Antenatal Prediction of Postpartum Depression with Blood DNA Methylation Biomarkers

ORIGINAL ARTICLE Antenatal prediction of postpartum depression with blood DNA methylation biomarkers

J Guintivano1, M Arad2, TD Gould2,3,4, JL Payne1 and ZA Kaminsky1

Postpartum depression (PPD) affects B10–18% of women in the general population and results in serious consequences to both the mother and offspring. We hypothesized that predisposition to PPD risk is due to an altered sensitivity to estrogen-mediated epigenetic changes that act in a cell autonomous manner detectable in the blood. We investigated estrogen-mediated epigenetic reprogramming events in the hippocampus and risk to PPD using a cross-species translational design. DNA methylation profiles were generated using methylation microarrays in a prospective sample of the blood from the antenatal period of pregnant mood disorder patients who would and would not develop depression postpartum. These profiles were cross-referenced with syntenic locations exhibiting hippocampal DNA methylation changes in the mouse responsive to long-term treatment with 17b-estradiol (E2). DNA methylation associated with PPD risk correlated significantly with E2-induced DNA methylation change, suggesting an enhanced sensitivity to estrogen-based DNA methylation reprogramming exists in those at risk for PPD. Using the combined mouse and human data, we identified two biomarker loci at the HP1BP3 and TTC9B genes that predicted PPD with an area under the receiver operator characteristic (ROC) curve (area under the curve (AUC)) of 0.87 in antenatally euthymic women and 0.12 in a replication sample of antenatally depressed women. Incorporation of blood count data into the model accounted for the discrepancy and produced an AUC of 0.96 across both prepartum depressed and euthymic women. Pathway analyses demonstrated that DNA methylation patterns related to hippocampal synaptic plasticity may be of etiological importance to PPD.

Molecular Psychiatry (2014) 19, 560–567; doi:10.1038/mp.2013.62; published online 21 May 2013 Keywords: biomarker; DNA methylation; estrogen; HP1BP3; postpartum depression; TTC9B

INTRODUCTION that is triggered by gonadal hormone withdrawal. DNA Postpartum depression (PPD) occurs in approximately 10–18% of methylation may represent the link between estrogen and its women and results in significant morbidity in both mother and effects on mood. Indeed, it has previously been demonstrated that child, with offspring risks including low self-esteem, low intellec- E2 administration in vitro can modify DNA methylation at multiple 12 tual skills, child abuse and infanticide.1–6 Women with mood locations downstream of an estrogen-response element. disorders are at an increased risk of PPD;7 however, the benefits of Given that fluctuations in estrogen coincide with PPD psychiatric treatment must be carefully weighed against the symptoms and can be antidepressant when administered as a 7,13–16 potential risks of in utero exposure of the offspring to treatment. treatment, we hypothesized that predisposition to PPD risk Antidepressant treatment during pregnancy can result in is due to an altered sensitivity to estrogen-mediated epigenetic increased miscarriage rates in early pregnancy and have been changes that act in a cell autonomous manner detectable in the associated with low birth weight, preterm birth and birth defects blood. In this study, we performed a multitiered translational with some classes of antidepressants.8 Limited information is approach to predict PPD status in a prospective cohort using DNA available on the long-term neurocognitive effects of in utero methylation from both the human blood and the hippocampus antidepressant exposure.8 of mice administered E2. We first define genomic regions of PPD occurs up to 4 weeks following parturition according to the E2-mediated epigenetic change in E2-treated mice and Diagnostic and Statistical Manual of Mental Disorders, Fourth investigated the relationship between E2-induced DNA Edition (DSM-IV) criteria and follows a dramatic drop in the methylation and PPD risk at syntenic regions in humans. Finally, circulating levels of estradiol (E2) and progesterone (P4). Although we use E2-induced methylation models generated in the mice to depression risk is not predicted by serum levels of gonadal predict PPD status in the humans. hormones in humans,9 rapid withdrawal from these hormones appears to be an important factor in establishing PPD. In an important experiment, women with a previous history of PPD MATERIALS AND METHODS subjected to supraphysiological doses of E2 and P4 experienced Experimental animals significantly depressed mood symptoms relative to controls upon C57BL/6J mice were ovariectomized at 8 weeks of age. At the time of 10,11 hormone withdrawal, suggesting that the at-risk population surgery, mice were randomized to receive (subcutaneous implantation) a exhibits a predisposition to PPD through unknown mechanisms Silastic capsule (i.d. 1.02 mm; o.d. 2.16 mm) containing 5 mm of dry-packed

