Genome-wide association identifies diverse causes of common variable immunodeficiency

Abstract

Background

Common variable immunodeficiency (CVID) is a heterogeneous immune defect characterized by hypogammaglobulinemia, failure of specific antibody production, susceptibility to infections, and an array of comorbidities.

Objective

To address the underlying immunopathogenesis of CVID and comorbidities, we conducted the first genome-wide association and gene copy number variation (CNV) study in patients with CVID.

Methods

Three hundred sixty-three patients with CVID from 4 study sites were genotyped with 610,000 single nucleotide polymorphisms (SNPs). Patients were divided into a discovery cohort of 179 cases in comparison with 1,917 control subjects and a replication cohort of 109 cases and 1,114 control subjects.

Results

Our analyses detected strong association with the MHC region and association with a disintegrin and metalloproteinase (ADAM) genes (P combined = 1.96 × 10⁻⁷) replicated in the independent cohort. CNV analysis defined 16 disease-associated deletions and duplications, including duplication of origin recognition complex 4L (ORC4L) that was unique to 15 cases (P = 8.66 × 10⁻¹⁶), as well as numerous unique rare intraexonic deletions and duplications suggesting multiple novel genetic causes of CVID. Furthermore, the 1,000 most significant SNPs were strongly predictive of the CVID phenotype by using a Support Vector Machine algorithm with positive and negative predictive values of 1.0 and 0.957, respectively.

Conclusion

Our integrative genome-wide analysis of SNP genotypes and CNVs has uncovered multiple novel susceptibility loci for CVID, both common and rare, which is consistent with the highly heterogeneous nature of CVID. These results provide new mechanistic insights into immunopathogenesis based on these unique genetic variations and might allow for improved diagnosis of CVID based on accurate prediction of the CVID clinical phenotypes by using our Support Vector Machine model.

Common variable immunodeficiency (CVID) is manifested by insufficient quantity and quality of immunoglobulin, leading to susceptibility to bacterial infections.¹ CVID is a primary immunodeficiency disease because it is believed to result from intrinsic deficits affecting immunologic functions. CVID is heterogeneous, can present early or later in life, and is associated with specific comorbidities.² Efforts to subcategorize CVID to predict outcomes and comorbid conditions both clinically and based on immunologic phenotypes are ongoing.³ B cell–activating factor of the TNF family receptor,⁴ transmembrane activator and calcium modulator and cyclophilin ligand interactor (TACI),^5–7 and certain HLA haplotypes^8,9 have been identified as potential gene candidates for susceptibility to CVID. Inducible costimulator, ^10,11 CD81,¹² CD19,^13,14 and CD20¹⁵ harbor disease-causing mutations that presently explain only a small percentage of cases.¹⁶ The heterogeneous presentations of patients with CVID complicate clinical management, and searches for novel genetic predictors of disease or comorbidity have been of limited success.

Genome-wide single nucleotide polymorphism (SNP) arrays have enabled high-throughput genotyping of genomic DNAwith tagging of the whole genome based on linkage disequilibrium. Given both intensity and genotype content, copy number variations (CNVs) can also be detected with the same genome-wide association study (GWAS) platforms and are predicted to be responsible for generation of rare phenotypes. We genotyped 363 CVID cases and 3031 healthy control subjects on the Illumina Infinium Human Hap610K BeadChip (Illumina, San Diego, Calif). We sought to associate SNPs and CNVs with CVID, as well as with the characteristic clinical subphenotypes of this syndrome, to address disease heterogeneity. Our hypothesis was that although complex, CVID is sufficiently genetically distinct from the healthy population to allow genetic prediction of this uncommon but clinically relevant disease.

METHODS

Patients

The diagnosis of CVID was established in concordance with existing diagnostic criteria.^16,17 All patients were enrolled in institutionally approved research protocols to enable genetic analysis and collection of clinical data. Subsets of the patients reported here have been previously included in published studies.^2,18,19

Illumina Infinium assay for SNP genotyping and CNV discovery

We performed high-throughput, genome-wide SNP genotyping with the InfiniumII HumanHap610 BeadChip technology at the Center for Applied Genomics at the Children’s Hospital of Philadelphia. The genotype data content together with the intensity data provided by the genotyping array provides high confidence for CNV calls. Importantly, the simultaneous analysis of intensity and genotype data in the same experimental setting establishes a highly accurate definition for normal diploid states and any deviation from the norm. To call CNVs, we used the PennCNV algorithm,²⁰ which combines multiple sources of information, including log R ratio and B-allele frequency at each SNP marker, along with SNP spacing and population frequency of the B allele to generate CNV calls. Rare recurrent CNVs were the focus of our study.

