Genome-wide DNA methylation variability in adolescent monozygotic twins followed since birth

Abstract

DNA methylation patterns are characterized by highly conserved developmental programs, but allow for divergent gene expression resulting from stochastic epigenetic drift or divergent environments. Genome-wide methylation studies in monozygotic (MZ) twins are providing insight into the extent of epigenetic variation that occurs, irrespective of genotype. However, little is known about the variability of DNA methylation patterns in adolescence, a period involving significant and rapid physical, emotional, social, and neurodevelopmental change. Here, we assessed genome-wide DNA methylation using the 450 K Illumina BeadChip in a sample of 37 MZ twin pairs followed longitudinally since birth to investigate: 1) the extent of variation in DNA methylation in identical genetic backgrounds in adolescence and; 2) whether these variations are randomly distributed or enriched in particular functional pathways. We also assessed stability of DNA methylation over 3–6 months to distinguish stable trait-like and variable state-like genes. A pathway analysis found high within-pair variability in genes associated with development, cellular mechanisms, tissue and cell morphology, and various disorders. Test-retest analyses performed in a sub-sample of 8 twin pairs demonstrated enrichment in gene pathways involved in organismal development, cellular growth and proliferation, cell signaling, and particular disorders. The variability found in functional gene pathways may plausibly underlie phenotypic differences in this adolescent MZ twin sample. Furthermore, we assessed stability of methylation over 3–6 months and found that some genes were stable while others were unstable, suggesting that the methylome remains dynamic in adolescence and that dynamic sites tend to be organized in certain gene pathways.

Abbreviations:

CBMC

cord blood mononuclear cells

IPA

Ingenuity Pathway Analysis

HLA

human leukocyte antigen

HUVEC

human umbilical vascular endothelial cells

K-SADS

Kiddie-Schedule for Affective Disorders and Schizophrenia

MHC

major histocompatibility complex

MZ

monozygotic

QNTS

Quebec Newborn Twin Study

Introduction

Although each individual's genome is fixed throughout life and from one cell type to another, epigenetic mechanisms are implicated in the regulation of gene expression. Cell type DNA methylation patterns emerge during development^1,2 and are postulated to play a role in gene expression by directing the configuration of inactive chromatin³ or interfering with the binding of transcription factors.⁴ The involvement of DNA methylation in controlling cell identity implies that DNA methylation patterns should show little variation during the life span. However, emerging evidence suggests that DNA methylation is responsive to both physical and social environments during pregnancy⁵ and early in life.^6,7 Indeed, many studies have shown that environmental events are associated with epigenetic modifications, including DNA methylation.^7-10 For example, a study by Heijmans and colleagues demonstrated that individuals who were exposed to famine in the perinatal period had, 6 decades later, altered DNA methylation patterns compared to their siblings.¹¹ Furthermore, Borghol and colleagues⁸ found an association between methylation levels in key cell-signaling pathways and low socioeconomic status during childhood.

These data raise the questions of how much variability is present in DNA methylation and whether this variation is stochastic or reveals some level of functional organization. Most of the epigenome must be well conserved for an organism to be viable, but some variability is possible.^12,13 Since genetics can influence DNA methylation, genetically identical monozygotic (MZ) twins have been examined in order to differentiate between genetically and externally driven DNA methylation variation.^14-24 Several studies have shown high genome-wide within-twin pair similarity in DNA methylation, although the level of similarity varies depending on tissue, gene, and twin pair.^{14,16,17,19-22,25} More specifically, Gordon and colleagues¹⁷ found that the most discordant DNA methylation sites across co-twins were associated with genes that are associated with the immune system and responding to the environment. Similar findings of DNA methylation discordance on genes associated with immune function were found in MZ twins discordant for psoriasis²² and autoimmune inflammatory diseases.²⁶ Moreover, both Gordon and colleagues¹⁷ and Saffery and colleagues²⁷ found that the most discordantly methylated genes from cord blood mononuclear cells (CBMCs) and human umbilical vascular endothelial cells (HUVECs) were those shown to be involved in responding to the environment. Additionally, studies have found within-pair DNA methylation discordance in association with autism,^28,29 bipolar disorder,³⁰ risk taking behavior,³¹ Alzheimer disease,³² intestinal disease,²³ diabetes,^33-35 and even birth weight.²⁵ See Table S1 for a summary of epigenetic findings in MZ twins. Overall, studies show high similarity within twin pairs across tissues; however, differences are also found, particularly when phenotypes diverge across twins.

In addition to within-twin pair variability in methylation, a related question is whether DNA methylation is responsive to external factors throughout the lifespan. If the methylome is indeed dynamic throughout life, then differences in DNA methylation profiles in identical twins should increase through life. As expected, studies demonstrate that although within-pair discordances are present from birth,³⁶ they increase with age,^37-45 particularly when twins experience divergent medical histories/environments.^46–49 In fact, Novakovic and colleagues⁵⁰ found differences during gestation, which increased with gestational age. The direction of change is complex: DNA methylation increased with age at some loci and decreased at others.^37,51 Furthermore, methylation does not vary with age in all genes equally, suggesting some specificity.^52,53 Genes found to be associated with age are enriched for functions including DNA binding and regulation of transcription,⁴⁴ molecular and cellular characteristics of skin tissue development,⁴² and aging-related conditions including Alzheimer disease, cancer, tissue degradation, DNA damage and oxidative stress.⁴³ See Table S2 for a summary of findings associating epigenetic marks with age.

