Association of a Reproducible Epigenetic Risk Profile for Schizophrenia With Brain Methylation and Function

Key Points

Question

Can a blood marker of epigenetic risk for schizophrenia be derived that is specific for the disease and predicts epigenetic changes in brain and disease-associated brain function?

Findings

In this case-control study, machine learning was used to identify a reproducible schizophrenia blood DNA-methylation signature that was associated with functional dorsolateral prefrontal cortex hippocampal connectivity, mapped to methylation differences found in dorsolateral prefrontal cortex hippocampal connectivity postmortem samples, and indexed biological pathways associated with synaptic function. No interactions with polygenic schizophrenia risk were found.

Meaning

These findings support the presence of a systemic methylation profile in schizophrenia that is associated with established intermediate functional phenotypes as well as with epigenetic signatures in brain and should thus be useful to capture the biological effects of gene-environment interactions.

Abstract

Importance

Schizophrenia is a severe mental disorder in which epigenetic mechanisms may contribute to illness risk. Epigenetic profiles can be derived from blood cells, but to our knowledge, it is unknown whether these predict established brain alterations associated with schizophrenia.

Objective

To identify an epigenetic signature (quantified as polymethylation score [PMS]) of schizophrenia using machine learning applied to genome-wide blood DNA-methylation data; evaluate whether differences in blood-derived PMS are mirrored in data from postmortem brain samples; test whether the PMS is associated with alterations of dorsolateral prefrontal cortex hippocampal (DLPFC-HC) connectivity during working memory in healthy controls (HC); explore the association between interactions between polygenic and epigenetic risk with DLPFC-HC connectivity; and test the specificity of the signature compared with other serious psychiatric disorders.

Design, Setting, and Participants

In this case-control study conducted from 2008 to 2018 in sites in Germany, the United Kingdom, the United States, and Australia, blood DNA-methylation data from 2230 whole-blood samples from 6 independent cohorts comprising HC (1238 [55.5%]) and participants with schizophrenia (803 [36.0%]), bipolar disorder (39 [1.7%]), major depressive disorder 35 [1.6%]), and autism (27 [1.2%]), and first-degree relatives of all patient groups (88 [3.9%]) were analyzed. DNA-methylation data were further explored from 244 postmortem DLPFC samples from 136 HC and 108 patients with schizophrenia. Neuroimaging and genome-wide association data were available for 393 HC. The latter data was used to calculate a polygenic risk score (PRS) for schizophrenia. The data were analyzed in 2019.

Main Outcomes and Measures

The accuracy of schizophrenia control classification based on machine learning using epigenetic data; association of schizophrenia PMS scores with DLPFC-HC connectivity; and association of the interaction between PRS and PMS with DLPFC-HC connectivity.

Results

This study included 7488 participants (4395 men [58.7%]), of whom 3158 (2230 men [70.6%]) received a diagnosis of schizophrenia. The PMS signature was associated with schizophrenia across 3 independent data sets (area under the curve [AUC] from 0.69 to 0.78; P value from 0.049 to 1.24 × 10⁻⁷) and data from postmortem DLPFC samples (AUC = 0.63; P = 1.42 × 10⁻⁴), but not with major depressive disorder (AUC = 0.51; P = .16), autism (AUC = 0.53; P = .66), or bipolar disorder (AUC = 0.58; P = .21). Pathways contributing most to the classification included synaptic processes. Healthy controls with schizophrenia-like PMS showed significantly altered DLPFC-HC connectivity (validation methylation/magnetic resonance imaging, t < −3.81; P for familywise error, <.04; validation magnetic resonance imaging, t < −3.54; P for familywise error, <.02), mirroring the lack of functional decoupling in schizophrenia. There was no significant association of the interaction between PMS and PRS with DLPFC-HC connectivity (P > .19).

Conclusions and Relevance

We identified a reproducible blood DNA-methylation signature specific for schizophrenia that was correlated with altered functional DLPFC-HC coupling during working memory and mapped to methylation differences found in DLPFC postmortem samples. This indicates a possible epigenetic contribution to a schizophrenia intermediate phenotype and suggests that PMS could be of interest to be studied in the context of multimodal biomarkers for disease stratification and treatment personalization.