CNV quality control

We calculated quality control measures on our HumanHap610 GWAS data based on statistical distributions to exclude poor-quality DNA samples and false-positive CNVs. The first threshold is the percentage of attempted SNPs that were successfully genotyped. Only samples with call rates of greater than 98% were included. The genome-wide intensity signal must have as little noise as possible. Only samples with an SD of normalized intensity (log R ratio) of less than 0.35 were included. All samples must have white ethnicity based on principle components analysis, and all other samples were excluded. Furthermore, case and control matching was ensured by calculating a genomic inflation factor between groups. Wave artifacts roughly correlating with guanine-cytosine content resulting from hybridization bias of low full-length DNA quantity are known to interfere with accurate inference of CNVs. Only samples in which the guanine-cytosine-corrected wave factor of the log R ratio −0.02 <×< 0.02 were accepted. If the count of CNV calls made by using PennCNV exceeds 100, it is suggestive of poor DNA quality, and those samples were excluded. Thus only samples with CNV call counts of less than 100 were included. Any duplicate samples (eg, monozygotic twins or repeats on the same patient) were identified, and as a result, 1 sample was excluded.

Statistical analysis of CNVs

CNV frequency between cases and control subjects was evaluated at each SNP by using the Fisher exact test. We only considered loci that were nominally significant between cases and control subjects (P < .05) for which patients had the same variations that were observed in multiple cohorts or were not observed in any of the control subjects and were validated with an independent method. We report statistical local minimums to narrow the association in reference to a region of nominal significance, including SNPs residing within 1 Mb of each other. Resulting nominally significant CNVs were excluded if they met any of the following criteria: (1) residing on telomere or centromere proximal cytobands; (2) arising in a “peninsula” of common CNVs arising from variation in boundary truncation of CNV calling; (3) being genomic regions with extremes in GC content, which produces hybridization bias; or (4) contributing to multiple CNVs. Three lines of evidence establish statistical significance: an independent replication P value of less than .05, permutation of observations, and no loci observed with control-enriched significance. We used the Database for Annotation, Visualization, and Integrated Discovery (DAVID) to assess the significance of functional annotation clustering of independently associated results into InterPro categories.

RESULTS

Our CVID case cohort was composed of 223 patients from Mount Sinai School of Medicine, 76 patients from the University of Oxford, 37 patients from the Children’s Hospital of Philadelphia, and 27 patients from the University of South Florida. The diagnosis in each case was validated against the European Society for Immunodeficiencies/Pan-American Group for Immunodeficiency diagnostic criteria.¹⁶

We first evaluated the quality and suitability of the data for a case-control study. Seven samples had call rates of less than 98% and were excluded. Based on principle components analysis with EIGENSTRAT software,²¹ the cases were stratified into 3 clusters: 288 cases were of confirmed European ancestry, whereas 30 samples were significant outliers from white race and were omitted from this analysis. A random split of 179 cases was matched with 1917 white control subjects. We identified an additional 1114 control samples that clustered well with the balance of 109 white cases used for replication. We excluded subjects with observed cryptic relatedness of genotypes. For CNV calling, samples with high noise (SD of log R ratio >0.35) in intensity signal were additionally removed.

SNP association analysis

We first performed a GWAS using the SNP genotype data. Genotype frequencies were compared between cases and control subjects by using a χ² test statistic applied in PLINK²² for SNPs with an at least 90% call rate and 1% minor allele frequency. The discovery cohort of 179 cases and 1917 control subjects yielded a low genomic inflation factor of 1.02783 (Fig 1, A). The most significant association was to the MHC, with the most significant SNP, rs3117426, having a P value of 8.62 × 10⁻¹⁰ (Table I). The second best association was to a locus on 8p21.2 harboring a disintegrin and metalloproteinase 28 (ADAM28), ADAM7, ADAMDEC1, and STC1. However, this locus did not meet genome-wide significance criteria (P = 6.24 × 10⁻⁶ for rs4872262) and results, which is in keeping with modest power from the relatively small sample sizes.

