Bidirectional genetic overlap between bipolar disorder and intelligence

GWAS summary datasets of BD (n = 41,917 cases and 371,549 controls) [8] and intelligence (n = 269,867 individuals) [29] were retrieved from published studies. As described in the original study, the BD GWAS sample included 41,917 cases from 57 cohorts collected in Europe, North America, and Australia, and 371,549 controls from European countries [8]. The intelligence GWAS data included 14 datasets from Europe and North America, and a common latent g factor underlying multiple dimensions of cognitive functioning was calculated and applied to operationalize the cohorts with intelligence measured using distinct approaches [29]. The information about the effect allele, effect size (beta or odds ratio), standard error, and P value were obtained from the GWAS studies.

The genome-wide genetic correlation (r_g) of BD with intelligence was calculated using LDSC [30]. MiXeR (version v1.3) was used to construct a bivariate causal mixture model to estimate the total number of shared and trait-specific causal variants between BD and intelligence based on GWAS summary statistics [21, 31]. MiXeR results are presented as a Venn diagram of the shared and unique polygenic components across BD and intelligence, and the dice coefficient score (polygenic overlap measure in the 0~100% scale) were also computed [18].

The conditional quantile-quantile (Q-Q) plots were generated to provide a visual pattern of enrichment in single nucleotide polymorphism (SNP) associations between BD and intelligence. As described in the previous study [32], the Q-Q plots computed the empirical cumulative distributions of P values in the primary phenotype for all SNPs and for subsets of SNPs (e.g., P ≤ 0.1, P ≤ 0.01, P ≤ 0.001, respectively) in the secondary phenotype. Increased degree of leftward deflection, from the expected null line as the association significance increases in a phenotype, suggests enrichment of associations of the other phenotype [32].

We next conducted the conjFDR analyses to characterize the genomic loci and SNPs jointly associated with both BD and intelligence. As described in the previous studies [32, 33], this method re-adjusted the GWAS summary statistics in the primary phenotype (e.g., BD) by leveraging pleiotropic enrichment with the GWAS summary statistics in the secondary phenotype (e.g., intelligence). The conditional FDR (condFDR) estimates were calculated for each variant in the primary phenotype using the stratified empirical cumulative distribution function. This process was then performed again with the primary and secondary phenotypes switched, and conjFDR was defined as the maximum of the two condFDR values [32, 33]. During the conjFDR analyses, all P values in the original GWAS datasets were corrected for genomic inflation, since the empirical null distributions of SNP associations in GWASs might be affected by population stratification [32]. Random pruning of SNPs was performed throughout the 500 iterations in both the conditional Q-Q plots and conjFDR analyses to minimize inflation resulted from linkage disequilibrium (LD) dependency, and one randomly-returned representative SNP for each LD-independent block was retained after every pruning iteration (cluster of SNPs with r² > 0.1). As recommended in the previous study [18], we remained only one signal in the highly extended major histocompatibility complex (MHC) region (hg19, chr6:26M-34M) to minimize the impacts of complicated LD patterns in this genomic area. SNPs with a conjFDR < 0.01 were considered statistically significant.

To identify genes associated with the risk loci, we conducted expression quantitative trait loci (eQTL) analyses of all the significant SNPs in each independent genomic region (conjFDR < 0.01 and at r² < 0.2), to identify genes associated with the risk loci in multiple datasets. We used datasets of postmortem postnatal human brain tissues (dorsolateral prefrontal cortex (DLPFC) or hippocampus) such as psychENCODE (n = 1387 for DLPFC) [34], BrainMeta (n = 2865 for cortex) [35], BrainSeq (n = 477 for hippocampus) [36] and Genotype-Tissue Expression (GTEx; n = 175 for DLPFC and n = 165 for hippocampus) [37], datasets of prenatal cortex tissues from O’Brien et al. (n = 120) [38] and Walker et al. (n = 201) [39] studies, as well as single-cell eQTLs resources (including mature midbrain dopaminergic neurons (from 175 donors) and serotonergic-like neurons (from 161 donors) differentiated from human induced pluripotent stem cells (iPSC) lines [40]; and excitatory neurons and inhibitory neurons from 196 individuals by single nuclei RNA-sequencing [41]). Detailed information about the sample characteristics, expression quantification, and normalization, as well as statistical analysis in each eQTL dataset, can be found in the original publications. The psychENCODE and BrainSeq datasets only provided SNP associations passing genome-wide level of statistical significance (false discovery rate (FDR) < 0.05 in psychENCODE and FDR < 0.01 in BrainSeq), we hence directly retrieved the significant eQTL associations. For the other eQTL datasets, we empirically considered the genes with eQTL P < 0.001 to be significant. We did not perform eQTL analysis in the MHC region (hg19, chr6:26M-34M) given its complicated LD patterns and potential mapping problems using short-reads sequencing.

