Placental gene network modules are associated with maternal stress during pregnancy and infant temperament

Abstract

Maternal psychosocial stress during pregnancy (MPSP) is a known contributor to maladaptive neurobehavioral development of the offspring; however, the underlying molecular mechanisms linking MPSP with childhood outcome remain largely unknown. Transcriptome-wide gene expression data were generated using RNA-seq from placenta samples collected in a multi-ethnic urban birth cohort in New York City (n = 129). Weighted gene co-expression network analysis (WGCNA) was used to characterize placental co-expression modules, which were then evaluated for their associations with MPSP and infant temperament. WGCNA revealed 16 gene coexpression modules. One module, enriched for regulation of chromosome organization/gene expression, was positively associated with MPSP and negatively associated with Regulatory Capacity (REG), a component of infant temperament. Two other modules, enriched for cotranslational protein targeting and cell cycle regulation, respectively, displayed negative associations with MPSP and positive associations with REG. A module enriched with oxidative phosphorylation/mitochondrial translation was positively associated with REG. These findings support the notion that the placenta provides a functional in utero link between MPSP and infant temperament, possibly through transcriptional regulation of placental gene expression.

1 |. INTRODUCTION

The prenatal period represents a sensitive window to both external and internal environmental cues, which may dramatically impact fetal/child development, potentially extending throughout life.¹ Maternal psychosocial stress during pregnancy (MPSP) is an established risk factor for suboptimal birth outcomes and developmental disorders² Both preclinical and clinical studies have provided evidence that MPSP contributes significantly to adverse child neurobehaviors. Animal studies in particular, have demonstrated the long-term impact on attention and enhanced anxiety in the offspring³; human studies have further confirmed the link between maternal stress/anxiety and behavioral problems in children.^4,5

The placenta, a transient fetal organ made up of its parenchyma, chorion, amnion, and umbilical cord, situates at the interface between the mother and the developing fetus and regulates fetal development by facilitating the transfer of nutrients and signaling factors between the two. As maternal insults disrupt this interplay, the ensuing alterations in the in utero environment can be conveyed to the fetus via the placenta.⁶ In addition to its passive role in influencing fetal development through the transmission of maternal signals, evidence suggests that the placenta also actively modulates fetal brain development. For instance, it has been shown to actively synthesize and deliver neurotransmitters, such as serotonin, to the fetal brain to support neuronal differentiation.⁷ The molecular mechanisms underlying the impact of MPSP on neurobehavioral traits, such as infant temperament, are very complex and likely involve a large number of genes. While traditional univariate or agnostic transcriptomic approaches may highlight individual genes underlying the MPSP—temperament axis, they fail to take the coregulation among genes participating in common signaling pathways into account. Weighted gene co-expression network analysis (WGCNA) is an approach to study biological networks by taking those patterns in gene expression levels into account⁸ as WGCNA-defined modules can reveal more than simple coincident patterns; instead, they reflect functional links between the molecular players where coregulation is indicative of common pathways/processes.

In this study, we evaluated inter-relationships of placental transcriptome-wide changes at the module level with MPSP and infant temperament to identify functional networks/pathways that link MPSP and offspring’s temperament in early childhood.

2 |. MATERIALS AND METHODS

2.1 |. Study population

The study population is part of the Stress In Pregnancy (SIP) project that enrolled pregnant women in New York City, as previously described.⁹ Briefly, expectant mothers in the second trimester were recruited from obstetric clinics at Mount Sinai Hospital and New York-Presbyterian/Queens. Exclusion criteria included multiple pregnancies, HIV infection, maternal psychosis, maternal age <15 years, life-threatening maternal medical complications, and congenital or chromosomal abnormalities in the fetus. Demographic information, such as maternal ethnicity, education, marital status, and age were collected. Birth-related information was collected at birth and included, among other variables, infant gender, birthweight, and gestational age at birth. Mothers completed self-report scales of normative prenatal stress during the second trimester and infant behavior questionnaire at 12 months after the birth of the child prospectively. The current study represents a subsample (n = 129) of the SIP cohort from whom placenta tissues were collected at the time of delivery.

All methods were carried out in compliance with CUNY Human Research Protection Program policies and procedures. Protocols were approved by the Integrated Institutional Review Board of CUNY (IRB File #339130). Written informed consent was obtained from all eligible women, or from a parent and/or legal guardian of the subjects whose age are less than 18, for all study procedures.

