The genetic overlap between schizophrenia and height

Abstract

Epidemiological studies suggest that height and schizophrenia risk are inversely correlated. These findings might arise because i) height and schizophrenia share genetic variants and ii) the effects of these shared variants are in opposite direction for the two traits. We use genome wide association data to empirically evaluate these hypotheses. We find that variants which impact on height and risk for schizophrenia are distributed across several genomic regions and the directions of effect vary, some consistent and others inconsistent with the direction expected from the phenotypic data. Moreover, signals that were in and not in accord with the phenotypic data aggregated in distinct biological pathways.

1. Introduction

Schizophrenia (SCZ) is a severe psychiatric disorder with largely unknown etiology and a lifetime prevalence of 4–5 per 1000 (McGrath et al., 2008). However, the characteristics of subjects with SCZ have been extensively studied. Clinical and epidemiological studies dating back to the beginning of the 20th century suggest that there is a link between height (H) and SCZ (Burchard, 1916; Gunnell et al., 2005; Kemali et al., 1976; Nopoulos et al., 1998; Perrin et al., 2007; Zammit et al., 2007). The largest investigation, a study of 720,000 Swedish men and women, found an inverse relationship between height (H) and SCZ (Gunnell et al., 2005). The reported hazard ratio is substantial: 0.8 for every 10cm height increase. These findings were later confirmed by the largest study to date to look at the relationship between risk to SCZ and H, i.e. a Swedish epidemiological study of 1.35 million males which included the males from the study described above (Zammit et al., 2007). Coupled with the high heritability of SCZ (Sullivan et al., 2003) and H (Yang et al., 2010), these epidemiological findings suggest that i) H and SCZ share common genetic variants and ii) genetic variants which increase H are likely to decrease SCZ risk, i.e. the effect of the variants on the two phenotypes is in opposite directions (i.e. discordant).

To evaluate the discordance hypothesis for common genetic effects of H and SCZ, we need a statistical method to assess the genetic overlap between any two phenotypes. One possible approach is to look at the overlap of significant findings for genome-wide association studies (GWASs) of these phenotypes. However, GWASs often fail to provide many significant signals, even more so for psychiatric disorders (Ripke et al., 2011; Sklar et al., 2011; Major Depressive Disorder Working Group of the Psychiatric GWAS Consortium, 2012). Thus, because, with current sample sizes, many positive signals are likely to exhibit p-values at only “suggestive” levels, we can increase the power to detect genetic overlap pathways by analyzing the shared suggestive signals. The definition for suggestive signals is subjective: it requires the p-values to fall below a reasonably small threshold. However, the findings from pathway analyses are sensitive to the choice of p-value threshold. For instance, in the paper describing ALIGATOR software (Holmans et al., 2009), authors noted that increasing the p-value threshold from 10⁻⁴ to 10⁻² improved the significance of different bipolar disorder pathways. Consequently, an objective, data driven method to set the p-value threshold for suggestive signals would be a useful tool in uncovering causal pathways for any phenotype/overlap of phenotypes.

In this paper, we use available GWAS data to obtain the likely overlapping genetic variants/pathways of H and SCZ. This goal is achieved by i) using a novel data driven method to obtain suggestive signals for each phenotype, ii) establishing the existence of a statistically significant overlap between H and SCZ suggestive signals and iii) performing pathway analyses for these overlapping suggestive signals. We use pathway analysis results to make inferences regarding possible SCZ etiologies.

2. Methods

To estimate the genetic overlap between H and SCZ, we use the normally distributed summary statistics, i.e. Z-scores, from the largest meta-analyses available (see Web Resources) for these traits (characteristics of the two meta-analyses are detailed in the Supplementary Material, henceforth abbreviated SM). For SCZ, the meta-analyses were performed by the Psychiatric Genetic Consortium (Ripke et al., 2011). The H meta-analysis was performed by the GIANT consortium (Lango et al., 2010).

To obtain the genetic overlap between two phenotypes, e.g. H and SCZ, we retain SNPs which harbor suggestive GWAS signals for both phenotypes. To objectively select the suggestive signal SNPs for each trait, we use a novel data-driven method (Bacanu and Kendler, 2012). This method utilizes a two-step procedure. First, an original shrinkage method is employed to accurately estimate the means of the GWAS Z-scores (see SM). Second, for a desirable sensitivity–specificity profile, we select as suggestive only SNPs having an estimated (adjusted) mean (of the Z-scores) above a threshold of three. (Given that the estimated means are obtained by shrinking the GWAS Z-scores toward zero, the original Z-scores often have a magnitude much larger than three.) The genetic overlap of the two phenotypes consists of SNPs which are deemed suggestive for both traits (see Table S1 for relevant characteristics of these signals).

