Polygenic prediction of the risk of perinatal depressive symptoms

Perinatal depression carries adverse effects on maternal health and child development, but genetic underpinnings remain unclear. We investigated the polygenic risk of perinatal depressive symptoms.

group displaying consistently high compared with consistently low depressive symptoms through out the prenatal and postpartum periods.
Conclusions: Polygenic risk scores for major depressive disorder, schizophrenia, and cross-disorder in non-perinatal populations generalize to perinatal depressive symptoms and may afford to identify women for timely preventive interventions.
K E Y W O R D S depression, epidemiology, genetics, mood disorders, pregnancy and postpartum

| INTRODUCTION
Maternal perinatal depression, defined as depression during pregnancy or within 12 months of childbirth, affects over 1 of 10 women in childbearing age (World Health Organizaton, 2015). Reported clinically relevant symptoms are even more common, affecting 1 of 5 women (Kumpulainen et al., 2018;Lahti et al., 2017). It causes a major burden on the health and well-being of the women and is associated with problems such as gestational diabetes and hypertension in pregnancy (Bansil et al., 2009) and obesity (Kumpulainen et al., 2018). It is also a major risk factor of suicide, one of the most common causes of death among women during the perinatal period (Chang, Berg, Saltzman, & Herndon, 2005;Nock et al., 2008). In addition to the women themselves, perinatal depression affects their children: prenatal depression is associated with developmental adversities, including increased risk of low birth weight and preterm birth (Pesonen et al., 2016). Both prenatal and postpartum depression (PPD) are associated with lower rates and shorter duration of breastfeeding (Figueiredo, Canário, & Field, 2014), less secure mother-child attachment (Martins & Gaffan, 2000), and higher risk of child neurodevelopmental, emotional, and behavioral problems Toffol et al., 2019;Tuovinen et al., 2018;Wolford et al., 2017).
While the risk of perinatal depression is modified by psychosocial (Yim, Tanner Stapleton, Guardino, Hahn-Holbrook, & Dunkel Schetter, 2015) and hormonal (Schiller, Meltzer-Brody, & Rubinow, 2015) factors, twin (Treloar, Martin, Bucholz, Madden, & Heath, 1999;Viktorin et al., 2016), sibling (Murphy-Eberenz et al., 2006;Viktorin et al., 2016) and family (Forty et al., 2006;Pearson et al., 2018) studies indicate that heritable factors also play a part. Heritability estimates of perinatal depression have varied between 25% (Treloar et al., 1999) and 54% (Viktorin et al., 2016) in twin studies, and 44% (Viktorin et al., 2016) in a sibling study. Furthermore, the daughters of prenatally depressed women have over the threefold risk of depression during their own pregnancy (Pearson et al., 2018). While perinatal depression and nonperinatal major depressive disorder (MDD) may represent at least partly distinct disorders, perinatal depression at least partially shares its genetic component with MDD (Viktorin et al., 2016) and bipolar disorder (BD; Payne et al., 2008). Moreover, the genetic component may also be shared with schizophrenia (SCZ), anorexia nervosa, attention-deficit/hyperactivity disorder (ADHD), autism spectrum disorder (ASD), obsessive-compulsive disorder (OCD) and Tourette syndrome, due to the partially shared heritability between MDD and these disorders Lee et al., 2019). Furthermore, candidate gene studies have linked single nucleotide polymorphisms (SNPs) with the risk of these disorders. However, it is unlikely that the heritabilities of these complex disorders are assigned to individual SNPs. Indeed, a recent meta-analysis concluded that no candidate genes or gene sets reliably predict depression phenotypes (Border et al., 2019).
A promising alternative is the polygenic risk score (PRS) approach, which exploits findings from genome-wide association studies (GWAS) using an aggregate measure of weighted genetic variants associated with the phenotype. One study showed that PRSs for BD, but not MDD, were associated with postpartum depression (Byrne et al., 2014).
However, the study used polygenic profile scoring methodology instead of association analysis and exploited summary statistics from an older GWAS. In addition, to our best knowledge, no PRS study has sufficiently accounted for the other risk factors of perinatal depression and have focused only on PPD. However, women suffering from PPD often have elevated depressive symptoms already during early pregnancy (Evans et al., 2012;Kumpulainen et al., 2018;Tuovinen et al., 2018;van der Waerden et al., 2017), indicating that PPD is often a continuation of prenatal depression. Accordingly, we investigated PRSs based on the most recent and largest GWAS on MDD (Howard et al., 2019;Wray et al., 2018), BD (Ruderfer et al., 2018;Stahl et al., 2019), SCZ (Ripke, Neale, & Spiker, 2014;Ruderfer et al., 2018)

