and enrichment.score.GoGe was 1.698). We tested to see if there was a significant difference in CpG enrichment between MAM-treated and saline controls for each score and for each method. No significant difference was seen in enrichment score for MAM-treated versus control rats (p=0.257-0.920; Supplementary Table 1), indicating that these methods enriched for CpG sites, regardless of treatment.

3.3 Regions of Differential Methylation between MAM-treated and Saline-control Rats

DMRs between treatment and control samples defined within 100bp windows were ascertained using the edgeR statistical functionality for differential expression analysis within MEDIPS. It is important to note that the DMR intervals reported here are based on an arbitrary bin size (100bp, based on the length of sequences), rather than by CpG density, gene locations, or any other biologically relevant metric. All 100bp windows that were significantly differentially methylated, using a false discovery rate (FDR) of 5% for multiple testing correction, are shown in Supplementary Tables 2 and 3 (MBD-seq and MeDIP-seq, respectively). DMRs with negative log fold changes represent gains of methylation due to treatment, and positive log fold changes represent loss of methylation within the region. After merging of adjacent 100bp DMRs, the MeDIP-seq data identified 1,771 DMRs between saline-control and MAM treated F2 rats, 57% more than the 769 identified from MBD-seq (Table 2).

Table 2.

Characteristics of differentially methylated regions (DMRs) detected by MBD-seq and MeDIP-seq assays.

		Total DMRs	Intersected CGIs	DMRs Intersecting Genes*
MBD-seq		769	43	135
	Methylation loss	599	43	89
	Methylation gain	170	0	46
MeDIP-seq		1771	57	424
	Methylation loss	1101	55	219
	Methylation gain	670	2	205
Overlap		224	25	23
	Methylation loss	198	25	18
	Methylation gain	26	0	5

DMRs: Differentially methylated regions. These include adjacent 100bp windows in which DNA methylation levels are significantly different between MAM-treated rats and Saline-control rats at FDR<0.05.

CGI: CpG island. Numbers shown indicate how many DMRs fall within a CGI; in some cases multiple DMRs fall within the same CGI.

^*The number of DMRs that intersect with one (or more) annotated genes (including those of uncertain function, e.g. LOC genes)

The total length of DMRs for each chromosome, relative to chromosome length, was not even across the genome (p<0.00001 for both MeDIP-seq and MBD-seq; Figure 1, Supplementary Table 4). Specifically, across both methods we see consistent decreased representation (>2-fold) of DMRs on chromosomes 4, 9, 15 and X, and increased representation (>1.5-fold) of DMRs on chromosomes 1 and 14. In addition, MBD-seq detected overrepresentation of DMRs on chromosome 5 and underrepresentation of DMRs on chromosomes 11 and 19, whilst MeDIP-seq detected underrepresentation of DMRs on chromosome 16.

Figure 1.

Genome-wide coverage of DMRs detected by MeDIP-seq and MBD-seq. Total length of all DMRs across each chromosome is shown on the left Y axis, chromosome length is shown on the right Y axis. Y chromosome not represented due to only partial alignment in the current genome build (Rn6).

Supplementary Table 5 shows gene annotations for all merged DMRs detected by MBD-seq and MeDIP-seq – overlapping, upstream, and downstream genes are shown, as well as overlapping CpG islands. In total, 135 of 769 DMRs (17.56%) detected by MBD-seq overlapped with transcripts (expressed at any level), slightly less than the 424 of 1,771 DMRs (23.94%) detected by MeDIP-seq.

In our previous publication,¹⁶ we reported on 181 genes (representing upstream and downstream genes of 92 DMRs). Although the updated rat genome and our alternate analysis protocol increased the number of DMRs detected, there was high concordance between DMRs identified in our previous publication and those identified here; 71% of the DMRs representing upstream and downstream genes in our initial study had overlapping DMRs that were also detected here, all of these showed the same direction of effect (data not shown). The increased detection of DMRs afforded by our new pipeline, leading to an increased burden for multiple testing, likely resulted in some (29%) DMRs no longer being detected.