1The Mood Disorders Center, Department of Psychiatry and Behavioral Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, USA; 2Department of Psychiatry, University of Maryland School of Medicine, Baltimore, MD, USA; 3Department of Pharmacology, University of Maryland School of Medicine, Baltimore, MD, USA and 4Department of Anatomy and Neurobiology, University of Maryland School of Medicine, Baltimore, MD, USA. Correspondence: Dr Z Kaminsky, The Mood Disorder Center, Department of Psychiatry and Behavioral Sciences, Johns Hopkins University, 720 Rutland Avenue, Ross Research Building 1070, Baltimore, MD 21205, USA. E-mail: [email protected] Received 8 February 2013; accepted 2 April 2013; published online 21 May 2013 Postpartum depression prediction with DNA methylation biomarkers J Guintivano et al 561 17b-E2 (n ¼ 5 per group per timepoint). Controls received an empty the mean DNA methylation values from two replicates of the same sample capsule. Analysis of serum demonstrated consistent levels of E2 in the run in the prepartum depressed cohort (batch 2). blood and at 1, 2 and 4 weeks, which was predictive of an increase in uterus weight over those time points (Supplementary Figure 1). Sodium bisulfite pyrosequencing Bisulfite conversion was carried out using EZ DNA Methylation Gold Kit Affymetrix DNA methylation profiling (Zymo Research, Irvine, CA, USA) according to the manufacturer’s instructions. Nested polymerase chain reaction amplifications were DNA methylation was assessed in mice using methods described performed with a standard polymerase chain reaction protocol in 25-ml previously17,18 using HpaII and HinP1I enzymes. Following quality control volume reactions containing 3–4 ml of sodium-bisulfite-treated DNA, 0.2 mM assessment through Agilent BioAnalyzer-based visualization, the primers and master mix containing Taq DNA polymerase (Sigma-Aldrich, unmethylated fraction of genomic DNA was hybridized to Affymetrix St Louis, MO, USA). Primer sequences can be found in Supplementary GeneChip Mouse Tiling Promoter 1.0R Arrays at the JHMI Deep Sequencing Table 2. Polymerase chain reaction amplicons were processed for and Microarray Core facility (Baltimore, MD, USA). Affymetrix CEL files were pyrosequencing analysis according to the manufacturer’s standard protocol background corrected and quantile normalized using the AffyTiling (Qiagen, Gaithersburg, MD, USA) using a PyroMark MD system (Qiagen) with package in R, yielding normalized log -transformed M values represen- 2 Pyro Q-CpG 1.0.9 software (Qiagen) for CpG methylation quantification. tative of the DNA hypomethylation profile per sample. Differentially methylated regions (DMRs) were calculated using the BioTile algorithm (http://psychiatry.igm.jhmi.edu/kaminsky/software.htm).19 Identified DMRs Statistical analysis were refined by filtering out those not flanked within 1 kb of the DMR All statistical tests were performed in R (http://www.r-project.org/). Using boundary by either an HpaII or HinP1I restriction site based on the mouse an Anderson–Darling test from the nortest package, all distributions of mm8 genome build sequence. Microarray data is located under GEO data that rejected the null hypothesis of normality were subsequently accession no.: GSE43460. evaluated with non-parametric tests. All statistical tests performed were two-tailed and a Po0.05 is considered significant. Unless otherwise specified, ‘±’ denotes the standard error of the mean. Human sample We recruited 93 pregnant women with a history of either major depression Weighted genome coexpression network analysis or bipolar disorder (I, II or not otherwise specified) and prospectively 21 followed them during pregnancy and after delivery to identify genetic and Weighted genome co-expression network analysis (WGCNA) was clinical characteristics that precede the development of a postpartum performed using the WGCNA package in R. In the mouse comparisons, depressive episode. Approximately one-third of the sample had bipolar 3606 mean DMR values were used with a power of 20 and minimum disorder. The average age of the participants was 30.6 years and 70% of module size of 10. For all human analyses, 13 091 nominally significant loci the sample was Caucasian. Participants were managed by their treating in the combined comparison of PPD (N ¼ 11) to non-PPD (N ¼ 21) psychiatrist as clinically indicated and were evaluated during each euthymic cohort women were used for correlation with a power of 10 trimester for pregnancy, and then 1 week, 1 month and 3 months and minimum module size of 10. postpartum. Women were classified as being depressed if they met the DSM-IV criteria for a major depressive episode based on a psychiatric interview at each time point (first, second and third trimester and 1 week RESULTS and 1 month postpartum). We analyzed a subgroup of 32 women Identification of hippocampal targets of E2-mediated DNA euthymic during the third trimester (prepartum euthymic), 34.4% of this methylation change subsample (N ¼ 11) became depressed within the first 4 weeks postpartum We sought to identify hippocampal DMRs in the mouse associated and met the DSM-IV criteria for major depressive episode. A second with E2 exposure to model the molecular changes occurring subgroup of 19 women depressed during pregnancy (prepartum depressed) was assessed in subsequent analyses as an independent during heightened estrogen levels in pregnancy. We chose to replication cohort, of which N ¼ 12 remained depressed within the first 4 utilize hippocampal tissue because effects of E2 on mood are weeks postpartum and met the DSM-IV criteria for major depressive believed, in part, to be localized in the hippocampus, based on episode. The trimester of blood draw is depicted in Supplementary Table 1. numerous studies including knockout experiments,22 17b-E2 administration experiments23 and selective estrogen receptor antagonists and agonists13,24,25 demonstrate anxiolytic and Illumina DNA methylation profiling antidepressant effects of E2 exposure in rodents. Furthermore in Samples quality assessment and microarray analysis were conducted at rodent models, E2 administration has been shown to increase The Sidney Kimmel Cancer Center Microarray Core Facility at Johns synaptic plasticity and dendritic spine density within the Hopkins University using Illumina’s Infinium Human Methylation450 hippocampus,14,15 while withdrawal from pregnancy levels of E2 Beadchip Kit (WG-314-1001; Illumina Inc., San Diego, CA, USA) according to the manufacturer’s manual. Images were processed in Illumina’s iScan results in decreased hippocampal brain-derived neurotrophic factor (BDNF) expression16 and suppressed hippocampal scanner (Illumina Inc.) and data were extracted using Methylation Module 26 of GenomeStudio v.1.0 Software (Illumina Inc.). Illumina probe type was neurogenesis. We identified 891 significant DMRs before corrected using the Beta2M function in the wateRmelon package in R. correction for multiple testing. Of these, 380 DMRs exhibited a Methylation status of each CpG site was calculated as b (beta)-value based decrease and 511 exhibited an increase in DNA methylation in on following definition: response to E2 (Figure 1 and Supplementary Table 3). Gene 27 b-Value ¼ (signal intensity of methylation À detection probe)/(signal ontology (GO) analysis using GOstat identified a number of intensity of methylation À detection probe þ signal intensity of non- significantly enriched GO categories within genes proximal to the methylation À detection probe þ 100). identified DMRs (Table 1). Motif enrichment analysis of genomic Microarray data is located under GEO accession no.: GSE44132. sequences of the top 100 significant E2 DMRs as well as estrogen receptor-b promoter methylation-correlated regions identified an enrichment for the SP-1 and estrogen receptor transcription Cell subtype analysis factor-binding motifs (Supplementary Figure 2). We quantified cell subfraction percentages for CD8 T cells, CD4 T cells, B cells, monocytes and granulocytes by inputting DNA methylation PPD DNA methylation differences are correlated with E2-mediated signatures of 473 loci into an algorithm designed for quantification of the epigenetic change above cell types using DNA methylation proxies from HM450 arrays.20 Before cell-type proportion calculation for the prepartum depressed cohort, We split the human sample into a discovery sample and DNA methylation values at the 473 loci were transformed by subtracting replication sample consisting of N ¼ 6 and 5 women who would the residuals from a linear model of the mean DNA methylation values of and N ¼ 12 and 9 who would not develop PPD, each with B35% three cross-batch controls from the prepartum euthymic cohort (batch 1) vs PPD to 65% non-PPD samples. In the discovery sample, we