Notably, little is known about methylation patterns in adolescence, even though this developmental transition is a time of increased independence and physiological maturation, and therefore may potentially be a period of increased variability in epigenetic mechanisms within twin pairs. The few studies that have examined DNA methylation during adolescence have produced similar findings. Kaminsky and colleagues⁵⁴ assessed DNA methylation in MZ and DZ twin pairs aged 12–15 and found significant within-twin pair variability across different tissues (e.g., white blood cells, buccal epithelial cells, and gut biopsies). Variability was greater within DZ than MZ twin pairs, and among MZ twins, variability was higher among dichorionic than monochorionic twins. Furthermore, Essex and colleagues⁹ assessed methylation in buccal epithelial cells and found an association between parental stress in children's early lives and methylation of several genes involved in biosynthetic and metabolic processes during adolescence. However, to our knowledge, no study has used saliva to assess genome-wide methylation differences in adolescence—which may be a useful non-invasive means to acquire DNA in this population—and only one study has assessed the hypervariability across the widest range of CpG sites currently possible (> 480,000 methylation sites) in adolescents. By specifically examining DNA methylation from buccal cells in monozygotic twin preadolescents (8–10 y.o.) and young adults (18–19 y.o.), van Dongen and colleagues⁵⁵ found that most twin pairs clustered together. However, the short-time stability of hypervariable genes is still unknown. Distinguishing stable genes from those that are highly dynamic among MZ twins is necessary in order to identify genes that may be responsive in a stable trait-like manner to the immediate environment.

The present study examined MZ twin pairs through a whole-genome approach to determine: (1) whether within-twin pair differences in DNA methylation are present during adolescence and; (2) whether these differences reflect a level of functional organization. We then assessed (3) whether the DNA methylation pattern in adolescence exhibits dynamic features independently of their genetic background. By limiting ourselves to short time intervals, we were able to directly examine how dynamic the methylome is in adolescence.

Results

We used the 450 K Illumina BeadChip to assess whole-genome DNA methylation profiles from saliva at one or 2 time-points in a sample of 37 adolescent MZ twin pairs. Following data filtering (see methods section), we were left with a final dataset of 179,408 temporally stable probes and 241,211 temporally unstable probes. Z-scores of absolute twin differences were calculated and probes more than 3 standard deviations above the mean were considered to be hypervariable. This resulted in 258 temporally stable probes, which mapped to 226 unique genes and 47 temporally unstable probes, which mapped to 46 unique genes. For a list of these genes along with locations and types, please see Table S3 for trait-like sites and Table 1 for state-like sites.

Table 1.

Hypervariable genes across MZ twins and time

Gene Symbol	Gene Name	Location	Type(s)
ADAM3A	ADAM metallopeptidase domain 3A (pseudogene)	Other	other
ADORA3	adenosine A3 receptor	Plasma Membrane	G-protein coupled receptor
AGAP1	ArfGAP with GTPase domain, ankyrin repeat and PH domain 1	Cytoplasm	enzyme
AMACR	α-methylacyl-CoA racemase	Cytoplasm	enzyme
APITD1/APITD1-CORT	apoptosis-inducing, TAF9-like domain 1	Nucleus	other
B4GALNT3	β-1,4-N-acetyl-galactosaminyl transferase 3	Other	enzyme
BAIAP3	BAI1-associated protein 3	Extracellular Space	other
BRD2	bromodomain containing 2	Nucleus	kinase
CACNA1A	calcium channel, voltage-dependent, P/Q type, α 1A subunit	Plasma Membrane	ion channel
CDH20	cadherin 20, type 2	Plasma Membrane	other
CLDN11	claudin 11	Plasma Membrane	other
CNNM4	cyclin M4	Plasma Membrane	other
DDR2	discoidin domain receptor tyrosine kinase 2	Plasma Membrane	kinase
DNAJB6	DnaJ (Hsp40) homolog, subfamily B, member 6	Nucleus	transcription regulator
EGFR	epidermal growth factor receptor	Plasma Membrane	kinase
ENPP7	ectonucleotide pyrophosphatase/phosphodiesterase 7	Plasma Membrane	enzyme
FAM20C	family with sequence similarity 20, member C	Extracellular Space	enzyme
GALNT9	polypeptide N-acetylgalactosaminyltransferase 9	Cytoplasm	enzyme
GNA12	guanine nucleotide binding protein (G protein) α 12	Plasma Membrane	enzyme
HBE1	hemoglobin, epsilon 1	Cytoplasm	transporter
HLA-DQB1	major histocompatibility complex, class II, DQ β 1	Plasma Membrane	other
HLA-DRB6	major histocompatibility complex, class II, DR β 6 (pseudogene)	Other	other
KCTD2	potassium channel tetramerization domain containing 2	Other	ion channel
KDM1A	lysine (K)-specific demethylase 1A	Nucleus	enzyme
LPP	LIM domain containing preferred translocation partner in lipoma	Nucleus	other
LRWD1	leucine-rich repeats and WD repeat domain containing 1	Nucleus	other
MCF2L	MCF.2 cell line derived transforming sequence-like	Cytoplasm	other
MCF2L	MCF.2 cell line derived transforming sequence-like	Cytoplasm	other
METTL9	methyltransferase like 9	Other	other
mir-548	microRNA 548c	Cytoplasm	microRNA
OR52N5	olfactory receptor, family 52, subfamily N, member 5	Plasma Membrane	G-protein coupled receptor
PCGF3	polycomb group ring finger 3	Nucleus	other
PKDCC	protein kinase domain containing, cytoplasmic	Cytoplasm	kinase
PLCH2	phospholipase C, eta 2	Cytoplasm	enzyme
PTPRN2	protein tyrosine phosphatase, receptor type, N polypeptide 2	Plasma Membrane	phosphatase
RANBP6	RAN binding protein 6	Cytoplasm	other
RTN2	reticulon 2	Cytoplasm	other
SCAMP1	secretory carrier membrane protein 1	Cytoplasm	transporter
SLC45A4	solute carrier family 45, member 4	Other	other
TP63	tumor protein p63	Nucleus	transcription regulator
USP42	ubiquitin specific peptidase 42	Other	peptidase
UVSSA	UV-stimulated scaffold protein A	Nucleus	other
VGLL2	vestigial-like family member 2	Nucleus	transcription regulator
VPS13B	vacuolar protein sorting 13 homolog B (yeast)	Nucleus	transporter
YTHDF3	YTH domain family, member 3	Cytoplasm	other
ZNF155	zinc finger protein 155	Nucleus	transcription regulator
ZNF665	zinc finger protein 665	Other	other