A total of 224 overlapping DMRs were detected across both assays (note, in some cases a single DMR overlapped two DMRs from the alternate assay), and all showed concordance in their methylation status between the two sequencing assays (Table 2, Figure 2, Supplementary Table 6). In addition, 19 DMRs detected by each sequencing assay were directly adjacent to each other, and also showed concordance between the two assays. Of the 556 non-overlapping DMRs detected by MBD-seq, 208 (37%) of these were proximal to one or more DMRs detected by MeDIP-seq: in addition to the 19 adjacent DMRs, 86, 51 and 52 DMRs were within 1,000bp, 2,000bp and 5,000bp, respectively (Figure 2A, Supplementary Table 6). Of the 1,562 non-overlapping DMRs detected by MeDIP-seq, 293 (19%) were proximal to one or more DMRs detected by MBD-seq; this included the 19 adjacent DMRs, as well as 102, 73 and 99 DMRs that were within 1,000bp, 2,000bp and 5,000bp, respectively (Figure 2A, Supplementary Table 6). All proximal MBD-seq-MeDIP-seq DMRs were concordant for methylation status, except for six regions where the direction of effect differed between the two assays. For each of these regions, the distance between DMRs for the two assays were >1,600bp (Supplementary Table 6). Figure 2B shows the UCSC Genome Browser track for a DMR in Tenm4, which was detected by both MeDIP-seq and MBD-seq. This DMR is of particular interest for this project as it is within the Tenm4 gene, is downstream of Mir708, and is upstream of Nars2; these three genes have all been implicated in neuropsychiatric and neurodegenerative disorders, including schizophrenia.^52–54 The majority of identified DMRs from both experiments represented CpG methylation loss, with a particularly notable proportion of loss among the overlapping DMRs from both assays. Few DMRs intersected to CGI (5.59% and 3.22% of MBD-seq and MeDIP-seq identified DMRs, respectively), and those that did so primarily indicated loss of methylation (Table 2) in the MAM treated group.

Figure 2.

Concordance of DMRs detected by MeDIP-seq and MBD-seq. A. Venn diagram showing overlap of DMRs (n=224), adjacent DMRs, and those within 1,000bp, 2,000bp, and 5,000bp for MeDIP-seq and MBD-seq. B. UCSC Genome Browser track, showing a DMR within the Tenm4 gene (chr1) detected by both MeDIP-seq and MBD-seq; Mir708 (upstream of DMR) and Nars2 (downstream of DMR) are also represented CGI tracks are not shown because no CGI was within that region of interest.

The most recent Rattus norvegicus genome release is the first to include the Y chromosome (in scaffolds, from a SRY rat strain), which is of interest for schizophrenia models. We find that both the MBD-seq and MeDIP-seq assays record a large number of significant DMRs (FDR threshold 0.05) falling within the Y scaffolds despite being the smallest chromosomal contributor to genome size (18.4% of the DMRs for MeDIP-seq and 41.1% for MBD-seq). However, since the Y chromosome is currently only partially assembled and not yet annotated, it is difficult to draw conclusions about the true number of DMRs across the Y chromosome and their importance to the MAM rodent model. Nonetheless, given the earlier age of onset for schizophrenia in men⁵⁵ and an increased risk associated with advanced paternal age, ^56–58 which is not explained by de novo mutations,⁵⁹ it is feasible that alterations in DNA methylation of the Y chromosome may partially contribute to the pathophysiology of schizophrenia.