& 2014 Macmillan Publishers Limited Molecular Psychiatry (2014), 560 – 567 Postpartum depression prediction with DNA methylation biomarkers J Guintivano et al 562

Figure 1. Estradiol (E2)-mediated DNA methylation change is associated with postpartum depression (PPD) risk. (a) Volcano plot depicting the difference in DNA methylation between women who suffered PPD vs those who did not (x axis) against the negative natural log of the P-value of association between groups (y axis). (b) A volcano plot depicting DNA methylation differences between the ovariectomy (OVX) and OVX þ E2 groups per DMR (x axis) and the—natural log of the P-value for each comparison. Horizontal red lines depict the signiﬁcance threshold of 5%. (c) Scatter plot of the –log of the P-value of association to discovery sample PPD status and the effect size of E2-mediated DNA methylation change at 103 overlapping loci nominally signiﬁcant in both humans and mice. (d) Scatter plot of the difference between PPD and non-PPD women in the discovery sample (y axis) as a function of that in the replication sample (x axis).

Table 1. Over-represented GO categories in E2-responsive DMRs

GO category N identiﬁed N in group Corrected P-value Category

Methylation decrease GO: 0001505 10 88 0.005 Synaptic transmission: regulation of neurotransmitter levels GO: 0003001 9 87 0.020 Generation of a signal involved in cell–cell signaling GO: 0007267 17 325 0.027 Cell–cell signaling GO: 0051640 5 27 0.037 Organelle localization GO: 0045941 15 281 0.041 Positive regulation of nucleobase

Methylation increase GO: 0043231 144 6207 8.6 Â 10 À 5 Intracellular membrane-bound organelle GO: 0005622 187 8708 8.6 Â 10 À 5 Intracellular GO: 0043227 144 6212 8.6 Â 10 À 5 Membrane-bound organelle GO: 0005737 126 5314 2.0 Â 10 À 4 Cytoplasm GO: 0044424 179 8417 3.4 Â 10 À 4 Intracellular part GO: 0043229 154 7027 5.9 Â 10 À 4 Intracellular organelle GO: 0043226 154 7032 5.9 Â 10 À 4 Organelle GO: 0016192 18 401 2.3 Â 10 À 3 Vesicle-mediated transport GO: 0044444 78 3058 2.4 Â 10 À 3 Cytoplasmic part GO: 0005794 22 552 3.0 Â 10 À 3 Golgi apparatus GO: 0005515 109 4751 4.6 Â 10 À 3 Protein binding GO: 0043283 93 4055 0.020 Biopolymer metabolic process Abbreviations: DMR, differentially methylated region; E2, estradiol; GO, gene ontology.

performed a probe-wise Student’s t-test between PPD and non- microarray (Figure 1b). Synteny was calculated based on the PPD cases. We cross-referenced genomic locations of the E2 DMRs relative position of the implicated DMR (Mouse array) or individual from the mouse data with syntenic loci located on the human CpG locus (Human array) from the closest proximal transcription