2.2 |. Maternal psychosocial stress and infant temperament

MPSP was measured using a set of scales: Edinburgh Postnatal Depression Scale (EPDS), Pregnancy-related anxiety questionnaire-revised (PRAQ-R), 14-item Perceived Stress Scale (PSS-14), State-Trait Anxiety Inventory STATE (STAI-S), State-Trait Anxiety Inventory TRAIT (STAI-T), and Life Events with negative valence (LEneg). Latent profile analysis (LPA) using the full maximum likelihood estimation, as described in the previous paper,¹⁰ was applied to incorporate these five measures and extract stress (i.e., MPSP) membership. LPA results indicated that the three-class model provided the best solution since both the Bayesian Information Criterion (BIC) and adjusted BIC values continued to decrease, and the Lo-Mendell-Rubin test indicated that the three-class model provided a significantly better fit than the two-class model (p < .001). As a result, the more parsimonious three-class model was selected, and the participants were divided into low-, medium-, and high- normative stress categories.

Mothers completed the Infant Behavior Questionnaire-Revised (IBQ-R)¹¹ about their offspring at 12 months post-partum. The IBQ-R includes 91 questions about the frequency of specific child neurobehaviors, comprising 14 domains of temperament (activity, distress to limitations, fear, duration of orienting, smiling and laughter, high-pleasure seeking, low-pleasure seeking, soothability, falling reactivity, cuddliness, perceptual sensitivity, sadness, approach, and vocal reactivity). The mothers reported weekly behavior frequency on a scale ranging from 1 (never) to 7 (always), with an option to indicate that the behavior was not observed. Those 14 domains of temperament formed three higher-order composite scores^12,13: Negative Affect (NEG), Surgency/Positive Affectivity (SUR), and Regulatory Capacity (REG). NEG refers to discomfort, fear, anger, sadness, and low soothability¹⁴ which underlies the behavioral inhibition system and leads to heightened vigilance, emotional and physiological arousal, and restricted behavior.¹⁵ NEG is indicated by the subcomponents including Fearfulness, Distress to Limitations, and Sadness. SUR refers to sociability, sensation seeking, and activity¹⁴ which underpins the behavioral approach system and leads to both reactive approach in reward situations and frustration or irritability in non-reward situations.¹⁵ SUR is indicated by subcomponents including Activity Level, High-Intensity Pleasure, and Approach. REG refers to processes such as allocating attention, which modulates the expression of negative emotionality and reactivity.^12,13 REG consists of duration of orientation, cuddliness, soothability, and low-pleasure seeking. Higher NEG scores and lower REG scores are considered to contribute to a profile of “difficult” temperament.¹⁶ The IBQ-R is a widely used measure with high internal consistency and good interrater reliability among reporting caregivers of the infant.¹¹

2.3 |. Tissue collection and RNA extraction

Placenta tissues were sampled as previously described.⁹ Briefly, placenta biopsies, free of maternal decidua, were excised from each placenta quadrant, snap frozen in liquid nitrogen for 24 h and then stored at −80°C. RNA extraction was carried out by first grinding frozen tissue in a liquid nitrogen-cooled mortar. RNA was extracted with the Maxwell 16 automated DNA/RNA extraction equipment (Promega: Madison, WI) using the proprietary extraction kits following the manufacturer’s protocol. RNA was quantified with Nanodrop spectrophotometer (Thermo Electron North America: Madison, WI) and stored at −80°C. RNA integrity number (RIN) was determined with an Agilent 2100 bioanalyzer. Among 129 samples included in the final dataset, three samples had RIN < 6, and the remaining samples had RIN ranged 6 to 8.3. The mean of RIN was 6.6 with standard deviation of 0.6.

