Identifying drug targets for schizophrenia through gene prioritization

Abstract

Background:

Schizophrenia genome-wide association studies (GWASes) have identified >250 significant loci and prioritized >100 disease-related genes. However, gene prioritization efforts have mostly been restricted to locus-based methods that ignore information from the rest of the genome.

Methods:

To more accurately characterize genes involved in schizophrenia etiology, we applied a combination of highly-predictive tools to a published GWAS of 67,390 schizophrenia cases and 94,015 controls. We combined both locus-based methods (fine-mapped coding variants, distance to GWAS signals) and genome-wide methods (PoPS, MAGMA, ultra-rare coding variant burden tests). To validate our findings, we compared them with previous prioritization efforts, known neurodevelopmental genes, and results from the PsyOPS tool.

Results:

We prioritized 62 schizophrenia genes, 41 of which were also highlighted by our validation methods. In addition to DRD2, the principal target of antipsychotics, we prioritized 9 genes that are targeted by approved or investigational drugs. These included drugs targeting glutamatergic receptors (GRIN2A and GRM3), calcium channels (CACNA1C and CACNB2), and GABA_B receptor (GABBR2). These also included genes in loci that are shared with an addiction GWAS (e.g. PDE4B and VRK2).

Conclusions:

We curated a high-quality list of 62 genes that likely play a role in the development of schizophrenia. Developing or repurposing drugs that target these genes may lead to a new generation of schizophrenia therapies. Rodent models of addiction more closely resemble the human disorder than rodent models of schizophrenia. As such, genes prioritized for both disorders could be explored in rodent addiction models, potentially facilitating drug development.

Introduction

Schizophrenia is a highly-heritable and heterogeneous disorder characterized by positive symptoms (e.g. delusions and hallucinations), negative symptoms (e.g. blunted affect), and cognitive impairment¹. Schizophrenia patients are often also diagnosed with neurodevelopmental disorders^1,2 (e.g. intellectual disability and autism spectrum disorder) and other psychiatric conditions^3,4 (e.g. substance use disorder [SUD] and depression). Antipsychotic medications antagonizing the dopamine receptor D2 are currently the first-line treatment for schizophrenia. However, approximately 34% of patients are considered treatment-resistant⁵, and especially cognitive deficits and negative symptoms often persist^6,7. These unmet clinical needs, as well as the high burden of antipsychotic side effects^8,9, clearly underline the necessity for pharmacotherapies with novel mechanisms of action.

Only 6.2% of psychiatric drug programs that enter Phase I trials are ultimately approved—well below the average success rate of 9.6% across all medical areas¹⁰—and investment in psychiatric drug development programs have decreased in recent years¹¹. This low success rate likely reflects the complex nature of mental disorders, limited knowledge of disease mechanisms, and sparsity of validated animal models. Given that 63% of drugs approved by the FDA from 2013–2022 were supported by human genetic evidence¹², pursuing targets that are genetically-linked to disease may lead to increased success rates^13,14. A major source of this human genetic evidence comes from genome-wide association studies (GWASes)^13,14. For instance, schizophrenia GWASes^15–17 have identified a robust association near DRD2, which encodes the dopamine receptor D2. It is estimated that only 1.9% of genetically-supported drug targets for psychiatric disorders have been clinically explored¹⁸, suggesting that follow up of other schizophrenia GWAS findings may eventually lead to the design of new medicines.

The largest published schizophrenia GWAS identified 287 significant loci and prioritized 120 genes for follow up using fine-mapped credible sets¹⁹, summary data-based Mendelian randomization²⁰ (SMR), and Hi-C interactions between enhancers and promoters²¹. However, these methods only use information within a given locus, ignoring information from other significant loci and the rest of the genome. The polygenic priority score (PoPS)²² is a gene prioritization tool that incorporates genome-wide information from MAGMA²³ gene-level association tests and more than 57,000 gene-level features (i.e. gene expression, biological pathways, and protein-protein interactions). The original PoPS publication²² reported that it was possible to predict “probable causal genes” (defined using fine-mapped coding variants) with 79% precision and 39% recall simply by selecting genes that 1) were the nearest gene to a GWAS lead variant, and 2) had the top PoPS value in that same GWAS locus. This combined approach substantially increased precision compared to either individual approach, but with a moderate loss of recall (nearest gene: 46% precision and 48% recall; top PoPS value: 50% precision and 50% recall).

Here, we prioritized genes likely to play an important role in schizophrenia (SCZ) risk by combining PoPS and nearest gene results with additional high-precision prioritization methods—fine-mapped coding variants and ultra-rare coding variant burden tests²⁴. We nominated 62 genes, 10 of which are targets of approved drugs (7 genes) or drugs that have been tested in clinical trials (“investigational drugs”, 3 genes). We discuss the potential for repurposing these drugs for schizophrenia and highlight an additional 3 prioritized genes that may be tractable via small molecule drugs.

Methods and Materials

Ethics statement

This research was conducted in accordance with the ethical standards of the institutional and national research committees. Informed consent was obtained from all participants. Details on Institutional Review Board approvals of the individual studies included in the presented work are provided in the original publication¹⁷.