List of 47 genes, locations, and types that were found to be hypervariable both across individuals and across time-points 3–6 months apart.

Twin similarity

Correlations among mean methylation levels across all samples were very high, suggesting conservation of DNA methylation states in humans. Each individual predicted at least 95.6% of the variability in every other individual (r > 0.978). Twins were the best predictors of each other's mean methylation. Indeed, twins predicted between 95.85% and 99.57% (r values ranged from 0.958 to 0.998) of the variance in one another's DNA methylation patterns. Twin correlations were also assessed using hierarchical clustering. In almost every case, an individual's data was best predicted by their twin's data. This association is displayed in a clustergram (Fig. 1).

Figure 1.

Twin correlations as assessed by hierarchical clustering Clustergram representing the discordance between all participants. Blue represents the least amount of discordance; red, the greatest amount. This demonstrates that, in almost every case, an individual's DNA methylation was best predicted from his or her twin's DNA methylation.

Pathway analyses of hypervariable genes stable in time

In spite of the strong conservation of DNA methylation states across individuals, hypervariable DNA methylation sites were observed between genetically identical twins, indicating that these differences are not genetically predetermined. A question that has remained unanswered is whether these differences are functionally organized or randomly distributed in the genome. We first focused on the most variable, but also temporally stable, DNA methylation sites within twin pairs, as potential representatives of “early life” differences in DNA methylation that remain stable throughout life. Ingenuity Pathway Analysis (IPA) of the 226 genes showing highly variable DNA methylation sites identified enrichment of 16 networks or pathways involved in several diseases and disorders: neurological, metabolic, reproductive system and hematological diseases, as well as psychological, developmental, hereditary, and endocrine system disorders. Developmental networks, including organismal, embryonic, cellular, tissue, skeletal, muscular, and cardiovascular system, were also prominent, as were cellular mechanisms involving cell-to-cell signaling, cellular assembly and organization, cell cycle, small molecule biochemistry, cell death and survival, as well as cell morphology. The top 11 networks with a score of 15 or greater are presented in Table 2 and the top network (cell-to-cell signaling and interaction, tissue development, and cardiovascular system development and function) is visually represented in Figure 2. Scores were used to rank networks according to fit between the biological pathway and the number of eligible molecules found in our analyses, and were calculated using the right-tailed Fisher's Exact Test using the following formula: Network score = -log(Fisher's Exact Test result). A greater score represents a better fit. Finally, sites can also be categorized into top diseases and functions. Here, the top diseases and functions of hypervariable genes were cancer, with 162 molecules (P = 2,81E-06), and organismal survival, with 54 molecules (P = 2,41E-05). See Table S4 for a complete list.

Table 2.

Networks of hypervariable genes (stable in time).