| Participants
The participants come from the Prediction and Prevention of Pre-eclampsia and Intrauterine Growth Restriction (PREDO) study (Girchenko et al., 2017). We enrolled 1,079 pregnant women to the clinical subsample: 969 had one or more, and 110 had none of the known risk factors for pre-eclampsia and intrauterine growth restriction.
The women were recruited on their first ultrasound screening at 12-14 gestational weeks from 10 hospitals in Southern and Eastern the 1,079 women, 997 donated blood for DNA in early pregnancy (median = 13.0; interquartile range = 12.6-13.4 weeks). Of them, 742 (74.4%) had data available on prenatal or postpartum depressive symptoms; these women formed our analytic sample. Of them, data were available on prenatal depressive symptoms for 721 (97.2%), postpartum depressive symptoms for 726 (97.8%), and both for 705 (95.0%).
Compared with the rest of the clinical subsample, the women in the analytic sample more often had attained tertiary education (42.7% vs. 55.0%; χ 2 = 15.29; p < .001), but did not significantly differ otherwise.

| Ethics statement
All participants signed informed consent forms. The PREDO study protocol has been approved by the ethical committee of the Helsinki and Uusimaa Hospital District, and aligns with the Declaration of Helsinki.
Population outliers were excluded based on visual inspection. There were 15,544,584 variants after imputation.

| PRS of MDD, BD, SCZ, and CD
PRS for MDD (Howard et al., 2019;Wray et al., 2018; hereafter referred to as MDD2018 and MDD2019), BD (Ruderfer et al., 2018;Stahl et al., 2019;BD2018 andBD2019), SCZ (Ripke et al., 2014;Ruderfer et al., 2018;SCZ2014 andSCZ2018), and CD (Lee et al., 2019) were calculated by taking genetic variants from the imputed best guess genotypes up to varying significance thresholds from the GWAS discovery sample and applying a score from these variants, weighted by the associations in the discovery sample, to predict a trait in an independent target sample. Clumping and calculation of PRSs were performed using PRSice-2 (Choi & O'Reilly, 2019), using LD threshold (R 2 ) 0.1 and clumping window width of 500 kilobases. Before clumping, we excluded all SNPs with INFO score < 0.90. Other filtering criteria in the PREDO sample included HWE p < 1 × 10 −6 and missingness > 0.05. PRSs were created using p-value thresholds 5 × 10 −8 , .001, .01, and .05. The number of SNPs in each PRS is in Table 1. The correlations between the PRSs are in Tables S1-S4.

| Depressive symptoms
The women completed the Center for Epidemiologic Studies Depression Scale (CES-D; Radloff, 1977) biweekly up to 14 times throughout the perinatal period starting from 12 0/7 -13 6/7 until 38 0/7 -39 6/7 or delivery, and twice during the postpartum period (median 2.1 weeks and 6.4 months, interquartile range 2.0-2.4 weeks and 6.1-7.3 months, respectively). The 20 CES-D questions were rated on a scale from none (0) to all of the time (3). Higher scores indicate more depressive symptoms during the past week.

| Covariates
These included age at delivery (years), family structure (cohabiting/ married vs single parent), body mass index (BMI) in early pregnancy (kg/m2) and cigarette smoking (smoked through pregnancy/quit in early pregnancy vs no) with data from the Medical Birth Register.
Education (secondary, tertiary vs basic) and alcohol use (yes vs no) were reported in early pregnancy. We conducted multidimensional scaling analyses on the whole genome genotypes with Plink v0.64 to control for population stratification. Four main components depicted the population substructure.

