Synaptic and Gene Regulatory Mechanisms in Schizophrenia, Autism, and 22q11.2 CNV Mediated Risk for Neuropsychiatric Disorders

Abstract

Background

22q11.2 copy number variants (CNVs) are among the most highly penetrant genetic risk variants for developmental neuropsychiatric disorders such as schizophrenia (SCZ) and autism spectrum disorder (ASD). However, the specific mechanisms through which they confer risk remain unclear.

Methods

Using a functional genomics approach, we integrated transcriptomic data from the developing human brain, genome-wide association findings for SCZ and ASD, protein interaction data, and gene expression signatures from SCZ and ASD post-mortem cortex to: 1) organize genes into the developmental cellular and molecular systems within which they operate; 2) identify neurodevelopmental processes associated with polygenic risk for SCZ and ASD across the allelic frequency spectrum; and 3) elucidate pathways and individual genes through which 22q11.2 CNVs may confer risk for each disorder.

Results

Polygenic risk for SCZ and ASD converged on partially overlapping neurodevelopmental modules involved in synaptic function and transcriptional regulation, with ASD risk variants additionally enriched for modules involved in neuronal differentiation during fetal development. The 22q11.2 locus formed a large protein network during development that disproportionately affected SCZ- and ASD-associated neurodevelopmental modules, including loading highly onto synaptic and gene regulatory pathways. SEPT5, PI4KA, and SNAP29 genes are candidate drivers of 22q11.2 synaptic pathology relevant to SCZ and ASD, and DGCR8 and HIRA are candidate drivers of disease-relevant alterations in gene regulation.

Conclusions

The current approach offers a powerful framework to identify neurodevelopmental processes affected by diverse risk variants for SCZ and ASD, and elucidate the mechanisms through which highly penetrant multi-gene CNVs contribute to disease risk.

Introduction

Copy number variants (CNVs) at chromosome 22q11.2 are among the genetic variants most strongly associated with developmental neuropsychiatric disorders. The hemizygous deletion, 22q11.2 deletion syndrome (22q11DS), typically spans a gene-rich ~2.5 Mb region(1) and is the greatest known single genetic risk factor for schizophrenia (SCZ). Approximately 25% of 22q11DS patients develop a psychotic disorder by adulthood(2); conversely, ~0.3% of patients with SCZ have 22q11DS(3,4). Interestingly, the reciprocal duplication (22q11DupS) may protect against SCZ(4,5), whereas both 22q11DS and 22q11DupS confer risk for autism spectrum disorder (ASD;6). As psychiatry increasingly looks towards precision medicine approaches to therapeutic development, clarifying the mechanisms through which 22q11.2 CNVs confer risk for SCZ and ASD, and how these converge with risk from broader variants associated with each disorder, is a critical goal.

Our knowledge of the genetic architectures of both SCZ and ASD has advanced considerably in the last decade. It is now clear that SCZ and ASD are both highly polygenic and genetically heterogeneous disorders, with risk variants for each disorder spanning common (i.e., present in >1% of the population), rare (i.e., present in <1% of the population), and de novo variants (i.e., new mutations in offspring), in addition to CNVs(7). Despite this polygenicity, risk variants for SCZ and ASD each appear to converge on a subset of biological pathways. Neuronal and synaptic genes have been repeatedly implicated in SCZ, including the N-methyl-D-aspartate receptor (NMDAR) and activity-regulated cytoskeleton-associated protein (Arc) complexes, and targets of the fragile X mental retardation protein (FMRP;4,8–12). Similarly, neuronal and synaptic genes have been implicated in ASD, along with chromatin and transcription regulators, and FMRP targets(6,13–17). However, many genes in these gene-sets are developmentally regulated, and SCZ and ASD are both neurodevelopmental in origin. Indeed, brain development is tightly controlled through changes in transcription factor activity and chromatin condensation that alter gene expression patterns at precise times in development(18). These changes program the progression of stem cell proliferation, neuronal differentiation, and synaptogenesis during prenatal development, and myelination and synaptic refinement during postnatal development. Examining the convergence of SCZ and ASD risk variants on groups of developmentally co-expressed genes may therefore yield fundamental insights into the neurodevelopmental pathogenesis of each disorder (e.g., 16,17), and provide a powerful lens to understand how 22q11.2 CNVs confer risk for each disorder.

Here, we used network analyses of transcriptomic data from the developing human brain(18) to organize genes into the developmental cellular and molecular systems within which they operate(19). We then conducted a comprehensive examination of neurodevelopmental processes associated with risk variants for SCZ and ASD across the allelic frequency spectrum. Finally, we identified the brain developmental 22q11.2 protein network, and developed a framework leveraging this disease-informed lens to predict pathways and individual genes underlying 22q11.2 CNV-mediated risk for SCZ and ASD.