This case-control study assesses the association between a blood DNA methylation signature with a reproducible differentiation of patients with schizophrenia from healthy controls and several other major neuropsychiatric disorders.

Introduction

Schizophrenia is a severe brain disorder thought to be caused by a complex interplay of genetic predisposition and environmental exposure.^1,2,3,4,5 In the context of gene-environment interplay and developmental programming,^6,7 epigenetic mechanisms, such as DNA methylation, have received substantial attention in schizophrenia^8,9 and other neuropsychiatric disorders.^10,11,12,13 As epigenetic mechanisms may account for heritability not captured by other current methods, such as polygenic risk scores (PRSs), potential interactions between genetic susceptibility and epigenetic changes are of particular interest. To our knowledge, to date, much of this work has studied single genes. For example, methylation at the single-nucleotide polymorphism rs6265 within the neurodevelopmentally important brain-derived neurotrophic factor gene shows a genotype-dependent association with working memory (WM), hypoxia-associated early life events, and a WM-associated schizophrenia intermediate phenotype.¹⁴ Similarly, methylation differences in dopamine receptor D4 predict WM, suggesting that the dopaminergic methylation status could affect cognitive functions in a dissociable manner.¹⁵ Likewise, membrane-bound catechol-O-methyltransferase promoter methylation is associated with dorsolateral prefrontal cortex (DLPFC) activity during WM¹⁶ and methylation of the COMT Val(158) allele with lifetime stress, WM performance, and prefrontal activity during WM.¹⁷

While these studies have explored targeted hypotheses, methylation differences in single genes^14,17 are only weakly associated with schizophrenia. To capture systems-level effects, methylation differences have been explored across the genome in whole-blood^8,18 and postmortem brain tissue samples.^19,20 Joint analyses of genome-wide methylation and genotyping data provided evidence that the methylation changes found in schizophrenia differ from those associated with polygenic risk but overlap with previously identified genetic susceptibility loci.^18,20,21 This supports the notion that the methylation differences are associated with schizophrenia and not merely a result of disease-unrelated environmental factors.^22,23,24 An important question is whether, similar to polygenic scores, there is a combined contribution of these methylation differences associated with illness risk. A machine-learning study provided evidence for a schizophrenia polymethylation score (PMS) that could be validated in independent data.²⁵ However, we do not yet understand whether this PMS is relevant in the brain and how it is associated with genomic risk with regard to neural effects.

Components of the risk architecture of schizophrenia have been successfully interrogated using a strategy termed imaging genetics,²⁶ an approach that has facilitated the identification of so-called neural intermediate phenotypes, illness-associated, heritable traits that reflect a manifestation of illness liability.²⁷ One of the best established intermediate phenotypes of schizophrenia is dorsolateral prefrontal cortex hippocampus (DLPFC-HC) connectivity during WM performance.^28,29 The DLPFC-HC connectivity is altered in healthy first-degree relatives of patients with schizophrenia and is associated with risk alleles of established genome-wide significant schizophrenia gene variants, such as ZNF804A.³⁰ Working memory is impaired in schizophrenia, associated with genetic risk,³¹ and affected by environmental factors that contribute to illness risk, including childhood trauma and socioeconomic status,^32,33 strengthening the argument that this intermediate phenotype reflects risk-associated pathophysiological processes.

Using this approach, we investigated genome-wide DNA methylation data from 2041 whole-blood samples from 4 independent cohorts comprising 1238 healthy controls (HCs) and 803 patients with schizophrenia. We aimed to identify and validate a PMS differentiating schizophrenia from HC. Subsequently, we assessed whether the PMS was reproducibly associated with DLPFC-HC connectivity in HC and explored potential interactions with a schizophrenia PRS. Finally, we used genome-wide DNA methylation data from postmortem DLPFC samples from 136 healthy donors and 108 donors with schizophrenia to assess whether the peripheral PMS was mirrored by analogous changes in the brain. This investigation aimed at characterizing the systems-level association between genetic and epigenetic risk for schizophrenia and to test the association of these parameters with schizophrenia-relevant neural functioning.

