Copy-number variants differ in frequency across genetic ancestry groups

Summary

Copy-number variants (CNVs) have been implicated in a variety of neuropsychiatric and cognitive phenotypes. We found that deleterious CNVs are less prevalent in non-European ancestry groups than they are in European ancestry groups of both the UK Biobank (UKBB) and a US replication cohort (SPARK). We also identified specific recurrent CNVs that consistently differ in frequency across ancestry groups in both the UKBB and SPARK. These ancestry-related differences in CNV prevalence present in both an unselected community population and a family cohort enriched with individuals diagnosed with autism spectrum disorder (ASD) strongly suggest that genetic ancestry should be considered when probing associations between CNVs and health outcomes.

The authors report that CNVs annotated as recurrent or deleterious vary in frequency across genetic similarity groups in both the UK Biobank and SPARK. They also identify individual recurrent CNVs previously associated with neuropsychiatric phenotypes that consistently differ in frequency across genetic similarity groups in both cohorts.

Main text

Copy-number variants (CNVs), which are duplicated or deleted genomic segments larger than 1,000 bp,¹ are associated with a wide range of human complex traits^2,3,4,5,6,7 and diseases.^8,9,10,11 They are especially well studied in neurodevelopmental^12,13,14,15 and psychiatric disorders,^{4,15,16,17,18} most notably autism spectrum disorder (ASD)^19,20,21,22 and schizophrenia.^23,24,25 Unfortunately, most CNV association studies have been limited to European (EUR) ancestry groups^{2,5,6,7,8,9,11} or pooled across ancestry groups.^{10,13,14,15,17,19,20,21,23} While some studies have characterized CNVs in African (AFR)^26,27,28,29 and other non-EUR ancestry groups,^15,30,31,32 efforts to compare CNV frequencies across ancestry groups have been inconclusive due to limited sample sizes.^{14,33,34,35,36,37,38,39,40} Since genetic variation among human populations^41,42 contributes to differential disease risk,^43,44,45,46 immune response,^47,48 and pharmacogenomics,^49,50 it is plausible that ancestry-related differences in CNV frequency could impact precision medicine. Hence, we took advantage of previously overlooked diversity within the UK Biobank (UKBB) to compare CNV frequency across four genetic ancestry groups: individuals with inferred EUR (n = 51,334), south Asian (SAS; n = 8,848), or AFR (n = 8,447) ancestry plus a subset of EUR-ancestry individuals who self-identified as “white British” (WB; n = 385,636) (Figure S1). Even though the SAS and AFR groups each represent about 2% of the UKBB, they are orders of magnitude larger than the non-EUR groups included in previous CNV studies.

We called autosomal CNVs consisting of at least 50,000 bp from microarray data using PennCNV and QuantiSNP (supplemental methods). When we considered all CNVs, we found that the carrier frequency was generally similar across ancestry groups. Nonetheless, there were more deletion (DEL) carriers in the EUR and SAS groups and duplication (DUP) carriers in the WB group than expected given the group size (Figure 1A). For all ancestry groups, DEL carriers were more common than DUP carriers.

Figure 1.

CNV carrier prevalence by ancestry group in the UK Biobank

(A) When CNVs are not filtered based on recurrence or burden, deletion (DEL) carrier prevalence is higher than expected for the South Asian ancestry group (SAS), and duplication (DUP) carrier prevalence is lower than expected for the African ancestry group (AFR) (χ² = 146.7).

(B) When considering only recurrent CNVs, there were fewer AFR DEL and DUP carriers than expected (χ² = 48.145).

(C) Filtering instead on burden (total 1/LOEUF ≥ 5.7), there were fewer DEL and DUP carriers than expected for both AFR and SAS (χ² = 126.89).

(D) Limiting to carriers of recurrent CNVs with total 1/LOEUF ≥ 5.7, DEL carrier prevalence was lower than expected for both AFR and SAS, but DUP carrier prevalence was only lower than expected for AFR (χ² = 61.397). Plus signs indicate significantly higher than expected carrier prevalence, and minus signs indicate significantly lower than expected carrier prevalence. Ancestry-specific carrier prevalence was computed as the number of carriers of at least one DEL (or DUP) divided by the total number of individuals in that ancestry group. Simulated p values (2,000 replicates) were <0.0005 for all four χ² tests of independence. ∗∗False discovery rate (FDR)-corrected p < 0.005; ∗∗∗FDR-corrected p < 0.0005. An additional category (“neither”) that was included in each χ² test is not shown in the bar charts.