For the risk genes identified by eQTL analyses, we conducted enrichment analyses of Gene Ontology (GO), Reactome pathway [42], UniProt keywords [43], and Monarch Human Phenotype Ontology (HPO) [44] using the STRING dataset (version 11.5) [45].

The polygenic overlap (i.e., shared causal variants) between BD and intelligence was assessed using MiXeR based on GWAS summary statistics. The bivariate MiXeR analysis revealed that approximately 8.6K (standard deviation (SD) = 0.2K) variants influenced BD, and approximately 10.3K (SD = 0.3K) variants influenced intelligence. Intriguingly, 7.6K (SD = 0.5K) variants influenced both BD and intelligence, and the overall measure of polygenic overlap between BD and intelligence was 80.3% (standard error (SE) = 4.9%) on a 0–100% scale (quantified as the Dice coefficient) (Fig. 1A).

We then generated conditional quantile-quantile (Q-Q) plots conditioning BD on intelligence and vice versa, to further describe the pleiotropic enrichment of SNP associations between BD and intelligence. The obvious leftward deflection of the curves for both BD-given intelligence and intelligence-given BD suggested strong enrichment of the significant variants for one trait in the other. Notably, SNPs with higher significance for the conditional trait exhibited greater leftward deflection in both plots, confirming the substantial polygenic overlap between BD and intelligence (Fig. 1B).

Both MiXeR and the Q-Q plots indicated significant polygenic overlap between BD and intelligence, but no significant genetic correlation was observed between them (LDSC r_g = −0.066, P = 0.062, Fig. 1A). Since LDSC analyses only reported significant genetic correlations with the premise that there were numerous SNPs showing associations with both phenotypes in concordant direction of allelic effects, we suspected that the shared variants between BD and intelligence had mixed directions of allelic effects. Indeed, MiXeR analysis showed that the “concordant variants” took up only 47% (SE = 0.4%) of the shared genetic components (7.6K variants) between BD and intelligence.

We hence applied the genetic pleiotropy-informed conjFDR method [33], which identifies loci associated with both BD and intelligence regardless of the allelic effect directionality with boosted statistical power. We identified 37 distinct genomic loci (r² < 0.2) that were jointly associated with BD and intelligence with a conjFDR < 0.01 (Fig. 2 and Table 1), including 24 independent loci that were not identified in the original BD GWAS [8]. Further computation of the z-scores of these loci revealed that 16 SNPs exhibited the same directions of effect between BD and intelligence (i.e., one allele predicted a higher risk of BD and greater intelligence), while 21 SNPs had the opposite directions (Fig. 2 and Table 1). This finding is consistent with a previous study describing loci jointly associated with BD and intelligence [25], albeit that study reported fewer joint risk loci.

We then performed a detailed genomic mapping analysis of the 37 shared loci between BD and intelligence. The locus with the strongest concordant effect on both the risk of BD and intelligence was in MIR2113 on chromosome 6 (rs1487445, conjFDR = 7.78×10⁻¹¹), and the T-allele of rs1487445 predicted both increased risk of BD (odds ratio (OR) = 1.077, P = 1.48×10⁻¹⁵) and higher IQ scores (beta = 0.031, P = 3.27×10⁻²⁹) (Fig. 3A). The variant with the strongest opposite effects on BD and intelligence is in the chromosome 3p21.1 region (rs12487445, conjFDR = 1.54×10⁻⁶), and the A-allele of rs12487445 predicted increased risk of BD (OR = 1.062, P = 2.44×10⁻¹⁰) but lower IQ scores (beta = −0.018, P = 3.31×10⁻¹¹) (Fig. 3B). The locus showing the second strongest opposite effects was in the highly extended MHC region (rs3749971, conjFDR = 3.27×10⁻⁵), and the G-allele of rs3749971 predicted increased risk of BD (OR = 1.106, P = 8.91×10⁻¹⁰) while lower IQ scores (beta = −0.030, P = 3.06×10⁻⁹) (Fig. S1). Another locus with the opposite directions of allelic effects between BD and intelligence was in the gene ADCY2 (rs17826816, conjFDR = 1.04×10⁻⁴), and its G-allele was linked to higher risk of BD (OR = 1.065, P = 1.32×10⁻⁸) but lower IQ scores (beta = −0.018, P = 1.68×10⁻⁸) (Fig. S2).

