Effects of gene dosage on cognitive ability: A function-based association study across brain and non-brain processes

Knowledge gap

Overall, it has been difficult to investigate the broad landscape of ultra-rare CNVs potentially involved in neurodevelopmental traits, such as cognitive ability. As a result, we have a limited understanding of the full range of gene-dosage-sensitive biological processes linked to cognitive ability.

To circumvent the issue of power, research groups, including ours, have implemented alternative approaches aggregating rare variants disrupting genes with similar constraint scores in order to perform “constraint burden” association studies^{11,12,19–21,24}. These burden analyses showed that genes with increasing intolerance to haploinsufficiency were associated with increasing effect sizes on cognitive ability and risk for psychiatric illnesses, such as ASD, schizophrenia, and bipolar disorder^19,21. Similarly, studies have developed methods to aggregate common variants²⁵, demonstrating that a robust association with a condition (e.g., ASD) can be established at the group level when individual single nucleotide polymorphisms (SNPs) do not meet genome-wide criteria for association.

In this study, we aimed to investigate the full range of gene-dosage-sensitive biological processes linked to cognitive ability. To this mean, we analyzed all CNVs >50 kilobases in 258k individuals across 6 cohorts from the general population. The CNV-level GWAS identified the first CNV associated with higher cognitive ability. We then performed functional-burden associations with cognitive ability across 6,502 gene-sets corresponding to gene functions at the tissue, cell type, and molecular levels. Findings reveal that most biological functions have preferential effects on cognitive ability when either deleted or duplicated. As a result, we observed a negative correlation between the effects of deletions and duplications across all levels of biological observation independently of intolerance to haploinsufficiency.

Cohorts

We analyzed 258,292 individuals from six general population cohorts^26–31, which can be further divided into 9 sub-cohorts based on cognitive assessment (Table1). Three additional autism cohorts^32–34 were only used for sensitivity analyses (Supplementary table 1, figure 1, and Methods). Each cohort received approval from their local institutional review boards. Parents/guardians and adult participants gave written informed consent, and minors gave assent.

• Download figure
• Open in new tab

Figure 1. CNV GWAS on general cognitive abilities at the gene level.

(A) Proportion of genes deleted (red) or duplicated (blue) at least once in the dataset, among all genes in the human genome. Deleted or duplicated genes observed in less than 30 carriers (light color) and 30 or more (dark color) as well the proportion of genes not observed in any CNV (grey),(n=13,747). (B) The majority of deleted or duplicated genes were observed in more than 30 carriers. (C) Venn diagram illustrating the overlap between genes deleted and-or duplicated at least once in the dataset. (D) The Miami plot illustrates the −log10 transformed p-value of the association with cognitive ability for each gene included in deletions (red) at the top, and duplications (blue) at the bottom, along the genome. Adjacent chromosomes are shown in alternating light and dark colors. Triangles represent significant genes after FDR correction, while circles represent non-significant genes. The direction of the triangle indicates the effect size. The dash line represents the nominal significant pvalue threshold. (E) A specific duplication, chr2:109,510,927-110,376,563 (including EDAR, LOEUF=0.91; SH3RF3, LOEUF=0.53; SEPT10, LOEUF = 1.17; SOWAHC, LOEUF=0.77), exhibited a previously unobserved positive effect on cognitive ability in the CNV GWAS. (F) To further investigate this positive effect, we conducted a post hoc analysis using a two-sided t-test on a homogeneous cohort with consistent technology, ancestry, and phenotype, aiming to eliminate biases. Our focus was specifically on individuals of white British ethnicity in the UK Biobank (UKBB) with adjusted fluid intelligence. In the left part of the analysis (F), individuals were categorized into three groups: carriers of the CNV of interest (green), non-carriers of this specific CNV but carrying other exonic CNVs (light orange), and non-carriers of any exonic CNV (blue). The t-tests were performed on fluid intelligence (FI) adjusted for sex, 1 to 10 PC for ancestry, and age. In the right part of the analysis (F), two groups were defined: carriers of the CNV of interest (dark pink) and non-carriers (light blue). The t-tests were conducted on fluid intelligence adjusted for sex, ancestry, age, and the burden of 1/LOEUF for deletions and duplications. For the duplication of chr2:109,510,927-110,376,563 observed in the CNV GWAS (F), the carriers exhibited significantly higher cognitive ability measures compared to both other groups. Furthermore, when we weighted the cognitive ability by the burden of 1/LOEUF for deletions and duplications, a positive effect was also observed among carriers of the CNV of interest.