Methods and Materials

Defining Neurodevelopmental Co-Expression Modules

Modules of developmentally co-expressed genes were constructed using weighted gene co-expression network analysis (WGCNA) of BrainSpan data, the largest, publicly-available developmental brain transcriptome dataset derived from healthy human donors(18). Samples spanning the full period of brain development (i.e., ~6 post-conception weeks-30 years; 1061 samples; Table S1) were included. Briefly, correlations were computed for the expression of each gene with every other gene, and genes were clustered based on connectivity to identify modules of co-expressed genes. Modules were functionally annotated using: 1) brain cell-type markers(20); 2) gene ontology (GO) using g:Profiler(21); and 3) genes expressed in specific brain regions and developmental periods using the Specific Expression Analysis tool(22). The first principal component of each module (i.e., module eigengene) was used to summarize the developmental trajectory of each module.

Identifying Neurodevelopmental Modules Associated with SCZ and ASD

Summary statistics, gene-lists, and/or variant lists from genome-wide association studies of SCZ and ASD were compiled and tested for module enrichment using methods appropriate to each data type (see Table 1). Briefly, when summary statistics for common(9,12,23) and rare variants(24) were available, they were tested for enrichment using MAGMA(25). De novo mutations (DNMs) in SCZ(10,26–33) and ASD(34,35) were tested for enrichment using DenovolyzeR(36). Genes in SCZ-(4) and ASD-associated(15) CNV loci were tested for enrichment using logistic regression, controlling for gene length. Genes containing more deleterious ultra-rare variants in SCZ patients than controls(11); manually curated high-confidence ASD genes according to the SFARI database(37); and genes associated with ASD across rare, de novo, and copy number variants(15) were tested for enrichment using Fisher’s Exact tests with the GeneOverlap package(38).

Table 1.

Genome-wide SCZ and ASD Risk Variant Datasets Tested for BrainSpan Module Enrichment.

Dataset	Reference	Variant Type	Description	Patient N	Control N	Analysis
SCZ
SCZ PGC-deCODE*	12	Common	PGC + deCODE meta-analysis	36,989	113,075	MAGMA
SCZ CLOZUK-PGC*	9	Common	CLOZUK + PGC meta-analysis	40,675	64,643	MAGMA
SCZ Richards	24	Rare	Rare variants (MAF<1%)	5,585	8,103	MAGMA
SCZ Genovese	11	Ultra-rare	43 genes enriched for deleterious ultra-rare variants (p<0.01)	4,877	6,242	Fisher’s Exact Test
SCZ DNM	10, 26–33	De novo	Compilation of DNMs	1,139	NA	DenovolyzeR
SCZ CNV	4	CNV	13 unique loci (FDR<0.05)	21,094	20,227	Logistic Regression
ASD
ASD iPSYCH-PGC	23	Common	iPSYCH + PGC meta-analysis	18,381	27,969	MAGMA
ASD SFARI	www.sfari.org	Gene list	80 literature-curated genes (score 1 or 2)	NA	NA	Fisher’s Exact Test
ASD Sanders	15	Rare	65 genes (FDR<0.1)	3,982	1,911	Fisher’s Exact Test
ASD DNM	34, 35	De novo	Compilation of DNMs	4,424	NA	DenovolyzeR
ASD CNV	15	CNV	9 unique loci, +3 nested loci, in SSC and AGP cohorts (FDR<0.05)	4,687	2,100	Logistic Regression

^*The PGC-deCODE meta-analysis includes 1,513 SCZ cases and 36,989 controls from the deCODE study not in the CLOZUK-PGC study and the SCZ CLOZUK-PGC meta-analysis includes 5,220 SCZ cases and 18,823 controls not in the SCZ PGC-deCODE study.

Characterizing the 22q11.2 Locus

To examine whether any BrainSpan module was over-represented among 22q11.2 genes, given the number of genes in each module, or whether SCZ- and ASD-associated modules were over-represented overall, logistic regression controlling for gene length was conducted for each module, and for all disease-associated modules together, respectively.

To characterize the total protein-coding gene content of the 22q11.2 locus relative to all regions of the genome, and per module, the rank of the 22q11.2 locus was compared to the empirical distribution from all equal-sized genomic regions, generated using the GenomicRanges package(39) in 250 kb sliding steps (i.e. 12,149 genome-wide ~2.57 Mb regions). Extreme-high rank values for the 22q11.2 locus relative to all genomic regions (P_{GENOMIC-REGION}<0.05) were considered statistically significant. A Mann-Whitney U test also examined whether 22q11.2 ranks were higher for SCZ- and ASD-associated modules, overall, compared to non-associated modules.

Characterizing the 22q_BD-PPI Network

Genes form protein-protein interaction (PPI) networks that fulfill critical biological roles, and perturbing protein networks is a common characteristic of disease-associated mutations(40). However, protein interactions depend on the tissue and temporal availability of each protein(41). We therefore identified and characterized the brain developmental PPI (BD-PPI) network for the 22q11.2 locus (i.e., 22q_BD-PPI network) by first compiling a comprehensive catalogue of pair-wise protein interactions from the BioGRID(42) and InWeb3 PPI databases(43,44); PPI pairs were then thresholded for co-expression during at least one stage of brain development (Spearman’s ρ>0.7) to identify PPI likely to exist in the brain during development.

