Early Life Adversity is Associated with Differential Gene Expression in Immune Cells: A Cluster-Based Analysis across an Acute Psychosocial Stressor

Abstract

Elucidating mechanisms by which early-life adversity (ELA) contributes to increased disease risk is important for mitigating adverse health outcomes. Prior work has found differences in immune cell gene expression related to inflammation and mitochondrial activity. Using a within-person between-group experimental design, we investigated differences in gene expression clusters across acute psychosocial stress and no-stress conditions. Participants were young adults (N = 29, aged 18 – 25 years, 62% female, 47% with a history of ELA). Gene expression was assessed in peripheral blood mononuclear cells collected at 8 blood draws spanning two 5-hour sessions (stress vs. no-stress) separated by a week, 4 across each session (number of observations = 221). We applied two unsupervised gene clustering methods – latent profile analysis (LPA) and weighted gene co-expression analysis (WGCNA) – to cluster genes with similar expression patterns across participants. LPA identified 11 clusters, 7 of which were significantly associated with ELA-status. WGCNA identified 5 clusters, 3 of which were significantly associated with ELA-status. LPA- and WGCNA-identified clusters were correlated, and all clusters were highly preserved across sessions and time. There was no significant effect of acute stress on cluster gene expression, but there was a significant effect of time, and significant differences by ELA-status. ELA-associated clusters related to RNA splicing/processing, inflammation, leukocyte differentiation and division, and mitochondrial activity were differentially expressed across time: ELA-exposed individuals showed decreased expression of these clusters at 90-minutes while controls showed increased expression. Our findings replicate previous work in this area and highlight additional mechanisms by which ELA may contribute to disease risk.

1. Introduction

Exposure to early-life adversity (ELA) increases the risk of early-onset morbidities, an outcome that may be driven in part by dysregulation of the immune system (Agorastos et al., 2019; Ganguly & Brenhouse, 2015). ELA-exposed individuals evince higher levels of inflammation, impaired latent virus control, and altered responses to immune challenge (Baldwin et al., 2018; Baumeister et al., 2016; Elwenspoek et al., 2017; Thurston et al., 2014). Mechanistic work linking ELA to immune dysregulation has shed light on intracellular processes, such as the regulation of genes controlling inflammatory processes and differential methylation profiles of leukocytes (Bower et al., 2020; Dieckmann et al., 2020; Hao et al., 2018; Schwaiger et al., 2016).

Altered gene expression in leukocytes may serve as a mechanism linking immune dysregulation and poor health outcomes often seen in ELA-exposed individuals (Kuhlman et al., 2022; Maes et al., 2022). A widely studied gene expression profile altered by ELA is the Conserved Transcriptional Response to Adversity (CTRA; Cole, 2019). The CTRA is characterized by a heightened inflammatory profile coupled with downregulated antiviral responses and antibody production in individuals with ELA (Marie-Mitchell & Cole, 2022; Slavich et al., 2023). The CTRA operates through beta-adrenergic signaling pathways, translating sympathetic nervous system (SNS) responses into alterations in transcription factor activities and the production of immune cells, including monocytes and dendritic cells (Levine et al., 2017; Powell et al., 2013).

Over time, these altered SNS and immune responses to stress may culminate in increased disease risk for ELA-exposed individuals; however, disentangling the impact of ELA on disease risk prior to disease onset presents a methodological challenge. Extant literature in this area is limited by methodological barriers (e.g., the application of either between- or within-person designs, measuring gene expression changes over short time frames, measurement of relationships in old age) that may obscure directionality of impacts or alternative mechanistic pathways among ELA, immune system dysregulation, and disease risk (Keren, 1993; Shalev et al., 2020). Functional differences in these processes (e.g., altered gene expression) may arise earlier in life and be detectable by within-person repeated measurements across perturbations to the stress response and immune systems.

Using a within-person, between-group laboratory-based experimental design, we aimed to overcome the aforementioned methodological barriers and probe for novel mechanisms linking ELA to disease risk via immune dysregulation. We used repeated gene expression data to identify clusters of genes with differential expression activity over time for ELA-exposed individuals compared to controls. Drawing on two methods for data-driven discovery of clustered gene activity, latent profile analysis (LPA) and weighted gene co-expression network analysis (WGCNA), we examined patterns of change in gene clusters across time, in response to acute stress, and by ELA-status. We hypothesized that gene cluster activity would be altered in individuals with a history of ELA with the goal of identifying novel mechanistic pathways for future hypothesis generation on the biological embedding of early adversity.