3.4 Gene Expression Differences between Treatment and Control

To examine potential downstream effects of MAM-induced methylation changes, RNA was also collected from the same MAM-treatment and saline-control rat ventral hippocampal samples, and pooled into two groups (treatment/control) for sequencing. Alignment and differential analysis were performed within the Tuxedo suite.^{46, 47} Gene expression differences between MAM-treated and Saline-control F2 pooled rat samples are shown in Supplementary Table 7. To determine whether DNA methylation changes might induce gene expression changes at each DMR, we looked for overlap between DMRs identified by either MBD-seq or MeDIP-seq (defined in Supplementary Table 5) and transcripts that were at least moderately expressed (10 normalized sequence counts or more across both pools) in either the MAM-treated or Saline-control pools (derived from Supplementary Table 7). MeDIP-seq once again demonstrated slightly better concordance between data types, although the number of genes that had DMRs and were at least moderately expressed was reasonably low. Of the 769 DMRs identified by MBD-seq, only 68 (8.84%) overlapped with one or more moderately expressed transcripts and of the 1,771 DMRs identified by MeDIP-seq, only 209 (11.80%) overlapped with one or more moderately expressed transcripts (Supplementary Table 8). Given the crude design of our pooled RNA-seq experiment to generate this pilot data, it is difficult to make true inferences about the role of DNA methylation on gene expression within the same transcript and the directionality of such effects. Furthermore, we have not made any attempts at defining the effects elicited by DMRs at transcripts that might be upstream or downstream, rather than overlapping the physical location of the DMR. Thus, our pooled RNA-seq design does not provide adequate statistical power to truly understand the effects of differential methylation on gene expression. Because Y chromosome data was now available, specific consideration was given to Y-linked differentially expressed genes. The only large (2.3-fold increase) expression change detected in a Y-linked gene was in Usp9y, which is expressed in the human brain and implicated in sex-related brain differences, neuronal differentiation,^60–62 and autism,⁶³ however expression (sequence counts) was low in both sample pools.

3.5 Gene Ontology Enrichment Analysis

We performed GO enrichment of biological processes using gene lists for DMRs (intersecting, upstream and downstream genes detected by both MeDIP-seq and MBD-seq) and differentially expressed genes (fold change >±1.5, at least 10 normalized read counts between the two pools; defined in Supplementary Table 9). For DNA methylation, a total of 248 genes were included in the analysis (Supplementary Table 10) and for gene expression, a total of 40 differentially expressed transcripts were included in the analysis. Of note, none of the 40 differentially expressed transcripts from our filtered RNA-seq list matched any of the genetic loci that directly intersect DMRs, likely due to DMRs exerting their effect on more distant transcripts, or eliciting a more subtle gene expression change than dictated by our filtering specifications (e.g., between 1.1- and 1.5-fold changes in gene expression). GO enrichment of DMRs and differentially expressed gene lists identified a number of biological processes relevant to the MAM rodent model and schizophrenia (Table 3, Supplementary Table 10, Supplementary Table 11).

Table 3.

Summary of GO enrichment analysis for DMR-associated and differentially expressed genes

GO Biological Process	Fold Enrichment	P-value
DMR-associated biological processes
Positive regulation of cell-cell adhesion	6.79	8.14x10−4
Central nervous system neuron differentiation	6.04	2.43x10−2
Positive regulation of multicellular organismal process	2.52	1.73x10−3
Anatomical structure morphogenesis	2.25	5.23x10−3
Regulation of multicellular organismal development	2.25	2.00x10−2
Positive regulation of cellular process	1.72	4.32x10−3
Negative regulation of cellular process	1.70	4.49x10−4
Differential expression-related biological processes
Prepulse inhibition	>100	1.60x10−2
Striatum development	>100	3.25x10−2
Activation of adenylate cyclase activity	91.82	4.37x10−2
Response to amphetamine	57.13	6.74x10−3
Transmission of nerve impulse	39.55	2.87x10−2
Regulation of behavior	38.26	1.82x10−3
Regulation of synaptic transmission, glutamatergic	37.26	3.63x10−2
Adenylate cyclase-activating G-protein coupled receptor signaling pathway	36.73	3.84x10−2
Response to nicotine	34.77	4.74x10−2
Learning	19.72	4.59x10−2
Regulation of ion homeostasis	18.28	7.49x10−3
Positive regulation of ion transport	14.07	3.35x10−2
Regulation of cation transmembrane transport	13.92	3.56x10−2

^*Only the most specific biological processes are shown. For full results, see Supplementary Tables 10 (DMR-associated genes) and 11 (differentially methylated genes).