ID	Genes	Score	# Focus molecules	Top Diseases and Functions
1	ALOX12, amylase, ATF1, CaMKII, Caveolin, CCAR2, Creb, DEPTOR, DMBT1, ERK1/2, GALNT2, HTATIP2, ITPR3, JPH3, LDL, MCF2L, Mek, MGRN1, Mlc, MTA1, NTN1, PCGF3, PDGF BB, PI3K (family), Pkg, PP1 protein complex group, PP2A, PROCR, RCAN1, RPS6KA2, SALL4, SLC6A3, TRIM9, VASP, YAF2	38	22	Cell-To-Cell Signaling and Interaction, Tissue Development, Cardiovascular System Development and Function
2	ABCC1, ADAP1, Akt, Alp, ALPPL2, AMPD2, AMPK, BRSK2, Collagen type I, CUX1, Cyclin A, estrogen receptor, EXOC7, Fgf, Fgfr, FGFR1, growth factor receptor, GTPase, IGSF9B, Integrin, Laminin, LPCAT1, MAGI2, N-cor, NCF2, PBX1, PLC gamma, PRKAA, Proinsulin, PTPRN2, RARB, SCD, TBL1XR1, TGFA, WNT5A	31	19	Organismal Development, Embryonic Development, Skeletal and Muscular System Development and Function
3	ADCY, Ap1, APOBEC3G, ARHGAP26, Calcineurin protein(s), calpain, Collagen type IV, CSK, DUSP22, FBL, Fibrinogen, FSH, G protein alphai, GNA12, GNAS, GST, GSTM1, GSTT1, HBP1, IGF2BP3, Igm, Lh, MAP2K1/2, Mapk, MT1L, NFIC, NFkB (complex), Pdgf (complex), PLC, PRKCA, Rac, Sos, SPIB, TFRC, VAV	25	16	Drug Metabolism, Glutathione Depletion In Liver, Cellular Development
4	AKR1C1/AKR1C2, CD3, CHMP3, CLDN4, COL17A1, Cpla2, ERK, FBLN2, FCGR2C, HLA-C, HLA-DQA1, HLA-DQB1, HLA-DRB1, HLA-DRB5, Hsp70, Hsp90, IgG, IgG1, IgG2a, IL1, IL12 (complex), IL4I1, Immunoglobulin, Interferon α, MHC, MHC Class II (complex), P38 MAPK, PDIA6, PRX, SRC (family), STAT5a/b, TCR, Tgf β, VIPR2, XCL1	25	16	Neurological Disease, Psychological Disorders, Developmental Disorder
5	26s Proteasome, Actin, AGAP1, Alpha catenin, ANO1, ASAP1, Calmodulin, caspase, Clathrin, DNAJC6, Focal adhesion kinase, G-protein β, GGA1, Gpcr, HCN2, Hdac, HMOX2, Insulin, IQGAP2, ITGA8, Jnk, LPHN1, MOV10, NMDA Receptor, PI3K (complex), Pka, Ras, Ras homolog, RFX4, RIMBP2, SHANK2, Shc, SLC9A3R2, SRC, voltage-gated calcium channel	25	16	Cellular Assembly and Organization, Cell Morphology, Cell-To-Cell Signaling and Interaction
6	ATP10D, ATP11B, ATP11C, ATP4B, ATP8A1, ATP8B2, ATP8B3, ATP9B, CHTF18, DENR, FNDC3B, IBA57, MARCH5, MRPL3, NCLN, NDUFA11, NDUFA12, NDUFAF2, NDUFB4, NDUFB11, NDUFS6, NDUFV3, NXN, PCDHA6, PLEKHA7, POTEM (includes others), RBMS1, SLC4A10, SLC4A11, TST, TUT1, TXNRD3, UBC, WDR37, ZCCHC6	25	16	Metabolic Disease, Developmental Disorder, Hereditary Disorder
7	ABCA6, ALKBH6, ARL17A/ARL17B, BAI2, BTBD11, BTN2A1, C20orf195, C5orf30, CCDC33, CCND1, FHOD1, FUK, FUT5, GPR137, GRIK3, HHLA2, HNF1A, HNF1α dimer, HNF4A, HRAS, Ins1, LHX4, MDFI, MRO, PAMR1, RPH3AL, SLC38A4, TBC1D16, TRAF2, TSH, VN1R1, VPS54, ZAN, ZNF155, ZNF707	23	15	Energy Production, Cell Cycle, Cellular Development
8	ACE, ACSF3, ADA, AIFM3, ANKRD37, AUP1, B3GNT6, B3GNTL1, C1orf52, CREB3L1, DALRD3, DDHD1, DDIT4, DENND3, DIP2C, FBXL18, FEM1B, HIF1AN, KIAA0319, LRR1, LUZP1, MARCKS, MUC2, PPM1F, RAD9A, SIM2, SLC27A2, SLC27A3, SYT2, UBC, UBE3B, USP9Y, ZFYVE28, ZMAT2, ZNF506	21	14	Cell-To-Cell Signaling and Interaction, Cellular Development, Tissue Development
9	AATK, AHR, Ahr-aryl hydrocarbon-RelA, APP, ARL6IP6, BCL2L1, CASR, CLDN14, CLDN20, COL6A1, DGCR6/LOC102724770, DNALI1, EBF3, EZH2, FMO2, GALNT10, GIMAP5, GPR35, GPR61, GPR78, ITK, LEP, LRRC8D, LYPD6B, MRAP2, OPLAH, POPDC2, PYDC2, RELA, Slpi (includes others), ST8SIA2, TNF, TRIM35, TTPA, TWIST1	21	14	Lipid Metabolism, Molecular Transport, Small Molecule Biochemistry
10	AIM1, ALB, AMN, ANKRD32, ARHGAP11A, CCDC57, CROT, DNAJC7, GABRP, GALNTL5, GMNN, GPR83, GSTP1, Gstt3, HSPA12B, KIAA1324, KIAA1804, KRTAP1-3, LOC100133315, MAPK1, MYADML2, MYT1L, NEIL3, NUMA1, NUPR1, OSBPL6, PCYOX1, PDX1, PLA2G4F, SCAF1, SHPK, SMARCA4, SPATA24, TNNC2, TP53BP1	17	12	Endocrine System Disorders, Organismal Injury and Abnormalities, Reproductive System Disease
11	C1q, DEAF1, Gpcr, GPR37, GPR62, GPR82, GPR85, GPR97, GPR111, GPR112, GPR128, GPR133, GPR139, GPR144, GPR149, GPR150, GPR152, GPR157, GPR162, GPR174, GPR137C, GPRC5D, HCRTR1, HTR1D, IFNB1, LAIR2, MAPK14, MAS1L, MCOLN1, MYOM2, NMUR1, OXGR1, RAC1, VN1R5, ZDHHC14	17	12	Cell-To-Cell Signaling and Interaction, Cell Signaling, Cell Death and Survival

Computed by IPA. Genes in bold were hypervariable in our sample. Genes not in bold were implicated in the network but not hypervariable in our sample. Only networks with scores greater than 15 are presented.