Molecular Psychiatry (2014), 560 – 567 & 2014 Macmillan Publishers Limited Postpartum depression prediction with DNA methylation biomarkers J Guintivano et al 563 start site within conserved sequence regions, as established by the cg21326881 and cg00058938, corresponding to the promoter UCSC Genome Browser Liftover tool. Owing to the nature of the regions of the HP1BP3 and TTC9B genes, respectively, which enzymatic enrichment used in the mouse array experiment, a CpG resulted in an AUC of 0.92 in the discovery sample and 0.9 in the locus was considered overlapping if it was adjacent within 200 bp replication sample (Supplementary Figure 3b). A genome-wide of the implicated DMR. In total, 1578 human CpGs were located significance for this biomarker set of P ¼ 0.041 was determined by within the nominally significant mouse E2 DMRs. Pathway analysis Monte Carlo permutation analysis. of genes associated with overlapping loci using the g.Profiler 28 analysis suite identified a single significant GO category Pyrosequencing validation of identified biomarkers (GO: 0010646, frequency observed ¼ 0.19, expected ¼ 0.024, We performed sodium bisulfite pyrosequencing to validate the P ¼ 0.046) for ‘regulation of cell communication’ and an microarray findings in the human sample at CpG dinucleotides enrichment of SP-1 (M00196_4, frequency observed ¼ 0.51, located within the region chr 1: 20 986 692–20 986 676 (strand À , expected ¼ 0.021, P ¼ 0.0084) and AP-2 (M00800_3, frequency human genome build hg18) and chr 19: 45 416 573 (strand þ , observed ¼ 0.54, expected ¼ 0.021, P ¼ 0.0029) transcription human genome build hg18), located upstream of HP1BP3 and factor-binding motifs. TTC9B, respectively. PPD status was significantly associated with We next attempted to correlate the mean DNA methylation the HP1BP3 microarray and pyrosequencing data and was difference between PPD and non-PPD samples and E2-mediated significantly correlated between methods (Figures 2a–c and DNA methylation fold change. No correlation was observed across Table 2). DNA methylation for TTC9B was significantly associated the 1578 overlapping loci (Spearman’s r ¼À0.028, P ¼ 0.27). We with PPD status for both the microarray and pyrosequencing data refined the interrogated data set to 103 loci exhibiting nominally and was significantly correlated between the two methods significant association to PPD status and observed significant (Figures 2e–g and Table 2). correlations in both the discovery sample (Spearman’s r ¼ 0.21, Using HP1BP3 and TTC9B pyrosequencing values in the P ¼ 0.030) and the replication sample (Spearman’s r ¼ 0.2, prediction linear discriminant model, we obtained an AUC of P ¼ 0.042). The P-value of association to PPD in the discovery 0.87 for the prepartum euthymic sample, which included three sample was also correlated with E2 DMR effect size (Spearman’s additional women not assessed via microarray (PPD N ¼ 13, non- r ¼À0.19, P ¼ 0.05) (Figure 1c), suggesting that more robust PPD PPD N ¼ 22). AUC values did not vary significantly when associations occur at targets of larger E2-mediated DNA methyla- determined for blood collected in each trimester separately tion change. Furthermore, the mean PPD minus non-PPD value (AUC 1st ¼ 0.86, AUC 2nd ¼ 0.80, AUC 3rd ¼ 1). We next evaluated was significantly correlated across the discovery and replication the performance of the biomarker loci on blood taken from the cohorts (Spearman’s r ¼ 0.32, P ¼ 0.0011) (Figure 1d). Permutation prepartum depressed sample. While the relative direction of TTC9B testing (20 000 iterations) demonstrated that randomly selected association with PPD status was similar to the prepartum euthymic groupings of 103 loci did not correlate better between cohorts women, it was not significantly different (Figure 2h and Table 2). (P ¼ 5 Â 10 À 5) nor with E2 DMR fold changes in either the For HP1BP3 the direction of association was significant but in the discovery or replication samples (P ¼ 0.016 and 0.02, respectively). opposite direction to that observed in the prepartum euthymic This analysis suggests that the degree to which the discovery cohort (Figure 2d and Table 2). Linear discriminant model and replication cohorts agree is strongly influenced by their prediction of PPD status in this cohort returned an AUC of 0.12, localization to syntenic regions of E2-mediated epigenetic which represents an 88% prediction accuracy of PPD status in the reprogramming. reverse direction. We evaluated the mean PPD minus non-PPD DNA methylation status at the nominally significant PPD associations in the prepartum depressed cohort (N ¼ 103 loci) and identified a trend Biomarker replication is influenced by blood cellular heterogeneity for a positive correlation with the fold change at syntenic E2 DMRs We hypothesized that the discrepancy between the prepartum (Spearman’s r ¼ 0.19, P ¼ 0.054). A positive correlation of mean euthymic and depressed cohorts may be related to differences in methylation difference between the 1578 loci marked as E2 blood cell-type counts between the two groups. Various experi- responsive was also observed between the prepartum depressed ments have identified elevated granulocytes and decreased CD8 and euthymic cohorts (Spearman’s r ¼ 0.078, P ¼ 0.002). Cumula- and CD4 T-cell and associated cytokine profiles in individuals tively, these results support our previous hypothesis that PPD risk exhibiting depressed mood.29,30 Using DNA methylation proxies in may be mediated by an enhanced sensitivity to E2-mediated the 19 prepartum depressed and 32 prepartum euthymic women, epigenetic reprogramming. we determined that cell-type proportions of CD8 T cells, CD4 T cells, B cells and monocytes were significantly reduced in the depressed prepartum group, whereas cross-batch controls Identification of DNA methylation biomarkers predictive of PPD exhibited nonsignificant differences in the opposite direction We next reasoned that if estrogen is important for PPD risk, we (Supplementary Table 4). Pyrosequencing DNA methylation values should be able to predict PPD status based on the degree to for HP1BP3 were evaluated against all cell types in an additive which E2 reprograms DNA methylation in the mouse. For each of linear model and identified a trend with monocyte proportions the 1578 mouse E2 DMRs that overlapped with the human data (b ¼À1.11±0.6, P ¼ 0.07). We subsequently evaluated the ratio set, we modeled the mean DNA methylation signature per DMR of monocytes to the summed proportions of CD8 T cells, CD4 T against the E2 treatment status. In a locus-specific manner, we cells, B cells and granulocytes, and observed a significant inputted the human DNA methylation levels per individual in the association with prepartum depression status (cell ratio, discovery sample and attempted to predict PPD status using depressed ¼ 0.021±5.2 Â 10 À 4, euthymic ¼ 0.032±3.3 Â 10 À 4, logistic regression. For each locus, the area under the curve (AUC) P ¼ 2.1 Â 10 À 4) but not PPD status (cell ratio, PPD ¼ 0.028±5 metric was used to measure prediction accuracy. We then Â 10 À 4, non-PPD ¼ 0.028±4.2 Â 10 À 4, P ¼ 0.86) (Figure 3a), the attempted to combine biomarkers to increase predictability using distribution of which was similar but opposite to that of HP1BP3 the following algorithm (Supplementary Figure 3a). Linear DNA methylation (Figure 3b). This monocyte to non-monocyte discriminant analysis was used to combine loci in a forward cell-type ratio was negatively correlated with DNA methylation of step-wise manner such that model included loci were those that HP1BP3 (Spearman’s r ¼À0.37, P ¼ 0.0074) (Figure 3c), whereas increased the AUC of the discovery sample until the value was TTC9B was not associated (Spearman’s r ¼À0.22, P ¼ 0.11). Linear maximized. This set of loci was then used to predict PPD status in regression modeling was performed for PPD diagnosis against an the replication sample. The algorithm returned two loci at CpGs interaction of HP1BP3 DNA methylation with the cell-type ratio,

