The full results for the ICE-UKB meta-analysis are available online (see ‘Data availability’). The quantile–quantile plots for both meta-analyses (Supplementary Fig. 1) show that our tests are well calibrated. The results for the top 20 genes from both the ICE-UKB and the ICE-UKB-BipEx LOF burden meta-analyses (Supplementary Tables 1 and 2) are labeled with gene symbols on the Manhattan plots (Figs. 1 and 2); brief descriptions of the genes involved are provided in the Supplementary Note.
The plot shows −log10 two-sided P values against the chromosomal position of each of the genes studied. The top 20 findings are indicated by the green dots; the corresponding gene name is represented by the most proximal gene symbol. Logistic regression was used to test for association in ICE and UKB; P values were obtained from a likelihood ratio test adjusted for cryptic relatedness and stratification. The results, including ORs, for the top 20 genes for the meta-analysis and for each study (ICE, UKB and BipEx) are provided in Supplementary Table 2. To obtain the overall P values, the results of the ICE-UKB fixed-effect meta-analysis (Fig. 1) were combined with the BipEx results on the z-score. The P values presented were not corrected for multiple testing, but the red line represents a Bonferroni threshold of P ≤ 3.6 × 10−6.
HECTD2 and AKAP11 are associated with bipolar disorder
We determined that combined rare LOF variants in HECTD2 were significantly associated with bipolar disorder based on the ICE-UKB analysis (frequency = 0.011%, odds ratio (OR) = 9.1, P = 4.7 × 10−7, 95% confidence interval (CI) = 3.85 to 21.49). We replicated the association of LOF variants in HECTD2 with bipolar disorder using data from the BipEx study (13,933 cases and 14,422 controls, OR = Inf, P = 0.0069); the significance was greater when combining the two studies (overall P = 4.1 × 10−8) (Table 1 and Figs. 1 and 2). The ICE-UKB analysis included 45 LOF variants in HECTD2 (140 carriers) and 38 in AKAP11 (76 carriers) (Table 1 and Supplementary Table 3).
Systematic meta-analysis of our study and BipEx revealed one additional association with bipolar disorder: AKAP11 (OR = 11.8, P = 7.4 × 10−9) (Table 1 and Fig. 2). For both HECTD2 and AKAP11, we observed at least a nominal association in the same direction in all three groups tested; however, gene-burden-wide significance was not reached in any of them individually (ICE, UKB and BipEx) (Table 1).
HECTD2 and AKAP11 are constrained in LOF variants
Genes can be classified according to how constrained they are with regard to LOF variants, estimated by comparing the number of observed versus expected very rare LOF variants, with genes scored according to the probability of being LOF intolerant (pLI) and the ratio of LOF observed/expected upper bound fraction; these metrics are available from the Genome Aggregation Database (gnomAD) for most genes. Enrichment of ultra-rare LOF variants in pLI genes (pLI ≥ 0.9) has been observed for bipolar disorder12, although this pattern is far weaker than that observed for schizophrenia11,13. Both AKAP11 and HECTD2 are pLI genes (Supplementary Table 4).
Effects of LOF variants on cognition in HECTD2 and AKAP11
A general intelligence factor (g-factor) was derived from the results of cognitive tests available for parts of both the ICE and UKB samples, as outlined previously14. When analyzed jointly for ICE and UKB, individuals with bipolar disorder, compared to controls, exhibited significantly lower cognitive performance (effect = −0.40, P = 1.0 × 10−18, 95% CI = −0.49 to −0.31). We next looked at carriers of LOF variants in HECTD2 and AKAP11, focusing on carriers not diagnosed with bipolar disorder because cognitive measures were available for only one LOF carrier diagnosed with bipolar disorder (AKAP11 in the UKB sample). In the combined analysis, there was nominal evidence for a lower g-factor among carriers of HECTD2 LOF variants who had not been diagnosed with bipolar disorder (effect = −0.38, P = 0.030, 95% CI = −0.72 to −0.04). Cognitive tests were not available for carriers of AKAP11 LOF variants in ICE, but there was no evidence for lowered g-factor among carriers of LOF variants in AKAP11 (not diagnosed with bipolar disorder) in the UKB sample (effect = −0.09, P = 0.76, naff = 12, nctrl = 74,558, 95% CI = −0.67 to 0.49) (Supplementary Note).