Figure 2.

Top trait-like network: Cell-to-cell signaling and interaction, tissue development, and cardiovascular system development and function Network 1 as assessed by IPA. Trait-like genes are genes whose state of methylation are hypervariable within twin pairs but remain stable over time. Genes in gray were hypervariable in our analyses. Genes in white were part of the network but not hypervariable.

Pathway analyses of hypervariable genes unstable in time

We then examined whether DNA methylation states are fixed early in life and remain stable during adolescence onwards, or whether certain DNA methylation states remain dynamic later in life. By examining the most variable sites within 8 adolescent twin pairs at 2 time points 3–6 months apart we were able to discover dynamic DNA methylation changes during adolescence that are independent of genetics. We identified 47 such sites. We then examined whether these dynamic DNA methylation sites in adolescence were functionally organized or whether they were randomly scattered across the genome. Our analysis revealed 3 significant networks. Network 1 (score of 43) contained 18 molecules involved in organismal development, cellular growth and proliferation, as well as digestive system development and function. Network 2 contained 15 molecules (score of 34) involved in connective tissue disorders, dental disease, and developmental disorders. Finally, network 3 contained 8 molecules (score of 15) involved in cancer, organismal injury and abnormalities, as well as reproductive system disease. See Table 3 for details and Figure 3 for a visualization of the top network (organismal development, cellular growth and proliferation, digestive system development and function). See Table S5 for a complete list of top diseases and functions.

Table 3.

Networks of hypervariable genes (hypervariable in time).

ID	Genes	Score	# Focus molecules	Top Diseases and Functions
1	ADORA3, AGAP1, ART1, BRD2, CACNA1A, Calmodulin, Calmodulin-Camk4-Ca2+, caspase, DDR2, DNAJB6, EGFR, Focal adhesion kinase, GALNT9, GNA12, GPR55, HBE1, Hdac, Histone h3, Histone h4, KDM1A, LPP, LRWD1, MCF2L, P38 MAPK, PI3K (complex), PKDCC, PLC, Plcd2, PLCH2, PLCZ1, Ras homolog, SCAMP1, TAAR5, TP63, Vegf	43	18	Organismal Development, Cellular Growth and Proliferation, Digestive System Development and Function
2	APITD1/APITD1-CORT, BAIAP3, Basp1, CECR5, CLDN11, CNNM4, COX11, ENPP7, FAM213A, FBRSL1, GIMAP1, GOLGA7, HLA-DQB1, HLA-DQB2, HRAS, KCTD2, METTL9, MGME1, Olfr1508, PCGF3, PRELID1, RANBP6, RASIP1, RPL39, RT1-A3 (includes others), SLC39A6, SLC45A4, TENM3, TOX2, UBC, UVSSA, VPS13B, YTHDF3, ZDHHC9, ZNF665	34	15	Connective Tissue Disorders, Dental Disease, Developmental Disorder
3	ADCK3, Alp, AMACR, AS3MT, CES2, ECD, ELMOD3, ETFDH, ETNK2, FAM20C, FASTKD2, FCAMR, GLIPR1, HNF4A, HPN, INSR, mir-548, miR-548c-3p (miRNAs w/seed AAAAAUC), MIS18BP1, MOCOS, MRPS14, MSRB1, NR3C1, PANK1, PTPRN2, RTN2, SH3BGRL2, SLC38A1, SLC43A1, TP53, USP29, USP42, VGLL2, ZNF155, ZNF175	15	8	Cancer, Organismal Injury and Abnormalities, Reproductive System Disease

Computed by IPA. Genes in bold are genes found to be hypervariable in our sample. Genes not in bold are related genes implicated in the network but not found to be hypervariable in our sample. The top 3 networks with scores of 15 and greater are presented.

Figure 3.

Top state-like network: Organismal development, cellular growth and proliferation, and digestive system development and function Network 1 as assessed by IPA. State-like genes are genes whose state of methylation varies among twin pairs within 3–6 months during adolescence. Genes in gray were hypervariable in our analyses. Genes in white were part of the network but not hypervariable.

Sex

Sex effects are not reported in this paper because MZ twins cannot vary in sex within a pair. While sex differences between twin pairs are certainly possible and scientifically interesting, in this paper we focused on the discordance between twin pairs. It is also possible that discordance would vary between male and female twin pairs, but we found no evidence for this proposition. In examining discordance at 174202 probes used to assess the trait-like probes, t-tests for the effect of sex revealed 8152 probes where the difference has an unadjusted P < 0.05. This represents 4.68% of the sample, slightly less than the 5% of the sample that would be expected by random chance. Adjusting these p-values for the false discovery rate suggested that only 2 of these probes should be considered significant, and they do not occur at probes with high discordance and do not affect our reported results. Results for the state-like probes were similar, with 4.91% of the sample showing sex effects at an unadjusted threshold of P = 0.05 and none of these probes surviving correction for the false discovery rate.