Methods

Cohorts

Genome-wide DNA methylation data from 7 cohorts, denoted as discovery methylation, validation methylation, validation methylation/magnetic resonance imaging (MRI), validation MRI, validation post mortem, specificity methylation, and relatives methylation, were analyzed in this work (demographic characteristics are summarized in Table 1 and eTable 1 in the Supplement). All participants (or their legal next-of-kin in case of brain tissue donors) gave written or audiotaped informed consent and all studies were approved by the local ethics committees (eMethods in the Supplement). Discovery methylation and validation methylation consisted of 2 independent cohorts 767 patients with schizophrenia [95.5%] and 755 HC [61.0%]). These data sets were used to identify and validate a PMS using machine learning. Validation methylation/MRI (36 patients with schizophrenia [4.5%] and 331 HC [26.7%]) was used for additional validation of the PMS, while MRI data from a subset of the HC (241 [19.5%]; n-back WM functional MRI paradigm^29,30) were used to test associations with DLPFC-HC connectivity. The MRI data from validation MRI (n = 152 HC) acquired during the Sternberg WM task³⁴ were used to validate the identified DLPFC-HC connectivity associations. The functional MRI face-matching task³⁵ was used in validation MRI to explore the specificity of findings for WM. Validation post mortem comprised genome-wide methylation data from postmortem DLPFC samples 108 schizophrenia [44.3%]; 136 HC [55.7%]) and was used to assess the overlap of the PMS with brain methylation changes. Genome-wide association study (GWAS) data in validation methylation/MRI and validation MRI were used to compute a schizophrenia PRS to test associations with the PMS and interactions with PMS on brain functional connectivity. For specificity testing, we explored DNA methylation data from a cohort (specificity methylation) of patients with bipolar disorder (BD; 39 [1.7%]), major depressive disorder (MDD; 35 [1.6%]) and autism (27 [1.2%]). To further characterize the effect of genetic schizophrenia risk, a cohort (relatives methylation) of unaffected first-degree relatives of participants with schizophrenia (27 [1.2%]), BD (15 [0.7%]), MDD (29 [1.3%]), and autism (17 [0.8%]) were analyzed. The machine-learning procedure used here was further applied to GWAS data from GWAS molecular genetics of schizophrenia (GWAS MGS) (n = 2718 HC and n = 2296 schizophrenia) to identify risk components not captured by the PRS that could potentially explain PMS-associated findings.

Table 1. Overview of Sample Characteristics.

Cohort	Tissue	Status	Sex, No. (%)		Age, Mean (SD), y	Methylationa	Genotypea	Imaginga
			Male	Female
Discovery methylation	Whole-blood	Control	142 (6.4)	180 (8.1)	37.7 (15.2)	322
		Case	254 (11.4)	99 (4.4)	43.7 (14.7)	353
Validation methylation		Control	319 (14.3)	114 (5.1)	45.0 (12.1)	433
		Case	283 (12.7)	131 (5.9)	46.6 (13.6)	414
Validation methylation/MRI		Control	168 (7.5)	163 (7.3)	27.8 (10.2)	331
		SubCtl	126 (5.7)	115 (5.2)	28 (10.7)	241	241	241
		Case	28 (1.3)	8 (0.4)	33.8 (10.4)	36
Specificity methylation		AUT	18 (0.8)	9 (0.4)	33.8 (9.6)	27
		BP	16 (0.7)	23 (1.0)	36.4 (10.5)	39
		MDD	11 (0.5)	24 (1.1)	37.2 (12.2)	35
Relatives methylation		relSCZ	8 (0.4)	19 (0.9)	37.2 (14.4)	27
		relAUT	8 (0.4)	9 (10.4)	44.9 (9.1)	17
		relBP	8 (0.4)	7 (0.3)	36.9 (12.3)	15
		relMDD	11 (0.5)	18 (0.8)	30.3 (10.4)	29
Validation MRI		Control	66 3.0%)	86 (3.9)	26.9 (8.8)	152	152	152
Validation post mortem	Brain	Control	90 (36.9)	46 (18.9)	46.5 (16.1)	136
		Case	59 (24.2)	49 (20.1)	52.7 (14.5)	108
GWAS MGS	Whole-blood	Control	1301 (25.9)	1417 (28.3)	50.6 (16.4)		2718
		Case	1606 (32.0)	690 (13.8)	43.4 (11.7)b		2296