Given that CNVs have a range of effect sizes, as judged by their association with health outcomes or by their degree of evolutionary constraint, we considered several subsets of CNV carriers. First, we limited our analyses to carriers of a pre-selected set of recurrent CNVs with common breakpoints that were previously observed in multiple individuals and found to be associated with neuropsychiatric phenotypes (Table S1). We observed 50 unique recurrent DEL (Table S2) and 60 unique recurrent DUP (Table S3) at these loci in the UKBB. Of the 110 unique recurrent CNVs observed in the combined sample, 106 were present in the WB. The four CNVs not observed in the WB group were present in the EUR group, where we observed 77 unique recurrent CNVs. The 40 unique recurrent CNVs observed in the AFR group and the 36 in the SAS group were subsets of those observed in the WB and EUR groups; none of the 110 recurrent CNVs were unique to the non-EUR ancestry groups in the UKBB. There were significantly fewer AFR-ancestry recurrent DEL and DUP carriers than expected under the null hypothesis that recurrent CNV carrier frequency is independent of ancestry group (Figure 1B).

Second, we filtered CNV carriers using the loss-of-function observed/expected upper bound fraction (LOEUF)⁵¹ constraint metric. When we limited our analyses to carriers of CNVs with constraint scores equivalent to disruption of at least two predicted loss-intolerant genes (i.e., total 1/LOEUF ≥ 5.7 summed across the genes within the CNVs carried by an individual), ancestry-related differences persist and become even stronger. Consistent with previous evidence that DELs in coding sequences are under stronger purifying selection than DUPs,⁵² the burden-filtered carrier frequency for DELs was substantially lower than that for DUPs for all four ancestry groups. The deleterious DEL carrier frequency for WB individuals was nearly twice that of AFR or SAS individuals, both when we filtered by 1/LOEUF summed across all CNVs (Figure 1C) and when we limited the summation to recurrent CNVs (Figure 1D). Likewise, carriers of deleterious DUPs were more prevalent in the WB group and less prevalent in the AFR group than expected. We obtained similar results when we excluded all individuals related at the third-degree or closer (Figure S2; Table S4).

We also examined the prevalence of individual recurrent CNVs. To maximize power, we limited our analysis to the 11 recurrent CNVs (5 DELs and 6 DUPs) that each had a total of ≥275 observations across the four ancestry groups. We found ancestry-related differences in the prevalence of six of these CNVs (Figure 2), with some CNVs being more prevalent in WB individuals but others occurring at higher rates in the AFR and SAS groups. While there were some related individuals among the carriers of some of the CNVs, the pattern of frequency differences was essentially unchanged when we excluded all individuals with third-degree or closer relatives from the dataset (Figure S3; Table S4). Hence, the observed differences cannot be explained by related individuals carrying the same CNVs.

Figure 2.

Prevalence of individual recurrent CNVs across ancestry groups in the UK Biobank

There were significant differences between expected and observed counts for 6 out of the 11 recurrent CNVs selected for analysis, χ² = 425.3, simulated p < 0.0005 (2,000 replicates). Carrier prevalence was calculated as the number of carriers of a given recurrent CNV divided by the total number of individuals in that ancestry group. Plus and minus signs indicate that the standardized residuals were statistically significantly higher or lower than zero, respectively, after FDR correction. ∗FDR-corrected p < 0.05; ∗∗FDR-corrected p < 0.005; ∗∗∗FDR-corrected p < 0.0005. An additional category (“none of the above”) that was included in the χ² test is not shown in the bar chart.

Ascertainment bias is another potential explanation for the lower rate of deleterious CNVs in some ancestry groups within the UKBB. Immigrants may be less likely to participate in the UKBB, especially if they are carriers of deleterious CNVs. While EUR, AFR, and SAS individuals were more likely than WB individuals to have been born outside the UK or Ireland, the proportions of SAS and AFR individuals who were immigrants did not significantly differ for recurrent CNV carriers compared to non-carriers (Figure S4), suggesting that differential immigration rates cannot explain the differences in CNV prevalence across ancestry groups.