FIG 1.

Case-control allele frequency significance. SNP-based case-control significance is shown in negative log base 10 genome-wide analysis for (A) discovery- and (B) replication-inclusive CVID cohorts. (C) Subsequently, CVID cases with specific disease subphenotypes were compared with CVID cases without the subphenotype. The single-SNP tests are shown as single points. Multiple neighboring SNPs of similar significance boost confidence in the association, as shown in the strong peak on 6p22.1-p21.32 of the discovery case-control cohort. Conversely, the median significance P value is kept low by minimizing population stratification, which minimizes the genomic inflation factor.

TABLE I.

Most significant associated regions based on genotype association

Region genomic span	Gene	Count SNPs top 1000	Distance from exon	P value	SNP	A1	F_A	F_U	OR	P value, replication	SNP, replication	F_A, replication	F_U, replication
chr6:27236785-32521295	MHC	110	0	8.62 × 10−10	rs3117426	T	0.3268	0.1907	2.061	.0004	rs2156875	0.5734	0.4475
chr8:23746576-24681608	ADAM28, ADAM7, ADAMDEC1, STC1	4	0	6.24 × 10−6	rs4872262	A	0.0391	0.0107	3.765	.0314	rs4872262	0.0321	0.0135
chr7:4270929-4287047	SDK1	5	0	4.70 × 10−5	rs895710	T	0.2905	0.4003	0.6135	.0235	rs4720301	0.3991	0.4793
chr1:20369369-20490791	FLJ32784, UBXN10	7	0	9.09 × 10−5	rs7514144	C	0.2654	0.1808	1.636	2.25E-8	rs6426636	0.4679	0.2857
chr1:59987116-59993733	FGGY	2	2458	1.52 × 10−4	rs11207520	G	0.3994	0.3026	1.533	.0015	rs11207520	0.2569	0.1706
chr2:65518304-65520560	FLJ16124	2	784	5.20 × 10−4	rs1194849	G	0.3659	0.4614	0.6737	.0119	rs1876518	0.3853	0.4744

Known immunologic function genes are shown in boldface and underlined. Additional details are shown in Tables E7 and E8 in this article’s Online Repository at www.jacionline.org.

A1, Allele 1, which is the minor allele, F_A, frequency of allele 1 in affected subjects; F_U, frequency of allele 1 in unaffected subjects.

In an attempt to replicate the 8p21.2 locus, we searched for direction of effects in our independent CVID cohort of 109 cases and 1114 control subjects (genomic inflation factor, 1.04789; Fig 1, B). Interestingly, 4 SNPs displayed significance (P < .05) and allele frequency in the same direction as the discovery cohort, including rs11207520, rs1194849, rs17790790, and rs4872262 (Table I). When all SNPs contributing to the discovery region were queried in replication rather than just the most significant SNP, a total of 8 SNPs displayed significance (P <.05) and allele frequencies in the same direction as discovery (Table I).

We next addressed the heterogeneity of the CVID phenotype by assessing the various CVID subphenotypes. On the basis of recent phenotypic characterization,¹⁸ we established 16 distinct clinical subgroups within the CVID cohort, including cancer, lymphoma, lymphadenopathy, nodular regenerative hyperplasia of the liver, lymphoid interstitial pneumonitis (LIP), bronchiectasis, biopsyproved granuloma, gastrointestinal enteropathy, malabsorption, splenectomy, cytopenias, organ-specific autoimmunity, low IgM level (<50 mg/dL), low IgA level (<10 mg/dL), low B cell number (CD19⁺ cells <1%), and young age at symptom onset (<10 years, see Fig E1 and Table E1 in this article’s Online Repository at www.jacionline.org).