Measures of cognitive ability

General cognitive ability was measured by either non-verbal intelligence quotient (NVIQ), fluid intelligence (fluid intelligence questions), or general intelligence factor (g-factor)¹⁵. Measures of cognitive ability were z-scored within each cohort based on sex and age (table 1, Supplementary table 1 and 2; Supplementary Methods).

Array processing and genotyping

All arrays used for genotyping and CNV calls are detailed (Supplementary table 1 and Supplementary method)¹². We used pLink³⁵ to check for duplicates, sex, and family relationships for each participant using a standard pipeline^11,12. Ancestry principal components and categories were computed using KING (www.kingrelatedness.com) with the 1000 Genomes Project^11,12 as the reference population.

CNV calling and quality control

For CNV calling, we used PennCNV³⁶ and QuantiSNP³⁷. For array and CNV quality control, we applied standard filtering parameters (Supplementary Methods, https://martineaujeanlouis.github.io/MIND-GENESPARALLELCNV/)^11,12. For the MSSNG dataset³⁴, we used published CNVs called on whole genome sequencing³⁸. To harmonize CNV calls across arrays with different probe densities, we only considered CNVs encompassing at least 10 probes present on all array technologies. We excluded from the analyses all individuals carrying a CNV ≥10Mb or a mosaic CNV.

Gene annotation

We annotated autosomal CNVs using bedtools (https://bedtools.readthedocs.io/en/latest/) with Gencode V19 (hg19, Ensembl gene name, https://grch37.ensembl.org/index.html).

LOEUF-based and function-based gene sets

We defined 38 overlapping gene-sets based on LOEUF values (loss-of-function observed/expected upper bound fraction, gnomAD version 2.1.1)³⁹ (Supplementary methods). LOEUF range from 0 to 2, and values below 0.35 are suggestive of intolerance.

We defined 269 gene-sets based on relative gene expression (Z-score >1SD) in 13 adult^40,41 and 16 fetal⁴² brain cell types⁴³, as well as bulk tissue from 215 brain regions (Human Protein Atlas, HPA v.22)⁴⁴ and 25 non-brain organs (GTEx^44,45, Supplementary table 3). As a sensitivity analysis, we defined the same gene-sets based on a previously published “Top Decile Expression Proportion” (TDEP)⁴⁶ method. The former and the latter methods favour relative and specific expression respectively. Both methods exclude 1,370 and 5,369 genes that are not assigned to any tissue in GTEx. We also used 6,233 functional gene-sets based on 6,130 GOterms^47,48 (Ensembl v.109, April 2023), and 103 Synapse ontology terms (SynGO⁴⁹). Throughout this study, we only considered gene-sets meeting the following 3 criteria: i) those with more than 10 genes, ii) those disrupted by ≥ 30 CNV carriers, and iii) those with at least 20% of their genes affected by CNVs.

Statistical analyses

Analyses were performed using R version 4.0.1(http://www.R-project.org.), with “meta”(https://cran.r-project.org/web/packages/meta/index.html) and “metafor”(https://cran.r-project.org/web/packages/metafor/index.html) packages for meta-analyses. Python 3.10.2(https://www.python.org) with “scipy 1.11.2”(https://pypi.org/project/scipy/), “statsmodels 0.13.5” (https://www.statsmodels.org) and “word-cloud 1.9.2” (https://amueller.github.io/word_cloud).

