The genomic landscape of familial glioma

Abstract

Glioma is a rare brain tumor with a poor prognosis. Familial glioma is a subset of glioma with a strong genetic predisposition that accounts for approximately 5% of glioma cases. We performed whole-genome sequencing on an exploratory cohort of 203 individuals from 189 families with a history of familial glioma and an additional validation cohort of 122 individuals from 115 families. We found significant enrichment of rare deleterious variants of seven genes in both cohorts, and the most significantly enriched gene was HERC2 (P = 0.0006). Furthermore, we identified rare noncoding variants in both cohorts that were predicted to affect transcription factor binding sites or cause cryptic splicing. Last, we selected a subset of discovered genes for validation by CRISPR knockdown screening and found that DMBT1, HP1BP3, and ZCH7B3 have profound impacts on proliferation. This study performs comprehensive surveillance of the genomic landscape of familial glioma.

Genes are identified that predispose to familial glioma, a rare, genetic and deadly disease.

INTRODUCTION

Glioma is the most common type of malignant brain tumor, although it is still considered rare, with an incidence of 6.0 per 100,000 individuals. Despite being rare, glioma causes high morbidity and mortality, with a median survival of 14 months (1–4). Approximately 5% of glioma cases are thought to be familial (5, 6), which is typically defined by multiple affected individuals within a family who usually have early disease onset (<50 years) (7–9). The proportion of tumors that result from germline predisposition has been reported to vary from 1% (e.g., stomach tumors) to more than 50% (e.g., thyroid tumors) (10–12). Predisposition to cancer is typically inherited or caused by loss-of-function (LoF) variants in tumor suppressor genes with modest to high penetrance (25 to 90%; e.g., BRCA1 and TP53) (13, 14). Recently, genome-wide sequencing has helped identify previously undescribed cancer predisposition genes (CPGs) (15, 16).

Case-control studies of glioma show that having a first-degree relative with glioma increases the risk of glioma by twofold (17). Several large genome-wide association studies (GWAS) on the risk of inherited glioma have identified approximately 25 loci that are associated with a modest increase in glioma risk (odds ratio, 1.1 to 3.5) (18). Functional, high-penetrance variants in several genes have also been associated with glioma, including NF1/2, TSC1/2, TP53, Lynch syndrome genes, CDKN2A, and CHD1 (16–21). Previously, we performed exome sequencing of 90 individuals from 55 families and identified POT1 as a previously unidentified glioma predisposition gene (15, 22); however, many cases of familial glioma remain unexplained. Whole-genome sequencing (WGS) is an increasingly inexpensive method to screen for rare variants across the genome (23) and provides robust coverage of coding regions. It also allows for discovery of structural and noncoding variants, the latter of which are gaining appreciation for their functional roles in inherited disease. In this study, we used WGS to examine the genomes of two cohorts with a family history of glioma. To identify putative novel glioma predisposition genes, we compared the results with those from a control cohort consisting of unselected individuals whose genomes were sequenced on the same platform. Coding and noncoding variants were examined for associations with both known and potentially novel CPGs.

RESULTS

We performed WGS on 325 individuals with glioma from 304 families with a history of familial glioma (table S1). The vast majority of participants self-reported their race as non-Hispanic white. The participants were divided into exploratory and validation cohorts consisting of 203 subjects (189 families) and 122 subjects (115 families), respectively. Analysis of single-nucleotide variants (SNVs), indels, and larger copy number variants (CNVs) identified approximately 1 billion germline variants across both cohorts. Variants were filtered and prioritized as outlined in Methods and in Fig. 1. Concomitantly, we performed WGS of 1013 non-Hispanic white participants from the United States who participated in the Atherosclerosis Risk in Communities (ARIC) study (24). These individuals were 45 to 65 years of age at the time of recruitment and were randomly selected from the U.S. Census data (see Methods).

Fig. 1. Study analysis pipeline.

Variants were classified as CNVs and smaller events. SNVs and indels were selected according to rarity or presence using the ClinVar database. The resulting variants were then divided into four groups: highly deleterious variants, intronic cryptic splicing variants, variants in TFBSs, and known glioma-associated GWAS SNPs (see Methods). Affected genes were stratified on the basis of whether they were known or novel CPGs. Genes with identified variants are shown below each category. Discovered genes were then validated in a second cohort (bold), and a subset of genes was examined via in vivo screening (underlined). B, billion; SNV, single-nucleotide variant; MAF, minor allele frequency in gnomAD overall; K, thousand; P/LP, pathogenic or likely pathogenic; LoF, loss of function; TFBS, transcription factor binding site; GWAS, genome-wide association study; CNV, copy number variant; CPG, cancer predisposition gene.

Coding variants in known and novel CPGs

Glioma susceptibility has been associated with several known CPGs. Therefore, we examined whether our familial cohort had LoF or known deleterious variants in any of 32 commonly tested CPGs (table S2). We identified significant (P = 0.0071, Fisher’s exact test) enrichment of LoF and likely pathogenic/pathogenic variants in known CPGs in our cohort (12 of 189) compared with the control group (23 of 1013) (Fig. 2 and table S3) (25). Heterozygous BRIP1 and PMS2 mutations were the most common mutations in our cohort, accounting for three samples each, and POLE was mutated in two samples. In contrast, none of the control cases had BRIP1 or POLE variants, and only one sample had a pathogenic PMS2 variant (P = 0.004, P = 0.03, and P = 0.06, respectively, Fisher’s exact test). In the validation cohort, we identified additional pathogenic mutations in ATM, BRCA1, and CDKN2A, the latter of which was shared by two affected family members (table S3).

Fig. 2. Proportion of the glioma exploratory cohort (green) or control cohort (blue) with variants in seven cancer genes and overall.

P values were calculated with Fisher’s exact test and Bonferroni correction.