GWAS summary statistics

We analyzed the publicly-available “core dataset” of GWAS summary statistics from the largest published SCZ GWAS from the Psychiatric Genomics Consortium (hereafter we will refer to this study as “PGC3”)¹⁷, a meta-analysis of 90 cohorts of European (EUR) and East Asian (EAS) descent including 67,390 cases and 94,015 controls (effective sample size [N_eff] = up to 156,797). For analyses requiring data from a single ancestry, we used the EUR subset of the core dataset (76 cohorts, 53,386 cases, 77,258 controls, effective sample size [N_eff] = up to 126,282s) and the EAS-ancestry subset (14 cohorts, 14,004 cases, 16,757 controls, N_eff = up to 30,515).

Reference panels

Accordingly, we used external data from the Haplotype Reference Consortium release 1.1 (HRC) to construct three linkage disequilibrium (LD) reference panels: an EUR panel (N ≥ 16,860), an EAS panel (N ≥ 538), and an EUR+EAS panel that included both EUR and EAS individuals in the same proportions as the GWAS summary statistics—80% EUR and 20% EAS (N_EUR = 2,191, N_EAS = 538).

Variant quality control

We removed EUR+EAS GWAS variants with: 1) a minor allele count < 10 (minor allele frequency [MAF] < 0.0018) in the EUR+EAS reference panel (259 variants removed), 2) a reported allele frequency that differed from the reference panel frequency by > 0.1 (29 variants removed), and 3) a reported allele frequency that differed from the reference panel frequency by > 12-fold (11 variants removed). After quality control, 7,584,817 variants remained.

Isolating independent association signals

In order to disentangle statistically-independent genetic signals in the EUR+EAS dataset, we first clumped variants using PLINK v1.9²⁵ (P < 5×10⁻⁸, r² < 0.1, window size = 3Mbp) and our EUR+EAS reference panel, expanded the boundaries of each clump by 500kb on either side, and merged overlapping boundaries. Within each resulting region, we ran COJO²⁶ and removed hits with joint P > 5×10⁻⁸. If multiple independent hits in a region were found, we used COJO to isolate each signal by performing leave-one-hit-out conditional analysis. For each isolated signal, we computed credible sets (CSs) using the finemap.abf function in the coloc R package^27,28. Finally, we defined loci as ±300kb around each credible set.

MAGMA and PoPS

We performed gene-based association tests using MAGMA²³ (“SNP-wise mean model”) and all variants with MAF > 1%. We analyzed the EUR- and EAS-based GWASes separately using the corresponding ancestry-specific reference panel and MAFs. We mapped variants to protein-coding genes using Genome Reference Consortium Human Build 37 (GRCh37) gene start and end positions from GENCODE v44²⁹. We removed genes that had fewer than 3 variants mapped to them. For each gene, we meta-analyzed the resulting ancestry-specific MAGMA z-scores weighted by the square root of sample size³⁰. Using the ancestry-specific MAGMA results as input, we performed PoPS²² using all 57,543 gene-based features as predictors. These features were not available for chrX so we restricted our analysis to autosomal genes. The resulting ancestry-specific PoPS values were then also meta-analyzed weighted by the square root of sample size. We only used the meta-analyzed MAGMA and PoPS values for gene prioritization.

Gene prioritization criteria

Following the original PoPS publication, we prioritized genes that met both of the following criteria: 1) had the top PoPS value in a given locus and 2) were the nearest gene to the corresponding GWAS signal based on the posterior inclusion probability (PIP)-weighted average position of credible set variants. Under these criteria, however, it is possible that the top POPS value within a locus is relatively weak on a genome-wide scale, or that the nearest gene is nevertheless relatively distant. We therefore also required that genes have a PoPS value in the top 10% of all values genome-wide and the top MAGMA z-score in the locus. We also prioritized genes that had 1) PIP > 1% for non-synonymous credible set variants affecting the gene, or 2) false discovery rate-corrected P value (P_FDR) < 5% in a published SCZ burden test of ultra-rare coding variants²⁴. We used non-synonymous variants from the “baseline-LF 2.2.UKB model” (80,693 variants) and subsetted to those with an estimated per-variant heritability > 1×10⁻⁷ (removed 4,709 variants, all with estimated h² < 1×10⁻¹⁰: >1,000-fold smaller)³¹. We removed loci that contained more than 8 genes since larger loci are more challenging to resolve³², but we have included results for these large loci in Table S3.

Comparison with previous schizophrenia gene prioritization efforts

We compared our prioritized genes with those highlighted in the original PGC3 publication. Specifically, we extracted the “Symbol.ID” and “Prioritised” columns from Table S12. While the PGC3 study utilized the same core dataset, they restricted analysis to loci that retained genome-wide significance in the “extended GWAS”—a meta-analysis of the core dataset, 9 cohorts of African American and Latin American ancestry, and a dataset from deCODE genetics. They prioritized genes using a combination of FINEMAP, SMR, Hi-C interaction mapping, and non-synonymous or untranslated region credible set variants with PIP > 10%. The PGC3 study validated their list of prioritized genes by looking for overlap with genes expressed in brain tissue, genes with signatures of mutation intolerance in large-scale exome studies³³, or genes linked to schizophrenia through rare genetic variation in the SCHEMA study²⁴. Furthermore, they also found genetic overlaps in other neurodevelopmental conditions using sequencing studies from autism spectrum disorder³⁴ and developmental disorder³⁵. We incorporated a subset of this information by extracting the “ASD” and “DDD” columns from Table S12 of the PGC3 study. For full details, please refer to the original publication¹⁷.