Demographic differences between the UKBB ancestry groups could also contribute to the observed differences in CNV prevalence. On the whole, UKBB participants are more likely to be female, healthier, wealthier, and older than the general population.⁵³ However, individuals in the SAS and AFR ancestry groups tend to be younger than those in the WB and EUR groups, and the female-to-male ratios differed across the ancestry groups (Tables S5 and S6). Furthermore, median Townsend deprivation index scores suggest that there are socioeconomic differences that could be associated with ancestry, sex, and recurrent CNV carrier status (Table S7). Given that pathogenic CNVs have been linked to socioeconomic status,⁵⁴ sex,⁵⁵ and age at death,⁵⁶ we used propensity score matching to balance the sample sizes and control for the potential effects of these variables on CNV carrier rates by downsampling the WB group to create two subgroups matched to the AFR and SAS groups (Figure S5; Tables S8 and S9). The AFR-ancestry group had significantly lower odds of carrying both unfiltered and 1/LOEUF-filtered recurrent DELs and DUPs than its matched WB group (all two-sided Fisher’s exact test p <0.0000005; Figure 3). Indeed, the matched comparisons yielded more pronounced differences than the unmatched ones despite having smaller sample sizes. In contrast, differences between the SAS and matched WB group were somewhat attenuated. Nonetheless, the SAS group showed significantly lower odds of carrying 1/LOEUF-filtered recurrent DELs when compared to its matched WB group (p < 0.00001).

Figure 3.

Odds of carrying deleterious CNVs in AFR- and SAS-ancestry individuals compared to WB individuals in the UK Biobank after propensity-score matching on Townsend deprivation index, age, and sex

Odds were computed for each ancestry group as the number of carriers divided by the number of non-carriers of a given class of recurrent CNV. ORs were computed as AFR odds divided by WB odds (purple dots) or SAS odds divided by WB odds (blue dots). Error bars indicate 95% Fisher confidence limits for the OR computed using the matched data, and the p value is for the corresponding two-sided Fisher’s exact test for that OR. The red dashed lines indicate the expected OR of 1 (i.e., equal odds). AFR odds were lower than the WB odds at each level of filtering, but SAS odds were not consistently lower than the WB odds. For comparison purposes, ORs were also calculated using the data from all 385,636 white British (WB), 8,447 African (AFR), and 8,848 South Asian (SAS) individuals and are included as open circles.

We also used the matched datasets to compare the odds of carrying our 11 recurrent CNVs of interest (Figure 4) and found that most of the observed ancestry-related differences in CNV prevalence remained significant. Using two-sided Fisher’s exact tests of the carrier odds ratios (ORs), we found that the AFR group had significantly lower odds of carrying 2q13 NPHP1 DEL, CRYL1 DEL, 15q11.2 DEL, 15q13.3 BP4.5-BP5 CHRNA7 DUP, 16p13.11 DUP, and 22q11.2 proximal (with low-copy-number repeat A [LCR-A]) DUP, as well as higher odds of carrying 2q13 NPHP1 DUP compared to its matched WB group. The SAS group also had significantly lower odds of carrying 2q13 NPHP1 DEL, CRYL1 DEL, 15q11.2 DEL, 15q13.3 BP4.5-BP5 CHRNA7 DUP, and 22q11.2 proximal (with LCR-A) DUP, and higher odds of carrying 2q13 NPHP1 DUP compared to its matched WB group. Additionally, the SAS group had higher odds of carrying ZNF92 DEL and 15q11.2 DUP compared to its matched WB group.

Figure 4.

Odds of carrying individual recurrent CNVs for non-EUR compared to WB-ancestry groups in the UK Biobank

Odds were computed for each ancestry group as the number of carriers divided by the number of non-carriers of a given recurrent CNV. ORs were computed as AFR odds divided by WB odds (purple dots) or SAS odds divided by WB odds (blue dots). Error bars indicate 95% Fisher confidence limits for the OR computed using matched data, and the p value is for the corresponding two-sided Fisher’s exact test for that OR. The red dashed lines indicate the expected OR of 1 (i.e., equal odds). Compared to the WB odds, AFR odds were significantly lower for 6 and significantly higher for 1 of the 11 selected recurrent CNVs. SAS odds were significantly lower than WB odds for 5 of the same 6 recurrent CNVs and significantly higher than WB for 3 recurrent CNVs, including the same one observed to be higher for AFR. The two blue half circles correspond to ORs of zero; neither of those recurrent CNVs were observed in the SAS-ancestry group. For comparison purposes, ORs were also calculated using the data from all 385,636 WB, 8,447 AFR, and 8,848 SAS individuals and are included as open circles.