2. Materials & Methods

2.1. Participants

Detailed methods for participant recruitment and sampling have been previously reviewed (Etzel et al., 2022; Shalev et al., 2020). In brief, healthy individuals aged 18–27 were recruited from The Pennsylvania State University and surrounding vicinity by word of mouth and local advertisements on campus bulletin boards. All potential participants were screened for early adversity along 13 categories of events (e.g., child sexual and physical abuse, death of a parent or sibling, life threatening events) as assessed by clinical interview via the Stressful Life Events Screening Questionnaire (SLESQ; Goodman et al., 1998). Drawing on evidence that exposure to three or more early traumatic events confers higher risk for disease (Felitti et al., 1998), participants who responded to at least 3 incidents up to age 18 years were invited to participate in the ELA group. Respondents’ examples for adverse exposures in this study included (unsubstantiated) child abuse and neglect, severe violence exposure, parental loss, suicide of a close friend or a family member, severe illness of an immediate family member or car accidents. The SLESQ was also used to screen participants into the control group. Individuals endorsing zero traumatic exposures up to age 18 years were invited to the study as control participants. Other general selection criteria included no current medical or endocrine illness (confirmed by self-report and physical examination), current non-smokers and no current medication use on a regular basis, including psychiatric medications. The final sample included 29 participants (13 ELA-group, 16 control-group; mean age=21.8±1.4, 62% female). The study was approved by the Ethics Committee at the Pennsylvania State University, registered at ClinicalTrials.gov (Identifier: NCT03637751), and all participants provided written informed consent. Participants received a modest monetary incentive for participation.

2.2. General Procedure

Participants visited The Pennsylvania State University’s Clinical Research Center between 11:00 am – 4:15 pm during weekdays, one week apart, on the same day of the week, with randomized counter-balanced ordering of the experimental manipulations (e.g., stress and no-stress conditions). Thirty minutes after arrival and 30 minutes prior to the experimental condition, trained nurses completed a physical exam and inserted an IV catheter into the antecubital vein. In the stress condition, participants completed the Trier Social Stress Test (TSST), a free speech and a mental arithmetic task of 10 minutes duration performed in front of a panel of two committee members (mixed gender) with a camera and microphone situated between the interviewers (Kirschbaum et al., 1993). For the no-stress condition, participants sat in a room, read magazines, and refrained from stressful activities (e.g., cell phone use was restricted). A standardized low-calorie meal was provided after the third blood draw at 1:45 pm (see Figure 1 for study design [Shalev et al., 2020]). A total of eight blood draws, four across each session, were collected from each participant (total number of samples = 221).

Figure 1.

Study design for both sessions (TSST and no-stress condition). Sessions were separated by one week. Time is denoted as minutes measured from the 15-minute long TSST or resting condition (0=midway point in session).

2.3. RNA Extraction, Quantification and Preprocessing

From each blood draw, peripheral blood mononuclear cells (PBMCs) were immediately isolated and stored at 4°C in RNAlater solution. 1 ug of total RNA per sample was extracted the next day using QIAamp RNA Blood Mini Kit (Qiagen). RNA quality was verified using the Agilent Bioanalzyer, followed by library construction. Library preparation was done using the QuantSeq 3’mRNA-Seq Library Prep Kit FWD for Illumina (Lexogen) supplemented with UMI (unique molecular index) as per manufacturer’s instructions. Sequencing was performed on Illumina’s Nova-seq 6000. Reads obtained from sequencing were preprocessed using BBDuk (version 38.87) for quality trimming, polyA trimming, and read decontamination. Reads were aligned to the human reference genome (GRCh38) using STAR (version 2.7.9a; Dobin et al., 2013). Read counts were generated using featureCounts (version 2.0.1); mapped reads were collapsed according to UMI sequences of each read to remove PCR duplicates (Liao et al., 2014). After reads were quantified, subsequent data analyses were conducted in $R$ (version 4.1.2). Counts per million (CPM) values for each gene were calculated using edgeR (version 3.38.2; Robinson et al., 2010). Genes with a CPM value higher than 1 in 200 or more samples (13,381 genes) were kept, and all other genes were excluded (Sha et al., 2015). After filtering out lowly expressed genes, the original count matrices were normalized using the Trimmed Mean of M (TMM) method, which does not correct for gene length (Smid et al., 2018), as is recommended by Lexogen when using QuantSeq Library Prep Kits (Moll et al., 2014). After normalization, CPM-log2 transformed expression values were then recalculated using the filtered gene’s raw count values. Batch correction was performed using the ComBat function in SVA (version 3.44.0; Leek et al., 2012).

2.4. Leukocyte Proportion Estimations

To account for differing PBMC proportions (i.e., proportion of lymphocytes vs. monocytes) between samples, we used additional 4 mL blood tubes for complete blood count (CBC) estimates at each time-point (Quest Diagnostics). We calculated PBMC immune cell proportions by only considering the lymphocyte and monocyte proportions from CBC measurements.

To supplement immune cell proportion estimates for individuals whose CBC estimates were not available (29 samples out of 221), we used DNA methylation measurements to estimate immune cell proportions using the IDOL-extended algorithm (Salas & Koestler, 2018/2023). Previous work has shown that even under acute psychosocial stress, DNA methylation and CBC estimates of immune cell proportions can be used interchangeably (Apsley, Ye, et al., 2023). The collection of DNA methylation data (Infinium MethylationEPIC v1) has been reported previously (Apsley, Etzel, et al., 2023). We computed estimates of PBMC subtype proportions from DNAm data using the ProjectCellType_CP function in FlowSorted.Blood.EPIC (version 2.4.2; Salas & Koestler, 2018/2023), which is equivalent to the ProjectCellType function in minfi. The FlowSorted.BloodExtended.EPIC library was used as reference data for blood cell proportion estimates (Salas et al., 2022). Transcriptome Representation Analysis (TRA; Abbas et al., 2005) was also considered as a way to supplement the unavailable CBC estimates. TRA estimates were less well correlated with CBC estimates than DNA methylation estimates and thus we used methylation estimates for all analyses (see Supplemental File S1 for additional TRA information).