PsyOPS

We further validated our prioritized genes using the Psychiatric Omnilocus Prioritization Score (PsyOPS) tool³⁶. The original PsyOPS publication³⁶ found that PsyOPS achieved similar performance to PoPS in predicting causal psychiatric disease genes, but using only three predictors: probability of loss-of-function intolerance (pLI) > 0.99, brain-specific gene expression, and overlap with 1,370 known genes for neurodevelopmental disorders (autism, epilepsy, intellectual disability). PsyOPS treats the nearest gene to each GWAS hit as a proxy for the causal gene in the locus, trains leave-one-chromosome-out logistic regression models, and outputs the predicted probability that a given gene is causal. We determined a gene to be prioritized by PsyOPS if the predicted probability of being a causal gene exceeded 50%. We computed PsyOPS scores using all 257 independent schizophrenia GWAS hits.

Drug repurposing and tractability

We determined whether our prioritized genes were targeted by approved or investigational drugs using GraphQL API queries of the Open Targets platform³⁷, which in turn queries the EMBL-EBI ChEMBL database. For genes that were not targeted by approved or investigational drugs, we performed additional Open Targets API queries to extract evidence of drug tractability—the probability of identifying a drug that is able to bind and modulate a given target. We focussed on small molecule drugs, but results for other modalities can be found in Figure S1.

Colocalization with other studies

We prioritized several genes that have also been highlighted by recent GWASes for addiction³⁸ and Parkinson’s disease³⁹. Using the EUR reference panel, we processed EUR-ancestry GWAS summary statistics from these studies using the same pipeline described above. We identified loci that physically overlapped with schizophrenia loci and computed the posterior probability of colocalization (H₄) using all variants in the shared locus and the coloc.abf function in the coloc R package^27,28.

Results

We prioritized schizophrenia genes using the “core dataset” from the largest published schizophrenia GWAS meta-analysis¹⁷, “PGC3” (67,390 cases and 94,015 controls). We identified 257 independent associations with P < 5×10⁻⁸ (Table S1). Across these loci, we prioritized 62 schizophrenia genes (Figure 1, Table S2) based on their distance to the credible set, PoPS and MAGMA scores, number of genes in the locus, presence of non-synonymous variants in the credible set, and support from a published schizophrenia burden test of ultra-rare coding variants²⁴ (see Methods). To validate our findings, we compared them with prioritization efforts from the PGC3 study¹⁷, genes linked to autism spectrum disorder³⁴ (ASD) and developmental disorder³⁵ (DD) via sequencing studies, and results from the PsyOPS tool (Figure 2). Across all genes in GWAS loci, prioritized genes were also DD and/or ASD genes (Fisher’s exact test P = 6.7×10⁻¹⁴, odds ratio [OR] = 67) or PsyOPS genes (Fisher’s exact test P = 3.9×10⁻⁶, OR = 14) significantly more often than expected due to chance.

Figure 1.

Heatmap

An overview of the evidence supporting each prioritized gene, separated based on whether they were (left panel) or were not (right panel) previously prioritized in the PGC3 study¹⁷. Distance: distance in kilobases between gene and credible set. PoPS: PoPS percentile where 0 represents the smallest genome-wide value and 1 represents the largest. MAGMA: MAGMA z-score percentile. # genes: number of genes in the locus. SCHEMA: a binary indicator of whether ultra-rare coding variant burden in a given gene was also significantly associated (P_FDR < 5%) with schizophrenia in a study from the Schizophrenia Exome Sequencing Meta-analysis (SCHEMA) consortium²⁴. Coding: a binary indicator of whether the credible set contained non-synonymous variants with a summed posterior inclusion probability >1%. Genes are sorted first by distance, then by PoPS percentile.

Figure 2.

Venn diagram

Venn diagram showing the overlap between the number of genes identified by the present analysis (PoPS⁺), rare-variant studies of autism spectrum disorder (ASD) and/or developmental disorder (DD), the Psychiatric Omnilocus Prioritization Score (PsyOPS), and prior gene prioritization efforts (PGC3). Gene symbols are displayed for a subset of intersecting regions.

Overlap with previous schizophrenia gene prioritization efforts

Of our 62 prioritized genes, 31 (50%) were also prioritized in the PGC3 study (“overlapping genes”) and several sources of evidence suggest that these genes are likely to play a role in schizophrenia risk. Ultra-rare coding variant burden in two overlapping genes (GRIN2A and SP4) was significantly associated (P_FDR < 5%) with schizophrenia in the SCHEMA study²⁴. Similarly, four overlapping genes (GRIN2A, CACNA1C, BCL11B, and SLC39A8) were also identified by rare variant exome sequencing studies of DD³⁵ and/or ASD³⁴ (see Figure 2). Furthermore, the lead schizophrenia variant in the SLC39A8 locus is a non-synonymous variant (PIP = 99%) that has been investigated in detail elsewhere⁴⁰. WSCD2 was also prioritized due to a non-synonymous variant in the credible set (PIP = 53%). Four overlapping genes (GRIN2A, DLGAP2, GABBR2, and CSMD1) were nominated by PsyOPS (see Methods). Notably, CSMD1 is known to inhibit the complement cascade, has reduced expression in first-episode psychosis patients⁴¹, and knockout mice have exhibited behaviors resembling schizophrenia negative symptoms⁴².