Finally, we replicated our study using SPARK, a younger US cohort enriched with individuals diagnosed with ASD and intellectual disability (ID) (Tables S10–S12). We observed 51 unique recurrent DEL (Table S13) and 51 unique recurrent DUP (Table S14) across the three largest inferred ancestry groups (Figure S6) comprising 46,869 EUR, 7,870 admixed American (AMR), and 3,680 AFR individuals. We used propensity score matching to balance the sample sizes and control for potentially confounding effects of ASD and ID status, age, and sex on recurrent CNV carrier rates by downsampling the EUR group to create two subgroups matched to the AFR and AMR groups (Figure S7; Tables S15 and S16). Replicating our UKBB results, we found that the SPARK AFR group had significantly lower odds of carrying both unfiltered and 1/LOEUF-filtered recurrent DELs and DUPs than its matched EUR group (all two-sided Fisher’s exact test p < 0.005; Figure 5). The AMR group had significantly lower odds of carrying both unfiltered and 1/LOEUF-filtered recurrent DUPs (two-sided Fisher’s exact test p < 0.005; Figure 5).

Figure 5.

Odds of carrying deleterious CNVs in AFR- and AMR-ancestry individuals compared to EUR individuals in SPARK after propensity-score matching on ASD and ID status, age, and sex

Odds were computed for each ancestry group as the number of carriers divided by the number of non-carriers of a given class of recurrent CNV. ORs were computed as AFR odds divided by EUR odds (purple dots) or AMR odds divided by WB odds (blue dots). Error bars indicate 95% Fisher confidence limits for the OR computed using matched data, and the p value is for the corresponding two-sided Fisher’s exact test for that OR. The red dashed lines indicate the expected OR of 1 (i.e., equal odds). AFR carrier odds were lower than the EUR odds at each level of filtering, but AMR odds were significantly lower only for recurrent DUP carriers. For comparison purposes, ORs were also calculated using the data from all 46,869 EUR, 3,680 AFR, and 7,870 admixed AMR individuals and are included as open circles.

We also used the matched datasets to compare the odds of carrying the same 11 recurrent CNVs we analyzed for the UKBB along with 2 additional recurrent CNVs that had at least 75 copies each in SPARK despite not meeting our inclusion criteria for the UKBB (Figure 6). Replicating our UKBB results, the SPARK AFR group showed significant differences for 2q13 NPHP1 DEL, 15q11 DEL, 2q13 NPHP1 DUP, 15q13.3 BP4.5-BP5 CHRNA7 DUP, and 16p13.11 DUP carrier odds relative to its matched EUR group (two-sided Fisher’s exact test p < 0.05). Additionally, the SPARK AFR group had significantly lower odds of carrying NRXN1 DEL and higher odds of carrying 16p12.1 DEL relative to its matched EUR group, and the SPARK AMR group had significantly lower odds of carrying NRXN1 DEL, 15q11.2 DEL, and 1q21 TAR DUP relative to its matched EUR group (two-sided Fisher’s exact test p < 0.05).

Figure 6.

Odds of carrying individual recurrent CNVs for non-EUR- compared to EUR-ancestry groups in SPARK

Odds were computed for each ancestry group as the number of carriers divided by the number of non-carriers of a given recurrent CNV. ORs were computed as AFR odds divided by EUR odds (purple dots) or AMR odds divided by EUR odds (blue dots). Error bars indicate 95% Fisher confidence limits for the OR computed using matched data, and the p value is for the corresponding two-sided Fisher’s exact test for that OR. The red dashed lines indicate the expected OR of 1 (i.e., equal odds). The purple half circle corresponds to an OR of zero; this recurrent CNV was not observed in the AFR-ancestry group. For comparison purposes, ORs were also calculated using the data from all 46,869 EUR, 3,680 AFR, and 7,870 admixed AMR individuals and are included as open circles.

We demonstrated that ancestry-related differences in CNV carrier prevalence are present in both unselected community populations (UKBB) and cohorts enriched with ASD-diagnosed individuals (SPARK). We replicated the observed differences between the AFR and WB cohorts in the UKBB by comparing the AFR and EUR cohorts in SPARK, which is notable given the ascertainment differences, differing genotyping platforms, and presumed genetic differences between the homogeneous WB subset of the UKBB and an EUR-ancestry cohort from the United States.⁵⁷ Furthermore, SAS (UKBB) and AMR (SPARK) ancestry groups also exhibited unique patterns of CNV prevalence, demonstrating that differences in CNV carrier prevalence cannot be generalized as EUR vs. non-EUR differences.