2.5. Construction of Gene Expression Clusters using Latent Profile Analysis

Clusters of genes with similar expression patterns across participants were identified using gene expression measurements from T1 (baseline; −30 minutes) of both the no-stress and stress sessions as input for LPA. Latent cluster calculations were performed using tidyLPA (version 1.1.0; Rosenberg et al., 2018) on z-scored gene expression values. Two types of models based on variance and covariance structure of estimated clusters were used: EEI (equal volume, equal shape [and undefined orientation]), and VVI (varying volume, varying shape [and undefined orientation]). All models were fitted for 1–15 cluster solutions, keeping only models that successfully converged. Model fit statistics of Approximate Weight of Evidence (AWE), Bayesian Information Criterion (BIC), Sample Size-Adjusted Bayesian Information Criterion (SABIC) and Consistent Akaike’s Information Criterion (CAIC) were used to determine which model and cluster number best fit the data (see Supplemental File S2 for model fit statistics for all converging model iterations).

After determining the model and number of clusters to use, genes were assigned to clusters based on their maximum profile probabilities, with all genes being uniquely assigned to a cluster. The Gene Ontology (GO) aspect (Aleksander et al., 2023; Ashburner et al., 2000) of “Biological Processes” was tested for overrepresentation in LPA gene clusters compared to the background 13,381 genes included in our analysis (Wu et al., 2021), with statistical significance set at FDR-adjusted p-value < .05.

2.6. Construction of Gene Expression Clusters using Weighted Gene Co-Expression Network Analysis

Using a second method to identify clusters of co-expressed genes, an unsigned weighted gene co-expression network was constructed using the WGCNA package (Langfelder & Horvath, 2008). The gene co-expression network was formed using gene expression measurements from T1 of both the no-stress and stress sessions. To capture pairwise relationships between $i$th and $j$th genes, we first generated biweight midcorrelation coefficients ($s_{i,j}$), superior to Pearson and Spearman correlation procedures for gene co-expression module identification as it is defined by the median rather than the mean, which gives preferential weight to values closer to the median, and is more robust to outliers (Song et al., 2012; Zheng et al., 2014). Next, correlations between genes were transformed to measures of weighted network adjacency, $A_{i,j}=s_{i,j}^{\beta }$. A smoothing parameter ($\beta$) value of 7 was selected by visual analysis of the plotted logged proportion of nodes with connectivity $k$, $[\log (p(k))]$, against the distribution of connectivity $k$ for different $\beta$, and checking for an R² approximating 1 (see Supplemental File S3 for details). Selection of $\beta$ for this transformation functioned as a method of soft-thresholding to selectively filter out lower correlations and maintain scale-free topology, a fundamental property of biologically plausible gene networks wherein hub nodes are sparse and highly interconnected to other network features (Langfelder & Horvath, 2008; Song et al., 2012). The resulting adjacency matrix was used to calculate the topological overlap measure (TOM), a measure of interconnectedness between genes, and the final dissimilarity measure (1 – TOM) used as input for hierarchical clustering.

Using dynamicTreeCut in $R$, a robust cutting method that is semi-adaptive to shape of the tree (Langfelder et al., 2008), gene clusters were selected from the previously generated hierarchical cluster tree using default parameters. Cluster eigengenes (weighted average of expression of all genes in each module) were used to collapse highly correlated clusters (eigengene correlation > 0.85). Similar to LPA analyses, the GO aspect (Aleksander et al., 2023; Ashburner et al., 2000) of “Biological Processes” was tested for overrepresentation in WGCNA gene clusters, with statistical significance set at FDR-adjusted p-value < 0.05.

To determine whether LPA- and WGCNA-identified clusters generated from T1 of both the stress and no-stress sessions were conserved across the duration of each session (i.e., whether the genes that were correlated in expression at T1 remained correlated over the duration of the sessions), preservation statistics (Langfelder et al., 2011) were calculated for each cluster over all time-points and sessions (see Supplemental File S4 for detailed preservation statistics graphic). Clusters with an average summary Z-preservation statistic > 10, indicating that the cluster was statistically significantly conserved over all time-points, were included in further analyses (Langfelder et al., 2011).