2.4 |. Transcriptome analysis

RNA-seq was performed at Novagene Corp. (Sacramento, CA). Ribosomal RNA was removed using Ribo-Zero Kit. mRNA was fragmented randomly by adding fragmentation buffer, then the cDNA was synthesized using mRNA template and random hexamer primers, after which a custom second-strand synthesis buffer (Illumina), dNTPs, RNase H, and DNA polymerase I were added to initiate the second-strand synthesis. After a series of terminal repair, ligation and sequencing adaptor ligation, the double-stranded cDNA library was completed through size selection and PCR enrichment. Seven samples were sequenced twice and served as technical replicates for internal control (replicates clustered pairwise in PCA analysis, confirming robustness of our analysis). Read mapping was performed using STAR aligner¹⁷ with quantification by featureCounts.¹⁸ The value of 10 counts was used as a threshold to mark detectable genes. Additional filtering were applied to remove from the analysis genes that were detected in less than half of the samples. The final WGCNA dataset included 14 029 protein coding genes. The variability in our dataset due to RNA quality (RIN values) was removed using the sva R package, then Variance-stabilizing transformation (VST) from DESeq2 R package¹⁹ was applied to the dataset.

2.5 |. Weighted gene co-expression network analysis

The gene co-expression network was generated using the WCGNA R package.⁸ Briefly, a similarity matrix was generated from the residualized, VST-transformed RNA-Seq data using absolute values of correlation coefficients among all gene pairs s_ij = |cor(x_i, x_j)|; the similarity matrix is then transformed into an adjacency matrix using an adjacency function based on a weighted soft threshold: $a_{ij}=S_{ij}^{\beta }$. The components of the resulting adjacency matrix indicate connection strengths among gene pairs, with connections among strongly correlated genes emphasized and weakly correlated genes suppressed.⁸

As preliminary analysis, we tested several combinations of parameters, including different correlation functions (cor for Pearson correlation, bicor for robust biweight midcorrelation), types of network (signed, unsigned), and β values. WGCNA analysis is based on the scale-free topology assumption, and R² coefficient serves as the scale-free topology fit index; it is recommended to choose parameters that provide a higher fit index. The selected combination of parameters satisfied the minimum value required to generate a scale-free topology network (linear regression R² ≥ .9). The value of β = 12 provided a local maximum of fit index (R² = .92) in our study. This is in agreement with WGCNA authors’ recommendations where β = 12 is explicitly mentioned for the signed networks of sample size more than 40. For the final analysis, we selected a signed network with bicor function (following the recommendations of the WGCNA package) and β = 12.

The WGCNA algorithm groups all genes to a set of modules which were defined as branches of the resulting cluster tree. For each module, a module eigengene was constructed as the first principal component of the standardized expression of genes within the module and serves as a summary measure of the module. For each gene, the correlation between expression level and module eigengene value was defined as its eigengene-based connectivity, also known as module membership, of this gene. Genes with the highest eigengene-based connectivity were considered as hub genes (in this study, we considered top 10 genes ranked by highest connectivity).

2.6 |. Gene set enrichment and functional interaction analysis

Modules were analyzed for enrichment in Gene Ontology (GO) terms with the topGO R package²⁰ using the “classic” algorithm and Fisher statistics; FDR-adjusted p-values from topGO are reported. Functional interactions were visualized using the STRING tool (Search Tool for the Retrieval of Interacting Genes/Proteins).²¹ STRING is a database of known and predicted protein-protein interactions, including direct (physical) and indirect (functional) associations; they stem from computational prediction, knowledge transfer between organisms, and interactions aggregated from other external databases. Enrichment for biological pathways and transcription factors was performed using Ingenuity Pathway Analysis (IPA)²² and Enrichr.²³

2.7 |. Statistical analysis

All statistical analyses were conducted using R version 3.6. To examine the link between gene expression and the traits of interest (such as MPSP and infant temperament scores), we performed a generalized linear regression analysis (glm function in R). For the MPSP analysis, we modeled MPSP as the independent variable and module eigengene values as continuous outcomes. For the infant temperament analysis, we modeled module eigengenes as continuous independent variables and each infant temperament domain (SUR, REG, NEG) as continuous outcomes. To analyze differential expression of individual genes, we implemented the DESeq function of the DESeq2 R package, and p values were adjusted by FDR.¹⁹ All models (glm- and DESeq-derived) were adjusted for a priori determined covariates including maternal age, gestational age at birth, child’s gender, birthweight, ethnicity, and educational levels, which were potentially associated with maternal stress or infant temperaments in our previous publications.^10,16,24 Ethnicity and educational levels are not associated with module eigengenes so they were not included in the final models. Mediation analysis was performed using the mediation R package.²⁵ One-way analysis of variance (ANOVA) was used to compare the infant temperament scores across the 3 levels of MPSP, and the raw p values were used. All statistical tests were twosided, and p < .05 was considered statistically significant.