Genes that were not nominated by previous schizophrenia gene prioritization efforts

Of our 62 prioritized genes, 31 (50%) were not prioritized in the PGC3 study (“non-overlapping genes”). However, a similar proportion of these non-overlapping genes were supported by the same evidence sources as presented above (9/31 for overlapping genes vs. 10/31 for non-overlapping genes). Two non-overlapping genes (STAG1 and FAM120A) were significantly associated with ultra-rare coding variant burden in the SCHEMA study²⁴. Five non-overlapping genes (FOXP1, TBL1XR1, ZEB2, CUL3, and TCF4) were also identified by rare variant exome sequencing studies of DD³⁵ and/or ASD³⁴. Note that TCF4 was not prioritized in the PGC3 study because they only investigated regions containing three independent genetic associations or fewer and there were four associations near TCF4. We prioritized BRINP2 due to a non-synonymous variant in the credible set (r² with lead variant = 97%, PIP = 2.5%), but was not prioritized in the PGC3 study which required PIP > 10%. Three non-overlapping genes (ZEB2, HCN1, and RIMS1) were nominated by PsyOPS (see Methods). Perhaps most importantly, our analysis uniquely highlighted the dopamine receptor gene DRD2, which is targeted by most approved antipsychotic medications⁴³ (Figure 3A).

Figure 3.

Variant-level associations and PoPS results for selected loci

The prioritized genes in plots A-E are targets of approved drugs; the prioritized genes in plots E-F are in loci shared by an addiction GWAS³⁸. The upper portion of each sub-plot is a LocusZoom plot. Each point represents a different genetic variant, the x-axis represents physical position on the listed chromosome, the left y-axis represents −log₁₀-transformed P value, the right y-axis represents the recombination rate, colour represents linkage disequilibrium with the lead variant in the locus (as shown in the legend), and the horizontal dashed line represents the genome-wide significance P value threshold of 5×10⁻⁸. The lower portion of each figure is a PoPS plot. Genes are denoted as blue bars spanning from their transcription start site to their transcription stop site using the same x-axis as the LocusZoom plot, the y-axis represents the raw PoPS score, the dashed horizontal grey lines represent the top 10% and 1% of PoPS scores genome-wide, and the solid horizontal grey line represents a PoPS score of 0.

Drug repurposing and tractability

In addition to DRD2, we prioritized 9 genes that are targeted by approved (6 genes) or investigational drugs (3 genes, Table S4). Of these, 6 were also prioritized in the PGC3 study (GRIN2A, CACNA1C, PDE4B, GABBR2, AKT3, and DPYD) and 3 (CACNB2, GRM3, and SNCA) were uniquely prioritized in our analysis (Table 1, see Discussion). Our list of prioritized genes also included 3 genes (HCN1, VRK2, TRPC4) that belong to known druggable protein families⁴⁴ and are reported to bind to at least one high-quality ligand³⁷, suggesting potential as small molecule drug targets (Figure S1).

Table 1.

Prioritized genes targeted by approved and investigational drugs

Gene	Example approved drugs (indications)	Example investigational drugs (indications, phase)
DRD2	Clozapine (schizophrenia); haloperidol (schizophrenia)	Not applicable, already approved for schizophrenia
GRIN2A	Esketamine (major depressive disorder); amantadine (Parkinson’s disease/influenza)1	Apimostinel (schizophrenia, Phase II), rapastinel (major depressive disorder, Phase III)
CACNA1C	Verapamil (hypertension); amlodipine (hypertension/coronary artery disease)	Verapamil (bipolar disorder, Phase III); nimodipine (schizophrenia, Phase I early)
GABBR2	Baclofen (spasticity from multiple sclerosis/spinal cord injury/cerebral palsy); oxybate (narcolepsy-cataplexy syndrome)	Arbaclofen (autism spectrum disorder, Phase II); baclofen (nicotine dependence, Phase II)
PDE4B	Roflumilast (chronic bronchitis/airway obstruction); pentoxifylline (cardiovascular disease),	Roflumilast (schizophrenia, Phase I); ibudilast (alcohol dependence, Phase II)
CACNB2	Bepridil (hypertension); terodiline (polyuria/urinary incontinence)	Imagabalin (generalized anxiety disorder, Phase III); atagabalin (insomnia, Phase II)
AKT3	Capivasertib (breast cancer)1	Ipatasertib (breast cancer, Phase III)
GRM3	None	Pomaglumetad methionil (schizophrenia, Phase III)
SNCA	None	Cinpanemab (Parkinson’s disease, Phase II); prasinezumab (Parkinson’s disease, Phase II)
DPYD	None	Gimeracil (neoplasm, Phase III)

Note. For each gene, up to two approved drugs and up to two investigational drugs are presented (see Table S4 for the full list of drugs associated with each gene). Associated indications are displayed in parentheses. Underlined drugs increase the activity of their target gene product, all other drugs decrease activity.

¹These drugs all decrease activity, but increased activity is likely desired for schizophrenia (see Discussion).