Abbreviations: AUT, autism; BP, bipolar disorder; GWAS, genome-wide association study; MDD, major depressive disorder; MGS, molecular genetics of schizophrenia; MRI, magnetic resonance imaging; relAUT, first-degree relatives of patients with autism; relBP, first-degree relatives of patients with bipolar disorder; relMDD, first-degree relatives of patients with major depressive disorder; relSCZ, first-degree relatives of patients with schizophrenia; SubCtl, subset of healthy controls.

^aThe rightmost 3 columns indicate the sample numbers for which the respective data were available (methylation: DNA methylation; genotypes: whole-genome genetic association data; imaging: functional MRI data [validation methylation/MRI: n-back task; validation MRI: Sternberg task]; case: patients with schizophrenia; subset of healthy controls; autism; bipolar disorder; major depressive disorder; first-degree relatives of patients with schizophrenia; first-degree relatives of patients with autism; first-degree relatives of patients with bipolar disorder; and first-degree relatives of patients with major depressive disorder).

^bForty participants with missing age information.

The data, methods, and confounding correction are detailed in the eMethods in the Supplement. Most patients were taking medication. Neuroimaging analyses focused on HCs not taking medication to demonstrate that PMS associations were not associated with medication. Additionally, we explored the association of the PMS and DLPFC-HC connectivity with chlorpromazine equivalents.

Gene and Pathway Assignment of Genome-Wide DNA Methylation Data

For each gene, CpGs harbored by the gene with an extended window size of 20 kb downstream and upstream were used for analysis. CpG locations and gene annotations used for mapping were obtained from the R library IlluminaHumanMethylationEPICanno.ilm10b2.hg19 (R Foundation). Genes were binned into biological process categories (denoted as pathways) using the Gene Ontology database (data freeze in December 2018).³⁶ A total of 2846 pathways comprising between 10 and 200 genes with at least 10 CpGs per pathway were used for the analysis.

Machine Learning, Replication, and Validation

An updated biologically informed machine learning (BioMM) approach was used (eMethods in the Supplement).²⁵ BioMM is a 2-stage machine-learning approach that first builds separate machine learning models for methylation sites mapping to each of the 2846 pathways, yielding 1 machine-learning model per pathway (first-stage). This procedure compresses data from individual methylation sites into a pathway-level feature. Then, a second-stage algorithm integrates these pathway-level features into a systems-level classifier. BioMM was trained on discovery methylation and the algorithm then applied to all other data sets. In each data set, the output of BioMM was a score (PMS) that quantified the likelihood of a given participant being in the schizophrenia group. To assess predictive accuracy, we determined the area under the receiver operating characteristic curve (AUC) as well as Nagelkerke R².

Statistical Analysis

Associations between PMS and DLPFC-HC functional connectivity were assessed using a linear regression in SPM, version 12, using PMS as covariate of interest and age and sex as covariates of noninterest. Statistical significance was set at P < .05. The PMS and DLPFC-HC connectivity associations in the imaging space were corrected using a familywise small-volume correction in the hippocampus (eFigures 1 and 2 in the Supplement) based on the automated anatomical labeling template. Associations between PMS and the schizophrenia PRS, as well as the association of the PMS by PRS interaction with DLPFC-HC connectivity, were tested using a multiple linear regression that accounted for the effects of sex, age, and (for PRS-associated analyses) 10 genetic principal components. For details and analysis of potential confounding, see the eMethods in the Supplement.