Differences in genotype frequencies were assigned significance, and regions with multiple significant SNPs were then scored. Significant SNP associations were made for all subsets (Table II and Fig 1, C). The most significant association was observed with nodular regenerative hyperplasia of the liver on AK096081-AK124028 (P = 2.29 × 10⁻¹⁰). Lymphoma was associated with KIAA0834, PFTK1, and HAVCR1, with corresponding P values ranging from 1.69 × 10⁻⁸ to 3.62 × 10⁻⁸. An association was also observed between LIP and FGF14 and ZNF81, with corresponding P values of 5.76 × 10⁻⁸ and 7.70 × 10⁻⁸, respectively. Organ-specific autoimmunity also associated with SNX31 (P = 6.89 × 10⁻⁸), and low IgM levels (<50 mg/dL) in the case subjects were associated with LDLRAP1 (P = 6.02 × 10⁻⁸).

TABLE II.

Most significant associated regions based on subphenotype genotype association

Significant region	P value	SNP	A1	F_A	F_U	OR	Count SNPs	Gene	Distance	Distance from exon	Subphenotype
chr1:68153970-68161019	2.29 × 10−10	rs1926283	C	0.2308	0.01807	16.3	3	AK096081, AK124028	0	33408	NRH
chr20:55901032-55985359	3.21 × 10−9	rs8124301	T	0.4231	0.07631	8.877	5	C20orf85	174030	174030	Malabsorption
chr7:90245135-90245280	1.69 × 10−8	rs975004	G	0.3333	0.03953	12.15	2	KIAA0834, PFTK1	0	12544	Lymphoma
chr5:156402979-156407976	3.62 × 10−8	rs10038271	T	0.7222	0.1877	11.25	2	HAVCR1	0	651	Lymphoma
chrX:22422160-22428888	5.71 × 10−8	rs5925651	A	0.3462	0.1386	3.291	3	ZNF645	219665	219665	Bronchiectasis
chr13:101641990-101673539	5.76 × 10−8	rs1336698	G	0.65	0.168	9.197	6	FGF14	0	69139	LIP
chr1:25735669-25753653	6.02 × 10−8	rs2065970	G	0.06553	0.2411	4.529	5	LDLRAP1	0	0	Low IgM
chr8:101728102-101728344	6.89 × 10−8	rs7815950	G	0.2133	0.05615	4.559	2	SNX31	0	2334	OSAI
chr8:13834782-13842376	7.62 × 10−8	rs2682665	C	0.1818	0.01394	15.71	3	SGCZ	149368	149368	Low B cells
chrX:152295068-152303019	9.84 × 10−8	rs5987017	A	0.2143	0.04773	5.442	4	ZNF275	24829	24829	Young age
chr12:2531014-2538733	1.21 × 10−7	rs4765961	C	0.4737	0.142	5.439	3	CACNA1C	0	0	GI Enteropathy
chrX:30733771-30736604	1.44 × 10−7	rs11095197	T	0.1469	0.3431	3.034	2	MAP3K7IP3	18876	18876	Low IgA
chrX:53035117-53039801	1.90 × 10−7	rs10127016	G	0.3833	0.125	4.351	2	TMEM29	0	1409	Lymphadenopathy
chr14: 23168471- 23169215	4.57 × 10−7	rs222723	C	0.6154	0.1968	6.531	2	DHRS2	0	0	Cancer

Known immunologic function genes are shown in boldface and underlined. Additional details are shown in Table E9 in this article’s Online Repository at www.jacionline.org.

GI, Gastrointestinal; NRH, nodular regenerative hyperplasia of the liver; OR, odds ratio; OSAI, organ-specific autoimmunity.

Finally, patients with CVID manifesting enteropathy and their resulting subphenotype significance were additionally compared with patients with inflammatory bowel disease from a previously reported cohort²³ to assess whether specific loci contributed to a common cause. The SNP rs12889533 was significant in both the CVID cases with enteropathy and the inflammatory bowel disease cohort, with allele frequency difference in the same direction (see Table E2 in this article’s Online Repository at www.jacionline.org).

CNV analysis

We next performed a CNV analysis, in which rare (<1%) CNVs were called by using PennCNV.²⁰ We analyzed deletion and duplication CNV frequencies of 311 cases and 2766 control subjects (see Fig E2 in this article’s Online Repository at www.jacionline.org). We discovered 5 deletions and 11 duplications to be recurrent and significantly enriched in the CVID cases compared with the control subjects (Tables III and IV). In addition, 15 regions were exclusive to CVID cases. The most noteworthy duplication resided at 2q23.1 and affected multiple exons of origin recognition complex 4L (ORC4L) that were exclusive to 15 cases, and were not identified in control subjects (P = 8.66 × 10⁻¹⁶, see Fig E3 in this article’s Online Repository at www.jacionline.org).