The 22q_BD-PPI network was defined as its seed genes and corresponding BD-PPI pairs, and was first tested for over-representation in any BrainSpan modules, given the number of genes in each module, using Fisher’s Exact tests. We then examined the 22q_BD-PPI network’s rank relative to the BD-PPI networks of 2 null distributions: 1) all 12,149 ~2.57 Mb regions of the genome, described above (P_{GENOMIC-REGION NEWORK}); and 2) 10,000 random gene-sets, equal in number of seed BrainSpan genes to the 22q11.2 locus (P_{GENE-SET NETWORK}). For each null distribution, extreme-high rank values (P<0.05) were considered statistically significant. Mann-Whitney U tests additionally examined whether 22q_BD-PPI network ranks were higher overall for SCZ- and ASD-associated modules than for non-associated modules.

Prioritizing Disease-Relevant 22q11.2 Genes

To prioritize 22q11.2 genes likely to drive risk for SCZ and ASD, the number of genes in each individual 22q11.2 gene’s BD-PPI network overlapping disease-associated BrainSpan modules was compared to the null distribution of all BrainSpan genes’ BD-PPI networks. 22q11.2 genes with extreme-high BD-PPI network ranks (P_{GENE NETWORK}<0.05) were considered enriched for disease-associated modules and functionally annotated using g:Profiler.

See Supplementary Methods for details on all statistical analyses.

Results

SCZ and ASD Risk Variants Converge on Partially Overlapping Neurodevelopmental Modules

Eighteen BrainSpan modules were identified across 17,216 genes (Figure S1A; Table S2) with distinct developmental expression trajectories, cell-type enrichment patterns (Figure S1B), and functional characteristics, capturing broad aspects of neurotypical development (see Supplementary Results for detailed biological characterization and validation of BrainSpan data and modules; Tables S3–S5; Figures. S1–S11).

SCZ and ASD risk variant enrichment for BrainSpan modules is summarized in Figure 1. SCZ-associated variants converged on 3 modules. SCZ-associated common variants and CNVs were significantly over-represented in M15 (FDR-adjusted p<0.05, corrected for number of modules), a neuronal marker-enriched module that increased in expression through adolescence and was enriched for postnatal-expressed genes (Figures 2A–C, S3O). GO analysis indicated that M15 was involved in synaptic transmission, modulation of neurotransmission, and learning and memory. SCZ-associated common variants were additionally over-represented in M13, a neuronal module that increased in expression during fetal development, with relatively stable postnatal expression (Figure 2D–F). Consistent with this, M13 was enriched for genes expressed during late fetal and postnatal development (Figure S3M). M13 was enriched for synaptic transmission and regulation of the membrane potential, and its hub genes included ATP1B1 and KCNMA1, which encode components of the Na+/K+ pump and a voltage-gated K+ channel, respectively, that regulate the electrochemical gradient(45,46). Thus, M13 appears to be involved in regulating basic neuronal excitability. Finally, SCZ-associated common variants were significantly over-represented in M7, with rare variants nominally enriched. M7 was involved in RNA processing and binding, and its hub genes included POLR2B, which encodes a major subunit of RNA polymerase II, the central enzyme responsible for transcribing mRNA, and CREB1, which encodes the transcription factor CREB, known to regulate neuronal differentiation, survival, and plasticity(47; Figure 2G–I). M7 showed its highest expression during fetal development but was minimally enriched for genes associated with specific stages of development (Figure S3G). Thus, M7 appears to regulate gene expression and RNA processing across development.

Figure 1. Genetic Risk for SCZ and ASD Converge on Partially Overlapping and Partially Distinct Neurodevelopmental Modules.

BrainSpan module enrichment for SCZ (red) and ASD (blue) risk variant datasets, 22q11.2 genes (green), and the 22q11.2 brain-development PPI network (22q_BD-PPI network; here shown tested for module enrichment using Fisher’s Exact tests). Log p values are indicated by color intensity for all datasets with enrichment odds ratios >1.0. To show potential convergence of enrichment across datasets, all enrichment values for p<0.05 are annotated (+p<0.05; *FDR corrected p<0.05 for number of modules). Numerical values denote number of 22q11.2 genes in each module. ^Module enriched for neuronal cell-type markers.

Figure 2. Characterization of BrainSpan Modules Associated with Both SCZ and ASD.

Left: top 25 hub genes, Middle: developmental expression trajectory (median ± median absolute deviation of module eigengene in grey shading), and Right: top 5 significantly enriched biological process and molecular function gene ontology (GO) terms for the SCZ- and ASD- associated modules: (A-C) M15, (D-F) M13, and (G-I) M7. Hub genes and GO terms are colored by module. Edges between hub genes denote positive correlations between genes.