Results

Determination and Validation of a PMS in Blood Samples

Genome-wide DNA methylation data from discovery methylation (675 [30.3%]) were used for model training. In this data set the model showed a cross-validation accuracy of an AUC of 0.78 (P = 2.95 × 10⁻⁶, corrected for 20 potential confounders; R² = 21%). The model was then predicted into validation methylation (847 [38.0%]), yielding an AUC of 0.69 (P = 1.24 × 10⁻⁷; R² = 10.5%). For additional validation, the model was predicted into validation methylation/MRI (367 [16.5%]), yielding an AUC of 0.74 (P = .049; R² = 21.8%). This demonstrates the biological reproducibility of the PMS (Table 2 and Figure 1).

Table 2. Prediction Performance of BioMM on Different Cohorts Using Discovery Methylation as the Discovery Set.

Data Sets	AUC	R2
Discovery methylation	0.78	0.21
Validation methylation	0.69	0.105
Validation methylation/MRI	0.74	0.218
Validation post mortem	0.63	0.083

Abbreviations: AUC, area under the receiver operating characteristic curve; BioMM, biologically informed machine learning; MRI, magnetic resonance imaging.

Figure 1. Prediction Performance of Biologically Informed Machine Learning (BioMM) on Different Cohorts Using Discovery Methylation as the Training Data.

AUC indicates area under the curve; meth, methylation; MRI, magnetic resonance imaging.

Identification of Implicated Biological Pathways

To identify pathways with methylation changes contributing strongly to the epigenetic signature, pathway-level (second-stage) data from BioMM in discovery methylation were used. The 10 pathways most associated with schizophrenia are shown in Figure 2 (eTable 2 in the Supplement). Of the 2846 pathways, 917 (32.2%) were positively associated with diagnosis and 57 (6.2%) of these showed significance at P for familywise error <.05. They did not differ significantly from the remaining pathways regarding size (determined as the number of CpGs per pathway; β = 2.46 × 10⁻⁵; P = .61). Individual genes within the 10 pathways harboring methylation differences have been previously implicated in schizophrenia (eTables 3 and 4 in the Supplement).

Figure 2. Biological Pathways Contributing Most to the Polymethylation Score and the Significance of Individual Methylation Sites.

GO:0032088, negative regulation of NF-κ B transcription factor activity; GO:0010977, negative regulation of neuron projection development; GO:0030097, hemopoiesis; GO:0000083, regulation of transcription involved in the G1/S transition of the mitotic cell cycle; GO:2000311, regulation of α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptor activity; GO:0098962, regulation of postsynaptic neurotransmitter receptor activity; GO:0043154, negative regulation of cysteine-type endopeptidase activity involved in the apoptotic process; GO:1903427, negative regulation of the reactive oxygen species biosynthetic process; GO:0031648, protein destabilization; GO:1900273, positive regulation of long-term synaptic potentiation; and GO:0019395, fatty acid oxidation. The dots indicate the significance of individual CpGs within these pathways. The blue dotted line marks the uncorrected significance level of P < .05.

Identifying and Validating the Association Between PMS and Functional DLPFC-HC Coupling

The schizophrenia PMS predicted in data from HCs not taking medication from validation methylation/MRI (241 [10.8%]) was negatively associated with DLPFC-HC functional connectivity during WM (P for familywise error, <.04; F_1,237 = 14.55; t₂₃₇ = −3.81; bilateral hippocampus-corrected, peak voxel at 33, −22, −13) in the right posterior hippocampus (Figure 3; eFigure 3 in the Supplement). The negative association between PMS and DLPFC-HC connectivity was replicated in validation MRI (n = 152), with the Sternberg WM paradigm within right posterior hippocampus (P for familywise error = .02; t₁₄₈ = −3.54; peak voxel at 33, −37, −7) corrected for right posterior hippocampus (Figure 3; eFigure 3 in the Supplement). Post hoc 1-sample t tests revealed negative connectivity between DLPFC and HC in the validation methylation/MRI (t₂₄₀ = −14.09; P < .001) and the validation_MRI (t₁₅₁ = −6.94; P < .001) sample. Associations between PMS and DLPFC-HC connectivity were specific for the WM tasks compared with the faces tasks and were not confounded by brain-structural effects (eResults in the Supplement).