| Statistical methods
Linear regression analyses tested the associations between the PRSs and prenatal (mean of all 14 prenatal values) and postpartum (mean of the two postpartum values) depressive symptoms. We then tested if the associations between the PRSs and postpartum depressive symptoms were accounted for by prenatal depressive symptoms, by adding prenatal depressive symptoms to the regression equations.
We pursued mediation if all variables were significantly interrelated and the regression coefficient diminished after introducing the mediator into the model (Baron & Kenny, 1986). To study if prenatal depressive symptoms moderated the associations between the PRSs and postpartum depressive symptoms, we included a prenatal de- We present the associations as adjusted for the MDS and maternal age, and further for all covariates. We normalized CES-D values with square root transformation and standardized the CES-D scores and the PRSs (Mean = 0; SD = 1) to facilitate interpretation.
Unstandardized regression coefficients and odds ratios (OR) and 95% confidence intervals (95% CI) with two-tailed p-values present effect sizes.
We controlled for multiple testing with a false detection rate (FDR) procedure (Benjamini & Hochberg, 1995). We corrected for 5% FDR over 28 tests (seven PRSs with four p-value thresholds), separately for prenatal and postpartum CES-D scores and for each statistical model.
We conducted the analyses using SPSS v24 and PROCESS macro v3.3 (Hayes, 2017) with a custom model builder (Frank, 2018) using 10,000 bootsrapped samples. We corrected for FDR using the p.adjust function in R.

| Data availability statement
The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.
Data requests may be subject to further review by the Finnish national register authorities, and by the ethical committees.
The correlation between the means of prenatal and postpartum measurements was Pearson's r = .68 (p < .001). Of the covariates, only alcohol use and early pregnancy BMI were correlated with prenatal depressive symptoms (Pearson's r ≥ .09; p ≥ .022), and none were correlated with postpartum depressive symptoms (Table S5). .000
Note: Model 1 adjusting for maternal age at delivery and population stratification. Model 2 adjusting for maternal age at delivery, education, family structure, body mass index, smoking, alcohol use, and population stratification. Model 3 adjusting for prenatal depressive symptoms, maternal age at delivery, education, family structure, body mass index, smoking, alcohol use, and population stratification. B, unstandardized regression coefficient; R 2 , the proportion of variance in depressive symptoms explained by the PRS alone.
F I G U R E 1 Mediation of the association between polygenic risk scores and postpartum depressive symptoms by prenatal depressive symptoms. The association between polygenic risk scores of MDD2018 (a), MDD2019 (b), SCZ2014 (c), SCZ2018 (d), and CD (e), and postpartum depressive symptoms is mediated by prenatal depressive symptoms. Thep-values and regression coefficients shown are for the PRSs calculated using the p-value threshold p < .01, and adjusted for maternal age at delivery and population stratification. p-Values for indirect effects are from Sobel's test. PRS, polygenic risk score MDD2019, which was not associated with prenatal depressive symptoms in the fully adjusted model. The associations were similar but less consistent with the other p-value thresholds ( Table 3). The PRSs for BD were not associated with prenatal or postpartum depressive symptoms (Table S6).
The proportion of variance explained by the entire statistical models varied between 1.3% and 6.2% for prenatal and 1.5-5.9% for postpartum depressive symptoms; R 2 's of the PRSs alone are in Table 3. To further characterize the polygenic basis of perinatal depressive symptoms, we conducted stepwise linear regression analyses where we first entered Model 1 covariates and then tested using forward selection which of the PRSs with p-value threshold 0.01 improved model fit the most. The PRSs for SCZ2018 and SCZ2014 were the best predictors of prenatal and postpartum depressive symptoms, respectively, and MDD2018 significantly predicted additional variance in prenatal and postpartum depressive symptoms: for adding the second PRS, R 2 changes were .7% and 1.0%, F's for change 5.20 and 7.14, and p = .02 and p = .008, respectively.

| PRS and prenatal and postpartum depressive symptoms: Mediation and moderation
No PRSs were significantly associated with postpartum depressive symptoms in models including prenatal depressive symptoms (Table 3; Table S6). As the criteria were met (Table 3 and  and 88.0%; Figure 1). After accounting for mediation, the direct effects of the PRSs for MDD2018, SCZ2014, and SCZ2018 remained significant, but MDD2019 and CD did not (Figure 1).
Analyses testing if prenatal depressive symptoms moderated the associations between the PRSs and postpartum depressive symptoms revealed no significant interactions (prenatal depressive symptoms × PRS interactions p ≥ .152; data not shown).
F I G U R E 1 (Continued)