Risk variants for ASD showed overlapping enrichment for SCZ-associated modules, as well as enrichment for 4 distinct modules (Figure 1). Thus, loss-of-function (LoF) DNMs in ASD were significantly over-represented in M7; protein-altering DNMs in ASD were over-represented in M13; and LoF DNMs and ASD-associated CNVs (many of which are also associated with SCZ; Table S4) were over-represented in M15.

The additional modules significantly associated with ASD were M1, M4, M6, and M12. Thus, rare variants, broadly, and LoF and protein-altering DNMs in ASD were consistently over-represented in M1. M1 peaked in expression during fetal development and was highly enriched for genes with fetal-specific expression (Figures. 3A–C, S3A). M1 was involved in transcription, chromatin modification, and gene expression, and its hub genes included the transcription factors SOX11 and SOX4, which are critical for promoting neuronal differentiation(48). Thus, M1 appears to be important for regulating fetal gene expression patterns that direct neuronal differentiation. LoF DNMs and SFARI-defined ASD genes were additionally over-represented in M4. M4 was enriched for neuronal markers, peaked in expression during mid-fetal development, and was involved in neuron differentiation and projection development (Figure 3D–F). Consistent with this, M4 was enriched for genes with fetal-specific expression (Figure S3D), and its hub genes included DOK4, ARHGAP33, and TRIO, which are involved in axon and dendrite morphogenesis(49–51). LoF DNMs in ASD were also over-represented in M6, which showed maximal expression during fetal development but was not enriched for genes expressed in any specific developmental stage (Figure S3F). M6 was involved in translation and protein catabolism (Figure 3G–I). Finally, LoF DNMs in ASD were over-represented in M12, a neuronal module that increased in expression until plateauing in late fetal development, and was involved in regulating the membrane potential (Figures 3J–L, S3L). Interestingly, ASD-associated common variants were nominally over-represented across multiple modules significantly enriched for rare and/or de novo variants in ASD (M1, M4, M7, M15). Module enrichment for gene-sets implicated in prior genetic studies of SCZ or ASD were consistent with GO analyses (Figure S1C). See Figure S4 for summary of non-SCZ or ASD-associated modules.

Figure 3. Characterization of BrainSpan Modules Associated Only with ASD.

Left: top 25 hub genes, Middle: developmental expression trajectory (median ± median absolute deviation of module eigengene in grey shading), and Right: top 5 significantly enriched biological process and molecular function gene ontology (GO) terms for the ASD-associated modules: (A-C) M1, (D-F) M4, (G-I) M6, and (J-L) M12. Hub genes and GO terms are colored by module. Edges between hub genes denote positive correlations between genes.

As expected, LoF and protein-altering DNMs in control subjects were not over-represented in any SCZ- or ASD-associated modules, and no modules were significantly enriched for synonymous DNMs in ASD, SCZ, or controls, ps>0.05.

Thus, polygenic risk for SCZ and ASD each appear to coalesce on neurodevelopmental networks involved in postnatal neuronal excitability and synaptic transmission, and pan-developmental regulation of gene expression; however, risk variants for ASD additionally converge on networks involved in early neuronal differentiation and synapse formation.

The 22q_BD-PPI Network Loads Highly onto SCZ- and ASD-Associated Modules

Forty-six protein-coding genes are contained in the ~2.57 Mb 22q11.2 locus, 38 of which were indexed in BrainSpan (Table 2; see Figure S12 for expression trajectory of each gene). Given the number of genes in each module, there were nominally more 22q11.2 genes than expected in M14 (i.e., 1 gene; SEPT5; p=0.02; Figure 1), although this did not survive FDR correction. Similarly, there were not significantly more 22q11.2 genes across all SCZ- and ASD-associated modules than expected, p>0.05. Relative to all regions of the genome, the protein-coding gene content of the overall 22q11.2 locus was in the 93.6 percentile, and among individual BrainSpan modules, 22q11.2 gene content was enriched for 3 modules, including the ASD-associated M1 module and the SCZ- and ASD-associated M15 module (P_{GENOMIC-REGION}<.05; Table S6). However, the genomic-region ranks of the 22q11.2 locus were not significantly higher across all SCZ- and ASD-associated compared to non-associated modules, p>0.05.

Table 2. BrainSpan Module Membership of 22q11.2 Genes.

The 38 22q11.2 genes indexed in BrainSpan are annotated by module assignment, Kme score (i.e., correlation with parent module eigengene) and pLI score (i.e., loss-of-function function intolerance; bolded text indicates predicted pLI gene).