3 |. RESULTS

The demographic characteristics of our study population are shown in Table 1. Similar to the overall SIP cohort, participants were multiethnic (62% Hispanic and 23% Black), predominantly non-smokers (86%) during pregnancy in the 21–30 age group, and with high proportion (55%) of single mothers. Distribution of stress components that formed the MPSP and the three composite scores of temperament measured are displayed in Figure S1). The distribution of the three MPSP latent profiles (low, medium, and high stress) was 36%, 43%, and 21% respectively. As shown on Figure S1B, all three higher-order infant temperament composite scores (SUR, NEG, and REG) followed a normal distribution.

TABLE 1.

Demographic characteristics of the study population (N = 129)

Characteristics	Maternal psychosocial stress during pregnancy			p-values*
	Low (n = 47)	Medium (n = 55)	High (n = 27)
Infant gender = Male (%)	22 (46.8)	23 (41.8)	16 (59.3)	.330
Maternal age (%)				.358
<18	1 (2.1)	2(3.6)	2 (7.4)
18–21	12 (25.5)	7 (12.7)	6 (22.2)
22–30	19 (40.4)	34 (61.8)	13 (48.1)
31–40	13 (27.7)	12 (21.8)	5 (18.5)
41–50	2 (4.3)	0(0.0)	1 (3.7)
Maternal education (%)				.363
Less than or high school	21 (44.7)	19 (34.5)	12 (44.4)
Some college	16 (34.0)	15 (27.3)	5 (18.5)
Associate’s degree	3 (6.4)	6 (10.9)	5 (18.5)
Bachelor’s degree or above	7 (14.9)	15 (27.3)	5 (18.5)
Maternal race (%)				.587
White	3 (6.4)	8 (14.5)	3(11.1)
Black	13 (27.7)	12 (21.8)	4 (14.8)
Hispanic	26 (55.3)	24 (43.6)	17 (63.0)
Asian	2 (4.3)	6 (10.9)	2 (7.4)
Others	3 (6.4)	5(9.1)	1 (3.7)
Mother’s marital status (%)				.461
Married or common law	19 (40.4)	28 (50.9)	8(29.6)
Single	27 (57.4)	26 (47.3)	18 (66.7)
Divorce/Separated/Widowed	1 (2.1)	1 (1.8)	1 (3.7)
Smoking during pregnancy = Yes (%)	6 (12.8)	8(14.5)	4 (14.8)	.957
Birthweight [mean (SD)]	3444.15 (669.37)	3376.13 (547.30)	3016.19 (526.74)	.009
Gestational age [mean (SD)]	39.39 (1.89)	39.16 (1.89)	39.11 (1.99)	.777

^*p values were from Chi square test or the one-way analysis of variance.

Consistent with previously reported findings,¹⁶ we observed a positive association between MPSP and NEG component (ANOVA p < .01) and negative association between MPSP and REG component with borderline significance (ANOVA p = .06; Figure S2).

Coexpression network analysis of the placental transcriptome revealed 16 coexpressed gene modules (Supporting Information Data 1). Module sizes varied between 58 and 3123 genes; 1611 genes did not load onto any specific module (marked as the “gray” pseudomodule here; Figure 1). Some of the key GO terms associated with selected modules, are listed in Figure 1.

FIGURE 1.

Gene expression modules detected by WGCNA. Numbers and length of the bar correspond to the sizes of corresponding modules. On the left, selected GO enrichment categories are shown. On the right, top 10 hub genes for each module are shown

We assessed associations between the constructed WGCNA module eigengene values and MPSP, three dimensions of infant temperament scores and other biological characteristics (Table S1). Figure 2A illustrates the association of each module with the MPSP profiles (low, medium, and high stress). As shown in Figure 2B, the “red” module (regulation of gene expression and chromatin organization) displayed the strongest positive association with MPSP (p = .003), suggesting that maternal stress triggers major changes in gene expression, presumably through transcriptional regulation. Negative associations were observed between MPSP and the “tan” module (cell cycle) (p = .02) and “magenta” module (cotranslational protein targeting) (p = .03).