& 2014 Macmillan Publishers Limited Molecular Psychiatry (2014), 560 – 567 Postpartum depression prediction with DNA methylation biomarkers J Guintivano et al 564

Figure 2. Validation of biomarker loci. Boxplots of the percentage of DNA methylation in the non-postpartum depression (PPD) and PPD groups for HP1BP3 microarray (a) and pyrosequencing (b) and TTC9B microarray (e) and pyrosequencing (f) values. Scatter plots of the % DNA methylation difference between PPD minus non-PPD samples in the prepartum euthymic sample obtained by pyrosequencing (y axis) and microarray (x axis) is depicted for HP1BP3 (c) and TTC9B (g). Boxplots of the percentage of DNA methylation in the non-PPD and PPD groups for HP1BP3 pyrosequencing (d) and pyrosequencing (h) values obtained from the independent replication cohort of prepartum depressed women.

inﬂammation. This effect appeared to be driven by a positive Table 2. Descriptive statistics of biomarker loci correlation of WBC count with granulocyte proportion (Spear- À 16 Method PPD Non-PPD P-value man’s r ¼ 0.92, P ¼ 2.2 Â 10 ), which is consistent with the above-cited elevations in granulocyte levels with depression.29 Prepartum euthymic The ratio of CBC-derived monocyte to non-monocyte HP1BP3 Microarray 0.08±0.0012 0.067±0.00034 0.0012 (lymphocytes and granulocytes) ratio did not correlate with Pyrosequencing 0.063±0.0012 0.045±0.00095 0.046 Method Spearman’s 0.018 those derived by DNA methylation proxy (Spearman’s r ¼ 0.24, correlation r ¼ 0.41 P ¼ 0.36). We limited the analysis to only those 11 samples where TTC9B Microarray 0.42±0.0036 0.48±0.0021 CBC data was derived from within the same trimester as the blood Pyrosequencing 0.30±0.0055 0.38±0.0034 0.0046 Method Spearman’s 2.6 Â 10 À 8 draw used for microarray analysis and observed a signiﬁcant Correlation r ¼ 0.81 correlation (Spearman’s r ¼ 0.66, P ¼ 0.044). We attempted to predict PPD status via bootstrap analysis across all 17 individuals Prepartum depressed using the linear model generated above with CBC data-based HP1BP3 Pyrosequencing 0.05±0.0017 0.081±0.0024 0.0072 TTC9B Pyrosequencing 0.32±0.0035 0.33±0.0079 0.84 monocyte to non-monocyte ratios in place of proxy-based ratios and generated a highly accurate prediction of PPD status Abbreviation: PPD, postpartum depression. (AUC ¼ 0.96) (Figure 3d).

with TTC9B DNA methylation as a covariate. The model was Functional classification of HP1BP3 and TTC9B significantly associated with PPD (R2 ¼ 0.38, P ¼ 1.9 Â 10 À 4), as We attempted to ascertain the function of HP1BP3 and TTC9B loci were all model terms including DNA methylation of HP1BP3 bioinformatically by using the STRING database31 (Supplementary (b ¼À0.22±0.075, P ¼ 0.0044), TTC9B (b ¼À0.033±0.0081, Figure 4) and by performing weighted gene coexpression network P ¼ 1.6 Â 10 À 4), the cell-type ratio (b ¼À49.66±14.64, analysis21 against DNA methylation of the HP1BP3 CpG P ¼ 0.0014), and the interaction between HP1BP3 DNA methylation (cg21326881) and TTC9B CpG (cg00058938) as the target and cell-type ratio (b ¼ 8.03±2.4, P ¼ 0.0016). Using a boot- variables for correlation (Supplementary Figure 5). The resulting strapping method, we predicted PPD status for each individual networks of HP1BP3- and TTC9B-coregulated genes were strikingly using the linear model and obtained an AUC of 0.82 (Figure 3d). anticorrelated (Spearman’s r ¼À0.76, P ¼ 2.2 Â 10 À 16), suggest- Importantly, the cell proxy analysis only takes into account the ing that HP1BP3 and TTC9B are coregulated (Supplementary relative percentage of various cell types, but not the overall white Figure 5b). Resultantly, we limited networks to those demonstrat- blood cell (WBC) count. Where available, prepartum WBC counts ing significant non-parametric correlation between module and proportions of lymphocytes, granulocytes and monocytes membership and correlation significance per group and identified were obtained from complete blood count (CBC) data (N ¼ 17 two coregulated networks significantly associated across both women). CBC derived total WBC counts were negatively correlated genes (Module 1: HP1BP3 r ¼ 0.56, P ¼ 8.8 Â 10 À 4, TTC9B with the proxy-derived monocyte to non-monocyte ratio (Spear- r ¼À0.54, P ¼ 0.0015; Module 2: HP1BP r ¼ 0.45, P ¼ 0.0087, man’s r ¼À0.7, P ¼ 0.02), suggesting that the decreased cell-type TTC9B r ¼À0.46, P ¼ 0.0081) (Supplementary Figures 5c and d). ratio observed in the prepartum depressed group may be No significantly enriched pathways were identified by g.Profiler in indicative of elevated WBC counts and depression-associated Module 2; however, Module 1 contained a number of significantly

Molecular Psychiatry (2014), 560 – 567 & 2014 Macmillan Publishers Limited Postpartum depression prediction with DNA methylation biomarkers J Guintivano et al 565