One potential explanation for the difference in prevalence of specific CNVs across ancestry groups is that population variation in the flanking sequence may affect the probability of deletions and duplications in a region. For example, chromosomal regions with polymorphic inversions have been shown to be enriched for recurrent CNVs associated with developmental delay and neuropsychiatric disorders.⁵⁸

Given that AFR populations have been shown to have greater genetic diversity,⁵⁹ the finding of fewer rare CNVs in the AFR groups of UKBB and SPARK is somewhat surprising. One possible explanation for our ancestry-divergent results could be the Eurocentric focus of previous studies of genetic variants.⁴³ To focus on deleterious CNVs, we limited some of our analyses to recurrent CNVs that had been previously implicated in neuropsychiatric disorders and/or filtered our results based on the LOEUF metric. Our targeted recurrent CNVs were originally identified in cohorts dominated by EUR-ancestry individuals, so it is conceivable that there are as-yet undiscovered CNVs with medical relevance that are more common in other ancestry groups. Also, the LOEUF constraint metric was derived from the gnomAD version 2 reference database, which includes sequences from 64,603 individuals from EUR populations and only 12,487 individuals from AFR (or African American) and 15,308 individuals from SAS populations.⁵¹ We expect that the discovery of additional recurrent CNVs, especially those smaller than 50 kb, will be facilitated by the increasing availability of long-read whole-genome sequence data collected from diverse populations, such as All of Us.^40,60 In any case, it is likely that differing linkage disequilibrium patterns³⁵ and flanking sequences⁶¹ also contributed to the ancestry-related differences in CNV frequency that we found. Future CNV studies utilizing long-read sequences should clarify the extent to which polymorphic differences in the flanking sequence across populations contribute to the observed differences in CNV frequency.

These findings are limited by the sample sizes of the AFR, SAS, and AMR ancestry groups in the UKBB and SPARK, which are much smaller than those of the WB and EUR ancestry groups. Larger CNV studies of diverse samples are needed. Additional limitations include our focus on only CNVs >50 kb, a by-product of using genotype array data to call CNVs, and the choice of LOEUF as a metric of the deleteriousness of a CNV. Given that LOEUF is focused on coding genes, non-coding CNVs that impact important regulatory elements may have been missed in our filtered set of deleterious CNVs with 1/LOEUF ≥ 5.7. Also, while the recurrent CNVs we selected for analysis (Table S1) had all been associated previously with neuropsychiatric disorders,^{13,23,62,63,64,65,66,67,68,69,70,71} their ClinGen dosage sensitivity scores⁷² do not universally implicate them as being clinically relevant. Of the recurrent CNVs that we identified as having significant ancestry-related frequency differences (Figures 4 and 6), ClinGen currently classifies only the NRXN1 and 15q11.2 DEL as having sufficient evidence for haploinsufficiency, the 16p12.1 DEL as having emerging evidence for haploinsufficiency, the 22q11.2 proximal (with LCR-A) DUP as having sufficient evidence for triplosensitivity, and the 16.p13.11 DUP as having emerging evidence for triplosensitivity (Table S17). The strengths of the study include the use of propensity score matching to address potential sources of bias and the consistency of results across samples with very different demographic profiles and ascertainment schemes.

Although classifying individuals into continental ancestry groups imposes discrete categories onto what is actually a continuum of human genetic variation,^73,74 it has nonetheless been a useful approach for considering the potential effects of population structure on genome-wide association analyses.^75,76 To the extent that there is admixture between the groups we have defined, it would be expected to reduce the genetic differences between groups. Thus, the differences we have described should be viewed as conservative estimates. We recommend that future CNV association studies also be adjusted for population structure, ideally in a quantitative way that reflects the continuous spectrum of genetic variation, to limit the risk of spurious discoveries.

Data and code availability

This paper solely uses publicly accessible, deidentified data. UK Biobank and SPARK data are available to qualified researchers upon application. We provide our custom Python code and a detailed description of our CNV calling pipeline at https://martineaujeanlouis.github.io/MIND-GENESPARALLELCNV/.

Acknowledgments

This work used the UK Biobank, a major biomedical database, under application number 40980. We are grateful to all the families in SPARK, the SPARK clinical sites, and the SPARK staff, and we appreciate obtaining access to phenotypic and genetic data on SFARI Base. Our research was funded by the National Institute of Mental Health (U01MH119690).

Declaration of interests

The authors declare no competing interests.

Web resources

Clinical Genome Resource (ClinGen), https://search.clinicalgenome.org/

CNV calling pipeline, https://martineaujeanlouis.github.io/MIND-GENESPARALLELCNV/

KING: Kinship-based Inference for GWAS, https://www.kingrelatedness.com/

The R Project for Statistical Computing, https://www.r-project.org

R package e1071, https://cran.r-project.org/web/packages/e1071/e1071.pdf

Simons Powering Autism Research (SPARK), https://www.sfari.org/resource/spark/

UK Biobank, https://www.ukbiobank.ac.uk/