Next, we identified genes with a novel association with familial glioma. We filtered variants in the case and control groups according to rarity [minor allele frequency (MAF) < 0.0001], LoF, and deleteriousness [Combined Annotation Dependent Depletion phred-like (CADD PH) score ≥ 30] and identified genes that were significantly enriched for these variants. We only considered genes with two or more variants in our cohort. After correction for multiple tests, we found 19 genes that were significantly enriched, with a false discovery rate of <0.1 (Table 1). Of these, HERC2 was the most significantly enriched, with four variants in our exploratory cohort and none in the control group (P = 0.0006, Fisher’s exact test). In addition, six of these genes contained at least one mutation in our validation cohort (bold).

Table 1. Genes enriched for variants in both cohorts.

Gene, number of variants in the exploratory and validation cohorts (comma separated) and the control cohort set, P value of enrichment (Fisher’s exact test) of variants, and families bearing the variant. Genes with variants in the validation cohort are shown in bold; variants validated by in vivo testing are underlined.

Gene	Families in cohort	Control counts	P value	Variants	Family
HERC2	4, 1	0	0.0006	p.R4293Q, p.F3704L, p.F3642L, p.D523N, p.L2384fs	FG101, FG118, FG136, FG140, VC23
TRPC4AP	3	0	0.0037	p.Q479X, p.R637Q, p.R307Q	FG104, FG121, FG139
KIAA1549	3	1	0.013	c.5248-2A>G, p.R1275X, p.V1469M	FG102, FG119, FG137
MYO7A	3, 2	1	0.013	p.R740Q, p.R830H, p.R63Q, p.E1579K, p.E602K	FG103, FG120, FG113, VC24, VC34
PIK3R4	2	0	0.024	p.V478I,p.P460L	FG105, FG122
ZC3H7B	2, 0	0	0.024	p.K854R, p.R923H	FG110, FG129
PCNXL4	2	0	0.024	p.P363L, p.A906V	FG106, FG123
THSD7A	2	0	0.024	p.E284X, p.R1253Q	FG107, FG124
PCYT1A	2	0	0.024	p.K186E, p.A99V	FG108, FG125
IP6K1	2, 1	0	0.024	p.R139Q, p.P71L, p.K53R (cryptic splice)	FG109, FG126, VC97
ADAMTS8	2, 1	0	0.024	p.R551H, p.S406C, p.M1I	FG110, FG127, VC54
TRPM1	2	0	0.024	p.L1021F, p.D82N	FG111, FG128
LRRK2	2, 1	0	0.024	p.R1334Q, p.F2059L, p.E2508X	FG112, FG129, VC84
FLOT2	2	0	0.024	p.A332V, p.A325V	FG113, FG130
PPP1R16B	2	0	0.024	p.R24Q, p.E175K	FG114, FG112
CTNND1	2	0	0.024	p.E73K, p.R653H	FG115, FG132
IFIH1	2	0	0.024	p.V453L, p.P330S	FG109, FG133
MAST4	2, 1	0	0.024	p.A1190S, p.P1439L, p.G1519S	FG116, FG134, VC17
TFAP2E	2	0	0.024	p.S229T, p.G367R	FG117, FG135

To further assess the role of HERC2 in gliomagenesis, we examined the Genomics England dataset and filtered all cases of glioma for rare (1 of 10,000 MAF) and deleterious (CADD score ≥ 30) variants in HERC2. We identified two additional HERC2 variants in the Genomics England dataset (p.R4587Q and p.S4607C) in two individuals, one male who was in his 60s at the time of glioma onset and one female with an unknown age of onset. At the time of sampling (6 June 2020), there were 611 cases of glioma in the Genomics England Clinical Interpretation Partnership (GeCIP) database. The glioma cohort in the GeCIP database is not focused on familial glioma but instead agnostically includes both familial and sporadic cases. On the basis of the frequency of familial glioma (approximately 5% of all gliomas), we would expect approximately 30 of 611 samples to originate from a familial predisposition to glioma. Under this assumption, we found significant enrichment of HERC2 mutations in this cohort (P = 0.0009, Fisher’s exact test), thereby confirming an association between HERC2 mutation and familial glioma in a second independent dataset. As a caveat, when considering the incidence of HERC2 variants across all glioma cases in the GeCIP database, enrichment of HERC2 mutation was not significant (P = 0.14), further emphasizing the role of HERC2 mutation in the incidence of familial, rather than sporadic, glioma.

Next, we sought to understand the role of CNVs in glioma susceptibility. We identified CNVs in known CPGs and large [>100 kilo–base pairs (kbp)] copy number losses across all genes (Fig. 3 and table S4). We identified deletions in three known cancer genes: a heterozygous deletion of four exons in CHEK2, a homozygous complex deletion in DMBT1, and a deletion of the terminal coding exon in ATM. Each deletion was specific to a family. We also identified internal tandem duplications in POT1 (exons 4 and 5) and HERC2 (exons 3 to 65), which are predicted to cause a frameshift insertion, leading to nonsense-mediated decay, or may otherwise be pathogenic and cause haploinsufficiency, bringing the total number of deleterious variants found in HERC2 to six (Fig. 4) (26, 27). We also identified four large deletions that did not involve known cancer susceptibility genes (table S4). No genes with deletions were shared across families, but for each deletion, at least one gene was both LoF intolerant and involved in cell growth or differentiation (KAT2B, DIP2A, PRKCA, and FOXJ2). Of these, FOXJ2 has been suggested to act as a tumor suppressor in glioma cells and other cancers (28) and lowered expression is associated with worse outcomes in glioma (P = 0.011) (29). Low expression of DIP2A is associated with decreased temozolomide sensitivity, but the role of DIP2A and PRKCA in tumor initiation is not clear and may be incidental. Last, KAT2B is a histone acetyl transferase putatively involved in many cellular processes, and lower expression is also correlated with worse outcomes (P = 0.03) (30). No large structural variants were identified in the validation cohort.

Fig. 3. CNVs identified in the glioma exploratory cohort.

The relative coverage of HERC2 (A), ATM (B), CHEK2 (C), DMBT1 (D), and POT1 (E) is shown in green; the average coverage is shown by a dashed black line. Exons (blue lines) and introns (thin lines) are shown below the coverage plots.

Fig. 4. Schematic of variants identified in HERC2.