Module	Gene	Definition	kME Score	pLI Score
M1	ZNF74	zinc finger protein 74	0.84	0.0004
	DGCR8	DGCR8, microprocessor complex subunit	0.76	0.9999
	SLC25A1	solute carrier family 25 member 1	0.75	0.6335
	ARVCF	armadillo repeat gene deleted in velocardiofacial syndrome	0.67	0.0000
	CLTCL1	clathrin heavy chain like 1	0.61	0.0000
	KLHL22	kelch like family member 22	0.56	0.7026
	TMEM191A	transmembrane protein 191A	0.18	NA
	SCARF2	scavenger receptor class F member 2	0.01	NA
M2	CRKL	v-crk avian sarcoma virus CT10 oncogene homolog-like	0.51	0.1635
M3	CDC45	cell division cycle 45	0.78	0.0691
M6	MRPL40	mitochondrial ribosomal protein L40	0.78	0.0201
	ESS2	DiGeorge syndrome critical region gene 14	0.75	0.0000
M7	UFD1	ubiquitin recognition factor in ER associated degradation 1	0.80	0.9974
	SNAP29	synaptosome associated protein 29	0.75	0.1517
	HIRA	histone cell cycle regulator	0.64	1.0000
	RTL10	retrotransposon Gag like 10	0.55	0.0000
	TRMT2A	tRNA methyltransferase 2 homolog A	0.35	0.0000
	THAP7	THAP domain containing 7	0.21	0.0293
	DGCR2	DiGeorge syndrome critical region gene 2	0.17	0.0011
M10	AIFM3	apoptosis-inducing factor, mitochondrionassociated, 3	0.94	0.0000
	P2RX6	purinergic receptor P2X, ligand-gated ion channel, 6	0.80	0.0000
M11	PRODH	proline dehydrogenase (oxidase) 1	0.86	0.0000
	ZDHHC8	zinc finger, DHHC-type containing 8	0.76	0.9890
	COMT	catechol-O-methyltransferase	0.63	0.0008
	TXNRD2	thioredoxin reductase 2	0.59	0.0000
	SERPIND1	serpin peptidase inhibitor, clade D (heparin cofactor), member 1	0.42	0.0000
	C22orf39	chromosome 22 open reading frame 39	0.24	0.0000
	DGCR6L	DiGeorge syndrome critical region gene 6-like	0.17	0.1286
	GSC2	goosecoid homeobox 2	0.15	0.0372
M14^	SEPT5	septin 5	0.80	0.9146
M15^	PI4KA	phosphatidylinositol 4-kinase alpha	0.88	0.0008
	TANGO2	transport and golgi organization 2 homolog	0.82	0.0000
	SLC7A4	solute carrier family 7 member 4	0.72	0.0001
	RTN4R	reticulon 4 receptor	0.57	0.5264
	MED15	mediator complex subunit 15	0.38	0.9610
	LZTR1	leucine zipper like transcription regulator 1	0.25	0.0000
M18	CLDN5	claudin 5	0.43	0.7405
	TSSK2	testis-specific serine kinase 2	0.38	0.0042

^{^}Module enriched for neuronal cell-type markers.

We next characterized the 22q11.2 brain developmental PPI network (22q_BD-PPI network), given that perturbation of protein networks is a common characteristic of mutations associated with many diseases(40). The 22q11.2 locus formed a large BD-PPI network, consisting of 239 genes from 38 seed genes (Figure S13). Given the number of genes in each module, the 22q_BD-PPI network was significantly over-represented in the SCZ- and/or ASD-associated modules M6, M7, and M15, as well as M2 and M3 (FDR p<0.05; Figure 1).

Some BrainSpan modules consist of genes that are more highly connected than other modules (ANOVA p<2.0×10⁻¹⁶; Table S7), and highly connected genes are more likely to be found in the BD-PPI network of any random set of genes. This may be biologically meaningful, with gene networks that are highly connected during brain development potentially more susceptible to perturbation. Nevertheless, to better understand the 22q_BD-PPI network, we next characterized the connectivity of the 22q_BD-PPI network relative to all regions of the genome, and to random gene-sets derived from an equal number of seed genes.

Relative to all genomic regions, the 22q_BD-PPI network was significantly enriched for 5 modules, (P_{GENOMIC-REGION NETWORK}<0.05), including the SCZ- and/or ASD-associated neuronal modules, M4, M13, and M15 (Figure 4A, Table S6). Further, the 22q_BD-PPI network ranks relative to all Genomic-Region_BD-PPI networks were significantly higher for SCZ- and/or ASD-associated modules, overall, compared to non-SCZ- or ASD-associated modules, p=0.03.

Figure 4. Characterizing the 22q_BD-PPI Network Relative to Genomic-Region and Gene-Set BD-PPI Networks.