Discussion

We prioritized 62 genes near 257 independent GWAS signals. Of these genes, 41 (66%) were also supported by evidence (Figure 2) from the PGC3 study (31 genes), DD/ASD sequencing studies (10 genes), and PsyOPS (7 genes). We prioritized DRD2 (Figure 3A)⁴³, 9 other genes targeted by approved drugs (6 genes) or drugs that have been tested in clinical trials (3 genes), and 3 other genes that may represent tractable small molecule drug targets. Our analyses do not predict whether the effect of these drugs (e.g. inhibitor) aligns with the effect that would be desired for schizophrenia. Therefore, we will now discuss literature supporting the potential for these drugs to be repurposed as treatments for schizophrenia.

Glutamate receptors: GRIN2A and GRM3

We prioritized GRIN2A, which encodes a subunit of the N-methyl-D-aspartate receptor (NMDA-R, Figure 3B). In addition to GWAS, there is evidence that decreased NMDA-R function increases schizophrenia risk from GRIN2A ultra-rare variant burden tests²⁴, GRIN2A mouse knockout models⁴⁵, and pharmacological antagonism of the NMDA-R⁴⁶. This raises the possibility that increasing NMDA-R activity may provide therapeutic benefit for schizophrenia patients. A meta-analysis of 4,937 schizophrenia patients from 40 randomized controlled trials found that NMDA-R modulator augmentation (e.g. via glycine or glycine transporter type I inhibitors) significantly improved total, positive, and negative schizophrenia symptoms versus placebo⁴⁷. These compounds have also been proposed as a therapeutic strategy for schizophrenia patients who are treatment-resistant or have impaired cognition⁴⁸. There are currently three Phase III clinical trials underway assessing the effect of iclepertin, a glycine transporter type I inhibitor, on cognitive impairment associated with schizophrenia⁴⁹. If ultimately approved, this may become the first medication indicated to treat the cognitive symptoms of schizophrenia.

We also prioritized GRM3, which encodes a different glutamate receptor: metabotropic glutamate receptor 3 (mGluR₃). Clinical trials of pomaglumetad methionil, an mGluR_2/3 agonist, have yielded inconclusive effects on positive symptoms^50,51–54. However, an analysis of clinical trial data suggested that specific patient subgroups may have benefited⁵⁵ and preclinical research has suggested that a cognitive endpoint may be more appropriate^56,57.

Voltage-gated calcium channels: CACNA1C and CACNB2

We prioritized CACNA1C (Figure 3C), which encodes the alpha-1 subunit of a voltage-gated calcium channel (Ca_v1.2). A Phase III clinical trial for bipolar disorder showed that 11 out of 13 non-responders to first-line therapy (lithium) showed a clinically-meaningful response to verapamil (a calcium channel blocker [CCB]), or verapamil + lithium⁵⁸. The genetic correlation between schizophrenia and bipolar disorder is approximately 70%² and a recent bipolar disorder GWAS also identified a significant association near CACNA1C⁵⁹, suggesting that verapamil may be a promising treatment option for schizophrenia. Other CCBs may also be effective—a large cohort study (N = 10,460) found that use of dihydropyridine CCBs was associated with reduced risk of psychiatric rehospitalization⁶⁰. CCBs may also improve certain cognitive functions^61,62. The use of CCBs for treating schizophrenia is further supported by the fact that we prioritized CACN2B, an auxiliary subunit of voltage-gated calcium channels.

Loci shared with addiction: PDE4B and VRK2

We prioritized PDE4B, which encodes phosphodiesterase 4B (Figure 3E). A recent GWAS of an addiction-related latent factor derived from four SUDs³⁸ also found a signal near PDE4B and highlighted PDE4B as the likely causal gene. SUDs are frequently comorbid with schizophrenia⁴ and there is significant genetic correlation between schizophrenia and several SUDs⁶³. While it is challenging to assess psychotic symptoms in rodents, high-quality rodent addiction models exist for a wide range of substances⁶⁴. Indeed, several drugs that are approved to treat alcohol use disorder (e.g. naltrexone and acamprosate) were originally pursued based in part on success in preclinical animal models^64,65. Administering ibudilast, a drug that inhibits PDE4B and other phosphodiesterases, has been shown to reduce alcohol intake by approximately 50% in rats⁶⁶ and decrease the odds of heavy drinking by 45% in a randomized clinical trial in humans⁶⁷. Given that both addiction and schizophrenia GWASes have suggested an important role for PDE4B in disease risk, PDE4B inhibitors may also benefit schizophrenia patients. A Phase I study in 15 schizophrenia patients found that roflumilast, an inhibitor of all four PDE4 phosphodiesterases, significantly improved verbal memory, but not working memory⁶⁸.

We prioritized VRK2, which encodes vaccinia-related kinase 2 (Figure 3F). While the role of VRK2 in schizophrenia remains unclear, it is expressed in microglial cells and a mechanism involving synaptic elimination by microglial cells has been proposed^69,70. Like PDE4B, the same addiction GWAS³⁸ also found an association near VRK2. The addiction and schizophrenia signals colocalize (H₄ = 92%), suggesting a shared causal variant. Therefore, modulating VRK2 activity might result in clinical benefit for people with SUD and/or schizophrenia. VRK2 is a member of the highly-druggable serine/threonine kinases group of enzymes⁴⁴ and has been co-crystallised with a small molecule ligand⁷¹. VRK2 modulation could be tested in rodent addiction models and, if successful, may warrant further testing in human clinical trials of SUD and SCZ patients.