Copy gain (blue bar) and variants (arrows: red, orange, and gray indicate CADD PH scores of 35, ≥30 but <35, and frameshift mutations, respectively) are shown relative to the regulator of chromosome condensation protein domains (green bar). Approximately 99% of all variants identified by WGS were noncoding; however, the functional consequences and MAF distributions of these variants are not as well understood as those of coding variants. To avoid numerous spurious associations, we restricted our analysis to deep-intronic (i.e., >5 bp from the exon) variants that were predicted to cause cryptic splicing and variants in the immediate upstream region of genes that overlapped a TFBS. We identified 16 deep-intronic variants in LoF-intolerant and novel CPGs that were predicted to cause cryptic splicing in our exploratory cohort (see table S5). Identical variants in two genes, NCAM1 and SMG6, were identified in the validation cohort.

In our exploratory cohort, we identified 17 rare variants that likely affect gene expression by altering transcription factor binding sites (TFBSs) (see Table 2). In five cases, these variants affected four known CPGs (CDK4, AXIN2, STK11, and MSH2), with MSH2 variants occurring in two families. We identified a TFBS variant upstream of HP1BP3 in one family with a notable history. However, TFBS variants were not considered significantly enriched compared with those in our control cohort. We also identified three genes (WDR7, DDN, and PC) that had TFBS variants in three or more families in our cohort and no variants in the control cohort (P = 0.011, Fisher’s exact test, Bonferroni-corrected). One family, FG110, had two TFBS variants, and FG128 had a CNV in POT1. We also identified upstream variants in the previously identified TRPC4AP (Table 1) gene and SESTD1, the primary docking gene for the TRPC4 complex. The SESTD1 variant also occurred in two of the families in the validation cohort. When examining the types of TFBSs with identified variants, we found that the most commonly mutated TFBS was the Myc-associated zinc finger protein, MAZ, followed by EGR1 and SP1, with two variants each.

Table 2. Variants identified in upstream Factorbook TFBSs.

The affected gene, transcription factor type, and family with the variant are shown. Genes shown in bold were validated in the second cohort. Genes underlined were validated by in vivo screening.

Gene	Variant	TF	Family
AXIN2	17:63557882T>A	EGR1	FG141
CDK4	12:58146267G>A	SP1	FG129
DDN	12:49393353G>A	EGR1, MAZ	FG143
DDN	12:49393365G>A	MAZ	FG110
DDN	12:49393561G>A	UA2	FG144
MSH2	2:47630140G>C	E2F4	FG128
MSH2	2:47629890GT>G	ZNF263	FG146
PC	11:66726049C>G	E2F1	FG147
PC	11:66726027G>C	E2F4	FG148
PC	11:66726101C>T	SP1	FG135
HP1BP3	1:21113282A>G	E2F4	FG198
SESTD1	2:180129654C>G	MAZ	FG150, VC46, VC65
STK11	19:1205343G>C	UA3	FG151
TRPC4AP	20:33680727G>C	ZNF143	FG152
WDR7	18:54318543T>C	MAZ	FG153
WDR7	18:54318485C>T	NRF1	FG110
WDR7	18:54318419G>C	YY1	FG154

Last, we examined 42 known glioma predisposition GWAS signals that were identified in our previous study (table S6) (18) and found that, after correction for multiple tests, five sites were significantly enriched in our exploratory cohort compared with their MAFs in the general population. The two strongest associations were found for telomerase reverse transcriptase (TERT) (40% versus 50%, rs2736100, P = 0.00014) and an intergenic single-nucleotide polymorphism (SNP) often associated with CCDC26 (12% versus 6%, rs55705857, P = 0.00011). The former variant occurred in an MITF TFBS. In our validation cohort, the CCDC26-associated SNP was observed at a similar rate (approximately 12%), but the TERT-associated SNP was even more depleted than it was in the discovery cohort, with only 18% of the cohort carrying this SNP.

Functional analyses

We prioritized 72 genes containing prospective LoF mutations for validation studies in an established mouse model of glioma. To examine the contribution of this gene cohort to gliomagenesis, we performed CRISPR-Cas–based LoF screening in an in utero electroporation (IUE) model of glioma (table S7). The potential for gliomagenesis was assayed using barcoded CRISPR-Cas9 screening in the IUE model (31). We generated barcoded guide RNAs (gRNAs) for each of the candidate genes and combined this library with our established IUE model, which contains gRNAs for TP53, PTEN, and NF1 (i.e., 3xCr), along with Cas9 and green fluorescent protein (GFP) expression constructs (Fig. 5A, see Methods). This DNA cocktail was introduced via IUE into the cortex, and after glioma formation, we used targeted sequencing of tumors to detect barcodes as a surrogate for enrichment of a given gRNA within the tumor. This analysis revealed that gRNAs for DMBT, ZC3H7B, and HP1BP3 were enriched in tumors (Fig. 6B), suggesting that loss of these genes promotes tumorigenesis. These gRNAs were subsequently validated by Sanger sequencing, which showed that they cut at the intended regions (fig. S5).

Fig. 5. In vivo functional screening identified regulators of glioma tumorigenesis.

(A) Schematic of barcoded screening, where the 3xCr glioma system was combined with 72 barcoded gRNAs and co-electroporated into the embryonic cortex. (B) Next-generation sequencing to determine barcode amplification. The barcode for each gRNA (red) and the input signal (black) are shown (n = 3 tumors). Data are indicated as the mean and SEM. (C) Kaplan-Meier curves from individual validation studies (n = 20, px330 control, median survival 104 days; n = 20, ΔZC3H7B, median survival 76.5 days; n = 20, ΔDMBT1, median survival 78 days; n = 20, ΔHP1BP3, median survival 86.5 days). (D) BrdU staining of end-stage tumors; quantification for each group is derived from five different tumors. *P < 0.05, **P < 0.01, and ***P < 0.005. Scale bar, 50 μm.

Fig. 6. ZC3H7B, DMBT1, and HP1BP3 are key regulators of glioma tumorigenesis–related gene populations.