Since BD and schizophrenia exhibit strong genetic correlations [46], and schizophrenia also has substantial polygenic overlap with intelligence [25], there is the possibility that the identified BD risk loci through leveraging intelligence are not exclusively associated with BD. We verified this by examining the aforementioned 37 BD risk SNPs in the largest schizophrenia GWAS in Europeans (74,776 cases and 101,023 controls) [47], we found that 15 of these loci were also associated with schizophrenia in the same allelic directions (Table 1). Notably, 13 of the 21 SNPs showing opposite directions of effect between BD and intelligence also exhibited significant associations with schizophrenia; by contrast, only 2 of the 16 SNPs showing the same directions of effect between BD and intelligence were associated with schizophrenia. Therefore, although both BD and schizophrenia have polygenic overlap with intelligence, the potential genetic mechanisms regulating intelligence in these two disorders are likely divergent.

Since majority of the risk SNPs for BD are in the noncoding genomic regions, functional annotations of them are urgently needed. Accumulating evidence suggests that noncoding SNPs affect gene transcription and splicing likely in a cell type- and developmental stage-specific manner [7, 48]. We thus examined the regulatory effects of the BD risk SNPs using datasets containing human eQTL results in postnatal DLPFC and hippocampal tissues, prenatal cortex tissues, as well as different types of brain cells. In summary, 25 of the 37 risk loci had eQTL associations in at least one dataset (Table 1); 9 of the 16 risk loci showing concordant effects on BD and intelligence contained eQTL-associated risk genes, while 16 of the 21 loci showing opposite effects on BD and intelligence contained eQTL associated risk genes. There were no overlapped eQTL risk genes between the “concordant loci” and “discordant loci”.

Specifically, 37 protein-coding genes were significantly affected by SNPs in the loci showing concordant effects on BD and intelligence (Table S1). Although GO analysis of these genes did not reveal any significant enrichment, Reactome pathway analysis found that their proteins were significantly enriched in the pathways of “Activation of AMPK downstream of NMDARs,” “Microtubule-dependent trafficking of connexons from Golgi to the plasma membrane,” “RHO GTPases activate IQGAPs,” “Selective autophagy,” “Transmission across Chemical Synapses,” and “Membrane Trafficking” (Fig. 4A and Table S2). UniProt keywords analysis suggested that these proteins were significantly enriched in the terms of “GTP-binding” and “Fatty acid biosynthesis” (Fig. 4A and Table S2). Monarch HPO analysis found that these genes were enriched in the terms “Intelligence,” “Cognition,” and “Self reported educational attainment.” Intriguingly, we found that these genes also showed significant enrichment in the term “Sleep measurement” (Fig. 4A and Table S2).

We identified 135 protein-coding genes affected by SNPs in the genomic loci showing opposite effects on BD and intelligence (Table S3). GO analysis of these genes revealed significant enrichment in the terms “Cell adhesion” and “Calcium ion binding”, whereas Reactome pathway analysis did not identify any significantly enriched pathways (Fig. 4B and Table S4). Similarly, UniProt keywords analysis found significant enrichment of these proteins in the terms “Cell adhesion” and “Calcium” (Fig. 4B and Table S4). Monarch HPO analysis found that these genes were enriched in the terms “Intelligence” and “Cognition.” It should be noticed that these genes were also significantly enriched in “Anxiety,” “Emotional symptom measurement,” and “Worry measurement” (Fig. 4B and Table S4).