Three other prioritized genes reside in loci shared with the addiction GWAS³⁸: DRD2, SLC39A8 (H₄ = 100%), and PLCL2 (H₄ = 74%). Although our analyses did not find evidence that SLC39A8 and PLCL2 are easily druggable by small molecule drugs, knockdown or overexpression of these genes in rodent addiction models may nevertheless improve our understanding of the shared biology of addiction and schizophrenia.

GABBR2

We prioritized GABBR2, which encodes the gamma-aminobutyric acid (GABA) type B receptor and is known to inhibit neuronal activity via downstream signaling cascades (Figure 3D). A Phase II clinical trial is currently testing whether arbaclofen, a GABA_B receptor agonist, can rescue ASD symptoms⁷². Both post-mortem and in vivo studies identified reduced GABA levels in schizophrenia patients compared to controls, and impaired gamma band oscillations—which are linked with GABAergic signaling—are associated with schizophrenia^73–77. If proven to be a successful therapy for ASD, arbaclofen may therefore represent an interesting drug repurposing candidate for schizophrenia, particularly for symptoms and socio-cognitive deficits that are shared between the two disorders^78,79.

AKT3

We prioritized AKT3, the member of the AKT serine/threonine-protein kinase gene family with the highest brain-specific expression. Capivasertib—an inhibitor of all three AKT kinases—was recently approved by the FDA to treat a subset of breast cancer patients⁸⁰. However, AKT inhibition can lead to adverse psychiatric side effects⁸¹ and AKT3 knockout or knockdown resulted in cognitive deficits and reduced brain size in mice^82,83. Further studies are necessary to determine whether overall or isoform-specific⁸⁴ increases in AKT3 activity would benefit schizophrenia patients without increasing cancer risk.

SNCA

We prioritized SNCA, which encodes α-synuclein (α-syn). α-syn aggregates are the pathological hallmark of Parkinson’s disease (PD) and antibodies targeting aggregated α-syn have been tested in two Phase II clinical trials for PD, although neither meet their primary endpoint^85,86. The schizophrenia association near SNCA colocalizes (H₄ = 85%) with an association from a recent European-ancestry PD GWAS³⁹. The schizophrenia risk allele was associated with increased PD risk, which is in turn linked to increased α-syn production⁸⁷. As such, interventions that decrease α-syn production may benefit both PD and schizophrenia patients.

Limitations

The PGC3 study prioritized 89 genes that were not prioritized in our study. The majority of these (52 genes) were prioritized via SMR. We did not include SMR because it demonstrated lower precision than other methods in predicting a “gold standard” dataset of causal and non-causal trait-gene pairs²², consistent with recent models for systematic differences between variants highlighted by GWAS and expression studies⁸⁸. The precision of SMR-nominated genes that failed to meet our gene prioritization criteria is likely to be lower still. The PGC3 study also prioritized 5 autosomal genes affected by non-synonymous credible set variants: ACTR1B, CUL9, IRF3, THAP8, and ZNF835. These genes resided in “large loci” (containing >8 genes), which are intrinsically harder to resolve³². However, these genes may warrant further attention given that coding variants have been shown to prioritize causal genes with high precision⁸⁹. An additional 10 genes met all of our prioritization criteria, but resided in large loci. Of these, 2 were prioritized by the PGC3 study (FURIN and ACE) and 8 were not (YWHAE, CACNA1I, CHRNA3, AGO3, KIF21B, PTPRF, SYNGAP1, and GATAD2B). CACNA1I, CHRNA3, and ACE may be particularly interesting since they are targeted by approved drugs and may represent drug repurposing opportunities.

The original PGC3 study performed gene prioritization analyses in the “core dataset”. This excluded individuals of African (AFR) or Latin American (LAT) ancestry found in the “extended dataset”. To ensure consistency with the original PGC3 study, we also analyzed the core dataset. Furthermore, the AFR and LAT datasets only included GWAS summary statistics, not individual-level genotypes, preventing us from identifying well-matched LD reference panels—something particularly important for admixed populations⁹⁰. Nevertheless, we stress the importance of expanding gene prioritization to include more ancestries to ensure that findings are generalizable to a broader range of people.

Conclusion

We have curated a high-quality list of 62 genes that likely play a role in the development of schizophrenia. Developing or repurposing drugs that target these genes may lead to a new generation of schizophrenia therapies. The highest-priority candidates nominated by our work and previous clinical trials are NMDA-R modulator augmentation (GRIN2A) and brain-penetrant calcium channel blockers (CACNA1C and CACNB2). We prioritized genes that likely also play a role in SUD, including PDE4B and VRK2. Drugs that modulate the activity of these genes should be tested in high-quality rodent models of addiction and, if shown to be safe and effective, should be considered for human clinical trials for SUD and/or schizophrenia. As new drug modalities continue to be invented and refined, more genes will become druggable. We hope that our list of prioritized genes will ultimately facilitate the development of new medicines for people living with schizophrenia.