(A) Venn diagram of DEGs affected by ZC3H7B-LOF, DMBT1-LOF, and HP1BP3-LOF. (B) GO analysis of the common DEGs. (C) Heatmap of GO-related genes from the RNA-seq results of 3xCr [control (C)], ZC3H7B-LOF (Z), DMBT1-LOF (D), and HP1BP3-LOF (H). (D and E) VEGF and FGF2 staining from a postnatal day 60 tumor. DAPI, 4′,6-diamidino-2-phenylindole. (F) Intensities of VEGF and FGF2 staining determined by ImageJ software. The quantification of each group is derived from three different tumors. *P < 0.05, **P < 0.01, and ***P < 0.005. Scale bars, 50 μm. a.u., arbitrary units.

To confirm that loss of these genes promotes tumorigenesis, we individually introduced the gRNAs in our established IUE glioma model and found that individual loss of each gene decreased overall survival (Fig. 5C). In parallel, we performed overexpression and gain-of-function (GoF) studies by introducing piggyBac (PB) constructs containing each open reading frame (ORF) in our established IUE glioma model (32). Complementary to our LoF results, we found that overexpression of each gene extended overall survival (fig. S1), further supporting the notion that these genes suppress glioma tumorigenesis. Next, we harvested tumors and performed immunostaining to confirm loss of each gene while also validating the histopathological features of glioma (fig. S2). The decrease in overall survival suggested that loss of these genes accelerates malignant progression of glioma and associated proliferation. Bromodeoxyuridine (BrdU) labeling of mice bearing end-stage tumors showed that the number of BrdU-labeled cells was significantly increased in tumors lacking ZCH7B, DMBT1, and HP1BP3 (Fig. 5D). Notably, loss of ZCH7B and DMBT1 had a more profound effect on proliferation than loss of HP1BP3 (Fig. 5D), which is consistent with the observed relative reductions in overall survival (Fig. 5C). Together, these data indicate that familial glioma-associated genes contribute to tumorigenesis and that DMBT1, ZC3H7B, and HP1BP3 play key roles in this process.

Expression profiling of ZC3H7B, DMBT1, and HP1BP3 LoF glioma tumors

To understand the molecular changes that occur in glioma tumors lacking ZC3H7B, DMBT1, and HP1BP3, we performed RNA sequencing (RNA-seq) of tumors expressing ZC3H7B-LOF (gRNA vector for LoF), DMBT1-LOF, HP1BP3-LOF, and 3xCr constructs. Bioinformatics analysis of the differentially expressed genes (DEGs) revealed that 277 genes were down-regulated and 250 genes were up-regulated by ZC3H7B-LOF, 178 genes were down-regulated and 252 genes were up-regulated by DMBT1-LOF, and 736 genes were down-regulated and 147 genes were up-regulated by HP1BP3-LOF compared with tumors expressing 3xCr (fig. S3 and table S8). Next, we performed Gene Ontology (GO) analysis of the DEGs from the RNA-seq data of the ZC3H7B, DMBT1, and HP1BP3 tumor models. GO analyses revealed that chemokine-binding, antigen receptor–mediated signaling pathways, regulation of cytosolic calcium ion concentration, regulation of cytokine production, and cytokine signaling pathways are regulated by ZC3H7B; extracellular matrix organization, cell proliferation, and cell-matrix adhesion are regulated by DMBT1; and neutrophil activation, cytokine production, and regulation of interleukin-6 are regulated by HP1BP3 (fig. S3 and table S9).

Because the loss of each of these genes accelerates glioma tumorigenesis and is associated with increased glioma risk, we hypothesized that these genes regulate a common set of pathways. Therefore, we evaluated the overlapping genes of each DEG set and found that 167 genes are common to each of the LOF tumors, and the GO terms were related to calcium ion homeostasis, inflammatory response, and regulation of angiogenesis (Fig. 6, A to C, and table S10).

To determine whether these conserved DEGs are altered in glioma by loss of ZC3H7B, DMBT1, and HP1BP3, we examined the vascular endothelial growth factor (VEGF) and fibroblast growth factor 2 (FGF2) genes, which are associated with angiogenesis and calcium ion homeostasis. By immunostaining our glioma tumors, we found that VEGF and FGF2 expression was significantly increased in glioma tumors lacking ZC3H7B, DMBT1, or HP1BP3 (Fig. 6, D to F). Together, these data indicate that the familial glioma–associated genes DMBT1, ZC3H7B, and HP1BP3 contribute to tumorigenesis by influencing the glioma tumor microenvironment through regulation of multiple processes, including angiogenesis.

DISCUSSION

Familial glioma occurs rarely and has few associated genes. Our two cohorts of 189 and 115 glioma families represent the largest cohorts to undergo WGS. Previously, we identified POT1 variants in three families, including a large deletion in MSH2 in one family and variants in TP53 in two families, in a smaller cohort of subjects who underwent exome sequencing (15, 33). Here, we identified 54 genetic variants in 28 genes, or unique genetic loci in the case of CNVs, that contained rare, deleterious variants, noncoding functional variants, or CNVs that occurred in both our exploratory cohort and our validation cohort or in known CPGs. In total, 37 families (20%) in our exploratory cohort and 13 families (10%) in our validation cohort had at least one such alteration.