Discussion

We used the 450 K Illumina BeadChip Kit to profile DNA methylation states across the genome in saliva at one or 2 time points in a sample of 37 adolescent MZ twin pairs. Similar to previous studies using other Illumina BeadChips that cover fewer sites,^{14,16,17,19-22,25} we found that DNA methylation profiles were highly conserved across unrelated individuals and that this conservation was enhanced in MZ twins, presumably because of both their identical genome and their similar environment. This finding suggests high conservation of DNA methylation states during human evolution, which is consistent with the critical role of DNA methylation in defining cellular identities. In addition to the use of a BeadChip with greater coverage, we focused on methylation during adolescence, an under-studied period of great change with significant consequences for the rest of our lives. Furthermore, we demonstrated the convergence in DNA methylation in saliva, which may be sampled non-invasively and at lower cost in a greater number of people. Finally, our test-retest samples across a short period of time allowed us to assess the state- vs. trait-nature of specific genes, the results of which will be highly relevant for future studies.

The high conservation of DNA methylation in humans and the fact that identical twins revealed a high level of conservation is consistent with the view of “innate” evolutionary conserved and predetermined factors delineating DNA methylation states. At the same time, if DNA methylation is implicated in physiological responses to the environment, there should be sites in the genome where the state of methylation varies within an identical genetic background. We addressed this question by examining sets of identical twin pairs and identified genetically independent variability in DNA methylation in a subset of 226 genes. The following functional analysis suggested that these variations were not randomly distributed across the genome but were rather associated with various diseases (i.e., neurological, reproductive system, hematological, and metabolic); developmental, hereditary, and psychological disorders; tissue and cell morphology; development (i.e., organismal, embryonic, cellular, tissue and cardiovascular, muscular and skeletal system); and cellular mechanisms involving cellular movement, cell-to-cell signaling, and cell death and survival. Hypergeometric tests indicated that the assignments of the variable DNA methylation sites to particular genomic pathways were not random. This supports the idea that the human genome contains sites that are responsive to different extraneous signals that are particularly involved in nodal regulatory pathways, and is thus consistent with the view that the DNA methylome is adapted to signals from the environment.⁵⁶ The fact that these variable sites were common to many twin pairs and stable over time may mean that many of these changes occurred early in life and were then maintained throughout life. These kinds of DNA methylation changes are hypothesized to play a role in stable phenotypes that emerge in response to early life exposures.

Another critical question is whether this putative process of DNA methylation variation is stable or dynamic over short periods of time. In the present twin study, those genes that were the most dynamic or unstable over time were associated with similar, albeit fewer, networks involved in organismal development and developmental disorders, cellular growth and proliferation, as well as cell signaling and different diseases. This provides support for the hypothesis that the DNA methylome is highly responsive in adolescence to experience and extraneous signals. It should be emphasized that the stable and dynamic sites identified in our study were likely a conservative estimate of such variation in adolescence, given the limited environmental variation within twin pairs. It stands to reason that the variation in DNA methylation would be larger in the general population given the wider range of environmental exposures and life course experiences. Nevertheless, the present study distinguished genetic-innate variations from others and thus established the plausibility of this hypothesis.

In regards to hypervariable genes that were stable over time, it is of particular interest that we found multiple sites on several major histocompatibility complex genes (MHC), also known as the human leukocyte antigens (HLA) in humans. These genes are involved in immune functions,⁵⁷ and may be divided into 3 different classes (MHC Class I, II, and III). Among our hypervariable genes, we specifically found the HLA-C from Class I and the MHC complex II, HLA-DQA1, HLA-DQB1, HLA-DRB1, and HLA-DRB5 from Class II. All are implicated in presenting foreign antigens to the immune system.⁵⁷ Other studies have found DNA methylation of such genes in association with gastric cancer (HLA-C⁵⁸) and type 1 diabetes (HLA-DQB1 and HLA-DRB1⁵⁹). Interestingly, Ye and colleagues⁵⁸ found that HLA-C promoter methylation patterns were also associated with age and gender (higher methylation rates negatively associated with age in males). The present twin study suggests that environmental epigenetic processes may drive some of the variation in HLA functions (irrespective of DNA sequence) that are already associated with inter-individual differences in susceptibility to disease in adolescence.

Remarkably, the HLA-DQB1 gene came up as both variable in a stable manner and responsive in adolescence, although different sites were associated with stability (trait-like) and variation in time (state-like). A member of the MHC Class II, HLA-DQB1 provides instructions for making a protein with a critical role for the immune system and assists the immune system in distinguishing foreign invaders from the body's own proteins (RefSeq, Sep 2011) and has been involved in both celiac disease⁶⁰ and narcolepsy,⁶¹ again pointing to putative epigenetic-environmental origins for some of these vulnerabilities.

In addition to its strong design and the fact that this study had a very narrow age range specifically focused at the mid-adolescent period, a strength of this study is the assessment of variability in epigenetic patterns over a short period of time, thereby allowing for the identification of state vs. trait epigenetic marks. As shown in Ziller and colleagues,¹³ it appears that parts of the epigenome may be quite stable, whereas others may be much more dynamic across short periods of time. This is relevant information when designing a study of epigenetic mechanisms, particularly through time.

A limitation of this study is the use of only one tissue type. It is now well known that epigenetic patterns differ across tissues.^25,62 Is it worthwhile to assess methylation in a peripheral tissue as a marker of less accessible tissue such as the brain? If it is shown that epigenetic patterns can be assessed non-invasively using saliva, this will increase the feasibility of doing methylation studies on a large scale, particularly in samples in which obtaining blood samples is difficult (e.g., youth, newborns). Research to date suggests that some epigenetic variation may be found across tissues. For instance, Gordon and colleagues¹⁷ found that the most discordant genes across MZ twins are consistently discordant across both HUVECs and CBMs, but more work is needed to replicate and extend this finding. What's more, even one tissue type can contain different cell types that may contain divergent epigenetic patterns. A study by Talens and colleagues⁶³ assessing whether cellular heterogeneity in whole blood might explain inter-individual variability in DNA methylation patterns found no effect of monocyte percentage, but this issue of assessing methylation from peripheral tissue merits further consideration.