Null distributions for number of BD-PPI network genes per module across all: (A) ~2.57 Mb genomic regions (i.e., 12,149 Genomic-Region_BD-PPI networks) and (B) random 38 seed gene gene-sets (i.e., 10,000 Gene-Set_BD-PPI networks), derived to characterize the 22q11.2 locus relative to all regions of the genome, and the connectivity of 22q11.2 genes within brain development protein networks, respectively. Boxplots mark the median number of BD-PPI network genes per module, with lower and upper hinges corresponding to the 25^th and 75^th quartiles, respectively, and the whiskers extending to 1.5 x interquartile range. The number of 22q_BD-PPI network genes in each module are marked in red. *22q_BD-PPI network P<0.05 relative to null Genomic-Region_BD-PPI or Gene-Set_BD-PPI network distribution.

Relative to random sets of 38 seed genes, the 22q_BD-PPI network remained significantly enriched for the SCZ- and ASD-associated M13 module, as well as the non-associated M2 module (P_{GENE-SET NETWORK}<0.05; Figure 4B; Table S6). The 22q_BD-PPI network ranks relative to Gene-Set_BD-PPI networks were also higher overall for SCZ- and/or ASD-associated modules than for non-associated modules, p=0.03.

This suggests that 22q11.2 CNVs confer risk for SCZ and ASD by not only spanning a genomic region that is rich in protein-coding genes, but by spanning genes that form protein networks in the developing brain that are highly and disproportionately connected to broader networks involved in SCZ and ASD pathogenesis.

SEPT5, PI4KA, and SNAP29 are Candidate Drivers of Synaptic Pathology, and DGCR8 and HIRA are Candidate Drivers of Dysregulated Gene Expression

To identify specific genes driving 22q_BD-PPI network loading onto SCZ- and ASD-associated modules, we next examined the rank of each 22q11.2 gene’s BD-PPI network per module, relative to all BrainSpan genes. The SEPT5 and PI4KA BD-PPI networks were enriched for 3 neuronal modules associated with SCZ and/or ASD: M4, M13, and M15 (P_{GENE NETWORK}<0.05; Figure 5); PI4KA was additionally enriched for the ASD-associated M12 module. SNAP29’s BD-PPI network was also enriched for M13 and M15, as well as M7 (P_{GENE NETWORK}<0.05). Indeed, the individual SEPT5, PI4KA, and SNAP29 networks were each enriched for cell parts and/or biological processes related to synaptic function (Figure 6A–F). HIRA’s BD-PPI network was also enriched for M7 (Figure 5D), and was enriched for genes involved in chromatin organization and gene expression (Figure 6G–H). Additionally, DGCR8’s BD-PPI network was enriched for M1 (Figure 5A), and was enriched for RNA binding and processing (Figure 6I–J). Finally, MRPL40 and ESS2’s BD-PPI networks were enriched for M6 (Figure 5C); the ESS2 network was not enriched for any GO terms, while MRPL40’s network was enriched for ribosomal components and mitochondrial elongation (Figure 6K–M).

Figure 5. Prioritizing Individual 22q11.2 Genes Based on BD-PPI Network Overlap with Each SCZ- and/or ASD-Associated BrainSpan Module.

Number of genes in each individual 22q11.2 gene’s BD-PPI network belonging to the: (A) M1, (B) M4, (C) M6, (D) M7, (E) M12, (F) M13, and (G) M15 SCZ- and/or ASD-associated modules. In each plot, 22q11.2 genes are shown grouped by module membership along the x-axis. The top, second, and third red dotted lines in each plot mark the number of BD-PPI network genes nearest to the 99^th, 95^th, and 75^th percentile (when above 0) of all BrainSpan genes for each module. Bar plots are colored by module color. *P_{GENE NETWORK}<0.05 and **P_{GENE NETWORK}<0.01 for each module, relative to the BD-PPI networks of all BrainSpan genes. ^Module enriched for neuronal cell-type markers.

Figure 6. Characterization of BD-PPI Networks for 22q11.2 Genes Prioritized to Drive Risk for SCZ and/or ASD.

Left: BD-PPI network genes and Right: significantly enriched GO terms (up to top 5 terms, where applicable) for the BD-PPI networks of: (A-B) SEPT5, (C-D) PI4KA, (E-F) SNAP29, (G-H) HIRA, (I-J) DGCR8, (K-L) MRPL40, and (M) ESS2. Individual gene nodes in each network are colored by BrainSpan module assignment. GO plots are colored by module membership of the 22q11.2 gene of interest.

Support for Leveraging BD-PPI Networks to Understand 22q11.2 CNV-Mediated Risk for SCZ and ASD

Investigation of gene expression signatures in SCZ and ASD post-mortem cortex(43) indicated that nearly all BrainSpan modules associated with genetic risk for SCZ and ASD were enriched for down-regulated genes in each disorder (Figure S14A); the 22q_BD-PPI network was also enriched for down-regulated genes in SCZ and ASD, given the size of the 22q_BD-PPI network and each gene expression gene-set (ps<0.05). Relative to all BrainSpan genes, the SEPT5 and PI4KA BD-PPI networks ranked highly for overlap with down-regulated genes in both SCZ and ASD (P_{GENE NETWORK}<0.05), with the SNAP29 and KLHL22 BD-PPI networks additionally enriched for overlap with down-regulated genes in ASD (Figure S15).

