High-throughput screening of glucocorticoid-induced enhancer activity reveals mechanisms of stress-related psychiatric disorders

Significance

Stress exposure is associated with increased risk for a wide range of psychiatric disorders. The individual long-term impact of stress may be moderated by genetic variants. Using a massively parallel enhancer assay, we identify genetic variants located in gene regulatory elements that modulate the transcriptomic response to glucocorticoids, a major stress hormone. These variants are enriched for variants associated with psychiatric disorders and neurobehavioral traits and influence transcripts that are altered in the postmortem brain of individuals with psychiatric disorders. Furthermore, these genetic variants are associated with changes in the physiological stress response. Together, this suggests that these functional genetic variants may be causally involved in psychopathology, possibly by influencing the physiological stress response and stress-responsive brain transcription.

Abstract

Exposure to stressful life events increases the risk for psychiatric disorders. Mechanistic insight into the genetic factors moderating the impact of stress can increase our understanding of disease processes. Here, we test 3,662 single nucleotide polymorphisms (SNPs) from preselected expression quantitative trait loci in massively parallel reporter assays to identify genetic variants that modulate the activity of regulatory elements sensitive to glucocorticoids, important mediators of the stress response. Of the tested SNP sequences, 547 were located in glucocorticoid-responsive regulatory elements of which 233 showed allele-dependent activity. Transcripts regulated by these functional variants were enriched for those differentially expressed in psychiatric disorders in the postmortem brain. Phenome-wide Mendelian randomization analysis in 4,439 phenotypes revealed potentially causal associations specifically in neurobehavioral traits, including major depression and other psychiatric disorders. Finally, a functional gene score derived from these variants was significantly associated with differences in the physiological stress response, suggesting that these variants may alter disease risk by moderating the individual set point of the stress response.

Trauma, chronic stress, and exposure to adverse life events are among the most robust risk factors for a large range of psychiatric disorders (1–3). Accordingly, adverse life events or traumatic experiences substantially increase the risk for major depressive disorder (MDD), posttraumatic stress disorder (PTSD), substance and alcohol use disorders (AAD), bipolar disorder (BPD), schizophrenia (SCZ), and other disorders (4). Despite their association with disease risk and disease presentation (more severe, more comorbidities, and more difficult to treat) (4), we still have insufficient knowledge about the underlying molecular mechanisms, thereby limiting our ability to develop therapeutic approaches targeting the consequences of exposure to adversity. Notably, not all individuals exposed to chronic stress or trauma develop disease, suggesting that underlying individual factors moderate the relationship between risk exposure and pathology (5), including genetic factors. Studies using SNP-based heritability measures in the UK Biobank cohort indicate a genetic component moderating adversity-associated psychiatric disorders with heritability estimates of 12% for MDD alone but of 24% in MDD patients with reported trauma (6). This supports the findings of another study in an independent cohort demonstrating context-specific heritability in complex traits, including gene x adversity effects in depression (7). Taken together, these studies suggest that genetic components moderate the risk of developing psychiatric disease following exposure to adversity.

However, identifying specific genetic variants moderating the effects of adversity has proven difficult and the investigation of such interactions at the candidate gene, genome-wide, and polygenic level has yielded inconsistent results (8, 9). One approach is to uncover adversity-moderating genetic components directly from genome-wide interaction studies or other approaches using large cohorts. However, these approaches are hampered by statistical requirements for large sample sizes that come with increased heterogeneity due to differences and limitations in defining and characterizing adversity, challenges in comprehensive mapping of the exposome, and differences in outcome definitions. An alternate approach to identify genetic variants that moderate the response to adversity is to dissect the genetic moderation of the more immediate impact of exposure to adversity, which elicits a complex multilevel physiological stress response. This response can be broken down into distinct biological components and offers the possibility to focus on specific potential biological mediator systems as well as on more proximal outcomes, such as stress-induced changes in gene expression.

One such component postulated to mediate the risk for psychiatric disorders is the stress hormone system (10, 11). Here, exposure to a threat or stressor leads to the activation of the hypothalamic-pituitary-adrenal (HPA) axis, culminating in the systemic release of glucocorticoids (GCs) from the adrenal glands. GCs bind with a high affinity to the mineralocorticoid receptor and with a lower affinity to the glucocorticoid receptor (GR). Both are cytoplasmic receptors that translocate to the nucleus once activated and elicit a transcriptional response by binding to specific gene regulatory elements (REs) associated with target genes (12). These transcriptional effects are essential for mediating a concerted stress response as well as for the negative feedback regulation of the HPA axis once the stressor has subsided. Importantly, the dynamic regulation of the HPA-axis is altered in patients with different psychiatric disorders, including MDD, as well as in those with exposure to early/chronic adversity (13, 14) indicating that dysregulated GC activity is an important pathomechanism in adversity-related disease. Genetic variants that alter GC-induced RE activity could thus be important moderators of the downstream, system-wide effects of stress hormones and hence disease risk.