Burden analysis combining LOF and damaging missense variants
We also performed a burden analysis jointly analyzing LOF variants combined with predicted deleterious missense variants or in-frame insertions (for the list of markers, see Supplementary Table 3b). Briefly, adding this group of variants to the analysis removed all association with AKAP11 (OR = 0.92, P = 0.42) and did not improve that of HECTD2 (OR = 6.35, P = 3.4 × 10−6). The results for the top 20 genes from this analysis revealed no new significant signals but some suggestive associations that may be of interest, although these require further study in larger samples (Supplementary Table 5).
Bipolar disorder is not especially enriched for variants in pLI genes
Enrichment of damaging variants in pLI genes has been reported for bipolar disorder12, and both HECTD2 and AKAP11 are pLI genes. Hence, we explored whether there was overall enrichment of LOF or LOF + missense variants in bipolar disorder for this group of genes in our data, but did not find unequivocal evidence for particular enrichment of the whole group (Supplementary Note).
Differences in methodology
Burden studies involve rare variants and can thus be sensitive with regard to variant inclusion, experimental error or some underlying confounders, raising questions regarding using ICE, UKB and BipEx in a meta-analysis. BipEx itself is a meta-analysis of samples from several different studies and is thus subject to batch effects; however, the study implemented methods to minimize them12. We note that the UKB sample and around 40% of the ICE sample were whole-genome-sequenced with the same technology, with a minimum depth of at least 20× coverage, while the BipEx samples were whole-exome-sequenced with a somewhat higher coverage (~55×)12. Almost all variants identified using whole-exome sequencing were found using WGS15, so these differences in methodology are not likely to affect the study. One weakness in our approach is that we may have overlooked a considerable fraction of de novo mutations in Iceland because the Icelandic sample was not fully whole-genome-sequenced. Another difference is that the summary statistics available from BipEx are from analyses confined to ultra-rare variants (minor allele count (MAC) ≤ 5)12, and this type of filtering may lead to differences for some of the genes studied.
Verification of imputed carriers
Our methodology for sequencing and imputations in ICE has been described previously16,17. For this study, we used a high cutoff of genotype imputation probability (0.9); the genotypes with this quality of imputation are correct more than 90% of the time. Nevertheless, we tested all carriers from imputations of the rare LOF variants in HECTD2 and AKAP11 in ICE using Sanger sequencing (Methods). There were no imputed carriers for AKAP11 and 14 for HECTD2. Sequencing did not work for one study participant, but the other 13 imputed carriers were all confirmed Sanger sequencing. In our Sanger sequencing analysis, we also included several individuals predicted as unlikely carriers (P < 0.7, n = 1, and P < 0.1, n = 15) and found no new, unexpected carriers.
Sensitivity to variant inclusion
In our analysis of LOF variants, we included start-loss and stop-loss variants in addition to stop-gain, essential splice and frameshift variants. We performed the LOF of the burden analyses in ICE and UKB excluding start-loss and stop-loss variants. The results indicated that variants in this category neither drive the signals nor significantly attenuate them. For AKAP11, we obtained an OR = 12.7 (95% CI = 3.8 to 42.7); for HECTD2, we obtained an OR = 9.1 (95% CI = 3. 9 to 21.3).
Sensitivity to psychiatric comorbidities
Comorbidity among individuals with bipolar disorder, particularly with schizophrenia and schizoaffective disorders, and the fact that our phenotype definition is based on at least one diagnosis of bipolar disorder, raises the concern that the results might be driven by study participants who have also been diagnosed with other psychiatric disorders. To evaluate the contribution of schizophrenia or schizoaffective diagnoses on the results, we conducted burden analysis removing individuals with at least one diagnosis of either schizophrenia or schizoaffective disorder from both cases and controls. We found that the observed ORs were robust to this removal (Supplementary Note). We also performed an analysis of schizophrenia in UKB and ICE, removing all individuals with at least one bipolar disorder diagnosis from cases and controls. For HECTD2, no risk was observed for this subgroup as carriers diagnosed with schizophrenia all had dual diagnoses of bipolar disorder; for AKAP11, we obtained an OR = 4.3 (P = 0.25), suggestive of a residual risk consistent with LOF variants in AKAP11 conferring a risk of schizophrenia13.
Thus, for both AKAP11 and HECTD2, the results presented in Table 1 represent estimates that are based on verified carriers of LOF variants and are driven by bipolar disorder without substantial contribution from carriers with dual diagnoses of bipolar disorder and schizophrenia or schizoaffective disorder. For AKAP11 and HECTD2, we found no evidence for differences in the prevalence of anxiety or depression among carriers unaffected by bipolar disorder (Supplementary Note and Supplementary Table 6); however, further studies using larger samples are required for strong inference on questions relating to risk of other psychiatric disorders.