FIGURE 2.

Association of coexpression modules with maternal stress. (A) Volcano plot showing eigengene values of each modules, correlation estimated using generalized linear model; X axis: GLM-statistics, Y axis: −log10(p-value). (B) Eigengene values (one dot corresponds to one sample) for red module among different MPSP levels. (C) Volcano plot depicting individual genes correlated with maternal stress (low vs. high maternal stress). Colors indicate module membership. The observed fold change and significance levels are indicated by the x and y axis respectively

With respect to infant temperament, the strongest association for the REG score was again with the “red” module (p = .0006), but in the opposite direction compared to MPSP. (Figure 3A,B). Lastly, the SUR score was positively associated with the “black” module (immune response, p = .007). No significant associations were observed between any module and NEG score.

FIGURE 3.

Association of coexpression modules with IBQ-R Regulatory Capacity/Effortful Control (REG). (A) Volcano plot showing eigengene values of each modules, correlation estimated using generalized linear model; X axis: GLM-statistics, Y axis: −log10(p-value). (B) Eigengene values (one dot corresponds to one sample) for red module among different MPSP levels. (C) Volcano plot depicting individual genes correlated with REG trait. Colors indicate module membership. The observed fold change and significance levels are indicated by the x and y axis respectively

Given that the “red” and “magenta” modules were associated with both MPSP and the REG score, we performed mediation analysis to evaluate whether these modules mediate the association between MPSP and the REG score (Figure S3, Table S2). We observed a borderline significant (p = .07) association between MPSP and the REG score. which fall further from the threshold of significance (p = .30) after including “red” module expression into the model, suggesting partial mediation of this placental module on the impact of MPSP on REG.

To gain mechanistic insights on these modules, we performed enrichment analyses using such tools as IPA²² and Enrichr,²³ that are developed by leveraging curated published datasets. For example, the “red” module was enriched in genes involved in regulation of chromatin organization and transcription (GO:1903508, FDR-adjusted q = 3.3·10⁻⁴, with top participants like transcriptional regulators ASH1L, NIPBL, EP300) and protein ubiquitination/deubiquitination (GO:0004842, q = 1.1·10⁻⁴, genes including USP34, UBE3A, UBR5). Pathway analysis by IPA revealed ubiquitination (FDR-adjusted p = 1.7·10⁻⁴) and estrogen receptor signaling (FDR-adjusted p = 1.2·10⁻³, with EP300 and NCOA2 as upstream genes) pathways as the top pathways for the “red” module. The “tan” module was highly enriched with GO terms for cell cycle (GO:0007049, p < 10⁻³⁰); accordingly, “Kinetochore Metaphase Signaling Pathway” was revealed by IPA (p = 8.9·10⁻¹⁵), and E2F4 found as a main transcription factor (ENCODE and ChEA Consensus from Enrichr, q < 10⁻³⁰). The “yellow” module was associated with oxidative phosphorylation and mitochondrial translation (GO:0006119 and GO:0032543, both p < 10⁻³⁰), and lastly the “magenta” with cotranslational protein targeting (GO:0006614 and GO:0006613, p < 10⁻³⁰) and EIF2 signaling pathway (IPA, p < 10⁻³⁰). Figure 4 depicts interactions among the top hub genes for selected modules. Lastly, we carried out a traditional univariate differential gene expression analysis using DESeq2 to identify differentially expressed genes associated with study traits. Only a handful of individual genes were significantly associated with our traits of interest:the top hits included DAAM1 from “red” module and GNAS from magenta module (association with MPSP, adjusted p = .023) (Figures 2C and 3C).

FIGURE 4.

“Cell-cycle” (tan), “oxidative phosphorylation” (yellow) and “Chromosome organization; Regulation of gene expression” (red) modules composition, with main hub genes shown and their connections according to the STRING database

4 |. DISCUSSION

In the current study, we performed transcriptome-wide profiling of placenta samples to evaluate the potential molecular underpinnings linking MPSP to postnatal infant temperament. To our knowledge, this is the first study to evaluate transcriptome-wide gene expression in the placenta under this specific paradigm of MPSP and infant temperament.