Figure 3. Cell proportion and biomarker DNA methylation predict postpartum depression (PPD). (a) Boxplot of the ratio of monocyte percentage over the sum of T-cell, B-cell and granulocyte percentages as a function of prepartum depression status and PPD diagnosis. (b) Boxplot of the HP1BP3 DNA methylation percentage as a function of prepartum depression status and PPD diagnosis. (c) Scatterplot of the ratio of monocyte percentage over the sum of white blood cell and monocyte percentages as a function of the HP1BP3 DNA methylation percentage. (d) Receiver operator characteristic (ROC) curve of the sensitivity (y axis) vs specificity (x axis) of PPD prediction from the linear model of the HP1BP3 DNA methylation and cell-type ratio interaction, with TTC9B DNA methylation as a covariate. The solid line represents the ROC curve from the proxy-based cell proportion measurement and the dashed line represents that of the complete blood count-derived subsample. enriched KEGG (Kyoto Encyclopedia of Genes and Genomes) important to consider that the sample sizes interrogated in the pathways that can be tied to the antidepressant functions of E2 in mouse experiments were small, and that higher powered the hippocampus (Supplementary Table 5). experiments may identify additional genomic regions of We applied weighted genome coexpression network analysis E2-responsive DNA methylation change in the hippocampus. The within the PPD and non-PPD women separately, as well as within findings of enriched SP-1 binding sites and increased evidence the mouse E2 DMR data to ascertain the normal coregulation for hippocampal long-term potentiation-associated genes in pattern of HP1BP3 and TTC9B genes. The pattern of gene E2-responsive DMRs is consistent with the known downstream coregulation was positively correlated between HP1BP3 and TTC9B transcription factor activation32–35 as well as antidepressant in non-PPD cases and mice, but anticorrelated in PPD cases functions of E2 exposure in the hippocampus,36 and adds (Supplementary Figure 6), and is consistent with the proposed confidence to the assertion that we are detecting true E2 DMRs. heightened sensitivity to E2-mediated epigenetic reprogramming CpG methylation levels at two loci within the HP1BP3 and TTC9B in the PPD group. genes were identified as biomarkers predictive of PPD. Both genes have ties to estrogen signaling, as HP1BP3 was identified to associate with estrogen receptor b based on tandem affinity DISCUSSION purification assays performed on MCF-7 breast cancer cells37 and We addressed the hypothesis that regions of E2-mediated TTC9B expression has been shown to be responsive to gonadal epigenetic change may predict PPD risk. Numerous correlations hormones.38 Owing to the circulating nature of estrogen, the linking E2-mediated epigenetic change with DNA methylation identification of these markers in peripheral blood may be a marker changes occurring in the PPD risk population were identified in of estrogen-mediated epigenetic changes occurring in the both the original prepartum euthymic cohort as well as in the hippocampus and potentially conferring risk to phenotype based independent replication cohort of women depressed during on its actions in the brain. The functional relevance of TTC9B may be pregnancy. Cumulatively, the results suggest a systematic increase linked to hippocampal synaptic plasticity as tetratricopeptide repeat in DNA methylation change occurs in the blood of the PPD group containing domains such as that found in TTC9B have been shown during a period where pregnancy hormones are at high levels. As to inhibit HSP90-mediated trafficking of AMPA (a-amino-3-hydroxy- gonadal hormone levels have been shown not to predict PPD risk, 5-methyl-4-isoxazolepropionic acid receptor) receptors critical for these data provide suggestive evidence that the underlying risk in hippocampal long-term potentiation/long-term depression.39 this group may be related to an increased sensitivity for epigenetic Although there have been numerous attempts to generate change in response to normal levels of circulating hormones. It is biomarkers for PPD,40–45 few studies report a high prediction