TABLE III.

Most significant CVID-associated regions based on CNV association: deletions

CNV deletion	Count SNPs	Distance from exon	P value, deletion	Cases, deletion	Control subjects, deletion	OR	Gene	MSSM	Oxford	USF	CHOP
chr11:85365857-85381622	5	0	.01	2	0	Infinity	PICALM, PICALM variant protein	0	2	0	0
chr20:57735790-57741780	6	5902	.026	3	4	6.73	PHACTR3	2	1	0	0
chr22:17396663-18417315	270	0	.029	2	1	17.90	ARVCF, COMT, TBX1, others	2	0	0	0
chr4:10256682-10264316	4	2915	.029	2	1	17.90	CLNK	0	1	1	0
chr10:46003146-46042543	8	0	.029	2	1	17.90	DKFZp566K0524, PTPN20A, PTPN20B	2	0	0	0

Known immunologic function genes are shown in boldface and underlined. Additional details are available in Table E10 in this article’s Online Repository at www.jacionline.org.

CHOP, Children’s Hospital of Philadelphia; MSSM, Mount Sinai School of Medicine; OR, odds ratio; USF, University of South Florida.

TABLE IV.

Most significant CVID-associated regions based on CNV association: duplications

CNV duplication	Count SNPs	Distance from exon	P value, duplication	Cases, duplication	Control subjects, duplication	OR	Gene	MSSM	Oxford	USF	CHOP
chr2:148396730-148433180	6	0	8.66 × 10−16	15	0	Infinity	ACVR2A, ORC4L	10	5	0	0
chr15:35053039-35063531	6	23146	5.72 × 10−13	19	10	17.93	MEIS2	17	2	0	0
chr7:91789778-91801963	4	0	1.20 × 10−9	12	4	27.71	ANKIB1	9	3	0	0
chr2:163316298-163316595	3	17100	1.00 × 10−7	7	0	Infinity	KCNH7	6	1	0	0
chr7:87180702-87191931	4	202	7.41 × 10−7	7	1	63.67	RPIB9	5	2	0	0
chr19:9256584-9277749	4	0	0.001	3	0	Infinity	ZNF699	2	0	1	0
chr4:39190766-39201960	6	0	0.004	3	1	26.93	UGDH	1	2	0	0
chr7:18903915-18905725	4	23171	0.009	3	2	13.46	HDAC9	1	2	0	0
chr5:170531269-170536993	4	374	0.01	2	0	Infinity	DKFZp666P032, RANBP17	1	1	0	0
chr7:34685030-34686172	4	4517	0.01	2	0	Infinity	AAA1, NPSR1	1	1	0	0
chr1:170400944-170403742	4	26274	0.01	2	0	Infinity	DNM3	0	1	1	0

Known immunologic function genes are shown in boldface and underlined. Additional details are available in Table E10 in this article’s Online Repository at www.jacionline.org.

CHOP, Children’s Hospital of Philadelphia; MSSM, Mount Sinai School of Medicine; OR, odds ratio; USF, University of South Florida.

We additionally observed large CNVs captured by 10 or more SNPs that were exclusive to CVID cases totaling 84 deletions and 98 duplications. (see Table E3 in this article’s Online Repository at www.jacionline.org). No significant overrepresentation was observed in the control subjects in this frequency range. Of clinical relevance, 2 patients were newly identified with deletions in the 22q11 region, suggesting an alternative presentation of the well-characterized congenital immunodeficiency associated with this microdeletion. Interestingly, 10 regions of homozygous deletion were observed exclusively in CVID cases (see Table E4 in this article’s Online Repository at www.jacionline.org).