The use of WGCNA is an important strategy to provide unsupervised detection of coexpression gene modules relevant to biological networks underlying the association between maternal stress and child neurodevelopment. Several studies, including our own, have applied a similar approach for evaluating placental gene networks in relation to birth outcomes. A study by Deyssenroth et al.²⁶ based on 200 placenta samples observed 17 modules, while another study by Buckberry et al based on 16 samples observed 13 modules.²⁷ Two other studies used microarray data: 33 modules reported by Zhao et al.²⁸ after analysis of a previously deposited dataset of 16 samples; 17 modules reported by Cox et al.²⁹ These differences in the number of modules reported across these studies maybe driven by study considerations (sample size, platform, etc) as well as true complexity in placental gene activity. The number of modules identified in our study falls within the range reported in the literature, and the enriched pathways are consistent with what are previously reported.^26,27 For example, a module enriched in cell cycle-related genes (tan module in our network) was reported in networks of Buckberry et al.²⁷ and Deyssenroth et al.²⁶ (with top genes such as TOP2A, TPX2, LMNB1, CDK1, DEPDC1, HMMR, CCNA2); similarly, all three networks reported a module with extracellular matrix- or cell adhesion-related genes (pink module in our network, module M6 of Ref. [27], tan module of Ref. [26]; with top genes like B3GNT7, NOTUM, ASAP3, NOG, PRG2). The coexpression network described by Deyssenroth et al had several other overlaps: cotranslational protein targeting (magenta in our network, magenta in Ref. [26]; top genes including RPL7A, RPL19, RPS5), gas transport (lightcyan module in our network, gray60 module of Ref. [26]; SLC4A1, HBA2, ALAS2). Strong conservation of these modules likely indicates their functional relevance in placental tissues. The robustness of our network was further confirmed as most of our modules were conserved with different combinations of network construction parameters such as power value, signed or unsigned network etc. (Table S3).

In our analysis, the “red” module, is positively associated with MPSP, is enriched in GO categories such as chromatin organization, nucleobase regulation,and DNA transcription. Changes in placental gene expression associated with MPSP have been previously reported in different studies^30,31; results from our study point totranscriptional regulation of gene expression as a mechanism underlying these associations. More specifically, pathway analysis displayed enrichment for estrogen receptor signaling as one of the pathways potentially mediated by the top hub genes of the “red” module. Importantly, the same red module is negatively associated with the REG domain of infant temperament. Regulation domain refers to the processes that modulate and control irritability and the expressions of negative emotions.^12,13 Reductions in this domain are an indicator of ”difficult” temperament.¹⁶ For example, Poor regulation during infancy has been shown to increase the risk for later adverse behavioral and emotional outcomes.^32–34 While the association between MPSP and infant temperament has been previously shown, our study demonstrated an important potential biological pathway relevant to this connection.

We acknowledge several limitations of this study. First, the sample size of the study is limited and only allows detection of moderate gene effects. However, implementing WGCNA, which is also a dimension reduction approach, significantly reduces multiple comparison and increases study power. Second, we acknowledge that the placenta consists of heterogeneous mix of cells with distinct cell-specific transcriptome profiles; however, our placenta biopsies are collected using a standardized protocol to reduce this variability.³⁵ Third, measures of prenatal depression, anxiety, and infant temperament all relied on maternal report which is a subjective assessment. Future studies will benefit from assessments of infant temperament using more objective scales, by multiple informants or via observation.^36–38

In summary, our study provides supporting evidence that the placenta serves as a conduit linking MPSP and the infant temperament domain of regulation, possibly through transcriptional regulation of placental gene expression.

Abbreviations:

ANOVA

analysis of variance

BIC

Bayesian information criterion

EPDS

Edinburgh postnatal depression scale

IBQ-R

Infant behavior questionnaire-revised

IPA

ingenuity pathway analysis

LEneg

life events with negative valence

LPA

latent profile analysis

MPSP

maternal psychosocial stress during pregnancy

NEG

negative affect

PRAQ-R

pregnancy-related anxiety questionnaire–revised

PSS-14

14-item perceived stress scale

REG

regulatory capacity

SIP

stress in pregnancy

STAI-S

state-trait anxiety inventory STATE

STAI-T

state-trait anxiety inventory TRAIT

SUR

surgency

WGCNA

weighted gene co-expression network analysis