In this study, we used synthetic self-transcribing active regulatory region sequencing (STARR-seq) to identify genetic variants that specifically moderate GC-dependent RE function and assess how such variants relate to disease risk and the physiological stress response. Synthetic STARR-seq allows for a direct and quantitative assessment of RE activity. By cloning putative REs downstream of a minimal promoter, the resulting mRNA transcripts carry the sequence of the RE itself, which eliminates the need for barcodes, which are required in other massively parallel reporter assays (15). To enhance the likelihood of identifying such variants, we screened all single nucleotide polymorphisms (SNPs) within 320 distinct linkage disequilibrium blocks that we previously showed to moderate GC-induced transcriptional responses in peripheral blood cells (eSNPs) (16). This study provides an approach to uncovering genetic variants that moderate the response to adversity and understanding the role of the stress hormone system in the pathogenesis of psychiatric disorders.

Results

Identification of Allele-Specific Regulatory Elements Modulated by a Glucocorticoid Receptor Agonist in Response to Stress Using STARR-seq.

We conducted an enhancer screen to study the role of SNPs in modulating the transcriptional response to the GR agonist dexamethasone (dex). We selected a set of 3,662 eSNPs that were implicated in the transcriptional response to dex by a previous genome-wide expression quantitative trait locus (eQTL) analysis (Fig. 1A and SI Appendix, Table S1) and shown to be enriched among genetic variants associated with psychiatric disorders (16). These eSNPs were associated with 296 dex-responsive transcripts and were organized into 320 independent genomic loci with strong linkage disequilibrium (eSNP LD bins). The eSNPs were enriched in enhancer regions and could modulate target gene expression via variant-specific enhancer activities. To perform fine-mapping for the identification of the true functional variants among the linked eSNPs, we used STARR-seq to identify which SNPs lie in REs that respond to GCs.

Fig. 1.

Identification of dex-responsive regulatory elements (DREs) using STARR-seq. (A) Overview of the method used to identify dex-responsive eQTLs. Individuals were genotyped and their gene expression was assayed before and after treatment with dex (3 h, 1.5 mg) resulting in 3,662 dex-responsive eQTLs (eSNPs) identified. (B) Experimental procedure to assay the regulatory activity of each allelic variant of the 3,662 eSNP regions using STARR-seq. In STARR-seq, candidate sequences are placed downstream of a minimal promoter, such that active enhancers induce their own expression and high-throughput sequencing reveals both the sequence identity and quantitative information regarding the activity of each sequence variant. For each eSNP, we assayed a 201 bp genomic region centered on the eSNP: one with the reference allele and the other with the alternative allele. (C) Volcano plot showing the log₂FC (dex/veh) and −log₁₀ FDR for each of the variants assayed by STARR-seq. The 547 significant (FDR < 0.1) DREs are highlighted as orange dots. (D) Regulatory activity for a representative inductive and repressive DRE is shown for both veh (blue) and dex (orange) conditions. (E) Distribution of inductive and repressive DREs within the eSNP bins (median = 2 DREs/bin). (F) Analysis of the STARR-seq data resulted in the identification of 79 DREs that exhibit allele-dependent activity exclusively in the dex condition, 61 exclusively in the veh condition, and 93 DREs with allele-dependent activity both in the veh and dex conditions. A representative example of each of these three types of variant-DREs is displayed. The activity of the REs is displayed as RNA/DNA ratios.

We generated a reporter library containing 7,324 unique 201 bp oligonucleotides encompassing the eSNP and control sequences. These libraries were transfected into two cell lines, an osteosarcoma cell line stably transfected with GR (U2OS-GR) and a brain glioblastoma (U138MG) cell line (SI Appendix, Fig. S1 A and B for additional details on their dex response). We then performed STARR-seq in these cell lines in a vehicle (veh) condition and following stimulation with 100 nM dex for 4 h to identify active enhancers and GC-induced changes in enhancer activity (Fig. 1B). From this analysis, we identified 547 unique dex-responsive regulatory elements (DREs) that changed their activity upon dex treatment (FDR < 0.1, Fig. 1 C and D and SI Appendix, Table S2). Half of the DREs showed increased activity after dex (inductive) while the other half showed reduced activity (repressive). Inductive DREs had a 1.8 greater mean fold change in activity in response to dex (mean log₂FC = 0.59, SD = 0.64) compared to repressive DREs (mean log₂FC = −0.32, SD = 0.19). We found that 44% (n = 142) of the 320 eSNP LD bins contained at least one DRE, with 43% (n = 61) of these containing a single DRE and 57% (n = 81) containing multiple DREs (median = 2, maximum = 69). Of the 81 multi-DRE LD bins, 12 contained exclusively inductive DREs, 15 exclusively repressive DREs, and the remaining 54 contained both inductive and repressive DREs within a single bin (Fig. 1E), supporting the idea of enhancers and silencers interacting within a locus to fine-tune regulation (17). We next looked at the association between the DREs and gene expression by generating a score summing the number of inductive (+1) and repressive (−1) DREs in each bin. Using publicly available RNA-seq data (18) from dex-treated U2OS-GR cells, we examined whether the score was linked to dex-induced differential expression in the predicted target genes. Of the 142 bins, 111 had available gene expression data in the U2OS-GR cells. We observed a significant correlation between the score and differential gene expression (Pearson correlation coefficient = 0.22, P-value = 0.022). When considering only genes significantly differentially expressed after dex (FDR < 0.1), the association became stronger (Pearson correlation coefficient = 0.41, P-value = 0.013). We then compared the STARR-seq activity of the reference to the alternative allele in the baseline (veh SNP-DREs) and dex condition (dex SNP-DREs) (Fig. 1B). We identified allele-dependent activity for 154 DREs in the veh condition and 172 DREs in the dex condition, with 93 showing allele-dependent activity in both conditions (dex/veh SNP-DREs) (FDR < 0.1, Fig. 1F and SI Appendix, Table S3). The effect sizes of the SNPs on DRE activity were not significantly different between the veh and dex SNP-DREs, with an absolute mean fold change between alleles of 1.70 (SD = 0.56) and 1.59 (SD = 0.56), respectively.