Figure 3. Association Between the Predicted Polymethylation Score and Dorsolateral Prefrontal Cortex Hippocampal Connectivity in Healthy Controls Not Taking Medication.

A, In the validation methylation/magnetic resonance imaging (validation methylation/MRI) group (n-back task), uncorrected results (P < .05, shown in hippocampus) with peak voxel (F_1,237 = 14.55; t₂₃₇ = −3.81; Montreal Neurological Institute [MNI] coordinates: 33, −22, −13) in the right hippocampus that were significant after bilateral hippocampus correction in healthy participants (241 [10.8%]) (P for familywise error = .04). B, In the validation MRI group (Sternberg task): uncorrected results (P < .05, shown in hippocampus) with peak voxel (t₁₄₈ = −3.54; peak voxel at 33, −37, −7 in the right hippocampus showing a significantly negative association after right posterior hippocampus correction in healthy participants (152 [6.8%]) (P for familywise error = .02). Age and sex were controlled as covariates of noninterest. For presentation purposes, imaging results are shown in P < .05 uncorrected threshold at a similar spatial plane (MNI: X = 32, Y = −33, and Z = −10).

Association Between PMS and PRS and Interactions on DLPFC-HC Coupling

The DLPFC-HC connectivity was not associated with the schizophrenia PRS in validation methylation/MRI (β = 126.6; P = .79) or validation MRI (β = 401.3; P = .67). Similarly, no significant interactions between PMS and PRS on DLPFC-HC connectivity were found in validation methylation/MRI (β = −7280; P = .19) or validation MRI (β = −10 540; P = .41). The BioMM model was used to identify a risk score from GWAS MGS data using the same pathways assignment as used for DNA methylation data. The resulting associations with PMS and DLPFC-HC connectivity are described in the eResults and eTable 5 in the Supplement. Analysis of relatives methylation demonstrated that none of the groups of relatives showed significant PMS differences compared with HC (eFigure 4 in the Supplement).

Prediction of PMS in DLPFC Postmortem Brain Samples

Predicted PMS in validation post mortem was significantly higher in schizophrenia compared with HC (AUC = 0.63; P = 1.42 × 10⁻⁴; R² = 8.3%). Notably, the reverse prediction (ie, building a PMS from postmortem brain data and testing this score in blood data) did not allow case-control differentiation (eTable 6 in the Supplement) and showed no association with DLPFC-HC connectivity (validation methylation/MRI: P = .89; validation MRI: P = .92).

Assessment of the Robustness, Diagnostic Specificity, and Residual Confounding

Permutation of diagnostic labels, as well as the random selection of pathways used by BioMM, supported the significance and robustness of the PMS finding (eResults and eTables 7 and 8 in the Supplement). Analysis of the specificity methylation cohort further showed that the PMS increase was specific for schizophrenia (eTable 9 and eFigure 5 in the Supplement). Because despite the confounding correction the PMS was associated with some of all potential variables, additional analyses regarding confounding can be found in the eResults and eTables 10 and 11 in the Supplement.

Discussion

In this article, we identified a blood DNA methylation signature that reproducibly differentiated schizophrenia from HC and several other major neuropsychiatric disorders. The underlying biological pathways implicated several synaptic processes as contributing most to the classification. A more schizophrenia-like epigenetic profile was associated with an established intermediate phenotype for the disorder, persistent DLPFC-HC connectivity, in HC during 2 different WM tasks. Notably, the epigenetic signature could also differentiate schizophrenia from HC in data from DLPFC postmortem samples, supporting the relevance of the identified blood-derived epigenetic signature for brain-associated phenotypes associated with schizophrenia in vivo and ex vivo.