LOEUF-based and Function-based burden association tests

To estimate the effect on cognitive ability of gene-sets (and their corresponding biological functions or LOEUF categories), we adapted a previously published model¹². We performed a linear model for each of the 38 LOEUF gene categories and each of the 6502 functional gene-sets. The outcome was cognitive ability measured in each individual. The explanatory variable was the sum of genes fully encompassed in a CNV for a gene-set of interest (Figure 2A). Since CNVs are multigenic, the effect size estimated for a given gene set may be inflated. Therefore, all models were adjusted for the total number of genes within the CNV but not members of the gene set of interest. These latter genes were categorized into three variables: ID genes (only for Function-based gene-sets), genes with LOEUF<1, and genes with LOEUF≥1. Other covariates included ancestry (10 PCs), age, and sex. Models were computed for deletions or duplications, separately (Supplementary method).

• Download figure
• Open in new tab

Figure 2. Effect sizes of autosomal coding genes on general cognitive abilities based on their LOEUF values.

(A) The functional burden test is a linear model estimating the mean effect size of all CNVs fully encompassing genes assigned to a biological function of interest. Because many CNVs are multigenic, the model is adjusted for genes included in a CNV but not assigned to the biological function of interest. (B) Sliding window (Supplementary method statistical model 2) estimating the effect size on cognitive ability of deletions (top) and duplications (bottom) for 38 LOEUF categories (we slide a window of size 0.15 LOEUF units, in increments of 0.05 units thereby creating 38 categories across the range of LOEUF values) and definitive ID genes curated by ClinGen. Estimates were computed using a meta-analysis (circles) as well as a pooled dataset (squares). The red dashed line defines intolerant genes (LOEUF<0.35). (C) Heatmap showing the effect size (color scale) on cognitive ability of deletions and duplications across a range of sensitivity analyses removing non-europeans, older participants, large multigenic CNVs (those with a sum of 1/LOEUF >60, >40, >20 corresponding to values of well known recurrent CNVs: 22q11.2, 16p11.2 and TAR respectively) as well as adding a neurodevelopmental dataset (autism spectrum disorder). All estimates were computed on the pooled dataset.

tagDS

Previous publications have reported that effect size on cognitive ability is U-shaped. Effects of deletions are 2–3 times higher than duplications¹². To test whether the deletion/duplication effect size ratio deviates from the expected average ratio (specifically 2.4 in our dataset), we developed the trait-associated gene-dosage sensitivity score (tagDS) (Supplementary method). tagDS was normalized (Z-scored) based on a null distribution of tagDS values. The latter distribution was obtained using effect sizes on cognitive ability computed for deletions and duplications across 570,000 randomly sampled gene-sets. A tagDS of 0 suggests that the deletion/duplication effect-size ratio is equal to the expected ratio. A gene-set with a tagDS > 2 indicates that its deletion/duplication effect-size ratio on cognitive ability is beyond 2 standard deviations of the null distribution (i.e., larger effect sizes are biased towards deletions).

Gene dosage may be associated with higher cognitive ability

We observed that 15.6% of the 258,292 individuals included in the analysis carried at least one autosomal CNV fully encompassing one or more coding genes. Among 18,451 autosomal coding genes with LOEUF values, 35% and 64.9% were fully encompassed in one or more deletion and duplication, respectively (75% across CNVs), with 40% observed in both deletions and duplications (Figure 1A, B, C). Most genes were in ultra-rare CNVs (<1/10,000) with fewer than 30 carriers (Figure 1C). We computed a linear regression model (gene-level GWAS, Supplementary method statistical model 1) on general cognitive ability for 241 and 596 genes covered by at least 30 deletions or duplications, respectively, in 258,292 individuals from the general population (Figure 1D). We identified 9 deletions encompassing a total of 69 genes and 10 duplications encompassing a total of 123 genes with previously published negative effects (Supplementary table 4) that persisted when we conducted a meta-analysis across 9 sub-cohorts defined by cognitive assessments (Table 1, Supplementary figure 2). We identified a novel association between a duplication at 2q12.3, and positive effects on cognitive ability (Figure 1E). This duplication observed in 36 individuals included 4 genes with a LOEUF≥0.35 (EDAR, SH3RF3, SEPT10, SOWAHC) and was observed at a similar frequency (1 to 2 in 10,000) across cohorts (Fisher’s exact p-value_FDR >0.05). Results were not related to ancestry, array platform, or cognitive assessment methods (Figure 1F). The positive effect remained significant when comparing 2q12.3 duplication carriers to individuals without any CNVs.