Members of the Schizophrenia Working Group of the Psychiatric Genomics Consortium

Vassily Trubetskoy, Antonio F Pardiñas, Georgia Panagiotaropoulou, Swapnil Awasthi, Tim B Bigdeli, Charlotte A Dennison, Lynsey S Hall, Max Lam, Oleksandr Frei, Alexander L Richards, Jakob Grove, Zhiqiang Li, Mark Adams, Ingrid Agartz, Elizabeth G Atkinson, Esben Agerbo, Mariam Al Eissa, Margot Albus, Madeline Alexander, Behrooz Z Alizadeha, Köksal Alptekin, Thomas D Als, Farooq Amin, Volker Arolt, Manuel Arrojo, Lavinia Athanasiu, Maria Helena Azevedo, Silviu A Bacanu, Nicholas J Bass, Martin Begemann, Richard A Belliveau, Judit Bene, Beben Benyamin, Sarah E Bergen, Giuseppe Blasi, Julio Bobes, Stefano Bonassi, Alice Braun, Rodrigo Affonseca Bressan, Evelyn J Bromet, Richard Bruggeman, Peter F Buckley, Randy L Buckner, Jonas Bybjerg-Grauholm, Wiepke Cahn, Murray J Cairns, Monica E Calkins, Vaughan J Carr, David Castle, Stanley V Catts, Kimberley D Chambert, Raymond CK Chan, Boris Chaumette, Wei Cheng, Eric FC Cheung, Siow Ann Chong, David Cohen, Angèle Consoli, Quirino Cordeiro, Javier Costas, Charles Curtis, Michael Davidson, Kenneth L Davis, Lieuwe de Haan, Franziska Degenhardt, Lynn E DeLisi, Ditte Demontis, Faith Dickerson, Dimitris Dikeos, Timothy Dinan, Srdjan Djurovic, Jubao Duan, Giuseppe Ducci, Johan G Eriksson, Lourdes Fañanás, Stephen V Faraone, Alessia Fiorentino, Andreas Forstner, Josef Frank, Nelson B Freimer, Menachem Fromer, Alessandra Frustaci, Ary Gadelha, Giulio Genovese, Elliot S Gershon, Marianna Giannitelli, Ina Giegling, Paola Giusti-Rodríguez, Stephanie Godard, Jacqueline I Goldstein, Javier González Peñas, Ana González-Pinto, Srihari Gopal, Jacob Gratten, Michael F Green, Tiffany A Greenwood, Olivier Guillin, Sinan Gülöksüz, Raquel E Gur, Ruben C Gur, Blanca Gutiérrez, Eric Hahn, Hakon Hakonarson, Vahram Haroutunian, Annette M Hartmann, Carol Harvey, Caroline Hayward, Frans A Henskens, Stefan Herms, Per Hoffmann, Daniel P Howrigan, Masashi Ikeda, Conrad Iyegbe, Inge Joa, Antonio Julià, Anna K Kähler, Tony Kam-Thong, Yoichiro Kamatani, Sena Karachanak-Yankova, Oussama Kebir, Matthew C Keller, Brian J Kelly, Andrey Khrunin, Sung-Wan Kim, Janis Klovins, Nikolay Kondratiev, Bettina Konte, Julia Kraft, Michiaki Kubo, Vaidutis Kučinskas, Zita Ausrele Kučinskiene, Agung Kusumawardhani, Hana Kuzelova-Ptackova, Stefano Landi, Laura C Lazzeroni, Phil H Lee, Sophie E Legge, Douglas S Lehrer, Rebecca Lencer, Bernard Lerer, Miaoxin Li, Jeffrey Lieberman, Gregory A Light, Svetlana Limborska, Chih-Min Liu, Jouko Lönnqvist, Carmel M Loughland, Jan Lubinski, Jurjen J Luykx, Amy Lynham, Milan Macek Jr, Andrew Mackinnon, Patrik KE Magnusson, Brion S Maher, Wolfgang Maier, Dolores Malaspina, Jacques Mallet, Stephen R Marder, Sara Marsal, Alicia R Martin, Lourdes Martorell, Manuel Mattheisen, Robert W McCarley, Colm McDonald, John J McGrath, Helena Medeiros, Sandra Meier, Bela Melegh, Ingrid Melle, Raquelle I Mesholam-Gately, Andres Metspalu, Patricia T Michie, Lili Milani, Vihra Milanova, Marina Mitjans, Espen Molden, Esther Molina, María Dolores Molto, Valeria Mondelli, Carmen Moreno, Christopher P Morley, Gerard Muntané, Kieran C Murphy, Inez Myin-Germeys, Igor Nenadić, Gerald Nestadt, Liene Nikitina-Zake, Cristiano Noto, Keith H Nuechterlein, Niamh Louise O’Brien, F Anthony O’Neill, Sang-Yun Oh, Ann Olincy, Vanessa Kiyomi Ota, Christos Pantelis, George N Papadimitriou, Mara Parellada, Tiina Paunio, Renata Pellegrino, Sathish Periyasamy, Diana O Perkins, Bruno Pfuhlmann, Olli Pietiläinen, Jonathan Pimm, David Porteous, John Powell, Diego Quattrone, Digby Quested, Allen D Radant, Antonio Rampino, Mark H Rapaport, Anna Rautanen, Abraham Reichenberg, Cheryl Roe, Joshua L Roffman, Julian