| PRS and prenatal and postpartum depressive symptoms: Fluctuating or persistently high level of symptomatology
In the LCA we compared two to six subgroups solutions, and based on the five criteria identified a solution with three latent classes as optimal (Table S7). Each of the three subgroups showed high depressive symptom stability throughout the prenatal and postpartum periods ( Figure 2). The groups differed in symptom severity, showing consistently low (n = 158; 21.3%), moderate (n = 397; 53.5%) and high, clinically significant (CES-D scores ≥ 16; Radloff, 1977;n = 187; 25.2%) levels of depressive symptoms. The PRSs for BD2018 and BD2019 were not significantly associated with the LCA subgroups.

| DISCUSSION
We found that the PRSs for MDD2018, MDD2019, SCZ2014, and SCZ2018 and CD, but not BD2018 or BD2019, predicted higher levels of prenatal and postpartum depressive symptoms. While all the PRSs were correlated, only MDD2018, SCZ2014, and SCZ2018 significantly predicted unique variance in depressive symptoms. The associations were significant across PRSs calculated using multiple p-value thresholds, but they were most consistent with the PRSs containing larger numbers of SNPs. This is likely related to the genetic variance reflected in the different PRSs, as the numbers of SNPs differ widely between different p-value thresholds (Table 1).
These findings thus suggest that the genetic risk of MDD, SCZ, and CD in nonperinatal populations may generalize to depressive symptoms in perinatal women.
The effect sizes that predicted prenatal and postpartum depressive symptoms were small, but of similar magnitude as those of the covariates (Table S5). Moreover, as previously reported (Evans et al., 2012;Kumpulainen et al., 2018;Tuovinen et al., 2018;van der Waerden et al., 2017), the best predictor of postpartum depressive symptomatology was prenatal symptomatology.
However, the effect size estimates were much larger when we took into account the high stability of depressive symptoms throughout the perinatal period. Our LCA indicated that the PRSs for MDD2018, MDD2019, SCZ2014, SCZ2018, and CD were associated with a trajectory of consistently high depressive symptoms across the perinatal period; one SD increase in these PRSs translated to 1.24-1.45 increase in the odds to belong to the group with consistently high compared with low depressive symptoms. Thus, genetic finding reflects that depressive symptoms were highly stable throughout the perinatal period.
One previous study has investigated the associations between PRSs for MDD and BD and depression during the perinatal period (Byrne et al., 2014). | 873 smallest increases in hazard ratios. These findings partially agree with and partially conflict with ours, though our study differs in depression phenotype and age, and was restricted to perinatal women. Ethnic differences may also explain why the PRSs for BD were not associated with depressive symptoms in our study. Alternatively, since the PRS for BD2018 predicted smaller increases in hazard ratios for than MDD2018 and SCZ2014 in the previous study (Musliner et al., 2019), it is possible that significant associations could be found in samples larger than ours. While our findings provide important evidence of the shared genetic components between perinatal depression and other psychiatric disorders, further studies are warranted to elucidate the relationships between their respective genetic bases.
The strengths of this study include prospective design, multiple measurements of depressive symptoms during pregnancy and after delivery, and a well-characterized sample with data available on multiple important covariates. Also, as most of the genotyped samples had data on depressive symptoms available during pregnancy and after delivery, sample attrition is unlikely to have influenced our findings. The limitations include the lack of measurement of depressive symptoms before pregnancy, and relatively modest statistical power to detect genetic effects. Also, the comparability of our findings to studies using diagnoses of depressive disorders and conducted in populations with different characteristics and ethnicities is limited.
Furthermore, our PRSs do not account for gene-environment interactions, as they were based on summary statistics from GWASs investigating main effects.

| CONCLUSIONS
We found that two different PRSs for MDD and SCZ and PRS for CD, but not two PRSs for BD, were associated with prenatal and postpartum depressive symptoms. We also found that the associations between the PRSs and postpartum symptomatology were accounted for, and hence mediated prenatal depressive symptoms and that the PRSs predicted persistently high depressive symptoms throughout the perinatal period. Our study is consistent with the earlier findings suggesting that perinatal and nonperinatal depressive symptoms (Viktorin et al., 2016), and MDD (Viktorin et al., 2016), and SCZ