A large proportion of intolerant and tolerant genes modulate cognitive ability

Even with the current large sample sizes, CNVs cover only 3-4% of coding genes in our previous analyses. We know that a much larger proportion of the coding genome is involved in cognitive ability¹², we used burden association methods to explore all coding CNVs. We partitioned genes into 38 sets based on overlapping LOEUF categories and an ID gene-set(as defined by ClinGen Supplementary table 5), then calculated their 39 mean burden effect sizes using linear models adjusted for CNVs outside each category to prevent effect inflation from multigenic CNVs(c.f. methods, Supplementary method statistical model 2, Figure 2A). The 39 estimates provided by the meta-analysis across the 9 sub-cohorts were not different from those provided by aggregating these datasets (Figure 2B, Supplementary table 6 and 7). Therefore, all subsequent analyses were performed on the aggregated dataset. The effects of deletions were, on average, 2.4-fold higher than duplications, and we observed a positive correlation between the effect sizes of deletions and duplications across LOEUF categories (Spearman’s r=0.5, p_permutation=0.02; Supplementary figure 3). Negative effects on cognitive abilities were observed in 8 and 11 non-tolerant categories (LOEUF<1) for deletions and duplication, respectively. More intolerant the LOEUF category was, the more negative the effect size, with the ID gene-set having the largest effects. Of note, 2 and 3 categories showed positive effects for deletions and duplications, respectively. In other words, the effect sizes of these categories were significantly higher than the average effect of gene categories used to adjust the model. Sensitivity analyses showed no biases related to ancestry, large multigenic CNVs, or low-quality control scores (Figure 2C, Supplementary figure 4). Effect sizes of intolerant genes were higher when removing older age groups (>60 years old) (Figure 2C). Because the most intolerant CNVs are depleted in the general population, we included 3 ASD cohorts in a sensitivity analysis. This resulted in larger effects and smaller p-values for highly intolerant LOEUF categories without changing the effects of other categories ≥0.35 (Figure 2C).

Negative correlation between deletion and duplication effects on cognitive ability across brain regions

Previously published functional enrichment analyses^50,51 have focused on recurrent CNVs. We, therefore, developed a functional burden test to systematically investigate gene functions that may underlie the pervasive association between CNVs (too rare to reach individual association) and cognitive ability. We first focused on gene sets assigned to 215 adult brain regions. 91 and 94 were associated with cognitive ability when deleted and duplicated, respectively, and only 25 were impacted by both(c.f. methods, Figure 3A). This suggests that cognitive ability is preferentially affected by one CNV type, depending on the brain region. These preferential effects were supported by the negative correlation observed between the effect sizes of CNV type across all brain regions (Spearman’s r=-0.43, p_permutation=9×10⁻³; Figure 3B). Stratifying these brain gene-sets into 3 independent LOEUF categories provided the same negative correlations (Figure 3C). Sensitivity analysis showed that the negative correlation was not due to unbalanced power between deletions and duplications or the relative expression threshold used to define gene-sets (Supplementary figure 5). To further investigate gene-sets driving this negative correlation, we developed the “trait-associated gene-dosage sensitivity score” (tagDS), a normalized value reflecting whether a gene-set shows preferential effects on cognitive ability when either deleted or duplicated (cf method, Figure 3D). TagDS indicated that cerebral cortex gene-sets affected cognitive ability preferentially when duplicated, while the opposite was observed for non-cortical (subcortical and midbrain) gene-sets and deletions (Figure 3A, Mann-Whitney p_permutation=1×10⁻¹⁵). The same cortical/non-cortical gene dosage sensitivity was also observed when removing genes with low tissue specificity (Supplementary figure 5)⁴⁴.

• Download figure
• Open in new tab

Figure 3: Effects on cognitive ability of genes assigned to brain regions and cell types.