2.7. Modeling Gene Expression Clusters Across Acute Psychosocial Stress

To test whether LPA or WGCNA clusters varied over time, in response to acute stress, and by ELA exposure, we extracted the absolute value of the first principal component of each LPA-identified cluster and the cluster eigengene from each WGCNA-identified cluster. Mixed linear models were constructed using the following equations:

$$\begin{gathered} y_{it}={\beta }_{0i}+{\beta }_{1}Tim{e}_{it}+{\beta }_{2}Sessio{n}_{it}+{\beta }_{3}EL{A}_i+{\beta }_{4}Tim{e}_{it}\times Sessio{n}_{it}+{\beta }_{5}Tim{e}_{it}\times EL{A}_i+ \\ {\beta }_{6}Sessio{n}_{it}\times EL{A}_i+{\beta }_{7}Tim{e}_{it}\times Sessio{n}_{it}\times EL{A}_i+{\beta }_{cov}Co{v}_{it}+{\epsilon }_{it} \end{gathered}$$

$${\beta }_{0i}={\gamma }_{00}+{\pi }_{0i}$$

$${\beta }_{2i}={\gamma }_{20}+{\pi }_{2i}$$

where:

$y_{it}$ denotes the specific cluster expression score for individual $i$ at time $t$,

${\beta }_{0i}$ denotes the cluster expression value of individual $i$ during the no-stress session at T1,

${\beta }_1$ denotes the average effect between T1 and T2, T3 or T4 within the no-stress session,

${\beta }_2$ denotes the average difference between stress and no-stress sessions,

${\beta }_3$ denotes the effect of ELA status,

${\beta }_4$ denotes the interaction effect of time and session,

${\beta }_5$ denotes the interaction effect of time and ELA,

${\beta }_6$ denotes the interaction of session and ELA,

${\beta }_7$ denotes the interaction of time, session and ELA,

${\beta }_{cov}$ denotes the vector of covariate effects,

${\epsilon }_{it}$ denotes the residual value of individual $i$ at time $t$,

${\gamma }_{00}$ denotes the grand mean of module eigengene value at T1 of the no-stress session,

${\gamma }_{20}$ denotes the grand mean of session effects when including random effects for session,

${\pi }_{0i}$ denotes the residual value of individual $i$ from the grand T1 mean,

${\pi }_{2i}$ denotes the residual value of individual $i$ from the session effect.

Covariates used in these models were cell count proportions, RNA-seq batch number, BMI, sex (male as the reference group), age, minority-status (non-minority status as the reference group), and SES. The best fitting model for each cluster was chosen based on likelihood ratio tests (LRT) and a p-value threshold of P<.05 (detailed model iterations are presented in Supplemental Files S5 and S6). To visualize the relationship between ELA-status and gene clusters, cluster expression z-scores were plotted across time stratified by session and ELA-status. Gene clusters that had significant estimates for time, session, or ELA-status were probed for biological significance using gene set overrepresentation methods.

3. Results

3.1. Sample Demographics and Descriptive Statistics

ELA and control groups did not differ with respect to age, SES, % minority, or % female. There was a significant difference in BMI, with the ELA group having a lower average BMI (21.5) compared to the control group (26.1) (Wilcoxon-Mann-Whitney U-test P<.001; Table 1).

Table 1.

Sample summary statistics.

Variable, mean (SD)	Total (N=29)	ELA (N=13)	Control (N=16)	P-value
Age	21.8 (1.4)	22.1 (1.7)	21.5 (1.1)	0.322
SES (average)	3.2 (0.8)	3.1 (1.0)	3.4 (0.6)	0.373
% Minority Status	34%	23%	44%	0.113
% Female	62%	69%	56%	0.740
Body Mass Index	24.1 (3.7)	21.5 (2.0)	26.1 (3.4)	<0.001

3.2. Validation of the Acute Stress Induction Procedure

As previously reported (Etzel et al., 2022; Shalev et al., 2020), the TSST was effective for induction of acute stress as validated through assessments of HPA-axis and SNS activation, specifically, salivary cortisol and mean arterial pressure (MAP). Exposure to the TSST induced higher cortisol and a steeper decline in cortisol over time compared to the no-stress session. There were no differences in cortisol response by ELA-status. For MAP, exposure to the TSST was associated with higher MAP at one minute prior to TSST exposure before falling towards baseline. Compared to the control group, the ELA group exhibited greater increases in MAP in response to the TSST and a prolonged increase in MAP after TSST exposure (Etzel et al., 2022).

3.3. LPA and WGCNA Approaches Identified Similar Clusters of Genes and Associations with ELA

LPA identified 11 distinct clusters. On average, each cluster contained 1,216 genes (range:768–2,085) and had an average posterior probability (i.e., likelihood of each gene belonging to the respective cluster) of 0.95 (see Table 2 for details). Seven out of 11 clusters were initially significantly associated with ELA-status (all P<.014; see Figure 2). Clusters 6, and 8 remained significant for ELA-status after controlling for covariates. None of the LPA-identified clusters had significant main effects of session (see Supplemental File 5). Within the clusters associated with ELA-status, there was a main effect of ELA-status in cluster 6 (b=1.08±.05, P=.037) such that participants with ELA had increased cluster 6 gene expression activity compared to controls (see Table 3). There was a significant ELA X Time interaction for clusters 6, 7, and 8 (all P<.08) such that participants with ELA had lower cluster gene expression at T3 (90-minutes post-session) compared to baseline (see Table 3 and Figure 3).