In line with this intra-tissue heterogeneity, the saliva samples we obtained contained a mixture of buccal epithelial cells and leukocytes, and DNA was extracted from both cell types. The proportion of these cell types can vary between individuals and over time, which introduced a known confound into our data. We attempted to remove this variability by comparing the methylation at each probe with a probe known to reliably distinguish the cell types, and to statistically remove the distinctive methylation patterns of the buccal epithelial cells, leaving us with methylation data that primarily represents the methylation of leukocytes.

We required that the variability of these probes be at least 3 standard deviations above the mean variability of the sample. Although another threshold could equally well have been used, the appeal of our chosen cutoff is twofold: 1) a threshold of 3 standard deviations is often used as a rule-of-thumb when assessing outliers in a data set and; 2) this threshold yields a list of hypervariable probes that are suitable for pathway analysis. A well-known limitation of pathway analysis software is that overlong lists of genes, even if they are selected at random, generate highly significant associations which are likely spurious. However, we think that our chosen cutoff maximized our chances of finding biologically relevant results.

Notwithstanding these limitations, this study extended observations from previous studies that DNA methylation patterns are highly similar in MZ twin pairs in the mid-adolescent period. It also demonstrated that this similarity is variable across pairs during adolescence, a period of great physiological maturation and psychosocial change. Furthermore, this study identified networks of genes that show the greatest discordance in adolescent MZ twin pairs, both in a trait-like (stable over a period of 3–6 months) and a state-like (variable across 3–6 months) pattern. Ideally, future studies should repeat this type of analysis across a range of tissues in order to simultaneously assess stability across tissue types.

Our study findings are consistent with the hypothesis that the human methylome evolved to consist of at least 3 classes of DNA methylation profiles. First, there are stable DNA methylation sites across individuals and time, which may be innately determined and are most probably involved in establishing cellular identity. Second, there were highly variable sites even in identical genetic backgrounds that may be responsive to external signals, but that remain stable through short periods of time and are presumably involved in establishing trait-like phenotypes. Third, there are highly variable sites in time that may respond to changes in external signals and experiences throughout the life course. These results are relevant for future studies assessing methylation variation in association with environmental events, as they identified stable sites that are likely to be of relevance and others that should be regarded with caution due to their dynamic nature.

Methods

Participants

Seventy-four MZ twins (37 pairs) who have been followed since birth as part of the Quebec Newborn Twin Study (QNTS;⁶⁴) cohort were recruited. Participants were 15 y old and consisted of 38 males and 36 females (19 and 18 same-sex twin pairs, respectively). All reported good current health, denied any history of medical or neurological illness, and were determined to be free of any current psychopathology. Presence or absence of current psychopathology was determined using the Dominic, a 15–20 minute computerized diagnostic interview designed for children and adolescents,⁶⁵ and the Kiddie-Schedule for Affective Disorders and Schizophrenia (K-SADS), a version of the semi-structured interview assessing DSM IV disorders designed for school-age children of 6–18 y.⁶⁶ The appropriate institutional ethics committees approved the study and all participants and parents signed informed assent and consent forms, respectively.

Saliva samples

Whole saliva was collected using the Oragene^TM DNA self-collection kit following the manufacturer's instructions (DNA Genotek Inc., 2004, 2006). Participants were asked not to eat, chew gum, or drink anything but water for 30 minutes before the samples were taken. Each participant was asked to provide 2 ml of saliva, which was mixed with 2 ml of the oragene solution, beginning the initial stage of DNA isolation and stabilizing the sample until extraction could be performed. Extraction was accomplished using the Promega Genomic DNA Purification kit, and sent to Genome Quebec for whole-genome analysis using Illumina. In a sub-sample of 8 twin pairs, we took a second saliva sample 3–6 months following the first in order to perform a test-retest analysis.

Illumina

We made use of the Illumina Infinium HumanMethylation450 BeadChip Kit, which covers more than 480,000 methylation sites per sample, including 96% of CpG islands, as well as additional coverage in island shores and surrounding regions, again at single-nucleotide resolution. Briefly, DNA was analyzed using the 450 K Illumina BeadChip Kit at the Genome Quebec Innovation Center. The manual protocol supplied by Illumina was followed for all steps except for Single Base Extension and Staining, which were conducted using the automated protocol. Briefly, the isolated DNA was first checked for quality with picogreen and then bisulfite-converted using the Zymo EZ-96 DNA Methylation-Gold Kit. Samples were transferred to BCD and then MSA4 plates, and neutralized before overnight amplification. MSA4 plates were fragmented, precipitated, and re-suspended before hybridization and transfer to Multi BeadChips. The Multi BeadChips then underwent washing, single-base extension, and staining, before imaging using the HiScan array scanner.

Data analysis

The raw Illumina output was processed using the R package minfi, a part of biocLite (http://bioconductor.org). The data were first read in and preprocessed (preprocessIllumina) by background correcting and normalizing the data. The main outcome measures were β-values at each probe, and a number ranging from zero to one, which represents the proportion of methylated samples, was detected. Next, each CpG was associated with a particular chromosome and gene based on the manifest files provided by Illumina (http://support.illumina.com/downloads/humanmethylation450_15017482_v1–2_product_files.ilmn). The β values and their positional information were then exported to MATLAB (http://mathworks.com, version 13a).