Characterizing the Genomic Context of DREs.

We next sought to understand the endogenous genomic context of the 547 DREs identified by STARR-seq since this is an episomal assay that assesses enhancer activity independent of chromatin context. To understand the endogenous context of the DREs, we examined GR occupancy using ChIP-seq data from U2OS-GR cells (19) which showed that only inductive DREs (±500 kB) were enriched in GR binding sites (twofold enrichment, P-value = 0.03) (Fig. 2 A–C). However, most DREs lacked GR occupancy (491/508), likely reflecting a closed chromatin state in the endogenous context of the cell lines used, as GR binding predominantly occurs at open chromatin (20). Therefore, we also performed a motif analysis on the DRE sequences to identify enriched sequence motifs (SI Appendix, Table S4). Consistent with the ChIP colocalizations results, inductive DREs were enriched in the GR motif (FDR = 3.4 × 10⁻¹¹) and androgen receptor motif (FDR = 7.2 × 10⁻⁸), which has a highly similar motif to that of GR, while repressive DREs showed enrichment of AP-1 (FDR = 5.6 × 10⁻³) and TEAD1 (FDR = 3.3 × 10⁻²), known tethering partners of GR (21, 22). In addition to chromatin accessibility, the 3D chromatin architecture has also been shown to influence RE activity, with evidence that dex-induced GR binding increases chromatin interactions between REs and the transcription start site of the regulated transcripts via 3D chromatin loops (23). Hi-C data from U2OS cells (24) showed overall that DREs (±5,000 kb from center) were enriched in chromatin looping anchor points compared to all 3,662 eSNPs (permutation P-value = 2 × 10⁻², fold enrichment = 1.45). Again, this held only for inductive (permutation P-value= 2 × 10⁻², fold enrichment = 1.8) but not repressive DREs (Fig. 2D). Next, we examined whether the DREs were located within enhancer regions in different tissues, including those relevant to neuropsychiatric disorders, such as the brain. The DREs were annotated to ChromHMM states (18-state model) across multiple tissues and showed significant enrichment in predicted enhancers (FDR < 0.1) including brain (fold enrichment = 2.0, permutation-based FDR = 9.0 × 10⁻³). The DREs were also enriched in enhancer regions of all ten available brain regions (FDR < 0.1) (Fig. 2E). Overall, these results show that the episomal STARR-seq assay identifies active DREs that are enriched in hallmarks of enhancers, including specific TF binding sites, chromatin loops, and histone marks in tissues relevant to neuropsychiatric disorders.

Fig. 2.

Understanding the endogenous context of the DREs. (A) Representative genomic context of an eSNP bin (highlighted yellow box). (B) Zoom-in of the eSNP bin showing GR ChIP-seq signal for the bin which contains eSNPs (black dots), a repressive DRE (blue dot), and an inductive DRE (red dot) which overlaps with a GR ChIP-seq peak. (C) Distribution of overlapping eSNPs and GR ChIP-seq peaks with vertical lines denoting the number of observed overlaps in the DREs (orange, P-value = 0.02), inductive DREs (red, P-value = 0.03), and repressive DREs (blue, P-value = 0.8). (D) Distribution of overlapping GR-eSNPs and chromatin anchor points with vertical lines denoting the number of observed overlaps in the DREs (orange, P-value = 0.02), inductive DREs (red, P-value = 0.02), and repressive DREs (blue, P-value = 0.2). (E) ChromHMM enrichment of DREs predicted to be located within enhancer regions across specific brain regions compared to randomly selected matched SNPs. *FDR < 0.05; permutation FDR.

Validation of STARR-seq-Identified DREs.