Table 2.

LPA and WGCNA Gene Cluster Information. Average posterior probabilities represent the average probability that genes belonged to their respective cluster. Average module memberships represent the average correlation of the expression of all genes in a cluster with that cluster’s first principal component. Similar LPA and WGCNA clusters are shown across from one another, with LPA clusters 6 and 7 corresponding to the WGCNA Blue cluster. Clusters with no significant GO “Biological Process” terms are denoted as “NA.”

LPA Cluster	Number of Genes	Average Posterior Probability	Biological Processes	WGCN A Cluster	Numb er of Genes	Average Module Members hip	Biological Processes
Cluster 1	1,891	0.95	NA	-	-	-	-
Cluster 2	932	0.95	Meiosis and sensory organ development	-	-	-	-
Cluster 3	2,085	0.95	NA	-	-	-	-
Cluster 4	1,582	0.96	NA	-	-	-	-
Cluster 5	1,219	0.95	Chromatin organization and leukocyte differentiation	Brown	1,419	0.32	Mitotic chromoso me segregation
Cluster 6	1,187	0.95	Ribosomal-related activity and innate immune response	Blue	2,772	0.40	Ribosomal-related activity and RNA processing
Cluster 7	806	0.95	RNA processing
Cluster 8	768	0.95	Translation and mitochondrial activity	Yellow	646	0.33	Translation and mitochondrial activity
Cluster 9	1,113	0.95	Intercellular and cell membrane-related activity	-	-	-	-
Cluster 10	817	0.97	Synapse activity and morphogenesis	-	-	-	-
Cluster 11	981	0.93	NA	-	-	-	-
-	-	-	-	Black	874	0.33	NA
-	-	-	-	Pink	284	0.14	B-cell receptor signaling

Figure 2.

Correlation heatmap of ELA-status, LPA Clusters, and WGCNA Clusters. Each cell contains the bivariate correlation between variables with a corresponding p-value in parentheses. Darker red colors indicate a stronger positive correlation and darker blue colors indicate a stronger negative correlation. Correlations between all clusters and ELA status are displayed in the first row.

Table 3.

Effects of Time and Acute Stress on LPA Cluster Expression. LRT indicates likelihood ratio test and is shown when a model including more covariates was a significantly better fit than the reduced model (see Supplemental File S5 for details on model specifications).

First Principal Component for LPA Cluster	Cluster 5	Cluster 6	Cluster 7	Cluster 8
	Estimate (SE)	Estimate (SE)	Estimate (SE)	Estimate (SE)
Fixed effects
Intercept	0.172 (0.40)	−0.217 (0.38)	−0.173 (0.45)	−0.411 (0.41)
Session	0.188 (0.21)	0.326 (0.25)	0.332 (0.33)	0.525 (0.33)
Time (30 mins)	−0.020 (0.22)	−0.055 (0.27)	−0.035 (0.22)	0.333 (0.36)
Time (90 mins)	0.200 (0.37)	0.415 (0.26)	0.373 (0.21) †	0.714 (0.35) *
Time (240 mins)	−0.139 (0.22)	0.107 (0.27)	−0.008 (0.22)	0.475 (0.35)
Interaction: Session x Time (30 mins)	−0.001 (0.28)	−0.036 (0.34)	−0.002 (0.28)	−0.446 (0.45)
Interaction: Session x Time (90 mins)	0.067 (0.27)	−0.162 (0.33)	−0.053 (0.27)	−0.320 (0.44)
Interaction: Session x Time (240 mins)	0.219 (0.26)	0.151 (0.32)	0.329 (0.27)	−0.084 (0.43)
ELA-status	0.589 (0.56)	1.082 (0.50) *	1.011 (0.62)	1.124 (0.54) *
Interaction: Session x ELA	−0.311 (0.53)	−0.620 (0.36) †	−0.520 (0.48)	−0.612 (0.47)
Interaction: ELA x Time (30 mins)	−0.086 (0.33)	−0.180 (0.40)	−0.184 (0.33)	−0.410 (0.54)
Interaction: ELA x Time (90 mins)	−0.545 (0.32) †	−0.803 (0.39) *	−0.865 (0.32) **	−1.329 (0.52) *
Interaction: ELA x Time (240 mins)	0.040 (0.33)	−0.053 (0.41)	0.000 (0.33)	−0.390 (0.54)
Interaction: ELA x Session x Time (30mins)	0.184 (0.40)	0.121 (0.50)	0.113 (0.41)	0.339 (0.66)
Interaction: ELA x Session x Time (90mins)	0.230 (0.39)	0.357 (0.48)	0.385 (0.39)	0.482 (0.64)
Interaction: ELA x Session x Time (240mins)	−0.050 (0.40)	0.064 (0.49)	−0.147 (0.40)	0.133 (0.66)
Random effects
Subject-level Variance	0.836 (0.92)	0.705 (0.84)	1.252 (1.12)	0.584 (0.76)
Session-Intercept Covariance	−0.104	--	−0.101	--
Session	0.638 (0.80)	--	0.448 (0.67)	--
Residual Variance	0.154 (0.39)	0.234 (0.48)	0.156 (0.40)	0.418 (0.65)
AIC/BIC	334.693/416.691	363.882/439.806	340.127/422.125	423.355/499.280
LRT p-value	P<.001	--	P<.001	−-

^†P<.10

^*P<.05

^**P<.01.