HERC2 and other enriched genes

We identified potential novel drivers of gliomagenesis that were statistically enriched in our familial glioma cohorts compared with the control cohort (Table 1). HERC2 showed the highest enrichment with deleterious variants compared with our control cohort. In total, six mutations of this gene were found: five rare coding variants that were predicted to be highly deleterious SNVs and one internal copy gain variant (see Fig. 4). Furthermore, we identified two additional, rare, deleterious variants in the Genome England dataset (p.R4587Q and p.S4607C). In accordance with our findings, HERC2 mutation was enriched in the putative familial glioma subset of this dataset (P < 0.001). All coding variants occurred in highly conserved regions of the gene; two occurred in a regulation of chromosome condensation protein domain, and two occurred in the homologous to the E6-AP C terminus (HECT) domain. HERC2 is highly conserved throughout the human population, and we found marked selection against LoF [predicted LoF intolerance (pLoFI) = 1.0] and missense variants (Z = 4.42). However, HERC2 is an extremely large protein (4834 amino acids) that frequently harbors many variants in normal individuals. Identification of HERC2 as a possible gliomagenesis gene in this study was only possible through aggressive filtering to determine the rarest and most deleterious variants. However, in future studies, less stringent filtering could be used to identify more causal variants, especially if these variants can be tested functionally. HERC2 conducts several important cellular functions. It regulates ubiquitin-dependent retention of repair proteins on damaged chromosomes, affecting both the double-stranded break and excision repair pathways, and modulates p53 activity (34, 35). It also indirectly regulates the insulin-like growth factor receptor pathway and cell cycle and may be involved in immune function (36). HERC2 is also commonly mutated and focally deleted in glioma (37). HERC2 was not detected in our CRISPR screening, possibly because cells with HERC2 knockdown become nonviable and are suppressed because of cell competition. This hypothesis is supported by literature demonstrating that HERC2 is essential for embryonic growth (38). Furthermore, HERC2 may only act as a tumor suppressor in the context of wild-type p53; thus, using a model system that has normal p53 levels may be necessary to observe the tumorigenic potential of HERC2 mutation (39). The second most enriched gene, MYO7A, was mutated three times in our exploratory cohort and twice in the validation cohort. This gene is an unconventional myosin with a very short tail (40). Mutations in this gene are associated with recessive and dominant deafness and Usher syndrome (MIM:276900). The role of MYO7A outside of hearing loss is not well understood, although it is ubiquitously expressed (40). LoF mutations in this gene lead to the “Shaker” mouse phenotype, and MYO7A appears to be essential in other mammals and fruit flies (41, 42).

Known CPGs

Known CPGs were found in 18 families. Our exploratory cohort showed significant enrichment of LoF or pathogenic mutations of BRIP1 and POLE, although this was not observed in our validation cohort. Heterozygous mutations of BRIP1 are commonly associated with a small increase in the risk of ovarian cancer susceptibility and may be associated with colon cancer (43). Gliomas have been identified in colorectal cancer families with POLE missense variants (44, 45). LoF variants of POLE are not generally considered pathogenic; however, both POLE variants occurred in the background of an additional variant (ATM and MLH1 in families FG189 and FG155, respectively), which may indicate an alternate mechanism of pathogenicity wherein haploinsufficiency of POLE exacerbates variants in other CPGs in a multigenic manner (46). Further CNVs were identified in CHEK2 and ATM. These results indicate that known cancer genes, not previously associated with glioma, play a role in familial glioma. However, it remains unclear why these families presented with familial glioma and not with the canonical cancers associated with the respective genes. This is true for both the personal history of the proband and the family history of all members in the family.

Glioma linkage and telomeres

Previous attempts to perform linkage analysis to identify glioma predisposition genes identified several linkage loci (17q12-21.32, 15q23-q26.3, and 1q23) but were unable to identify an underlying causative gene (47–49). Several genes (BRIP1, PMS2, PRKCA, ACACA, SMG6, MAP2K6, HCN4, PRTG, and IGSF9) identified in this study are in or proximal to these linkage sites.

This study reaffirms the importance of telomere maintenance in gliomagenesis. We identified two novel variants (a frameshift and an internal duplication variant) in the telomere gene POT1, a gene that we previously associated with glioma susceptibility (15). We further identified variants in SMG6, a component of the telomerase ribonucleoprotein complex responsible for the replication and maintenance of chromosome ends and identified enrichment of a known gene identified by a GWAS upstream of TERT. In addition to the variant upstream of TERT, we showed significant enrichment of another GWAS signal, rs55705857, an intergenic SNP associated with oligodendroglioma. This latter variant occurs in the MITF TFBS. Hi-C data from astrocytes (ENCSR011GNI) indicated that regions around this variant can associate with an upstream enhancer of MYC (50, 51). MYC dysregulation in tumors has been associated with individuals who carry this SNP (52).

Functional validation

We identified multiple novel genes as candidate CPGs. Of these, three were selected in a high-throughput in vivo screening assay and exhibited tumor suppressor–like function toward glioma tumorigenesis when individually manipulated in our model. These observations suggest that hypomorphic alleles or complete LoF of these candidates can promote tumorigenesis and may explain why variants in these genes predispose individuals carrying these alleles to glioma. Future studies will be geared toward understanding how these genes influence neural progenitor expansion during early tumor-initiating stages that better reflect cancer predisposition states in humans. In addition, we performed functional analysis of glioma tumors lacking DMBT1, HP1BP3, and ZC3H7B and identified large-scale dysregulation of multiple genes compared with those of the triple CRISPR knockout model alone. By focusing on common pathways dysregulated by each of these genes, we found that the angiogenesis pathway is conserved across these genes and validated the up-regulation of VEGF and FGF2 in each of these tumor models. These observations, coupled with the dysregulation of immune and cytokine signaling signatures from individual ZC3H7B and HP1BP3 knockout tumors, suggest an interaction between these genes and the tumor microenvironment. It will be important to further determine how these genes and associated variants influence the immune and vascular environment during tumorigenesis and during early tumor-initiating events.

Noncoding variants

Few studies have used WGS to examine the causes of familial cancer, and this study is one of the first to examine rare, functional, noncoding variants in depth. The noncoding portion of the genome is less well understood than the coding portion; therefore, we necessarily took a conservative approach to analyze these variants, only considering very high-quality (score > 0.8) cryptic splicing events and well-annotated, high-quality TFBSs in the immediate upstream region of a gene. This led to the identification of noncoding, potentially deleterious variants in several known CPGs. In both cohorts, we identified cryptic splicing variants of NCAM1 and SMG6, both of which are highly LoF intolerant (53). SMG6 is associated with telomere maintenance (54), and NCAM1 regulates neurogenesis, neurite outgrowth, cell migration, and immune surveillance (55, 56). We identified identical mutations in a cis-regulatory element upstream of SESTD1 in three unrelated families. This variant is absent in the Genome Aggregation Database (gnomAD), likely indicating a population frequency of much less than 1 in 10,000. We also identified a mutation in a cis-regulatory element in HP1BP3, a gene that maintains heterochromatin and influences the G₁ phase duration during the cell cycle (57). In general, noncoding variants may have a more subtle effect on gene function. Genes that are extremely important to cellular functions may contain few or no coding variants across the population, as these may be lethal. Noncoding variants that modulate expression may, therefore, be an important and unappreciated mechanism of disease for these genes.