To validate the STARR-seq findings, we tested four DREs and three SNP-DREs using qPCR. Results were consistent with STARR-seq data and the four DREs showed significant (P < 0.05) dex responsiveness in the expected direction (SI Appendix, Fig. S3B). Two out of three SNP-DREs showed allele-dependent activity (SI Appendix, Fig. S3A). Log₂ FCs of all individual REs from the validation assays were highly correlated to the magnitudes observed in the STARR-seq data (Pearson correlation coefficient of 0.98, P-value < 0.0001) (Fig. 3A). To examine whether SNP-DREs in the episomal STARR-seq assay map to functional enhancers in the endogenous genomic context, two dex SNP-DREs sequences were selected for CRISPR/Cas9 editing. The first SNP, rs35288741 (Fig. 3 B and C), is within a repressive DRE regulating NUAK2 transcription. The second SNP, rs12206258 (Fig. 3 E and F), regulates FOXC1, a gene induced by dex. Both KOs showed altered dex-responsiveness of target genes (Fig. 3 D and G) and specificity over other transcripts (SI Appendix, Fig. S4 A–D). To confirm that rs12206258 alleles within the DRE-sequence were driving the differential expression of FOXC1, CRISPR/Cas9-SNP editing was performed. The C allele was used for homology-directed repair, resulting in homozygously edited (C/C) cells. Compared to homozygous wild-type (T/T) cells, the transcriptional response to dex was significantly reduced (Fig. 3H). A DRE (rs11773628) and a dex SNP-DRE (rs9525228), both GTEX eQTL SNPs, could also be validated with STARR-qPCR in cerebral organoids (SI Appendix, Fig. S5), supporting the relevance of the enhancers in the brain.

Fig. 3.

Exogenous and endogenous validation of the DREs. (A) Correlation between the magnitude of log₂ fold changes observed in the STARR-seq data with those observed in the qPCR validation of select candidate regions. (B–G) To validate the effect of two dex SNP-DREs on the regulation of their predicted transcripts, we removed each of the candidate enhancers in U2OS-GR cells and found that their deletion resulted in a blunted transcriptional response to dex. (B) Schematic depicting the genomic locus containing an eSNP bin (highlighted orange) in relation to its predicted target gene, NUAK2, (red box), and (C) a zoomed-in view of the region containing the repressive SNP-DRE rs35288741 with the CRISPR/Cas9-deleted region highlighted in blue. The black dot represents a flanking eSNP not targeted. (D) Deletion of the rs35288741 region resulted in reduced repression of NUAK2 (P-value = 0.038, one-sided unpaired Student t test). (E) Schematic depicting the genomic locus containing an eSNP bin (highlighted orange) and its predicted target gene, FOXC1, and (F) a zoomed-in view of the targeted region containing the inductive dex SNP-DRE rs12206258, with the CRISPR/Cas9-deleted region highlighted in red. (G) For FOXC1, activation was weaker when the rs35288741 region was deleted (P-value = 0.0024, one-sided unpaired Student t test). (H) Homozygously edited (C/C) cells showed significant transcriptional blunting in response to dex as determined by one-way ANOVA (F = 10.28, df = 2,22, P-value = 0.0007). During editing, we introduced an additional mutation to destroy the seed and prevent further cutting at the locus once the desired edit was made. To test whether this additional change affected FOXC1 regulation, we generated seed-mutated reference (T/T) cells. A Tukey post hoc test revealed that dex induction in the alternative (C/C) cells was significantly lower compared to the wildtype and seed mutated reference (T/T) cells (P-value = 0.001 and 0.002, respectively). *P-value < 0.05 and **P-value < 0.01.

Exploring the Association of DREs with Psychiatric Disorders and Brain Function.

We next explored the possible relationship between gene expression and function by conducting a GO term enrichment analysis on transcripts associated with the DREs (n = 122) (SI Appendix, Table S5). While the targeted genes/eQTLs were initially identified in blood (16), our analysis showed that terms related to neural development, such as regulation of slits and ROBOs, axon guidance, and signaling by robo receptors, were highly enriched (P-values: 1.66 × 10⁻⁴, 1.84 × 10⁻⁴, and 5.57 × 10⁻⁴, respectively). This suggests that these DREs may also play a role in brain function and is consistent with previous results that found the eSNPs transcripts expressed in mouse brain and regulated in brain by GR activation as well as social defeat stress (16). Further expanding the connection to expression in the brain, DRE and dex SNP-DRE associated transcripts were enriched in genes differentially expressed in postmortem cerebral cortex of subjects with neuropsychiatric disorders [SCZ, autism spectrum disorder (ASD), MDD, BPD, and AAD] (25) compared to controls (DREs: fold enrichment = 4.2, permutation P-value = 0.01; dex SNP-DREs: fold enrichment = 4.1, permutation P-value = 0.01). However, this enrichment was not observed with veh SNP-DRE regulated transcripts (FC = 1.05, permutation P-value = 0.44). These results suggest that genetic variants that selectively alter the transcriptional response to stress hormones (dex SNP-DREs) could play a role in psychiatric disorders by altering brain gene expression after stress.

Dex SNP-DREs Show an Increased Association with Psychiatric Disorders.

We next determined whether the functionally validated SNP-DREs showed a stronger association with psychiatric disorders compared to the previously tested eSNPs representing all SNPs within the LD block (16) by comparing the overlap with variants associated with psychiatric disorders (P-value ≤ 0.05) from a GWAS meta-analysis of eight psychiatric traits [BPD, MDD, attention deficit hyperactivity disorder, anorexia nervosa (AN), obsessive–compulsive disorder, SCZ, Tourette’s syndrome, and ASD] (26). Using 100 size-matched permuted eSNP sets as a comparison, we found that dex SNP-DREs (fold enrichment = 1.5, P-value = 0.03) but not dex/veh nor veh SNP-DREs were enriched for variants associated with psychiatric disorders (Fig. 4A) even over the eSNPs, highlighting the utility of additional functional fine-mapping of SNPs beyond QTL analyses in the context of disease association.