Roth, Matthias Rothermundt, Bart PF Rutten, Safaa Saker-Delye, Veikko Salomaa, Julio Sanjuan, Marcos Leite Santoro, Adam Savitz, Ulrich Schall, Rodney J Scott, Larry J Seidman, Sally Isabel Sharp, Jianxin Shi, Larry J Siever, Kang Sim, Nora Skarabis, Petr Slominsky, Hon-Cheong So, Janet L Sobell, Erik Söderman, Helen J Stain, Nils Eiel Steen, Agnes A. Steixner-Kumar, Elisabeth Stögmann, William S Stone, Richard E Straub, Fabian Streit, Eric Strengman, T Scott Stroup, Mythily Subramaniam, Catherine A Sugar, Jaana Suvisaari, Dragan M Svrakic, Neal R Swerdlow, Jin P Szatkiewicz, Thi Minh Tam Ta, Atsushi Takahashi, Chikashi Terao, Florence Thibaut, Draga Toncheva, Paul A Tooney, Silvia Torretta, Sarah Tosato, Gian Battista Tura, Bruce I Turetsky, Alp Üçok, Arne Vaaler, Therese van Amelsvoort, Ruud van Winkel, Juha Veijola, John Waddington, Henrik Walter, Anna Waterreus, Bradley T Webb, Mark Weiser, Nigel M Williams, Stephanie H Witt, Brandon K Wormley, Jing Qin Wu, Zhida Xu, Robert Yolken, Clement C Zai, Wei Zhou, Feng Zhu, Fritz Zimprich, Eşref Cem Atbaşoğlu, Muhammad Ayub, Alessandro Bertolino, Donald W Black, Nicholas J Bray, Gerome Breen, Nancy G Buccola, William F Byerley, Wei J Chen, C Robert Cloninger, Benedicto Crespo-Facorro, Gary Donohoe, Robert Freedman, Cherrie Galletly, Massimo Gennarelli, David M Hougaard, Hai-Gwo Hwu, Assen V Jablensky, Steven A McCarroll, Jennifer L Moran, Ole Mors, Preben B Mortensen, Bertram Müller-Myhsok, Amanda L Neil, Merete Nordentoft, Michele T Pato, Tracey L Petryshen, Ann E Pulver, Thomas G Schulze, Jeremy M Silverman, Jordan W Smoller, Eli A Stahl, Debby W Tsuang, Elisabet Vilella, Shi-Heng Wang, Shuhua Xu, Rolf Adolfsson, Celso Arango, Bernhard T Baune, Sintia Iole Belangero, Anders D Børglum, David Braff, Elvira Bramon, Joseph D Buxbaum, Dominique Campion, Jorge A Cervilla, Sven Cichon, David A Collier, Aiden Corvin, Marta Di Forti, Enrico Domenici, Hannelore Ehrenreich, Valentina Escott-Price, Tõnu Esko, Ayman H Fanous, Anna Gareeva, Micha Gawlik, Pablo V Gejman, Michael Gill, Stephen J Glatt, Vera Golimbet, Kyung Sue Hong, Christina M Hultman, Steven E Hyman, Nakao Iwata, Erik G Jönsson, René S Kahn, James L Kennedy, Elza Khusnutdinova, George Kirov, James A Knowles, Marie-Odile Krebs, Claudine Laurent-Levinson, Jimmy Lee, Todd Lencz, Douglas F Levinson, Qingqin S Li, Jianjun Liu, Anil K Malhotra, Dheeraj Malhotra, Andrew McIntosh, Andrew McQuillin, Paulo R Menezes, Vera A Morgan, Derek W Morris, Bryan J Mowry, Robin M Murray, Vishwajit Nimgaonkar, Markus M Nöthen, Roel A Ophoff, Sara A Paciga, Aarno Palotie, Carlos N Pato, Shengying Qin, Marcella Rietschel, Brien P Riley, Margarita Rivera, Dan Rujescu, Meram C Saka, Alan R Sanders, Sibylle G Schwab, Alessandro Serretti, Pak C Sham, Yongyong Shi, David St Clair, Ming T Tsuang, Jim van Os, Marquis P Vawter, Daniel R Weinberger, Thomas Werge, Dieter B Wildenauer, Xin Yu, Weihua Yue, Peter A Holmans, Panos Roussos, Evangelos Vassos, Danielle Posthuma, Ole A Andreassen, Kenneth S Kendler, Michael J Owen, Naomi R Wray, Mark J Daly, Hailiang Huang, Benjamin M Neale, Patrick F Sullivan, Stephan Ripke, James TR Walters, Michael C O’Donovan

Data availability statement

ChEMBL Database: https://www.ebi.ac.uk/chembl/

HRC reference release 1.1: https://ega-archive.org/datasets/EGAD00001002729

Gencode release 44: https://www.gencodegenes.org/human/release_44.html

OpenTargets platform: https://platform-docs.opentargets.org/

The PGC3 GWAS core dataset is available through the PGC data access portal: https://pgc.unc.edu/for-researchers/data-access-committee/data-access-portal/

Summary statistics of the PGC3 GWAS are freely available for download: https://pgc.unc.edu/for-researchers/download-results/