We next examined the overall CNV burden in cases compared with control subjects for both exonic CNVs and for large (100 kb) and rare (<1%) CNVs. We found deletions to be significantly enriched in the CVID cases when considering CNV observations on all loci genome wide (see Table E5 in this article’s Online Repository at www.jacionline.org). All but 5 of the 182 CNVs exclusive to the CVID cohort were intraexonic. To assess the reliability of our CNV detection method, we reviewed the B-allele frequency (genotype) and log R ratio (intensity) values of our Illumina data and experimentally validated all the significant CNVs by using an independent method, the Affymetrix Cytogenetics Whole-Genome 2.7M Array (Affymetrix, Santa Clara, Calif), which provides high resolution with 400,103 SNP and 2,387,595 CN probes (see Fig E4 in this article’s Online Repository at www.jacionline.org).

CVID disease prediction

We used a Support Vector Machine (SVM) algorithm to determine how well we could predict the CVID phenotype in a pool of “unknown” samples (see the Methods section in this article’s Online Repository at www.jacionline.org). The SVM algorithm was trained on the discovery cohort of 179 CVID cases and 1917 control subjects. We identified 1000 SNPs from association analysis that accurately separated the CVID cases from the control subjects with 98.7% accuracy overall. When applied on the independent cohort of 109 cases and 1114 control subjects, the positive and negative predictive values amounted to 1.0 and 0.957, respectively (Fig 2), suggesting that these markers were robust in distinguishing the genetic architecture of CVID cases from that of healthy control subjects.

FIG 2.

SVM prediction of CVID status. Plotted are log base 2 SVM scaling values of cost C (the penalty parameter of SVMs) and γ (1/number of variables) from LIMSVM software. Data were coded as +1 for affected and −1 for unaffected subjects. Genotype was scaled as follows: AA, 0; AB, 0.5; and BB, 1. Lines represent modeling with increasing resolution, with the green line representing the final model solution. The model was generated by using the discovery cohort and control subjects, whereas the subjects represented (blue, CVID cases; red, control subjects) were from the separate replication cohort.

DISCUSSION

CVID was described more than 50 years ago, but aside from a small number of recessively inherited genes in a few families and the more prevalent but poorly understood contribution of mutations in TNFRSF13B,^5,6,7,24 other causes have remained obscure. CVID has thus been hypothesized to represent a diverse collection of genetic lesions resulting in a similar immunologic phenotype. The MHC region has been associated with a myriad of complex diseases,²⁵ including immune-related conditions²⁶ and CVID.^8,9 MHC was robustly associated with CVID in our study, further validating the prior work that was not performed at the genome-wide level.

The SNP genotype association analysis presented here revealed a suggestive association outside of the MHC region with a locus encompassing ADAM28, ADAM7, ADAMDEC1, and STC1 (P = 6.24 × 10⁻⁶) that we also replicated in an independent case-control cohort (P value range = 2.25 × 10⁻⁸ to .0314). The ADAM family proteins are zinc metalloproteases involved in diverse biologic processes, including immune responses. Interestingly, the metallopeptidase MMP27 was also nominally associated with CVID (P = 1.69 × 10⁻⁴). Proteins of the matrix metalloproteinase family are involved in the breakdown of extracellular matrix to promote routine physiological processes but might also facilitate disease pathogenesis. Related genes have demonstrated immunologic function involved in the regulation of cytokine release, T_H2 immune responses, and specific inflammatory processes. ADAM28 is also known as the lymphocyte metalloprotease MDC-L, which is expressed on the lymphocyte cell surface.²⁷ As such, it has been defined as a ligand for α4β1 integrin, which enables the adhesion of other leukocytes expressing this integrin. STC1, which is located in the same region, is a gene that encodes stanniocalcin 1 protein, which is involved in regulation of calcium, including in antioxidant pathways of marcophages. ²⁸ In this light calcium regulation within immune cells has been previously identified as aberrant in certain samples of patients with CVID.²⁹ UBXN10 might be a compelling candidate potentially involved in immunopathogenesis of primary antibody failure in that it encodes an ubiquitin-like protein, which, in addition to phosphorylation, has been shown to regulate nuclear factor κB activity.³⁰ Sphingosine-dependent protein kinase 1 (SDK1) is important in the survival of alveolar macrophages.³¹ DEPDC6 is a negative regulator of mammalian target of rapamycin signaling pathways, and RNA expression levels were found to be significantly different between those mice resistant to H5N1 influenza virus and those that were susceptible.³² No significant associations in the region of TACI were observed, and subjects with TACI mutations were not separately identified. We reviewed the reported amino acid changes and looked up their corresponding nonsynonymous SNP IDs (see Table E6 in this article’s Online Repository at www.jacionline.org). No imputation reference files, including these SNPs, were available to infer genotypes.