Fig. 4.

DREs are relevant to psychiatric disorders. (A) Enrichment analysis of dex (green), dex/veh (salmon), and veh (purple) SNP-DREs in GWAS variants (P-value < 0.05) for cross-disorder psychiatric risk compared to random eSNPs sets of the same size. eSNPs have previously been shown to be enriched in psychiatric GWAS (16). Vertical lines denote the number of observed overlaps between DREs/SNP-DREs and SNPs nominally significantly (P-value < 0.05) associated with psychiatric disorders. Estimated causal effect of (B) dex SNP-DREs, (C) veh SNP-DREs, and (D) dex/veh SNP-DREs on psychiatric disorders using Mendelian randomization. (E) Number of FDR corrected PheWAS associations per outcome trait category. Horizontal lines indicate the proportion of PheWAS association across all trait categories for dex (green), dex/veh (salmon), and veh (purple) SNP-DREs.

SNP-DREs Are Causally Associated with Psychiatric Disorders.

Given these enrichment results, we used Mendelian randomization (MR) to examine the putative causal role of the SNP-DREs in psychopathology. MR was performed on veh SNP-DREs (n = 61), dex SNP-DREs (n = 79), and dex/veh SNP-DREs (n = 93). The MR analysis was first conducted on eight psychiatric disorders using the psychiatric cross-disorder meta-GWAS (26). The results showed a higher percentage of dex SNP-DREs (20% of 55 available) and dex/veh SNP-DREs (16% of 62 available) to be causally associated with psychiatric disorders (FDR < 0.05), while this was only true for a small percentage of veh SNP-DREs (6.8% of 44 available) (Fig. 4 B–D and SI Appendix, Table S6). When the analysis was limited to MDD (27), seven dex SNP-DREs (10.3% of 68 available), but no vehicle or dex/veh SNP-DREs, were found to be causally associated. Interestingly, this dex SNP-DRE specificity was also true when examining SNPs causally associated with MDD in individuals with trauma exposure but not seen for MDD without trauma (6), although the limited number of implicated SNPs make interpretation difficult (5% of the dex SNP-DREs were associated with MDD with trauma and no veh or dex/veh SNP-DREs). There were no significant associations between the SNP-DREs and PTSD (28), likely due to the limited power of the currently available PTSD GWAS. Further MR analyses on nonpsychiatric traits like height (29), type 2 diabetes (30), BMI (31), fracture risk (32), heart failure (33), chronic kidney disease (34), and Crohn’s disease (35) found that no dex SNP-DREs significantly associated with these traits except BMI (27.2% of 33 available) and type 2 diabetes (5.8% of 53 available).

We next examined whether the causal relationship to the dex SNP-DREs was overrepresented in brain and behavior-related traits by scaling up the MR analyses in a phenome-wide association study (PheWAS) approach. Specifically, we curated a list of 4,439 phenotypes from the MRC IEU OpenGWAS platform and applied MR separately on each outcome using a reduced list of 113 statistically independent (clumping threshold of R² = 0.001 using 10,000 kb window) SNPs representing our three functional SNP groups (dex, dex/veh, and veh SNP-DREs). After correcting for the large number (313,308) of statistical tests, we observed 2,197 (0.7%) associations that passed our significance threshold (FDR < 0.05). We categorized GWAS outcomes with significant associations into one of seven phenotype groups (biomarker, body morphology, disease, immune system, neurobehavioral, nutrition, and other; SI Appendix, Table S7). This revealed differing proportions of causally associated dex, dex/veh, and veh SNP-DREs across outcome trait categories (χ²(12) = 61.24, P-value < 0.001). Dex SNP-DREs had the highest proportion of significant MR associations in the neurobehavioral trait category as compared to the seven other categories (Fig. 4E and SI Appendix, Table S8 and Fig. S6 for illustration of top associations). Dex SNP-DREs accounted for 50% of all significant associations in this category, supporting a relative selectivity of these functional variants to neurobehavioral traits, the majority of which have also been reported to be affected by stress exposure. To allow detailed exploration of our results, we created a dedicated Shiny app to assess PheWAS associations of individual SNP-DREs (https://nprct.shinyapps.io/GRE_PheWAS/).

SNP-DREs Modulate Physiological Stress Responses.