& 2014 Macmillan Publishers Limited Molecular Psychiatry (2014), 560 – 567 Postpartum depression prediction with DNA methylation biomarkers J Guintivano et al 566 accuracy. To our knowledge, the identified biomarkers represent NARSAD 2010 Young Investigator Award to ZK. The Sidney Kimmel Cancer Center the first prospective epigenetic biomarkers capable of predicting Microarray Core Facility at Johns Hopkins University was supported by NIH Grant P30 PPD status with over 80% accuracy from blood. Segregation of the CA006973. All animal treatments were approved by the University of Maryland, sample by the trimester of blood collection did not appear to Baltimore Animal Care and Use Committee and were conducted in full accordance affect prediction accuracy. These results suggest that epigenetic with the NIH Guide for the Care and Use of Laboratory Animals. Human subjects research was conducted under IRB protocol no. 00008149 and subjects were variation at biomarker loci is established early on during collected with funding from K23 MH074799-01A2 to JP. pregnancy and may represent a latent epigenetic status in the PPD risk group independent of pregnancy. The clinical implications of this finding are that early screening of those at risk for PPD may be possible, allowing an earlier direction of REFERENCES clinical treatment course. 1 O’Hara MW. Postpartum depression: what we know. J Clin Psychol 2009; 65: The high prediction accuracy of the identified biomarkers 1258–1269. replicated in an independent cohort of women who were 2 Soufia M, Aoun J, Gorsane MA, Krebs MO. SSRIs and pregnancy: a review of the depressed during pregnancy. In this group, the PPD status was literatur]. Encephale 2010; 36: 513–516. segregated with 88% accuracy; however, the prediction was in the 3 Field T. Prenatal depression effects on early development: a review. Infant Behav opposite direction, driven by differences at the HP1BP3 locus. An Dev 2011; 34: 1–14. analysis of cell subfraction distributions across cohorts identified a 4 Breese McCoy SJ. Postpartum depression: an essential overview for the practi- difference in the ratio of monocytes to lymphocytes and tioner. South Med J 2011; 104: 128–132. granulocytes significantly decreased in the depressed cohort that 5 Cuijpers P, Brannmark JG, van Straten A. Psychological treatment of postpartum depression: a meta-analysis. J Clin Psychol 2008; 64: 103–118. appeared to account for the discrepancy. Our data are consistent 6 Hirst KP, Moutier CY. Postpartum major depression. Am Fam Physician 2010; 82: with genome-wide expression studies of WBCs taken from women 926–933. after parturition that demonstrated an association of immune 7 Marcus SM. Depression during pregnancy: rates, risks and consequences— 45 system-related genes with depression scores. Incorporation of Motherisk Update 2008. Can J Clin Pharmacol 2009; 16: e15–e22. the DNA methylation biomarkers with cell count data enabled the 8 Pinette MG, Wax JR. The management of depression during pregnancy: a report prediction of PPD status in the entire cohort of 51 women with an from the American Psychiatric Association and the American College of Obste- AUC of 0.82. A potential confounding factor is that DNA tricians and Gynecologists. Obstet Gynecol 2010; 115: 188–189author reply 189. methylation between the prepartum euthymic and depressed 9 Studd JW. A guide to the treatment of depression in women by estrogens. cohorts was assessed in two separate batches, as all initial analyses Climacteric 2011; 14: 637–642. 10 Bloch M, Schmidt PJ, Danaceau M, Murphy J, Nieman L, Rubinow DR. Effects were performed on the euthymic cohort only. To control for this, of gonadal steroids in women with a history of postpartum depression. Am J we normalized DNA methylation levels at all 473 loci used for Psychiatry 2000; 157: 924–930. blood count proxy analysis using a cross-batch control. The 11 Bloch M, Rubinow DR, Schmidt PJ, Lotsikas A, Chrousos GP, Cizza G. Cortisol predicted cell-type proportions at these controls showed response to ovine corticotropin-releasing hormone in a model of pregnancy and moderate but nonsignificant batch effects between cohorts parturition in euthymic women with and without a history of postpartum (Supplementary Table 4); however, the effects were in the depression. J Clin Endocrinol Metab 2005; 90: 695–699. opposite direction to the prepartum mood status association 12 Kangaspeska S, Stride B, Metivier R, Polycarpou-Schwarz M, Ibberson D, observed, suggesting that this association is a true effect of Carmouche RP et al. Transient cyclical methylation of promoter DNA. Nature 2008; prepartum mood status. In addition, the significant correlation 452: 112–115. 13 Walf AA, Frye CA. ERbeta-selective estrogen receptor modulators produce observed with CBC-derived values adds confidence to assertion antianxiety behavior when administered systemically to ovariectomized rats. that the proxy-derived values are representative of actual cell sub- Neuropsychopharmacology 2005; 30: 1598–1609. type proportions. Finally, the linear model incorporating CBC- 14 MacLusky NJ, Luine VN, Hajszan T, Leranth C. The 17alpha and 17beta derived cell proportions generated a highly accurate prediction of isomers of estradiol both induce rapid spine synapse formation in the CA1 PPD status (AUC ¼ 0.96). Owing to the small size of the subsample hippocampal subfield of ovariectomized female rats. Endocrinology 2005; 146: used for this prediction, larger prospective cohorts will be required 287–293. to validate the predictive efficacy of this model. Cumulatively, our 15 ter Horst GJ. Estrogen in the limbic system. Vitam Horm 2010; 82: 319–338. data suggest that cell count information in combination with DNA 16 Suda S, Segi-Nishida E, Newton SS, Duman RS. A postpartum model in rat: methylation at HP1BP3 and TTC9B successfully and accurately behavioral and gene expression changes induced by ovarian steroid deprivation. Biol Psychiatry 2008; 64: 311–319. predicts PPD status independent of prepartum mood status. 17 Kaminsky ZA, Tang T, Wang SC, Ptak C, Oh GH, Wong AH et al. DNA methylation The results of this study suggest that an increased sensitivity to profiles in monozygotic and dizygotic twins. Nat Genet 2009; 41: 240–245. E2-based epigenetic reprogramming may represent a molecular 18 Schumacher A, Kapranov P, Kaminsky Z, Flanagan J, Assadzadeh A, Yau P et al. mechanism of predisposition to PPD risk. Future studies will be Microarray-based DNA methylation profiling: technology and applications. Nucleic needed to rigorously test this hypothesis and track epigenetic Acids Res 2006; 34: 528–542. changes through the course of pregnancy in women at risk and 19 Guintivano J, Arad M, Tamashiro KL, Gould TD, Kaminsky ZA. BioTile, A Perl based not at risk for PPD. The investigated population was in women tool for the identification of differentially enriched regions in tiling microarray with a previous history of mood disorders; however, studies data. BMC Bioinform 2013; 14:76. investigating the efficacy PPD prediction in the general population 20 Houseman EA, Accomando WP, Koestler DC, Christensen BC, Marsit CJ, Nelson HH et al. DNA methylation arrays as surrogate measures of cell mixture distribution. will need to be determined. Accurate prediction of PPD status will BMC Bioinform 2012; 13:86. enhance the clinical management of psychiatric treatment during 21 Langfelder P, Horvath S. WGCNA: an R package for weighted correlation network the course of pregnancy. analysis. BMC Bioinform 2008; 9: 559. 22 Walf AA, Koonce CJ, Frye CA. Estradiol or diarylpropionitrile decrease anxiety-like behavior of wildtype, but not estrogen receptor beta knockout, mice. Behav CONFLICT OF INTEREST Neurosci 2008; 122: 974–981. 23 Osterlund MK, Witt MR, Gustafsson JA. Estrogen action in mood and neurode- ZK, JP and TDG are coinventors listed on a patent application for DNA methylation at generative disorders: estrogenic compounds with selective properties-the next biomarker loci related to PPD. The remaining authors declare no conflict of interest. generation of therapeutics. Endocrine 2005; 28: 235–242. 24 Lund TD, Rovis T, Chung WC, Handa RJ. Novel actions of estrogen receptor-beta on anxiety-related behaviors. Endocrinology 2005; 146: 797–807. ACKNOWLEDGEMENTS 25 Walf AA, Rhodes ME, Frye CA. Antidepressant effects of ERbeta-selective estrogen We thank The Solomon R and Rebecca D Baker Foundation for their generous receptor modulators in the forced swim test. Pharmacol Biochem Behav 2004; 78: support of this research. This work was funded, in part, by MH093967 to TDG and a 523–529.