Figure 3.

Effects of time, stress, and ELA status on cluster expression. Z-scored expression of LPA- and WGCNA-identified clusters across time, stratified by both session and ELA-status. Points represent group mean cluster scores. The top row indicates LPA cluster scores and the bottom indicates WGCNA cluster scores.

Using WGCNA, we identified 5 distinct clusters of genes. On average, each cluster contained 1,199 genes (range: 284–2,772) and had an average module membership value (i.e., a measurement of the average correlation of each gene’s expression with the cluster’s first principal component) of 0.30 (see Table 2). Three out of 5 clusters were initially significantly associated with ELA-status (all P<.01; see Figure 2). None of the WGCNA-identified clusters had significant main effects of session (see Supplemental File 6). Within the clusters associated with ELA-status, there was a main effect of ELA-status in the Blue cluster (b=1.18±.56, P=.040) such that participants with ELA had increased Blue cluster gene expression activity compared to controls (see Table 4). There was a significant ELA X Time interaction for the Brown and Blue clusters (all P<.035) such that participants with ELA had lower cluster gene expression at T3 (90-minutes post-session) compared to baseline (see Table 4 and Figure 3).

Table 4.

Effects of Time and Acute Stress on WGCNA Cluster Expression. LRT indicates likelihood ratio test and is shown when a model including more covariates was a significantly better fit than the reduced model (see Supplemental File S6 for details on model specifications).

First Principal Component for WGCNA Cluster	Brown	Blue	Yellow
	Estimate (SE)	Estimate (SE)	Estimate (SE)
Fixed effects
Intercept	0.057 (0.42)	−0.823 (0.43) †	0.621 (0.39)
Session	0.302 (0.42)	0.677 (0.36) †	−0.049 (0.43)
Time (30 mins)	0.174 (0.33)	0.361 (0.39)	−0.115 (0.24)
Time (90 mins)	0.425 (0.32)	0.742 (0.38) †	0.012 (0.23)
Time (240 mins)	0.196 (0.32)	0.423 (0.39)	−0.162 (0.24)
Interaction: Session x Time (30 mins)	−0.157 (0.41)	−0.361 (0.49)	0.151 (0.30)
Interaction: Session x Time (90 mins)	−0.099 (0.40)	−0.402 (0.48)	0.133 (0.29)
Interaction: Session x Time (240 mins)	−0.063 (0.39)	−0.198 (0.47)	0.133 (0.29)
ELA-status	0.827 (0.58)	1.179 (0.56) *	−0.113 (0.53)
Interaction: Session x ELA	−0.290 (0.61)	−0.633 (0.52)	0.371 (0.60)
Interaction: ELA x Time (30 mins)	−0.441 (0.49)	−0.425 (0.59)	−0.120 (0.36)
Interaction: ELA x Time (90 mins)	−1.005 (0.47) *	−1.368 (0.56) *	−0.348 (0.34)
Interaction: ELA x Time (240 mins)	−0.279 (0.50)	−0.275 (0.59)	0.127 (0.36)
Interaction: ELA x Session x Time (30mins)	0.419 (0.60)	0.441 (0.72)	0.088 (0.44)
Interaction: ELA x Session x Time (90mins)	0.389 (0.58)	0.747 (0.69)	0.069 (0.42)
Interaction: ELA x Session x Time (240mins)	0.041 (0.60)	0.152 (0.72)	−0.214 (0.44)
Random effects
Subject-level Variance	0.685 (0.83)	0.568 (0.75)	0.627 (0.79)
Session-Intercept Covariance	−0.129	--	−0.115
Session	0.642 (0.80)	--	1.127 (1.06)
Residual Variance	0.345 (0.59)	0.498 (0.71)	0.183 (0.43)
AIC/BIC	412.995/494.992	442.062/517.985	358.310/440.307
LRT p-value	P<.001	--	P<.001

^†P<.10

^*P<.05

^**P<.01.

LPA and WGCNA clusters exhibited substantial overlap in biological processes identified via overrepresentation analysis (see Figures 4 and 5). LPA-identified cluster 5 and WGCNA-identified Brown cluster contained terms related to leukocyte differentiation including chromatin and chromosome organization (GO:0006325 and GO:0007059), LPA-identified cluster 6 and 7 and WGCNA-identified Blue clusters contained terms for ribosomal-related activity (GO:0042254 and GO:0022613) and RNA processing (GO:0006396), and LPA-identified cluster 8 and WGCNA-identified Yellow cluster contained terms related to protein translation (GO:0006412) and mitochondrial-activity (GO:0009060 and GO:0042775). In general, LPA and WGCNA clusters exhibited significant correlations with one another, with clusters that had overlapping GO terms having significantly higher correlations than clusters with no overlapping terms (two-sampled t-test P<.01; see Figure 2).