Limitations

Because of the rarity of mutation of any particular gene in familial glioma, validation of findings through a second cohort can be challenging, highlighting the important role of functional validation. Here, we selected a subset of identified genes for functional validation through CRISPR knockdown. This approach identified three genes that increased tumor proliferation in vivo—DMBT1, HP1BP3, and ZC3H7B—with the latter two having no previous association with cancer. Other genes that were not detected during screening may also be involved in gliomagenesis but may have a mechanism that is more complex than LoF or require wild-type versions of genes that are knocked out in our model, or complete knockout of the gene may be incompatible with life. Furthermore, we chose to only use a single guide for each gene; thus, it may be possible that some guides are not as effective as others, leading to an increase in our false-negative rate. Thus, genes that pass our screen should be considered good candidates for gliomagenesis, and those that fail the screen cannot be discounted entirely.

Another limitation of this study is the lack of multiple family members available for study because of the rapid and aggressive nature of glioma and because glioma may not be classified as familial until many relatives have already died from the disease. The availability of first-degree relatives can eliminate approximately half of all rare variants considered (100 to 400 coding variants depending on filtering criteria). This can greatly reduce the number of variants for consideration and has been used to great effect in studying other rare diseases (58).

Although we potentially identified causal variants in many families, most of the families had no identified causal variant. This may be due to several factors, including stringent filtering of variants and the small cohort size. Furthermore, some families may not have had familial glioma but only coincidentally had multiple family members with glioma. This aspect may be further exacerbated in our validation cohort because most of those families only had self-reported second cases that were not confirmed by independent examination of the relevant medical records. We are also extremely limited in our understanding of noncoding variants, and any such variants require functional testing to determine their effects.

In conclusion, few genes are currently associated with familial glioma. We showed that genes associated with other heritable cancers may also increase glioma susceptibility. Furthermore, we identified novel roles for some genes, including HERC2 and TRPC4AP, as possible causes of familial glioma. Last, we showed significant enrichment of noncoding variants in several families with glioma. Noncoding variants, while difficult to annotate functionally, are a largely unstudied source of variation that may account for numerous cases of glioma.

METHODS

Case cohorts

The Gliogene Consortium was established in 2004 to identify susceptibility genes in high-risk familial brain tumor pedigrees (59). The consortium assembled information on 376 glioma families in 14 centers in the United States, Israel, Sweden, and Denmark between 2007 and 2011. The current report presents data from 189 families who are not part of our previous analysis (15). The eligibility criteria for inclusion were that the family had at least two members with pathologically confirmed primary glioma according to the 2007 World Health Organization Classification of Tumors of the Central Nervous System (60). Families with a known inherited genetic syndrome such as NF1 were excluded. We collected extensive pedigree information from each family, including cancer diagnoses, age, age at death, degree of relationship to the proband, size of the family, and number of generations, as well as blood or saliva samples from first- and selected second-degree relatives of the proband. The validation cohort included 122 familial glioma cases selected from the Glioma International Case Control Study. All patients completed an interviewer-administered questionnaire that included a detailed family cancer history report. Staff were trained to probe and to exclude any participant who was likely to have a nonprimary brain tumor or brain metastasis. Included participants reported a family history of glioma in first-, second-, or third-degree relatives. All participating institutions received institutional review board approval.

Control cohort

A set of 1013 non-Hispanic white control individuals from the ARIC cohort, aged 45 to 64 years, were randomly selected from four communities: Forsyth County, North Carolina; Jackson, Mississippi; suburban Minneapolis, Minnesota; and Washington County, Maryland. These samples were not specifically chosen to be age- or sex-matched with the case cohort but were chosen because of identical sequencing and processing to the case cohort to serve as a filter for artifactual findings. Samples from these participants underwent WGS concurrently with those from case participants and were analyzed in an identical manner to act as a control cohort (61).

Sequencing and alignment

WGS of DNA extracted from the peripheral blood was performed using the HiSeqX (Illumina Inc.) platform (62). Genomes were sequenced to an average redundant depth of 30×. Reads were aligned to HG19 using Burrows-Wheeler alignment, with duplicate reads removed before variant calling using platypus (63). Reads were annotated with CASSANDRA (64).

Coding and noncoding small and structural variant annotation and filtering

Variants were filtered out if they were not shared by all affected, sequenced family members, where applicable. All variants were filtered on the basis of MAF < 1 of 10,000 as measured by gnomAD “overall” (65). In addition, known pathogenic or likely pathogenic variants (any level of evidence) in known CPGs in ClinVar (25) were selected regardless of the MAF. Variants were then selected on the basis of the following criteria:

1) Coding variants were selected for either LoF (1-methionine, frameshift, canonical splice, stop-gain) or highly deleterious nonsynonymous variants. Deleterious variants were defined as having a CADD PH score of 30 or greater, which is equivalent to the top 0.1% most deleterious variants (66).

2) Intronic variants were filtered for deep-intronic variants (>5 bp from the exon, i.e., excluding the canonical splice region) and based on predicted effects on splicing by generating a cryptic splice acceptor or donor (score > 0.8 by spliceAI) (67).

3) TFBS variants were selected if they were 1 kbp upstream of a transcription start site and occurred in a Factorbook high-quality site (score > 2.0) (68).

4) Noncoding variants that were previously identified by GWAS as associated with glioma predisposition were selected (18).

5) Structural variants were identified using both the DELLY suite of tools and an in-house developed tool that uses a normalized set of genomes as a background for identifying CNVs (69).