We additionally performed an SNP genotype association to the particular features of CVID to define potential common mechanistic links among the specific clinical and immunologic variants to potentially enable the prediction of CVID clinical phenotypes. Significant individual associations were made with all CVID variables studied (Table II). In this regard PFTK1 is a member of the CDC2-related protein kinase family found to be constitutively expressed at high levels in B-cell lymphomas³³ and also found to be associated with lymphoma in the subjects with this complication. Interestingly, SNPs in HAVCR1, which is also associated in the same subjects (P= 3.62 × 10⁻⁸), are genes that play a role in T_H cell development and the regulation of asthma and allergic diseases. ³⁴ FGF14, which is associated with LIP and low IgA levels, is a member of the fibroblast growth factor family, which has crucial roles in embryonic development, cell growth, morphogenesis, tissue repair, tumor growth, and invasion.³⁵ SNX31, which is associated with organ-specific autoimmunity, is a sorting nexin that might be involved in protein trafficking; SNX family proteins are subunits required for CD28-mediated T-cell costimulation.³⁶

Although these SNP association results are not powered for definitive conclusions, the CNV association analysis uncovered several novel genes that were significantly associated with CVID. In fact, 84 CNV deletions and 98 duplications were identified in 1 or more patients with CVID but were not found in any of the 3031 control subjects. Most were intraexonic and thus likely to disrupt the genes involved. Some of the genes potentially affected by the identified CNV were also discovered in the GWAS part of this study. Many others have direct or potential relevance to the immune system, and many were unique to individual patients, thus underscoring the great mechanistic diversity that is likely to underlie this collection of disorders, which is also reflected in the variability in clinical presentation and disease natural history.¹ Among those, we noted a highly significant number of subjects with duplications in ORC4L, a gene previously associated with B-cell lymphoproliferative disorders.³⁷ This gene is essential for initiation of DNA replication and potentially in rapidly proliferating immune cells.

In addition to the potential of these findings to provide mechanistic insight into how CVID and its subphenotypes arise, we used SVM to exploit the overall genetic uniqueness of patients with CVID compared with healthy control subjects. These results suggest that the genetic signatures of the disease might also allow for accurate prediction of CVID based on genetic profiling. This has the potential to greatly improve the clinical management of the CVID and address a major unmet need in the field of primary immunodeficiency. By using current diagnostic criteria, diagnostic delay of multiple years can occur³⁸ which can in part be because of the evolving immunoglobulin phenotype.³⁸ This can lead to repeated infection and irreparable organ damage during the period of diagnostic delay. The ability to determine increased likelihood for CVID at the time of initial clinical suspicion could allow for improved intervention and reduced morbidity.

This study represents the first genome-wide population-based study of CVID. The use of the relatively large cohorts of this rare disorder assembled here was essential both to discover and confirm the findings and demonstrates the potential of genome-wide association in complicated polygenic rare diseases. This type of unbiased study has discovered many novel loci that might underlie the development of CVID. It can provide clues to the pathogenesis of the heterogeneous clinical complications and subtypes of CVID, providing a solid foundation for further studies to understand the mechanism, interplay, and clinical manifestations of CVID, with the immediate potential of improving clinical management through enhanced diagnosis. Finally, the great diversity of genetic findings identified with regard to unique CNVs substantiates the hypothesis that CVIDs are a collection of diverse mechanisms leading to complex phenotypes.

Clinical implications.

We define CVID as a genetically heterogeneous condition through genome-wide analysis. Patients with CVID can be distinguished from healthy subjects by using multiparameter predictive algorithms that could prove useful in facilitating diagnosis.

Abbreviations used

ADAM

A disintegrin and metalloproteinase

CNV

Copy number variation

CVID

Common variable immunodeficiency

GWAS

Genome-wide association study

LIP

Lymphoid interstitial pneumonitis

ORC

Origin recognition complex

SNP

Single nucleotide polymorphism

SVM

Support Vector Machine

TACI

Transmembrane activator and calcium modulator and cyclophilin ligand interactor