Figure 4.

Biological significance of LPA clusters. Results of overrepresentation analyses for LPA identified gene clusters. GO processes were significantly overrepresented against a background of all 13,381 genes included in analyses. The x-axis indicates the FDR-adjusted p-value and the broken vertical line indicates a significance threshold of FDR-adjusted P< .05. Shading of GO processes is based on magnitude of -log FDR-adjusted p-value. See Supplemental File S7 for full gene list used in overrepresentation analyses and Supplemental File S8 for a full list of every GO process identified.

Figure 5.

Biological significance of WGCNA clusters. Results of overrepresentation analyses for WGCNA identified gene clusters. GO processes were significantly overrepresented against a background of all 13,381 genes included in analyses. The x-axis indicates the FDR-adjusted pvalue and the broken vertical line indicates a significance threshold of FDR-adjusted P<.05. Shading of GO processes is based on magnitude of -log FDR-adjusted p-value. See Supplemental File S7 for full gene list used in overrepresentation analyses and Supplemental File S8 for a full list of every GO process identified.

3.4. LPA and WGCNA Clusters Exhibited Temporal Stability in Cluster Identity and Temporal Variation in Cluster Expression

LPA- and WGCNA-identified clusters were preserved over all time-points and across both the no-stress and stress sessions, meaning genes that were co-expressed during T1 of the no-stress and stress sessions continued to be co-expressed across all time points and across both the no-stress and stress sessions (see Supplemental File S4 for preservation statistic summaries).

The preservation of module identity over all time-points and sessions enabled us to examine how ELA, time and acute stress may have interacted to influence gene cluster expression (see Table 3, Table 4, and Figure 3). An overall pattern emerged in which ELA-exposed participants generally had higher cluster expression across sessions and time (see Figure 3). However, at T3 (90 minutes) there was a crossover effect wherein ELA-exposed participants had a decrease in cluster expression compared to baseline (ELA=1 x Time [90min]; cluster 5: $\beta =-0.545$, P=.09; cluster 6: $\beta =-0.803$, P=.040; cluster 7: $\beta =-0.865$, P=.008; cluster 8: $\beta =-1.329$, P=.012; Brown cluster: $\beta =-1.005$, P=.035; Blue cluster: $\beta =-1.368$, P=.017) and control participants had an increase in cluster expression compared to baseline (ELA=0 X Time [90min]; cluster 7: $\beta =0.373$, P=.09; cluster 8: $\beta =0.714$, P=.044; Blue cluster: $\beta =0.742$, P=.06). These changes were identified across both sessions indicating that cluster expression differences due to ELA were temporally influenced regardless of acute psychological stress exposure.

4. Discussion

Identification of molecular pathways linking ELA to later life negative health outcomes is critical to intervention and clinical efforts. Extant research in this arena points to dysregulated immune function as a potential mechanism. Prior work applying WGCNA to monocytes of ELA-exposed older adults collected at a single time point after acute stress exposure found evidence of differential expression of gene clusters related to inflammation and wound healing (Dieckmann et al., 2020). Similarly, gene co-expression analyses in whole blood of adults with and without ELA and social anxiety disorder identified differential expression of genes involved in interleukin regulation and production (Edelmann et al., 2023). Recent work in adolescents with and without childhood maltreatment found differential expression of immune-related genes deriving predominately from non-classical CD16+ monocytes (Kuhlman et al., 2024), a subtype of monocyte that contribute to increased inflammatory cytokines IL-1$\beta$ and TNF-$\alpha$ (Mukherjee et al., 2015). These prior findings reinforce our understanding of ELA as a driver of systemic alterations to immune-related processes, particularly inflammation, but do not offer insight into the precise role stress may play in these associations, and could be contaminated by early disease processes within older adult study populations.

Here, we aimed to provide insight into this causal chain by identifying gene expression patterns that may be driving ELA-associated altered immune function in young adults before disease onset. Using gene expression data from PBMCs of participants with and without a history of ELA exposed to an acute stressor and a no-stress comparison condition, we constructed clusters of gene expression using two approaches (LPA and WGCNA). Both clustering techniques identified gene sets highly correlated in expression between approaches and overlapped in the biological processes they represented. Further, clusters identified by both methods were preserved across time and sessions, highlighting the stability of network construction and allowing for examination of changes in cluster expression across condition, time, and by ELA-status.

Though previous work in this cohort pointed to stress-induced differences in immune parameters (Etzel et al., 2022), here we found no significant effect of acute stress on gene expression clusters in either ELA-exposed or control groups. Across most clusters, participants with ELA-exposure appeared to have increased gene expression activity overall, but with a significant decrease in expression at 90 minutes regardless of session. This pattern was distinct from the control group which experienced increased expression across clusters at 90 minutes, again regardless of session. Our finding of a lack of differences between the acute stress and no stress sessions could imply diminished gene expression reactivity to acute psychosocial stress relative to natural circadian changes, or else could be due to stress experienced across both sessions such as the unavoidable additional stressor of venipuncture present at the beginning of both sessions (Shalev et al., 2020). Alternatively, gene expression changes may have been driven by repeated blood loss across both sessions, independent of stress (Clapes et al., 2016).