^*Bonferroni-corrected p-values shown

4. Discussion

The primary purpose of this work is to directly compare two affinity enrichment methodologies used to assess DNA methylation – MeDIP-seq and MBD-seq. In addition, the incorporation of RNA-seq data and inclusion of Y chromosome alignments, afforded us further opportunity to delineate the transgenerational epigenetic and genetic mechanisms by which treatment with MAM induces a schizophrenia-like phenotype in rodents. Similar to previous analysis by Bock et al,⁶⁴ we identified considerable overlap of DMRs between MeDIP-seq and MBD-seq, as well as a substantial number of proximal (including adjacent) DMRs between the two assays. Based on our CpG saturation in the data, the MBD-seq protocol better enriched for sequences containing CpG dinucleotides. However, MeDIP-seq indicated a larger number of DMRs and showed slightly better concordance with overlapping RNA expression data. Our pooled RNA-Seq approach is inadequate for determining the true impact of differential methylation on gene expression, with our current data suggesting that regional methylation is unlikely to influence expression of the immediate transcript. Analysis of independent samples will be key to determining whether methylation influences expression of the immediate transcript, or those upstream or downstream of the DMR.

We identified an uneven distribution of DMRs (total chromosomal DMR length relative to chromosome length) across the genome. Chromosomes 4, 9, 15 and X were underrepresented by DMRs and chromosomes 1 and 14 were overrepresented, according to both methods (Figure 1, Supplementary Table 4). This might be explained by the underlying topology of CpG sites (e.g., CpG clusters with fewer or less densely spaced CpGs are predominantly hyper-methylated⁶⁵), the tendency for smaller chromosomes to be more CGI dense⁶⁶ or the presence of specific genes on the chromosome. For example, chromosome 14 shows an overrepresentation of DMRs (as measured by length), but many of these intersect the Rn45s/Rn28s locus and represent lengths of more than 1,000bp, suggesting that differential methylation might be particularly prevalent in this gene only and drive the overrepresentation of DMRs (based on length) detected on chromosome 14. We do, however, entertain the possibility that chromosomes showing an overrepresentation of DMRs by either MBD-seq or MeDIP-seq may contain genetic loci important to the mechanistic action of MAM and, which are potentially relevant to schizophrenia. Our data also suggests that neither MeDIP-seq nor MBD-seq is ideal for looking at differential methylation within CGIs because under normal conditions, most CGIs are hypomethylated and therefore poor targets for affinity binding; this is further evidenced by our detection of only a small number of DMRs within CGIs (~5%). To improve genome coverage at CGIs and other poorly represented CpG regions, Li et al,¹² suggests combining MeDIP-seq and MRE-seq. This bipartite approach yields an accurate high-coverage high-resolution methylome profile while still being significantly more economical than WGBS. In our study, although both affinity assays identified a number of DMRs, we concur with Li et al, regarding the potential need for a second overlapping method to guarantee maximal methylome information and override CpG-rich biases in affinity enrichment methods.