The pathways contributing most to the schizophrenia classification comprised synaptic and neurodevelopmental processes. This is consistent with previous results showing a colocalization of epigenetic changes with genetic susceptibility variants of schizophrenia,¹⁸ which are overrepresented in synaptic pathways.³⁷ The important role of synaptic processes in schizophrenia is supported by findings from induced pluripotent stem cells^38,39 and changes in different omics modalities pointing toward a synaptic pathology (eg, as shown by Osimo et al⁴⁰). Furthermore, schizophrenia-associated neurodevelopmental processes show an overrepresentation of methylation changes in the schizophrenia prefrontal cortex, and these are enriched for sites undergoing epigenetic changes during fetal neocortex brain development.¹⁹ Notably, postmortem expression and methylation studies support that the neural effects of epigenetic and genetic risk factors for schizophrenia already occur during early brain development, rather than the typical onset age of the illness,⁴¹ highlighting their importance for altered neurodevelopment in schizophrenia.

We showed that during 2 WM tasks, higher, more schizophrenia-like scores of the identified DNA methylation signature were associated with stronger negative DLPFC-HC connectivity. The DLPFC-HC connectivity is altered in the same way in schizophrenia, unaffected first-degree relatives, and HC carrying specific risk genetic variants,^28,29,30,42 with more negative DLPFC-HC connectivity being associated with a higher risk for schizophrenia. Furthermore, functional and structural abnormalities in the hippocampus and prefrontal cortex are at the core of the schizophrenia pathophysiology^1,43 and disease processes of both areas are highly interassociated.^43,44 For example, altered DLPFC-HC microcircuits in postmortem brains of schizophrenia affect excitatory and inhibitory cells as well as interneurons.⁴⁵ Interestingly, while the general liability for schizophrenia and specific genetic risk variants do affect this phenotype, the common polygenic risk for schizophrenia has been repeatedly demonstrated not to,^46,47 suggesting either a restricted set of genes responsible and/or more complex gene-environment interactions to shape DLPFC-HC connectivity. We extend these findings by demonstrating that an epigenetic risk signature for schizophrenia correlates with DLPFC-HC coupling and we did not find sufficient evidence for a direct polygenic and interactive association, suggesting that epigenetic analysis provides pathophysiologically relevant information not captured by PRS. Also, a BioMM-derived PRS was not associated with PMS or DLPFC-HC and showed no interactions with PMS on DLPFC-HC connectivity. This supports that the PMS outperformed classifiers identified from genetic association data and the observed PMS effects were not primarily driven by underlying genetics. The lack of PMS differences in relatives of patients with schizophrenia, BD, MDD and autism supported this finding. Notably, the accuracy a classifier can achieve is limited by the clinical and biological heterogeneity of schizophrenia. Applying multimodal subgroup identification strategies may aid in deconstructing this heterogeneity and ultimately contribute to an alternative disease classification.

Epigenetic modifications, including DNA methylation, are strongly associated with lifespan environmental exposures, such as postnatal mother-infant interactions⁴⁸ and experiencing stress-associated events,⁴⁹ making the polyepigenetic signature a potential proxy on which cumulative environmental risk factors could converge. Moreover, studies in animals and humans indicated that several neurobiological processes at different stages of development can modify DLPFC-HC connectivity across the lifespan. This includes early neuronal formation^50,51 and synaptic plasticity–associated processes,^48,52 raising the possibility that the neural association of epigenetic and genetic schizophrenia risk with synaptic processes during early development may have a lasting association with DLPFC-HC connectivity. Such an interpretation is also consistent with results from animal studies showing that lesions in the hippocampus lead to delayed maturation and dysfunction of the DLPFC.^43,44 Taken together, the present results may suggest that DLPFC-HC connectivity is associated with environmental risk accumulation mediated by the association of altered DNA methylation with synaptic plasticity.^48,52

In this article, the PMS was found to differentiate schizophrenia from HC when predicted in DLPFC DNA methylation data, indicating that elements of the signature were represented in the brain and may have mediated the observed DLPFC-HC connectivity association. This finding is consistent with the previously observed correlation between postmortem brain and blood methylation.^53,54,55 Such cross-tissue correlation has been hypothesized to result from genetic association, casting doubt on the added value of peripheral epigenome-wide association studies in brain disorders.⁵⁶ However, the association of the PMS with schizophrenia and schizophrenia-relevant neural phenotypes based on epigenetic changes in pathophysiologically relevant pathways, as well as the lack of evidence for an association with polygenic susceptibility, supports their use for integrative, multimodal approaches toward disease stratification and potentially treatment personalization. Notably, a schizophrenia PMS derived from DLPFC samples did not predict case-control status in blood samples and was not correlated with brain functional connectivity. This may have been due to the larger biological and methodologic variability in brain samples (such as cell heterogeneity or postmortem effects) or the comparatively smaller sample size.