Exploratory analyses of broader CNVs and genes associated with SCZ and ASD confirmed that high BD-PPI network loading onto SCZ- and ASD-associated BrainSpan modules and down-regulated genes in SCZ or ASD cortex is a common characteristic of disease-associated CNVs (Figure S16) and genes (Figure S17). Similarly, using more inclusive co-expression thresholds to define the BD-PPI networks of each gene highlighted similar 22q11.2 genes as likely drivers of SCZ and ASD risk (Figures S18, S19). Further, using a machine learning approach to classify SCZ- and ASD-associated genes based on BD-PPI network ranks prioritized similar 22q11.2 genes (Tables S2, S12). Finally, study bias could not account for our prioritization of 22q11.2 genes likely to drive disease risk (Figure S20). Together, these results support the utility of BD-PPI networks for identifying neurodevelopmental mechanisms and specific genes through which multi-gene CNVs confer risk for SCZ and ASD. See Supplementary Results for details.

Discussion

Recent genetic discoveries in SCZ and ASD have brought renewed focus to understanding the mechanisms through which genetic variation confers risk for these disorders at both the polygenic and individual variant level. Here, using a developmental functional genomics approach, we identified overlapping and distinct neurodevelopmental processes associated with SCZ and ASD risk, and then leveraged this disease-informed lens to elucidate mechanisms underlying disease risk in highly penetrant multi-gene CNVs. Our analyses confirmed the capacity of 22q11.2 CNVs to disrupt disease-relevant neurodevelopmental modules and highlighted both well-studied (e.g., DGCR8 and HIRA) and relatively understudied genes (e.g., SEPT5 and PI4KA) as candidate 22q11.2 genes likely to drive disease risk.

Prior studies using transcriptomic data from the developing brain to understand how risk variants contribute to disease largely focused on rare variants in ASD. These studies suggested that ASD risk variants preferentially disrupt neuronal development and transcriptional and chromatin regulation during fetal development(16,17,35). One early study of DNMs in SCZ also implicated disruptions in neurogenesis and transcriptional regulation(26); however, this was a relatively small study. Here, by creating co-expression networks from brain samples spanning the full period of development and testing an expanded set of risk variants for SCZ and ASD simultaneously, we confirmed prior enrichment patterns in ASD and identified co-expression modules with distinct enrichment patterns for ASD versus SCZ. We found robust enrichment in SCZ for a neuronal module that increased in expression across development and was involved broadly in postnatal synaptic transmission and refinement of neurotransmission; a neuronal module involved in regulating neuronal excitability that stabilized in expression around birth; and a module involved in regulating transcription and splicing across development. ASD risk variants showed overlapping enrichment for these modules, as well as robust enrichment for distinct modules involved in neuronal differentiation, and in regulating chromatin organization and gene transcription during fetal development. While our findings thus confirm prior findings implicating neuronal and synaptic genes in both SCZ and ASD(7,52), risk variants for ASD appear to partially diverge from those for SCZ by loading highly onto gene networks that orchestrate neuronal differentiation during fetal development. It is tempting to suggest that the higher loading of ASD risk variants onto early neuronal differentiation networks may be a key mechanism underlying the earlier onset of ASD relative to SCZ; however, experimental validation is necessary to confirm this.

By integrating BrainSpan co-expression and PPI data, we identified protein networks likely to be strongly affected by 22q11.2 CNVs during brain development. One prior study focused on 16p11.2 CNVs integrated PPI and BrainSpan data and found that ‘neuropsychiatric’ CNV networks were enriched for varying developmental periods, with the overall 22q11.2 network showing particular enrichment during childhood(41). Here, using a disease-informed framework, we focused on understanding how 22q11.2 CNVs confer risk for SCZ and ASD, more specifically. We found that 22q11.2 genes formed a large BD-PPI network that: 1) loaded highly onto multiple SCZ- and ASD-associated modules, relative to all regions of the genome; 2) was more highly connected to genes in SCZ- and/or ASD-associated modules than genes in non-disease associated modules; and 3) overlapped with genes that are down-regulated in SCZ and ASD cortex, with overlapping genes involved in neuronal and synaptic function. Thus, 22q11.2 CNVs may confer risk for SCZ and ASD, at least partially, by perturbing synaptic networks that modulate neuronal excitability and signaling. Consistent with this, 22q11DS mouse models show altered neuronal excitability, activity, and dendritic spine stability(53,54).