Variant gene selection

All variants were further filtered if they affected a gene with the following properties: (i) known CPGs, (ii) novel CPGs as determined by an increase in pLoFI when considering The Cancer Genome Atlas (TCGA) germline samples versus the entire gnomAD cohort (see below), and (iii) genes that were LoF intolerant (pLoFI > 0.5) and missense intolerant (Z > 3.0).

Novel CPGs

gnomAD (r03 March 2016 release) data were downloaded for the entire dataset as well as a dataset without TCGA data. Genes with increased pLoFI values in the TCGA-cleaned dataset compared with those in the gnomAD dataset in its entirety were depleted of LoF variants, i.e., an increased number of variants existed among TCGA germline samples. We defined these cancer-specific, LoF-intolerant genes as any genes with pLoFI > 0.8 in the TCGA-cleaned data, a difference of >0.2 between the pLoFI in the TCGA-cleaned data, or a full ExAc OR difference of >0.4.

Validation

All insertion or deletion events were validated by Sanger sequencing. SNVs were validated by Sanger sequencing if they had fewer than five reads (or <25% of the total reads) supporting the variant.

Plasmid constructs

All gRNAs used for the in vivo barcode enrichment competition assay LoF screening were designed using CRISPOR (http://crispor.tefor.net/crispor.py). The gRNAs were generated with barcode sequences through a previously described HiTMMoB approach in a PB (PBCAG) vector by gateway cloning (1). For individual LoF studies with ZC3H7B, DMBT1, and HP1BP3 gRNAs, each gRNA was generated in a Px330-Crispr-mCherry vector by gateway cloning (ZC3H7B, 5′-GGTCGTTGCCCTCGGCAAAC-3′; DMBT1, 5′-CTCCACAGGCCACGTCATCC-3′; and HP1BP3, 5′-GAATTCACGGCTCGACGTAT-3′). For individual GoF studies, the ZC3H7B ORF was obtained from Origin (catalog no. RC219566). The DMBT1 and HP1BP3 ORFs were obtained from Thermo Fisher Scientific (Ultimate ORF clones). The efficiency of individual vectors was confirmed by staining the brain (LoF vectors) and NIH3T3 cells (GoF vectors) (fig. S4).

In vivo functional screening

All mouse gliomas were generated on the CD-1 IGS mouse background as previously described (31). IUE was performed on embryonic day 15. Previously generated CRISPR constructs were used to knock out NF1, PTEN, and p53 (1.5 μg/μl each). gRNAs were genomically integrated and overexpressed using the PB transposase (PBase) system. The pGlast-PBase plasmid (2.0 μg/μl) (34) was co-electroporated with a PBCAG-gRNA construct (1.0 μg/μl).

In vivo barcode enrichment competition assay

A PB transposable vector (PBCAG-EGFP-T2A-GWR1R4) was engineered from the PBCAG-EGFP construct (32). The EGFP STOP codon was removed, and an in-frame T2A sequence followed by attR1/attR4 Gateway cloning sites was inserted. These attR sites flanked chloramphenicol and ccdB selection cassettes. A V5 tag sequence was also inserted downstream of the attR4 site. gRNAs and barcode sequences were cloned using the HiTMMoB approach (70). The gRNAs were designed using standard approaches and cloned into a PB transposase system containing a unique barcode identifier. Pooled injection cocktails were assembled such that the pool of tested gRNA totaled 1.0 μg/μl. Briefly, an equal number of moles of each plasmid was mixed together, ethanol-precipitated, and redissolved/concentrated in double distilled H₂O. This pooled plasmid mixture was diluted to 1.0 μg/μl in the final IUE injection cocktail. After IUE of the pooled cocktail (with 3xCr, PBase, and GFPt2aLuc), mice were born and observed for symptoms suggestive of brain tumors (see below). Upon demonstration of symptoms, the animal was euthanized, and tumors were examined using a fluorescence reporter. All procedures were approved by the Institutional Animal Care and Use Committee at the Baylor College of Medicine and conformed to the U.S. Public Health Service Policy on Human Care and Use of Laboratory Animals.

Subsequently, genomic DNA (gDNA) was prepared with the EZNA Tissue DNA Kit (Omega Bio-tek, D3396), according to the manufacturer’s instructions. Samples were prepared in biological and technical replicates (3 ≤ N_biological ≤ 5; N_technical = 3). In addition, the IUE injection cocktail was used for input and prepared in technical duplicates. Following gDNA isolation, barcode libraries were prepared as previously reported (31, 71). Polymerase chain reactions (PCRs) were used to amplify the barcode pools from 50 ng of gDNA (experimental samples) or 2 ng of the plasmid pool (input control) using Platinum Super Mix (Thermo Fisher Scientific, 12532016) with primers targeting the T3 promoter site (directly upstream of the barcode) (5′-CAATTAACCCTCACTAAAGG-3′) and the V5 tag (5′-ACCGAGGAGAGGGTTAGGGAT-3′). The amplification parameters were as follows: 1× (94°C, 4 min); 35× (94°C, 1 min; 54°C, 1 min; 68°C, 1 min); 1× (68°C, 10 min); 10°C, hold. The PCR products were purified with the PureLink PCR Purification Kit (Thermo Fisher Scientific, K310001), processed using the Ion Plus Library Kit (Thermo Fisher Scientific, 4471252), subsequently purified, and ligated to unique Ion Xpress Barcode Adaptors (Thermo Fisher Scientific, 4474517). The resulting Ion Xpress barcoded libraries were amplified, purified, and pooled for Personal Genome Machine sequencing (318 V2 Chip) following the manufacturer’s recommendations. Raw data were concatenated into one “reference” file and indexed using the Burrows-Wheeler alignment tool40 for alignment of barcode sequences (with parameters “-l7 -t12 -N -n3”) to count the occurrence of each barcode. Barcode enrichment was assessed by quantitating the number of occurrences of each barcode sequence as a ratio to the total number of barcode reads in each sample. The standard error was calculated across replicates and plotted as error bars on the barcode enrichment graphs.