(A) Effects sizes on cognitive ability of gene sets assigned to 215 brain tissues/regions. Brain regions are color-coded and clustered (first row, Wards’method⁶⁷) based on the level of overlap (gray matrix) between their corresponding gene-sets. The average LOEUF value for each gene-sets is color-coded in the 2nd row. The mean effect sizes, on cognitive ability, of genes assigned to each brain region, are coded for deletions (3rd row) and duplications (4th row). TagDS values are represented in the 5th row. (B) Spearman correlation (black line) between the effect sizes of deletions and duplications across all gene-sets with FDR significant effects on cognitive ability for either deletions (downward triangle), duplications (upward triangle), or both (cross). P-values were obtained from permutations to account for the partial overlap between gene-sets. Gene-sets are color-coded based on their tagDS. The dashed line represents the average exome-wide duplication/deletion effect-sizes ratio. (C) The same negative correlations between deletion and duplication were observed across 3 independent LOEUF groups: <0.35 (intolerant; red), [0.35, 1.0[ (moderately intolerant; orange), and [1.0, 2.0] (tolerant; green). (D) Raw tagDS is the Euclidean distance to the whole-genome ratio of effect sizes. TagDS is normalized following the null distribution of random gene-sets of identical size. (E) Effect size of deletions and duplications encompassing genes assigned to 6 cortical layers, 7 adult brain cell types, and 16 fetal brain cell types. Clustering was calculated on the level of overlap between cell-type gene-sets (Wards’method⁶⁷). Purple and orange represent negative and positive effects on cognitive ability, respectively. Black edges indicate significant effects.

At the microstructure and cell-type level (6 cortical layers, 7 adult, and 16 fetal brain cell types based on their normalized gene expression, c.f. methods), we observed the same negative correlation (r=-0.70, p_permutation<1×10⁻³; Supplementary figure 6). The largest effects for deletions and duplications were observed in gene-sets assigned to fetal cell types. Deletions and duplications, respectively, showed preferential effects in non-neuronal (endothelial, glia) and neuronal (excitatory) cell types (Figure 3E).

Genes involved in non-brain tissues affect cognitive ability

There is a growing interest in whole-body health comorbidities among individuals with neurodevelopmental and psychiatric conditions as well as CNVs affecting cognition^18,52. We, therefore, asked if CNVs affecting genes assigned to non-brain tissues were also associated with cognitive ability. We used 37 gene-sets, defined by relative expression in 12 brain and 25 non-brain tissues (>1SD, Figure 4A). Many non-brain gene-sets showed effect sizes (Figure 4B) of similar magnitude to those observed for regional brain gene-sets. This was not explained by the level of overlap between brain and non-brain gene-sets (Figure 4A, B). We observe the same pattern of deletion-duplication negative correlation independently of the gene-set definitions (r=-0.64 p_permutation<1×10⁻³; Figure 4C; Supplementary figure 7). To understand how gene-set definitions influence these results, we first removed 8,194 genes with low-tissue specificity assigned to multiple gene-sets. The resulting effect sizes were correlated with the initial estimates (r=0.57, Supplementary figure 8). In fact, genes assigned to multiple tissues show higher intolerance (LOEUF) compared to tissue-specific genes (p=1×10⁻¹¹ to 3×10⁻¹⁶¹, figure 4D, Supplementary figure 9). To further investigate the impact on results of gene-sets definition, we tested 37 previously published gene-sets assigned to 37 GTEx tissues computed by the TDEP method (proportional gene expression)⁴⁶. This method, which emphasizes specificity, excludes 5,454 genes of which 1,586 and 696 are respectively moderately (LOEUF [0.35, 1[) and highly intolerant (LOEUF<0.35; Supplementary figure 9). Effect sizes were well correlated with our analysis excluding LTS genes (r=0.76), but TDEP gene-sets were unable to detect any effect for deletions across all tissues (Supplementary figure 8).

• Download figure
• Open in new tab

Figure 4: Effects on cognitive ability of CNVs affecting genes implicated in brain and non-brain tissues.