Limitations

The present study has several limitations. First, most investigated patients were taking medication and the association of medication with the PMS could not be excluded. However, the association between the PMS and DLPFC-HC connectivity in HC not taking medication contradicts the idea that the identified methylation signature is a consequence of medication. Second, because of limited available data, the associations between DLPFC-HC connectivity and the PMS could not be explored in patients. This could have identified a stronger contribution of genetic schizophrenia risk toward the PMS associations, which may have been affected by limited statistical power. Third, while the PMS replicated in data from postmortem DLPFC samples, the explained variance in the brain data was comparatively low. This necessitates further studies to isolate the brain-specific associations of methylation differences with brain function. Fourth, despite extensive efforts to correct findings for confounding effects, we cannot exclude the potential presence of residual confounding. However, the analysis of patients with BD and MDD as well as autism suggested that the PMS increase in schizophrenia was not driven by potential confounders with transdiagnostic relevance. These findings, and PMS effects in relatives, require validation in larger cohorts. Finally, the machine learning approach used was based on the annotation of CpGs to genes and biological pathways, which may be biased and not suitably index biological function and can lead to an overlap of annotated genes between pathways. As machine learning isolates the most predictive features mapped to a given pathway, this can lead to a loss of biological specificity if the most predictive CpGs are shared among different pathways. This necessitates follow-up experiments to more precisely characterize the role of the identified pathways for mediating epigenetic risk in schizophrenia.

Conclusions

This study shows that a reproducible and specific blood DNA-methylation signature of schizophrenia was correlated with functional DLPFC-HC coupling as an intermediate phenotype for schizophrenia, mapped to methylation differences found in DLPFC postmortem samples of schizophrenia, and indexed biological pathways associated with synaptic function. These results help to characterize the systems-level association between genetic, epigenetic, and environmental risk for schizophrenia. They further support the use of PMS for multimodal biomarker discovery strategies aimed at disease stratification and the development of novel, personalized therapeutic approaches.

Supplement.

eMethods.

eResults.

eFigure 1. Post-hoc partial regression plots of the association between PMS and DLPFC-HC connectivity

eFigure 2. PMS comparision in validation_meth/MRI andrelatives_Meth

eFigure 3. PMS comparision in validation_Meth/MRI and specificity_Meth

eFigure 4. DLPFC ROI and hippocampus mask in the validation_Meth/MRI sample

eFigure 5. DLPFC ROI and hippocampus mask in the validation_MRI sample

eTable 1. Top 10 schizophrenia-associated pathways in the discovery sample (discovery_Meth)

eTable 2. 30 most significant CpGs derived from top 10 pathways

eTable 3. The existing evidence for the top genes harboring top CpGs from eTable 2

eTable 4. Prediction performance of BioMM on different cohorts using validation_postmortem as the discovery set

eTable 5. Overall prediction performance of BioMM on different cohorts and specificity analysis.

eTable 6. Differences of subject demographics between patients and controls

eTable 7. The association between PMS and DLPFC-HC, PMS and PRS, PMS and PRS

eTable 8. Permutation of diagnostic label for both machine learning prediction and the subsequent testing of imaging associations

eTable 9. Permutation of pathway level features for both machine learning prediction and subsequent testing of imaging associations

eTable 10. The association between predicted PMS and the confounding variables in controls

eTable 11. Prediction performance of BioMM on different cohorts and specificity analysis following residualization of predicted PMS scores against all potential covariates using linear regression

eReferences