The MAM-E17 rat model²⁷ has become an increasingly accepted neurodevelopmental model for schizophrenia.^{26, 67–69} Methylazoxymethanol acetate is known to be a DNA-alkylating agent; administration of the drug to pregnant rat dams (gestational day 17) induces multi-generational¹⁶ schizophrenia-like behaviors and cognitive deficits, as well as anatomical and functional changes within the brains of offspring, culminating in aberrant hippocampal dopamine signaling and disruption of glutamatergic-GABAergic neurotransmission.^{68, 70–76} We previously investigated DNA methylation changes using MeDIP-seq within the ventral hippocampus in F1 and F2 generation offspring of MAM-treated dams and identified alterations associated with neurological disease, developmental disorders, and cell signaling, as well as those associated with schizophrenia.¹⁶ In this study, GO enrichment analysis of DMR-associated genes identified enrichment of biological process related to cell adhesion, neuron differentiation and positive regulation of multicellular organismal process (Table 3, Supplementary Table 10). Input genes within the central nervous system neuron differentiation biological process included Sema3a, Gnaq, Nrxn1 and BCL11B, which have all been implicated in schizophrenia.^77–81 GO enrichment of differentially expressed genes was more striking, identifying several biological processes related to schizophrenia, including prepulse inhibition, striatum development, response to amphetamine, regulation of behavior, regulation of glutamatergic synaptic transmission, and learning (Table 3, Supplementary Table 11). There are two major competing, but likely complementary, neurotransmitter hypotheses of schizophrenia development: the dopamine hypothesis suggests that hyperactivity of dopamine transmission leads to neural dysfunction that is central to schizophrenia pathogenesis;^{82, 83} whilst the glutamate hypothesis speculates that hypofunction of glutamatergic signaling via N-methyl-D-aspartate (NMDA) receptors evokes neural dysfunction leading to schizophrenia.^{84, 85} Not surprisingly, biological processes related to dopamine (e.g., response to amphetamine,^{86, 87} striatum development,⁸⁸) and glutamate (e.g., glutamatergic regulation of synaptic transmission) were enriched for in our differentially expressed gene dataset. Such pathways reflect stimulation of dopamine and disruption of hippocampal glutamatergic-GABAergic neurotransmission, which are representative of our MAM-induced schizophrenia rodent model. In the MAM-induced rodent model of schizophrenia, hyper-responsivity to amphetamine is evident, which appears to be mediated through the dopamine system.^{70, 72, 89} Enrichment of the “response to amphetamine” biological process in our analysis is consistent with this, and may also have relevance to the onset of psychosis.⁹⁰ Prepulse inhibition was also enriched for in our dataset, which is known to be deficient in patients with schizophrenia,^{91, 92} and may be more prominent in males.⁹³ With respect to our MAM-induced schizophrenia rat model, although one study investigating earlier MAM-treatment (injection on gestational day 9–15) identified no differences in prepulse inhibition in treated versus control rats,⁹⁴ studies utilizing an identical model to ours (injection on gestational day 17) found deficits in prepulse inhibition at adulthood.^{27, 95} In addition, we see enrichment for learning, memory and cognition, deficits of which are hallmarks of schizophrenia (reviewed in^96–98). Overall, our enrichment analysis identified biological processes related to schizophrenia and that are perturbed in in utero MAM-exposed rats, and our findings extend these disturbances to the F2 progeny of MAM-exposed rats, which also exhibit a schizophrenia-like phenotype.

In our gene expression data, increased transcriptional expression at a locus encompassing rno-mir-22 and rno-mir-6326 was seen. Although these miRNAs weren’t incorporated into the PANTHER GO enrichment algorithm (note, rno-mir-6232 does not have a human homolog), rno-mir-22 may be an additional potential biological candidate for schizophrenia pathogenicity. Reduced expression of miR-22 is seen in Alzheimer’s disease brains,⁹⁹ while upregulation has been seen in grade 1 Huntington’s disease and downregulation in grade 2 Huntington’s disease.¹⁰⁰ Upregulation of mir-22 improves neuronal viability,¹⁰¹ and although we see an opposite direction of effect than expected, this may suggest a compensatory mechanism or feedback loop.

Aside from determining male sexual organs and spermatogenesis, Y chromosome genes are expressed in other tissues and have been linked to behavioral and brain phenotypes, autoimmune diseases, neural tube defects, and blood pressure.¹⁰² The addition of Y-chromosome scaffolds to the newest Rattus norvegicus genome build provides us with possible new insights about how maternal dosing with MAM confers a schizophrenia-like phenotype in male F2 generation offspring. Large portions of the Y chromosome were differentially methylated among the offspring of MAM-exposed rats, and most interestingly, Usp9y expression was upregulated log2(fold_change = 2.3) among MAM-exposed F2 rats (although overall expression was low). The X chromosome homolog Usp9x, however, was highly expressed but not differentially expressed between the treatment and control pools. Usp9y (ubiquitin specific peptidase 9, Y-linked) is one of several Y-linked genes implicated in brain sexual dimorphism during human development^103–105 and its expression is increased during neural differentiation,¹⁰⁶ making it an interesting candidate for further study in schizophrenia pathogenicity.