A primary advantage of our approach was our ability to systematically survey 22q11.2 genes to prioritize genes likely to contribute disease risk. Our analyses consistently highlighted SEPT5 and PI4KA as candidate drivers of synaptic pathology relevant to SCZ and ASD, and DGCR8 as a candidate driver of dysregulated gene expression that may be most relevant to ASD. SNAP29 and HIRA were also prioritized as contributing to SCZ and ASD. Notably, both SEPT5 and SNAP29 have been found to negatively modulate neurotransmitter release. SEPT5 encodes the filamentous protein septin 5, which has been found to associate with synaptic vesicles to create a physical barrier that inhibits vesicle fusion with the membrane, and thereby, inhibits neurotransmitter release(55,56). Similarly, SNAP29 encodes a SNARE protein that competes with α-SNAP for binding to SNARE complexes after membrane fusion, thus reducing SNARE protein recycling and synaptic vesicle availability(57). Conversely, altered PI4KA expression may contribute to disease risk by perturbing the proteomic composition of the plasma membrane (PM). PI4KA encodes the kinase PI4KIIIα, which is involved in synthesizing the phosphoinositide, phosphatidylinositol 4,5-biphosphate (PI(4,5)P₂), that recruits and regulates PM proteins and thereby governs diverse processes including vesicular trafficking, ion transport, and actin cytoskeletal dynamics(58). PI4KA depletion was found to disrupt PI(4,5)P₂-associated protein localization, cellular responses to stimulation(59,60), and synapse maintenance(61). Our prioritization of PI4KA is consistent with prior associations of PI4KA polymorphisms in SCZ(62,63,64), and with growing evidence that PI4KA may play an important role in synaptic function(60,61).

DGCR8 is a core component of the microprocessor complex involved in processing miRNAs, which regulates gene expression at the protein level by binding target mRNAs to silence their translation(65). HIRA is a histone chaperone that influences gene expression by facilitating the deposition of the non-canonical histone variant, H3.3, into chromatin(66,67). Given these gene regulatory functions, DGCR8 and HIRA have been previously hypothesized to play important roles in 22q11.2 CNV phenotypes(67,68). Here, using our disease-informed functional genomic lens, we provide independent, unbiased evidence for prioritizing DGCR8 and HIRA in 22q11.2 CNV-mediated risk for neurodevelopmental disorders.

Some caveats to the current study should be noted. First, in our analyses testing SCZ and ASD risk variants for module enrichment, some studies shared overlapping samples. As the genetic architectures of SCZ and ASD include many variants with small effects, collaborative cross-study analyses are increasingly common to facilitate power. We utilized independent datasets where possible, and focused on modules showing consistent disease association. Second, generating BD-PPI networks identifies protein networks putatively affected by 22q11.2 CNVs, but cannot tell us the functional output of decreased versus increased expression of each gene. We therefore cannot determine how deletions versus duplications affect these networks, nor quantify the full consequences of mutations in genes whose actions involve regulating the expression of other genes. Similarly, while using individual gene BD-PPI network rankings is useful for prioritizing 22q11.2 genes predicted to be highly connected within disease-associated modules, experimental validation is necessary to confirm the contribution of specific 22q11.2 genes to disease-relevant phenotypes (e.g., see 1,69). As high-throughput functional assays become increasingly feasible, quantifying the transcriptomic and proteomic consequences of altered expression of each 22q11.2 gene will likely yield further insights into 22q11.2 CNV-mediated disease risk. Nevertheless, identifying convergent neurodevelopmental networks affected by polygenic risk for SCZ and ASD, and systematically prioritizing 22q11.2 genes that load highly onto these networks, offers a critical step towards understanding the mechanisms through which 22q11.2 CNVs confer risk for SCZ and ASD.

KEY RESOURCES TABLE.

Resource Type	Specific Reagent or Resource	Source or Reference	Identifiers	Additional Information
Add additional rows as needed for each resource type	Include species and sex when applicable.	Include name of manufacturer, company, repository, individual, or research lab. Include PMID or DOI for references; use “this paper” if new.	Include catalog numbers, stock numbers, database IDs or accession numbers, and/or RRIDs. RRIDs are highly encouraged; search for RRIDs at https://scicrunch.org/resources.	Include any additional information or notes if necessary.
Deposited Data; Public Database	BrainSpan microarray data	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE25219	RRID:SCR_004219	NA
Deposited Data; Public Database	dbVar CNV database	https://www.ncbi.nlm.nih.gov/dbvar/	RRID:SCR_003219	NA
Deposited Data; Public Database	BioGrid PPI 3.4 database	https://downloads.thebioarid.ora/BioGRID	RRID:SCR_007393	NA
Deposited Data; Public Database	InWeb3 PPI database	https://doi.org/10.1371/journal.pgen.1001273	NA	NA
Software; Algorithm	g:Profiler (rev 1741, built 2017-10-19)	https://biit.cs.ut.ee/gprofiler/gost	RRID:SCR_006809)	NA
Software; Algorithm	Specific Expression Analysis Tool	http://genetics.wustl.edu/jdlab/csea-tool-2/	NA	NA
Software; Algorithm	Ensembl Variant Effect Predictor version 90	http://grch37.ensembl.org/Homo_sapiens/Tools/VEP	RRID:SCR_002344	NA