Given the enrichment of dex SNP-DREs in causal associations with neurobehavioral traits, we constructed a functional gene score (FGS) for these variants to test their association with variability in the physiological stress response in a transdiagnostic, psychiatric, deep phenotyping study (36). The FGS was generated from 78 (of 79) dex SNP-DREs and weighted by the differential RE activity quantified by STARR-seq. We focused on two stress phenotypes: changes in salivary cortisol levels in response to a psychological stress task (Fig. 5A, available in n = 183) (37) and the eyeblink startle reflex elicited by a fear-conditioning paradigm (Fig. 5C, available in n = 171) (38). Both phenotypes are affected in stress-related psychiatric disorders including MDD (39–41). We found that the FGS was not associated with cortisol levels before the stress task (P-value = 0.99) but significantly positively correlated to changes in cortisol immediately after completing the task ( $\beta$   = 0.34, P-value = 0.02), and that this association persisted 30 min posttask ( $\beta$   = 0.26, P-value = 0.03, Fig. 5B). In the eyeblink startle reflex, a higher FGS was associated with an increased startle magnitude (Fig. 5D, $\beta$ = 0.20, P = 0.01), as well as a dampened startle habituation ( $\beta$ = 0.18, P = 0.03). This means that individuals with higher FGS scores, harboring more alleles with increased GC-regulatory strength, showed a higher and more prolonged cortisol response as well as a magnified startle reflex and reduced habituation. These data provide preliminary evidence that the functional SNPs that modulate the transcriptomic response to stress also associate with differences in the stress response at the system level.

Fig. 5.

Dex SNP-DREs modulate the physiological stress response. (A) Experimental setup of a psychosocial stress task. Salivary cortisol was measured at baseline, prior to the stress task, after completion of the task, and following a 30-min rest period. (B) Association of salivary cortisol changes compared to baseline with the FGS for the pretest, posttest, and postbreak measurements. (C) Experimental setup for a fear acquisition task eliciting an eyeblink startle response. (D) Association between startle amplitude and habituation and the FGS. Note that an increased habituation z-score indicates a persisting startle response with multiple startles.

Discussion

Genetic variants relevant to common psychiatric disorders have previously been shown to be enriched in gene REs and likely act in a cell-type and context-dependent manner. MPRAs now allow us to map the cell-type and context-specific effects of such regulatory variants (42). Using this method, we identified genetic variants that selectively modulate the activity of gene REs following stimulation by GCs to mimic a stress context. MR analyses uncovered a causal relation between SNPs moderating GC responsiveness (dex SNP-DREs) and risk for psychiatric disorders and neurobehavioral traits. While these variants were also causally associated with other traits - including traits with well-documented links to GCs activity such as asthma (43), BMI (44), and immune function (45)—neurobehavioral traits showed the most pronounced overrepresentation for dex SNP-DREs. This suggests a relevant and specific contribution of genetic variants moderating the molecular response to stress hormones to risk for psychiatric disorders and traits. These SNP-DREs may thus provide a point of molecular convergence between environmental risk in the form of stress-induced increased GCs and genetic risk variants for psychiatric disorders.

While previous studies have identified GC-responsive REs using genome-wide STARR-seq (46), ours specifically explores such REs in the context of genetic variation. We focused on over 3,600 SNPs located within 320 individual LD bins that were identified as GR-eSNPs in a genome-wide analysis in peripheral blood (16). In the two GC-responsive cell lines we studied, we find 40% of these preselected bins to contain such DREs. Notably, the activity of DREs in the endogenous genomic context is often cell-type specific. This is likely partially due to differences in chromatin accessibility which plays a role in dictating which genomic regions can be bound by GR (20). In addition, cell-type specific expression of transcription factors that act cooperatively with GR can influence the activity of DREs (20, 47) and therefore we expect that we would find more DREs by expanding our analysis to include additional cell types. Unfortunately, however, our efforts to establish STARR-seq in neural precursors or neurons in the 2D context failed due to a lack of robust GC-responsiveness (SI Appendix, Fig. S1 C–E). Instead, we validated a DRE and SNP-DRE (SI Appendix, Fig. S5) using STARR-qPCR in human cerebral organoids, which have been used as models of the developing brain (48). Future investigations to validate more SNP-DREs in cerebral organoids and to characterize their functional role in a human brain model could provide valuable insights into the mechanisms by which these variants may alter the risk of neuropsychiatric disorders. Overall, balancing feasibility and using the right cell type is a major challenge in large-scale enhancer screening studies. Despite these limitations, our study indicates that testing MPRA in an episomal setting in cell lines selected for robust GC-responsiveness yields results of relevance for psychiatry.

While 43% of the LD eSNP bins with a signal contain a single DRE, 57% have two or more, with up to 20 or more observed for specific bins (Fig. 1E). In addition, we identify both inductive as well as repressive DREs, with a relatively equal distribution. Interestingly, only inductive DREs are enriched for GR-ChIP peaks. This is in line with previous studies showing that over 90% of GR binding sites with GC regulated activity had increased reporter gene expression in response to dex (49). Consistent with the traditional model of GR repression that is independent of direct DNA binding (50), we confirm that repressive DREs lack consensus steroid receptor-binding elements whereas they are enriched for the sequence motifs such as AP-1 (SI Appendix, Table S4) which is implicated in GR-dependent transcriptional repression (22). Of note, in close to 40%, we observe both repressive and inductive DREs in the same eQTL bin, supporting the notion of extensive regulatory fine-tuning of the same target transcript (17). This fine-tuned genetic moderation of GC-induced transcriptional changes in brain tissue could be associated with disease risk, as transcripts containing dex SNP-DREs in their eSNP bins were enriched among transcripts differentially expressed in the cortex of individuals with psychiatric disorders and controls (25).