3. Results

The number of suggestive SNPs in the overlap of the two traits was 51 when the expected number under a random sharing was estimated to be 21.32. However, not all these overlap SNPs are independent. It is more useful to report the number of quasi-independent signals, e.g. similar to the “clumped” (r² ≤ 0.25) SNPs from the PGC paper (Ripke et al., 2011). Assuming i) 117,000 clumped SNPs, as reported in Ripke et al. (2011) and ii) the fractions of suggestive signals for the two phenotypes being, as estimated from our data (Table S1), 0.006 (H) and 0.0029 (SCZ), the expected number of clumped SNPs under independence is estimated to be 2.03. The number of clumped SNPs in the overlap, as obtained using NIEHS tag SNPs (see Web Resources), was estimated to be 9. Assuming a Poisson distribution for the number of overlap clumped SNPs, the observed number (9) is significantly larger (p = 2.6 × 10⁻⁴) than the number expected (2.03) under the assumption of a random sharing of suggestive signals between the two phenotypes. Thus, our results strongly suggest that most overlap signals are likely to be true positive associations rather than chance findings. The large fraction of the true positives makes it possible to uncover the relevant molecular pathways via a pathway analysis, as detailed later on.

There are 6 distinct genomic regions that contain variants that impact on both phenotypes (Table 1). For three of these regions, as predicted by the epidemiological studies, the directions of genetic effects on the two phenotypes are discordant. These regions appear to affect the overall functioning of individual cells or of the entire organism, e.g. transcription regulation and immune response. Perhaps surprisingly, the other three regions show positively correlated or concordant effects, i.e. genetic effects for H and risk for SCZ are in the same direction. In fact, one such region (chromosome 3) exhibited the strongest minimum signal for the two phenotypes. These concordant regions seem to change the balance between cell proliferation and cell upkeep/maintenance.

Table 1.

The regions of genetic overlap between height and psychiatric disorders. The function of the genes closest to the overlap regions was obtained from GeneCards (www.genecards.org). Concordant denotes that the genetic effects on risk to schizophrenia and height are in the same direction.

Chr	HG18 position (Mb)	No. SNPs	No. clumped SNPs (r2 < 0.25)	Genes	Gene function	Concordant
2	24.546	1	1	NCOA1	Hormone dependent transcription	N
				ITSN2	Clathrin-mediated endocytosis
2	233.130–233.146	5	1	EIF4E2	Translation initiation	Y
3	52.512–53.036	33	3	ITIH cluster	Hyaluronan carriers, involved in cell proliferation	Y
6	26.261–26.396	7	2	HIST cluster	Regulator of transcription	N
6	31.349–32.787	3	1	HLA cluster	Immune system	N
20	54.274–54.285	2	1	MC3R	Pigmentation, energy homeostasis	Y
				C20orf108 (FAM210B)	Not well known

To uncover the underlying molecular mechanisms for these overlapping signals, we used both Grail and Ingenuity Pathway Analysis (IPA) (www.ingenuity.com) to perform separate pathway analyses for SNPs with concordant and discordant effects on H and SCZ. Signals of discordant effects seem to be enriched in immune and cell function pathways (Table S2 for Grail and Fig. S1 for IPA). Overlapping SNPs with concordant effects on the two phenotypes appear to aggregate in cell division/proliferation pathways (Table S3 and Fig. S2).

4. Discussion

Classic epidemiological studies have found that H and SCZ risk are negatively correlated. Based on this fact, and given the high heritability of both traits, H and SCZ are likely to have mostly overlapping genetic causes of discordant effect. Our analyses confirmed that H and SCZ are likely to have overlapping genetic causes. However, their overlapping genetic regions seem to be evenly divided between those which have discordant and concordant genetic effects.

Our analyses suggest that SNPs with discordant and concordant effects on H and SCZ belong to distinct pathways which, in turn, might point to different etiologic paths to SCZ. SNPs with discordant effects on SCZ and H fall into pathways affecting the organism/cell fitness, e.g. immune system and DNA/chromatin packing. SNPs with concordant effects on H and SCZ aggregate into pathways likely affecting the equilibrium between cell growth and maintenance, e.g. cell proliferation and tumor suppression.

Appendix A. Supplementary data

Supplementary data to this article can be found online at http://dx.doi.org/10.1016/j.schres.2013.10.016.