To assay in vivo cell proliferation, 4 hours before harvesting, 100 μg of BrdU [in phosphate-buffered saline (PBS)] per gram body mass was intraperitoneally injected. Mouse brains were collected, frozen, and sectioned as described above. Before blocking, the sections were incubated in 2 N HCl at 37°C for 30 min and neutralized with 3.8% sodium borate for 10 min at room temperature.

Validation of gRNA vectors

For full methods, see Yu et al. (31). Briefly, for genomic editing validation of individual gRNA vectors, tumor tissues were collected for analysis at post-natal day 60, and tumors were resected and subjected to gDNA isolation using the DNeasy Blood and Tissue Kit (Qiagen) for sequencing. Genomic specific primers were used to amplify amplicons surrounding the putative CRISPR-Cas9–modified sites from gDNA-derived ΔZC3H7B, ΔDMBT1, and ΔHP1BP3 tumors. The targeted genomic areas flanking single guide RNA (sgRNA) binding sites were PCR-amplified using Phusion High-Fidelity DNA Polymerase (NEB M0530S). PCR products were purified using Ampure XP magnetic beads at 1.8× volume to separate amplicons from primers and were inspected using the dsDNA 900 Reagent Kit (DNF-900-K0500, AATI) on the Advanced Analytical 12-Capillary Fragment Analyzer. The verified product was then tagmented, amplified, and cleaned up using the Nextera XT DNA Library Prep Kit (FC-131-1024, Illumina). The regulating library was validated using the Standard Sensitivity NGS Fragment Analysis Kit on the 12-Capillary Fragment Analyzer, and samples were normalized to 2 nM using Ampure XP bead-based normalization. Samples were sequenced on Illumina NextSeq550 following the manufacturer’s instructions. Subsequently, reads were aligned to reference sequences of the PCR amplicons and evaluated for the presence of genomic alterations using custom script (fig. S5).

Tissue preparation

For immunostaining, mice were anesthetized and transcardially perfused first with saline solution and then with 4% paraformaldehyde in 0.1 M phosphate buffer (pH 7.2). Brains were stored at 4°C in 4% paraformaldehyde for 2 days and then stored in a 20% sucrose solution until they sank. Six separate series of 40-μm coronal brain sections were obtained using a Model CM3050S cryostat (Leica, Wetzlar, Germany) and stored in an anti-freeze stock solution (phosphate buffer containing 30% glycerol and 30% ethylene glycol, pH 7.2) at 4°C before use.

Immunohistochemistry and immunofluorescence staining

For immunofluorescence staining, every sixth serial section in each set (approximately eight sections) was collected, rinsed twice with PBS containing 0.2% Triton X-100 (PBST), and rinsed once with PBST. After nonspecific binding was blocked by incubating the sections with 1% bovine serum albumin in PBST, the sections were incubated overnight at room temperature with primary antibodies [GFP (Abcam, ab13970), ZC3H7B (Invitrogen, Carlsbad, CA, USA, PA5-106537), DMBT1 (Invitrogen, PA5-115313), HP1BP3 (Oroteintech, 24556-I-AP), FGF2 (Invitrogen, PA5-116495), and VEGF (Invitrogen, MA5-13182)]. Immunoreactive proteins were visualized using Alexa Fluor 488– or Alexa Fluor 555–conjugated secondary antibodies (1:600 dilution; Invitrogen). Images were captured using a confocal microscope (Carl Zeiss, Oberkochen, Germany).

For hematoxylin and eosin staining, 10-μm paraffin-embedded sections were processed using the following solutions: xylene (10 min ×2), 100% EtOH (10 min), 95% EtOH (10 min), 80% EtOH (10 min), 70% EtOH (10 min), ddH₂O (5 min), Harris hematoxylin (2 min; Poly Scientific R&D Corp., S212A), running tap water wash, 95% EtOH (30 s), eosin (2 min, Poly Scientific R&D Corp., S176), 95% EtOH (2 min x2), 100% EtOH (2 min ×2), and xylene (10 min ×2).

A BrdU assay was performed to confirm in vivo cell proliferation. Four hours before harvesting, 100 μg of BrdU (in PBS) per gram body mass was intraperitoneally injected.

RNA extraction, library preparation, sequencing, and bioinformatic analysis

Mouse brains were harvested from each experimental group at postnatal day 60. Tissues were rinsed with PBS and dissociated in TRIzol (Thermo Fisher Scientific, 15596018). Total RNA was extracted using the RNeasy Mini Kit (Qiagen, 74106) according to the manufacturer’s instructions. Illumina sequencing libraries with 6-bp single indices were constructed from 300 ng of total RNA using a TruSeq Stranded mRNA LT kit (Illumina, RS-122-2101). Equal concentrations (2 nM) of libraries were pooled and subjected to paired-end sequencing of approximately 20 million to 30 million reads per sample using a Mid Output v2 kit (Illumina, 20024904) on Illumina NextSeq550 according to the manufacturer’s instructions.

Sequencing files from each flow cell lane were downloaded in fastq files and merged. Quality control was performed using fastQC (v0.11.7) and MultiQC (v1.11). Reads were mapped to the mouse genome mm10 assembly using STAR (v2.5.0a) (72). Mapped reads were used to build count matrices using the Bioconductor packages GenomicAlignments (v1.26.0) and GenomicFeatures (v1.42.2) in R (v4.1.2) (73). UCSC transcripts were downloaded from Illumina iGenomes in GTF file format. DESeq2 (v1.30.1) (74) was used for normalization and DEG analysis. GOs were determined using Enrichr and the GO resource (http://geneontology.org/), and significant GO terms with P values of <0.05 were selected for visualization using ggplot2 (v3.3.5) and GOplot (v1.0.2). Gene expression heatmaps were generated using ComplexHeatmap (v2.6.2).

Supplementary Materials

This PDF file includes:

Figs. S1 to S5

Legends for tables S1, S8 to S10

Tables S2 to S7, S11 and S12

Other Supplementary Material for this manuscript includes the following:

Tables S1, S8 to S10

View/request a protocol for this paper from Bio-protocol.