ELA-exposed participants exhibited differential expression in LPA- and WGCNA-identified clusters enriched for genes related to broad biological processes. One such process is differential transcriptional control of leukocyte proliferation and differentiation. Specifically, LPA clusters 5 and 6, which were differentially expressed in ELA-exposed participants, were enriched for processes related to leukocyte differentiation and the innate immune response. Moreover, LPA cluster 5 and the WGCNA Brown cluster were enriched for processes related to chromatin organization and chromosome condensation and segregation. Chromosome condensation and segregation are crucial steps during immune cell proliferation (Alberts et al., 2002), and are critical regulators of innate and adaptive immune responses (Kan & Hodgkin, 2014; Laphanuwat & Jirawatnotai, 2019). Altered transcription of genes related to leukocyte differentiation and chromosome organization in ELA-exposed participants contributes to our understanding of previously reported differences for this sample in leukocyte counts between ELA-exposed and control participants (Etzel et al., 2022). Further, this aligns with expected differences under the CTRA model of gene regulation post-adversity exposure (Cole, 2019).

LPA-identified clusters 6 and 7 and WGCNA-identified Blue cluster were enriched for genes related to ribosomal activity and RNA processing. To our knowledge, no prior studies have linked ELA to altered immune cell gene expression related to ribosomal activity or RNA processing; however, previous research implicates dysregulation of these processes in immune cells as contributors to negative health outcomes (Bianco & Mohr, 2019; Cui et al., 2022; Jin et al., 2022; Zhou et al., 2015). Thus, dysregulated ribosomal activity and RNA processing in immune cells could serve as a potential mechanism by which ELA increases disease risk.

Lastly, LPA-identified cluster 8 and the WGCNA-identified Yellow cluster were enriched for genes related to mitochondrial-related activity and protein translation. ELA-exposed participants generally evinced higher expression of these clusters than control participants. Extant research has uncovered a potential mechanistic role for dysregulated bioenergetic processes in those with early adverse exposures (Ridout et al., 2018; Zitkovsky et al., 2021); our finding of differential cluster activity related to mitochondrial processes lends further support for this hypothesis.

We believe our findings contribute meaningfully to the formation of hypotheses related to ELA and increased disease risk, though we acknowledge limitations. The within-person between-group experimental design increases our power to detect differences by ELA-status, but the sample size is small. We were underpowered to conduct a genome-wide expression study and thus applied two methods of gene clustering (LPA and WGCNA) as a form of data reduction. WGCNA guidelines suggest no less than 15 measurements for construction of gene co-expression networks (Langfelder & Horvath, 2008), and we have satisfied this guideline with the inclusion of 8 repeated measures within 29 individuals. Further, all samples passed the preprocessing agnostic identification of outliers and were thus able to be included in our network construction. The gene clusters we have identified are informative, providing support for ELA-induced differential transcriptional control of RNA processing and bioenergetic control of immune cells, as well as corroboration for our prior work examining differences in leukocyte counts due to ELA-exposure, but we were unable to determine expression of specific genes that may be driving these effects. Gene-specific hypotheses must be examined in larger, more diverse samples. Finally, though we controlled for sex in all analyses (see Supplemental Files S5 and S6 for detailed model iterations including fixed effects estimates and standard errors for sex as a covariate), we were unable to explore sex-related differences given our sample size.

In conclusion, we applied two methods of gene clustering across acute stress and no-stress conditions to determine clusters of genes operating differently between ELA-exposed participants and a control group. Though we found no evidence for acute psychosocial stress-induced differences in gene cluster expression, ELA-exposed participants had dysregulated patterns of expression in gene clusters related to regulation of leukocyte differentiation and proliferation, chromatic organization, RNA processing, and mitochondrial activity that differed from the control group. These findings have implications for ELA-exposed individuals and future risk of disease. Future work should explore these possible mechanistic links between ELA and disease risk given that early prevention is often more effective than later intervention in reducing disease burden.

Highlights:

• Differential expression of gene clusters for young adults with early life adversity

• Reinforces prior findings of differences in inflammation and mitochondrial activity

• Novel clusters identified for RNA processing and leukocyte differentiation/division

• Early life adversity associates with these clusters regardless of stress conditions

Abbreviations:

BMI

body mass index

CBC

complete blood count

CTRA

Conserved Transcriptional Response to Adversity

ELA

early-life adversity

HPA

hypothalamic pituitary adrenal

LPA

Latent Profile Analysis

MAP

Mean Arterial Pressure

PBMCs

Peripheral Blood Mononuclear Cells

SES

socioeconomic status

SLESQ

Stressful Life Events Screening Questionnaire

SNS

Sympathetic Nervous System

TSST

Trier Social Stress Test

WGCNA

Weighted Gene Co-expression Network Analysis