(A) We defined 37 gene-sets based on z-scored expression > 1SD. Expression of each gene was normalized across 37 tissues provided by GTEx. Gene sets were clustered (orange for brain tissues and blue for non-brain tissues) based on their overlap which is shown in the grayscale matrix. High overlap was observed between brain gene-sets (Wards’method⁶⁷), and much lower overlap was present across non-brain tissues and between brain and non-brain tissue. The mean LOEUF of each gene-set is color-coded in row 2. Effect sizes on cognitive ability and tagDS across tissues are color-coded in the 3rd row as well as in the body map (B), adapted from GTEx. Genes with low tissue specificity were defined by the Human Protein Atlas. (C) Spearman correlation (black line) between the effect sizes of deletions and duplications on cognitive ability. Downward and upward triangles, as well as crosses, represent significant effects for deletions, duplications and both respectively. Gene-sets are color-coded based on their tagDS. (D) Distribution of LOEUF values for genes assigned to 1, 2, 3, 4, or ≥5 gene-sets. The 5 groups are non-overlapping. Black lines represent the median LOEUF score for each group. The p-values are provided by the Mann-Whitney test between each category and the genes assigned to one gene-set.

The effects of deletion and duplication on cognitive ability are negatively correlated across all levels of biological observations

We asked if the deletion-duplication negative correlations observed for tissue-level gene-sets were also present at the molecular and cellular components levels. We first investigated 293 synaptic gene ontologies using SynGO⁴⁹. We observed that post-synaptic genes showed the largest negative effects on cognitive ability when deleted, and in contrast, pre-synaptic genes when duplicated (Figure 5A, Supplementary figure 10). As a result, the effects of the 2 opposing CNVs were negatively correlation across SynGO terms (r=-0.39, p_permutation=1×10⁻³; Supplementary figure 11).

• Download figure
• Open in new tab

Figure 5. Effects on cognitive ability of gene sets based on gene ontologies.

(A) Effect sizes of synaptic molecular functions and cellular component gene-sets as defined by SynGO⁴⁹ on cognitive ability (More details in the supplementary figure 10). Blue and red represent negative and positive cognitive ability tagDS, respectively. Ontologies with black edges indicate significant effects (FDR). The results are shown only for SynGO terms with more than 10 genes, observed at least 30 times in our dataset, and with a coverage greater than 20%. Note panel (A): 1) Regulation of modification of postsynaptic actin cytoskeleton, 2) Regulation of calcium-dependent activation of synaptic vesicle fusion, 3) Presynaptic modulation of chemical synaptic transmission, 4) Integral component of postsynaptic density membrane, 5) Synaptic vesicle membrane. (B) There is a negative correlation (Spearman) between the effect sizes of deletions and duplication across 601 GO-terms. (C) The same deletion-duplication negative correlation was observed across 3 independent LOEUF groups (intolerant <0.35: red, moderately intolerant [0.35, 1.0[: orange, tolerant [1.0, 2.0]: green). (D) We adapted the word cloud package, which groups GOTerms based on shared terminology. Y-axis: Sum of associations of each word with significant deleted and duplicated GOTerms. X-axis: Proportion of significant GOTerms for a given CNV type used for the association. “Positive” and “Negative” refer to “Positive regulation” and “Negative regulation”.

We extended our analysis to 6,130 GO-terms (and corresponding gene-sets), 5.0% and 3.5% of the GO-terms had an effect size on cognitive ability for deletions and duplications, respectively. A minority (0.7%) of GO-terms showed significant effects for both. We observed again a deletion-duplication negative correlation across GO-terms effect sizes (r=-0.54, p_permutation<1×10⁻³; Figure 5B), which remained significant across three independent levels of LOEUF stratification (Figure 5C). We asked if tagDS was similar to pHI (probability of haploinsufficiency) and pTS (probability of triplosensitivity), 2 previously published metrics that are highly correlated with each other (0.78) and with LOEUF (r=0.90 and 0.77, respectively). tagDS was unrelated to pHI and pTS scores⁵⁰ across GO-terms (Supplementary figure 15). Finally, several GO-terms, such as neuronal, synaptic, and cardiac functions, showed preferential effects when deleted, while the opposite was observed for cellular response functions, transport, metabolic processes, and signaling pathways. Furthermore, “positive regulation” GO-terms were more sensitive to deletions, while “negative regulation” terms showed preferential effects when duplicated (Figure 5D, Supplementary figure 12).