The mineralocorticoid receptor is a nuclear receptor that binds to GCs with higher affinity than GR. While this study did not assess the mineralocorticoid receptor, it is believed to be occupied under basal conditions, while GR is occupied in response to higher GC levels, such as during stress. Recent evidence suggests that the mineralocorticoid receptor and GR can interact and bind as homo- or heterodimers at glucocorticoid-responsive elements in response to stress, with GR facilitating mineralocorticoid receptor binding (51). Thus, it is expected that the SNP-DREs are enriched in both GR and mineralocorticoid receptor binding sites. Although the focus of this study has been on understanding the mechanisms of SNP-DREs through GR, it is likely that the mineralocorticoid receptor also plays a role. Future studies using a mineralocorticoid receptor–specific agonist, like aldosterone, would help determine the relative contributions of the mineralocorticoid receptor and GR in understanding how SNP-DREs modulate the transcriptomic response to stress.

Although MR has been more conventionally employed in the field of epidemiology to estimate the effect of a risk factor on an outcome variable using genetic variants that associate with the risk factor (52), in recent years, it has also been applied to estimate causal effects of molecular QTLs on complex traits (53). In psychiatric studies, MR analyses revealed a causal association between increased signaling of an inflammatory marker, IL-6, and suicidality (54), and glycemic traits with anorexia nervosa (55). Here, we used SNPs with regulatory effects in DREs as instruments and quantitative molecular read-outs from the STARR-seq analysis as exposure variables to test for causal relationships with different traits using MR. Initial MR analyses revealed a preponderance of causal associations of dex SNP-DREs with cross-disorder psychiatric risk as well as metabolic traits, but not others, such as height. The link to metabolic traits is not surprising, given the role of GR in glucose metabolism (56) and evidence for genetic coheritability between psychiatric traits (MDD, SCZ, AN) and BMI (57, 58). PheWAS MR in over 4,000 phenotypes confirmed that dex SNP-DREs were causally related to a large number of traits, but with a significantly higher proportion of associations in neurobehavioral traits, including psychiatric disorders, than other trait categories. In addition to their association with disease risk, these variants also associate with features of the physiological stress response. Individuals with a higher FGS, i.e., more alleles with an enhanced response to GCs, showed an increased startle amplitude as well as decreased habituation to the startle reflex and also a heightened and prolonged cortisol response following a psychological stressor. These genetic variants may thus influence the risk for psychiatric disorders by altering an individual’s physiological stress response set-point.

This study highlights that genetic variants modulating the transcriptomic response to stress may be causally involved in psychopathology, including MDD, possibly by influencing the physiological stress response and stress-responsive brain transcription. This gives further mechanistic support to the well-established link between exposure to stress and psychiatric disorders. Future studies need to expand the number of tested loci and diversify and refine the cell systems used for functional mapping.

Materials and Methods

Please also refer to SI Appendix, Materials and Methods for more details.

Cell Culture.

Cell culture was performed using U138MG cells obtained from the German Collection of Microorganisms and Cell Cultures GmbH and U2OS-GR cells (59). U138MG cells were cultured in Minimum Essential Medium Eagle with 10% fetal bovine serum, 1% sodium pyruvate, and 1% antibiotic–antimycotic. U2OS-GR cells were cultured in Dulbecco’s Modified Eagle Medium with high glucose and 10% fetal bovine serum and 1% antibiotic–antimycotic. To activate GR, cells were treated with 100 nM dexamethasone (dex) or 0.001% ethanol (veh) for 4 h.

STARR-seq.

The DNA fragments for integration into the human STARR-seq plasmid were generated from the hg19 genomic coordinates of the eSNPs and synthesized by Twist Bioscience. The fragments were 201bp long and included positive dex-responsive REs, randomly selected genomic regions, and random sequences generated by a Markov model. A pool of oligonucleotides containing all fragments and controls was amplified, cloned into the STARR-seq vector using In-Fusion HD Cloning, and transformed into MegaX DH10B cells. The STARR libraries were generated by PCR and sequenced on the Illumina MiSeq, and the quality of the reads was assessed using FastQC. For the STARR-seq experiment, 5 million cells of U2OS-GR and U138MG were transfected with the input STARR-seq library and treated with dex or veh. Total RNA was collected and mRNA was isolated, followed by cDNA generation. The STARR libraries were sequenced, and data analysis was preprocessed according to previously published methods and analyzed using MPRAanalyze. We used publicly available RNA-seq data (18) in the U2OS-GR cell line treated with 1 μM dex for 90 min to determine whether there was an association between dex-induced differential expression and number of DREs/bin.

Functional Annotation.