Cellular composition of saliva

In this protocol, DNA samples were collected from saliva. This has the advantage of being non-invasive, particularly in an adolescent population. However, the resulting DNA comes from 2 major cell types, leukocytes and buccal epithelial cells, and these cell types may differ in DNA methylation. Importantly, individual samples may differ in the proportions of these 2 cell types, which can bias results. A method for removing this confound has recently been proposed.¹⁹ Briefly, Souren and colleagues identified CpGs, which differentiated whole blood samples (including leukocytes) from samples of buccal epithelial cells, and found that the 2 cell types were best discriminated by methylation at cg18384097 in the PTPN7 gene. They then used methylation at that site as an index of the cell-type proportion, and fit a regression model between that probe and every other probe on the chip. In cases where the correlations were high, the probe values were replaced by the regression residuals, giving a dataset that is linearly independent of this index of cell-type proportion. In our data set we fit a regression model 1 at every probe.

1. ${\displaystyle \text{CpGβvalues}={\text{β}}_1+{\text{β}}_2\text{cg}18384097}$

In cases where β₂ significantly contributed to the model (P < 0.05), the values at that probe were replaced by the raw regression residuals. This fitting was done twice. In the first case, the dataset included a set of technical replicates (3 samples processed 3 times each) and the values for these replicates could be more accurately estimated by developing the regression equation in a larger data set. In the second case, only one sample per subject was included, and this data set was used for further analysis.

Assessment of test-retest variability

After the data was adjusted for the ratio of leukocytes to buccal epithelial cells, the values from the replicated samples were isolated and test-retest differences were calculated for each pair of samples for a given individual (sample A - sample B, sample A - sample C, sample B - sample C). This allowed us to calculate both the maximum observed pairwise difference, and a standard deviation for this difference distribution. These numbers were used in the data filtering steps below.

Data filtering

After removing the replicates and technical control samples from our dataset, we had a matrix of 482,421 probes by 74 participants (37 twin pairs). As a first step we replaced or removed missing values. There were a total of 955 missing values in the data set; whenever possible, an individual's missing value was replaced with the value of their twin. In cases where data from both twins was missing the mean of the entire sample was used. This could introduce a slight bias, and cause us to over-estimate twin-similarity, but as the main purpose of the study was to assess within-twin pair variability, this method tends to weight against finding effects, and allowed us to keep those probes in the dataset (0.2% of probes). We next removed data from probes where the maximum observed test-retest difference was larger than the maximum difference between data points at that probe (max(probe)-min(probe)). This excluded 13,642 probes from further analysis. We then removed the 11,135 probes on the X chromosome and the 416 probes on the Y chromosome. Because our aim was to examine highly variable pathways and networks, we restricted our search to probes associated with known genes according to the Illumina manifest. This removed 117,778 probes from the data set. Because methylation is not necessarily stable over time, we took advantage of test-retest in 16 individuals (8 twin pairs), in which a second saliva sample was collected approximately 3–6 months following the first. After correction for differences in the buccal epithelial cell content of the sample, we compared the temporal stability of each probe. We were interested in distinguishing between temporally stable probes that might contribute to more trait-like phenotypes, and temporally unstable probes that might be more closely associated with state-like phenotypes. This analysis identified 179,408 temporally stable probes (37% of the original dataset) and 241,211 temporally unstable probes (50% of the original data set). Although it is common in analyses of Illumina microarrays to omit probes whose hybridization could be disrupted by common SNPs, our experimental design based on monozygotic twin pairs excludes that possibility, so these probes were not excluded.

Z-scores

The Z-scores of absolute twin differences were calculated, and probes more than 3 standard deviations above the mean were considered to be hypervariable. This yielded a list of 250 mapped probes, associated with 226 unique genes. These gene names were further processed using Ingenuity Pathway Analysis. The core analysis procedure was used with default options.

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed.

Acknowledgments

We thank Ms. Bianca Zuccarini with her help with the review of literature and Ms. Jennifer Gillies, BA, for editing the manuscript. We are grateful to the children and parents of the Quebec Newborn Twin Study (QNTS), and the participating teachers and schools. We thank Jocelyn Malo and Marie-Élyse Bertrand for coordinating the data collection, as well as Hélène Paradis and Nadine Forget-Dubois for managing the data bank of QNTS.

Funding

The study was supported by an operating grant from the Canadian Institutes of Health Research (CIHR) awarded to Drs. Booij, Szyf, Tremblay, and Boivin. Mélissa Lévesque, M.Sc was funded by doctoral awards from CHU Sainte-Justine - Foundation of Stars and from the Fonds de recherche en santé-Québec. Dr. Linda Booij was supported by a New Investigator Award from CIHR. Dr. Michel Boivin was supported by the Canada Research Chair Program. QNTS was supported by various grants received over the years from the Fonds Québécois de la Recherche sur la Société et la Culture (FQRSC), the Fonds de la Recherche en Santé du Québec (FRSQ), the Social Science and Humanities Research Council of Canada (SSHRC), the National Health Research Development Program (NHRDP), the Canadian Institutes for Health Research (CIHR), Ste. Justine Hospital's Research Center, Université Laval and Université de Montréal.

Supplemental Material

Supplemental data for this article can be accessed on the publisher's website.