5. Conclusions

We compared two affinity enrichment-based sequencing methodologies to examine DNA methylation changes in male offspring of in utero MAM-exposed rats and extended our previous findings to examine gene expression changes that might be important for schizophrenia pathophysiology. Similar to other studies, we found that MBD-seq and MeDIP-seq were fairly comparable and showed overlap of genes differentially methylated between MAM-treated F2 rats and controls. With the recent release of the rat chromosome Y sequence scaffolds, we identified several differentially methylated regions, however the true nature of this is difficult to determine due to current inadequacies in alignment and annotation of this chromosome. To investigate whether DNA methylation might impact gene expression, we used a pooled experiment to assess gene expression differences between MAM-treated F2 rats and controls. Although we did not see significant overlap between differentially methylated and differentially expressed genes (based on at least moderate expression), this does not preclude the possibility of DNA methylation regulating upstream or downstream genes. Our DMR and gene expression-based GO enrichment analysis identified several important biological processes relevant to schizophrenia, including those related to neuron differentiation, glutamatergic signaling, amphetamine response, prepulse inhibition and learning. An obvious caveat to our gene expression analysis is the use of pooled samples, and therefore potentially rudimentary inclusion of differentially expressed genes. As such, expanded sequencing analysis of independent samples, rather than a pooled approach, is recommended to further clarify the role of methylation-induced gene expression changes associated with MAM-treatment and its implications for schizophrenia pathophysiology.

Supplementary Material

1

Supplementary Figure 1: Comparison of MeDIP-seq and MBD-seq methods

Supplementary Tables 1-11

Supplementary Table 1: Saturation analysis and CpG enrichment

Supplementary Table 2: MEDIPS MBD-seq sequence count data

Supplementary Table 3: MEDIPS MeDIP-seq sequence count data

Supplementary Table 4: Chromosomal distribution of DMRs for MBD-seq and MeDIP-seq

Supplementary Table 5: Annotated data from DMRs ascertained by MBD-seq and MeDIP-seq analysis, showing intersecting genes, nearest upstream and downstrem genetic elements and overlapping CpG islands

Supplementary Table 6: Overview of overlapping and adjacent/proximal DMRs detected by MBD-seq and MeDIP-seq

Supplementary Table 7: Gene expression differences (full data) for pooled F2 generation male rats from MAM-treated granddams versus pooled saline-treated controls

Supplementary Table 8: Intersection of MBD-seq and MeDIP-seq DMRs with gene expression results based on overlapping intervals (for any gene expressed in either the Saline or MAM pool)

Supplementary Table 9: Differentially expressed genes filtered for fold-change and read counts for input into Function and Pathways Analysis

Supplementary Table 10: GO enrichment analysis of intersecting, upstream or downstream genes detected by DMR analysis of both MBD-seq and MeDIP-seq

Supplementary Table 11: GO enrichment analysis of differentially expressed genes detected by RNA-seq

Research Highlights.

• 56% of MBD-seq detected DMRs were also identified by or proximal to MeDIP-seq DMRs

• DMRs were not evenly distributed across the genome (based on length of DMRs)

• No significant overlap was seen between DMRs and differentially expressed genes

• GO enrichment analysis defined several schizophrenia-relevant biological processes

Abbreviations

BS-seq

bisulfite sequencing

CpG

cytosine-phosphate-guanine

CGI

CpG island

DMR

differentially methylated region

GO

gene ontology

MeDIP-seq

methylated DNA immunoprecipitation sequencing

MAM

methylazoxymethanol acetate

MBD-seq

methyl binding domain sequencing