We used publicly available GR ChIP-seq data (19) in the U2OS-GR cell line treated with 1 μM dex for 90 min to determine whether GR binding sites were enriched within the DREs. To this end, we generated a null distribution by permuting eSNPs into 100 sets of the same size and counting the overlap with GR ChIP peaks. Motif enrichment analysis was performed on the DREs using the online tool “TRAP” (60), and sequence motif analysis was performed using the Jaspar vertebrate database and human promoters as the background model. Publicly available processed Hi-C data were acquired (24) to analyze the 3D structure of chromatin in the U2OS parental cell line at baseline. A consensus interaction set was generated and a null distribution was generated by permuting eSNPs into 100 sets of the same size to determine whether the DREs were enriched within or near chromatin loops. The STARR-seq-identified DREs were annotated using the HaploReg R package and the core 18-state model to functionally annotate the variants in relevant brain regions and assess cross-tissue functionality (61). A null distribution was generated by permuting randomly selected SNPs and quantifying the number annotated as located within enhancer regions to determine whether the DREs were enriched within enhancers.

Validation of STARR-seq.

The exogenous validation of the regulatory elements involved cloning the fragments into a linearized STARR-seq plasmid using In-Fusion HD Cloning. The transformed plasmids were then sequenced using Sanger sequencing, and the activity of the cloned regions was tested by transfecting 2 million U2OS-GR cells with the STARR-seq vector and stimulating with dex or the veh. cDNA was generated using the QuantiTect Reverse Transcriptase kit, and RE activity was assessed by qPCR using primers for RPL19 and GFP. For the validation in the human cerebral organoids, the organoids were generated as previously described (48). Organoids were electroporated at day 40 as previously described (62) and treated with dex or veh for 16 h (100 nM). Transcriptional activity was measured as described for the validation in the U2OS-GR cells, with an additional qPCR to quantify plasmid DNA for transfection efficiency normalization. For each plasmid and treatment condition, five replicates were used containing one organoid each.

For the endogenous validation, CRISPR-Cas9 genome editing was used to create genomic deletions of the REs containing the variants rs12206258 and rs35288741. gRNAs and Cas9 were used to cut the DNA by transfecting one million U2OS-GR cells. Two days post-electroporation, cells were plated for analysis and DNA isolation. PCR was performed to confirm the presence of CRISPR/Cas9-mediated knockout, and RT-qPCR was performed for gene expression analysis. SNP-specific genome editing was performed for rs12206258, predicted to control FOXC1 expression. A 200-bp HDR template was synthesized as a duplex oligonucleotide, and a gRNA was designed to introduce a point mutation in the GR binding site. The gRNA was then used for genome editing, and the cells were analyzed for expression using RT-qPCR. We assessed gene expression in the KO cell lines using RT-qPCR.

Enrichment of Transcripts Differentially Expressed in Psychiatric Disorders.

An enrichment analysis was performed to determine whether the DRE-associated transcripts were overrepresented among transcripts shown to be differentially regulated in the prefrontal cortex of subjects with psychiatric disorders (25). The analysis compared a set of transcripts differentially regulated in psychiatric disorders to a randomly selected background set of transcripts of the same size.

Mendelian Randomization on Single Phenotypes.

To test for a causal effect of the differential DRE activity on psychiatric traits, a two-sample MR approach using the R package “TwoSampleMR” was employed. The instrumental variable was set as SNP-DREs, and the log-fold change between the two alleles was used as the beta estimate. The exposure variables were the corresponding DREs showing allele-dependent activity, and the outcome variables were defined using GWAS summary statistics for the PGC cross-disorder (26) and MDD variants (6, 27).

PheWAS Analyses.

A PheWAS was conducted to identify potential associations between SNP-DREs and a variety of phenotypes. The SNP-DREs were grouped by associated genes and clumped to retain statistically independent SNPs as genetic instruments. The outcome phenotypes were selected from a GWAS of psychiatric disorders and additional phenotypes from the MRC IEU OpenGWAS platform (63). The phenotype list was filtered to remove non-European GWAS and duplicate phenotypes, resulting in a final list of 4,439 phenotypes. MR analyses were conducted using the TwoSampleMR package and corrected for multiple testing using the Benjamini–Hochberg method. The results of these analyses were grouped into seven categories, and deviations from expected proportions of dex, dex/veh, and veh SNP-DREs per phenotype category were assessed using a chi-squared test.

Fear Acquisition Task and Startle Response Quantification.

The study sample consisted of participants who underwent in-depth genotyping and phenotyping and completed 2 d of in-house measurements, including a psychosocial stress task on the second day and a classical fear-conditioning task on the first day. The sample was recruited as part of the Biological Classification of Mental Disorders study at the Max Planck Institute of Psychiatry (registered on ClinicalTrials.gov: NCT03984084). Approval for the study was obtained by the local Ethics Committee of the Ludwig Maximilians University, Munich, Germany, and written informed consent was obtained from all participants (36). Genotyping was performed using the Illumina Global Screening Array. A final dataset of 9,651,000 SNPs was used for the calculation of functional gene risk scores based on the log2 fold change of 78 dex SNP-DREs (out of 79). The cortisol response to stress and the startle response during fear acquisition were measured as previously described (37, 38). Linear models were used to predict the endophenotypes based on the functional gene risk scores, taking into account the participants’ age, sex, and diagnosis of a mood or anxiety disorder.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Data, Materials, and Software Availability

STARR-sequencing data have been deposited in Zenodo (10.5281/zenodo.8273084 (64) and 10.5281/zenodo.8349334 (65)).