Molecular Psychiatry (2014), 560 – 567 & 2014 Macmillan Publishers Limited Postpartum depression prediction with DNA methylation biomarkers J Guintivano et al 567 26 Green AD, Galea LA. Adult hippocampal cell proliferation is suppressed with 36 Ramanan VK, Kim S, Holohan K, Shen L, Nho K, Risacher SL et al. Genome-wide estrogen withdrawal after a hormone-simulated pregnancy. Horm Behav 2008; 54: pathway analysis of memory impairment in the Alzheimer’s Disease Neuroima- 203–211. ging Initiative (ADNI) cohort implicates gene candidates, canonical pathways, and 27 Beissbarth T, Speed TP. GOstat: find statistically overrepresented Gene Ontologies networks. Brain Imaging Behav 2012; 6: 634–648. within a group of genes. Bioinformatics 2004; 20: 1464–1465. 37 Nassa G, Tarallo R, Ambrosino C, Bamundo A, Ferraro L, Paris O et al. A large set of 28 Reimand J, Kull M, Peterson H, Hansen J, Vilo J. g:Profiler—a web-based toolset for estrogen receptor beta-interacting proteins identified by tandem affinity pur- functional profiling of gene lists from large-scale experiments. Nucleic Acids Res ification in hormone-responsive human breast cancer cell nuclei. Proteomics 2011; 2007; 35: W193–W200. 11: 159–165. 29 Shimamiya T, Terada N, Wakabayashi S, Mohri M. Mood change and immune 38 Cao S, Iyer JK, Lin V. Identification of tetratricopeptide repeat domain 9, status of human subjects in a 10-day confinement study. Aviat Space Environ Med a hormonally regulated protein. Biochem Biophys Res Commun 2006; 345: 2005; 76: 481–485. 310–317. 30 Lutgendorf SK, Lamkin DM, DeGeest K, Anderson B, Dao M, McGinn S et al. 39 Gerges NZ, Tran IC, Backos DS, Harrell JM, Chinkers M, Pratt WB et al. Independent Depressed and anxious mood and T-cell cytokine expressing populations in functions of hsp90 in neurotransmitter release and in the continuous synaptic ovarian cancer patients. Brain Behav Immun 2008; 22: 890–900. cycling of AMPA receptors. J Neurosci 2004; 24: 4758–4766. 31 Snel B, Lehmann G, Bork P, Huynen MA. STRING: a web-server to retrieve and 40 Kuijpens JL, Vader HL, Drexhage HA, Wiersinga WM, van Son MJ, Pop VJ. Thyroid display the repeatedly occurring neighbourhood of a gene. Nucleic Acids Res 2000; peroxidase antibodies during gestation are a marker for subsequent depression 28: 3442–3444. postpartum. Eur J Endocrinol 2001; 145: 579–584. 32 Guido C, Panza S, Santoro M, Avena P, Panno ML, Perrotta I et al. Estrogen 41 Skrundz M, Bolten M, Nast I, Hellhammer DH, Meinlschmidt G. Plasma oxytocin receptor beta (ERbeta) produces autophagy and necroptosis in human seminoma concentration during pregnancy is associated with development of postpartum cell line through the binding of the Sp1 on the phosphatase and tensin homolog depression. Neuropsychopharmacology 2011; 36: 1886–1893. deleted from chromosome 10 (PTEN) promoter gene. Cell Cycle 2012; 11: 42 Yim IS, Glynn LM, Schetter CD, Hobel CJ, Chicz-Demet A, Sandman CA. Prenatal 2911–2921. beta-endorphin as an early predictor of postpartum depressive symptoms in 33 Bartella V, Rizza P, Barone I, Zito D, Giordano F, Giordano C et al. Estrogen receptor euthymic women. J Affect Disord 2010; 125: 128–133. beta binds Sp1 and recruits a corepressor complex to the estrogen receptor alpha 43 Licinio J. Potential diagnostic markers for postpartum depression point out to gene promoter. Breast Cancer Res Treat 2012; 134: 569–581. altered immune signaling. Mol Psychiatry 2010; 15:1. 34 Ruegg J, Cai W, Karimi M, Kiss NB, Swedenborg E, Larsson C et al. Epigenetic 44 Albacar G, Sans T, Martin-Santos R, Garcia-Esteve L, Guillamat R, Sanjuan J et al. regulation of glucose transporter 4 by estrogen receptor beta. Mol Endocrinol Thyroid function 48h after delivery as a marker for subsequent postpartum 2011; 25: 2017–2028. depression. Psychoneuroendocrinology 2010; 35: 738–742. 35 Vivar OI, Zhao X, Saunier EF, Griffin C, Mayba OS, Tagliaferri M et al. Estrogen 45 Segman RH, Goltser-Dubner T, Weiner I, Canetti L, Galili-Weisstub E, Milwidsky A receptor beta binds to and regulates three distinct classes of target genes. J Biol et al. Blood mononuclear cell gene expression signature of postpartum Chem 2010; 285: 22059–22066. depression. Mol Psychiatry 2010; 15: 93–100, 102.

Supplementary Information accompanies the paper on the Molecular Psychiatry website (http://www.nature.com/mp)

& 2014 Macmillan Publishers Limited Molecular Psychiatry (2014), 560 – 567