A polygenic risk score analysis of psychosis endophenotypes across brain functional, structural, and cognitive domains

This large multi‐center study investigates the relationships between genetic risk for schizophrenia and bipolar disorder, and multi‐modal endophenotypes for psychosis. The sample included 4,242 individuals; 1,087 patients with psychosis, 822 unaffected first‐degree relatives of patients, and 2,333 controls. Endophenotypes included the P300 event‐related potential (N = 515), lateral ventricular volume (N = 798), and the cognitive measures block design (N = 3,089), digit span (N = 1,437), and the Ray Auditory Verbal Learning Task (N = 2,406). Data were collected across 11 sites in Europe and Australia; all genotyping and genetic analyses were done at the same laboratory in the United Kingdom. We calculated polygenic risk scores for schizophrenia and bipolar disorder separately, and used linear regression to test whether polygenic scores influenced the endophenotypes. Results showed that higher polygenic scores for schizophrenia were associated with poorer performance on the block design task and explained 0.2% (p = 0.009) of the variance. Associations in the same direction were found for bipolar disorder scores, but this was not statistically significant at the 1% level (p = 0.02). The schizophrenia score explained 0.4% of variance in lateral ventricular volumes, the largest across all phenotypes examined, although this was not significant (p = 0.063). None of the remaining associations reached significance after correction for multiple testing (with alpha at 1%). These results indicate that common genetic variants associated with schizophrenia predict performance in spatial visualization, providing additional evidence that this measure is an endophenotype for the disorder with shared genetic risk variants. The use of endophenotypes such as this will help to characterize the effects of common genetic variation in psychosis.


| INTRODUCTION
Psychotic illnesses, including schizophrenia and bipolar disorder, constitute the most severe forms of mental illnesses (WHO, 2013).
However, it is still largely unknown exactly how these genetic risk variants lead to the illness, and an important goal of psychiatric genetic research is to clarify the effects and mechanisms of these variants (Carter et al., 2017;Geschwind & Flint, 2015;Glahn et al., 2014;Hall & Smoller, 2010;Harrison, 2015). Endophenotypes-biological markers that are heritable, quantitative traits associated with the illness, and observed in unaffected relatives of patients-could help us to understand the pathways from genes to the illness (Braff & Tamminga, 2017;Geschwind & Flint, 2015;Iacono, Vaidyanathan, Vrieze, & Malone, 2014;Gottesman & Gould, 2003;Meyer-Lindenberg & Weinberger, 2006;Munafò & Flint, 2014). As endophenotypes are thought to be related to the genetic factors underlying disorders, it is likely that a subset of psychosis associated SNPs also influence them (Lencz et al., 2014;Toulopoulou et al., 2015).
Past research has found a genetic overlap between these endophenotypes and psychosis. This includes a genetic correlation (overlap due to genetic factors) between the P300 amplitude and bipolar disorder (−0.33) (Hall, Rijsdijk, Kalidindi, et al., 2007) and schizophrenia (−0.48)  . Brain volume has also been shown to be genetically correlated with psychosis, including whole brain volume (−0.36) (Rijsdijk et al., 2005), and white matter volume (−0.20) (van der Schot et al., 2009). Similarly, findings indicate a genetic correlation between psychosis and cognition. Toulopoulou et al. (2007) found genetic correlations between schizophrenia and working memory  The aim of this study is to test whether polygenic risk scores for schizophrenia and bipolar disorder influence these multi-modal psychosis endophenotypes, in a large international sample of patients with psychosis, their unaffected first-degree relatives, and healthy controls.

| Sample and clinical assessment
The total sample in this study included 4,242 participants of European ancestry: 1,087 patients with psychotic illnesses (see Table 1 for breakdown of diagnoses), 822 unaffected first degree relatives of probands (with no personal history of a psychotic illness), and 2,333 unaffected controls (with no personal or family history of a psychotic illness). Relatives and controls were not excluded if they had a personal history of non-psychotic psychiatric disorders (such as depression or anxiety), provided they were well and off psychotropic medication at the time of testing and for the preceding 12 months. This was to avoid recruiting a biased "super healthy" control group, unrepresentative of the general populations.
To confirm or rule out a DSM-IV (APA, 1994) diagnosis, all participants underwent a structured clinical interview with either the Comprehensive Assessment of Symptoms and History (CASH; Andreasen, Flaum, & Arndt, 1992), the Structured Clinical Interview for DSM Disorders (SCID; Spitzer, Williams, Gibbon, & First, 1992), the Schedule for Affective Disorders and Schizophrenia (SADS; Endicott & Spitzer, 1978) or the Schedule for Clinical Assessment in Neuropsychiatry, Version 2.0 (SCAN; Wing et al., 1990). Participants were excluded if they had a history of neurologic disease or a loss of consciousness due to a head injury.
Recruitment occurred across 11 locations in Australia and Europe (Germany, The Netherlands, Spain, and the United Kingdom). See Supplementary Materials for a summary of the data collected from each site. Participants provided written informed consent, and the study was approved by the respective ethical committees at each of the 11 participating centers.

| EEG data collection and processing
Electrophysiological data were obtained from three sites (Table S1).
EEG data acquisition and processing varied slightly between sites and are summarized below. The full methods for each site are reported elsewhere (Bramon et al., 2005;Hall et al., 2006;Price et al., 2006;Waters et al., 2009;Weisbrod et al., 1999).
In summary, EEG was collected from 17 to 20 electrodes placed according to the International 10/20 system (Jasper, 1958). The P300 event related potential was obtained using a standard two-tone frequency deviant auditory oddball paradigm, with standard ("nontarget") tones of 1,000 Hz and rare ("target") tones of 1,500 Hz. The number of tones presented varied from 150 to 800, the tones were 80 or 97 dB, lasted for 20-50 ms, and the inter-stimulus interval was between 1 and 2 s. The majority of participants (90%) were asked to press a button in response to "target" stimuli, but a subset were asked to close their eyes and count "target" stimuli in their heads. Excluding the 10% of participants receiving different instructions does not change the results.
The data were continuously recorded in one of three ways: 500 Hz sampling rate and 0.03-120 Hz band pass filter; 200 Hz sampling rate and 0.05-30 Hz band pass filter; or 400 Hz sampling rate and 70 Hz low-pass filter. Linked earlobes or mastoids were used as reference and vertical, and in most cases also horizontal, electro-oculographs were recorded at each site and used to correct for eye-blink artefacts using regression based weighting coefficients (Semlitsch, Anderer, Schuster, & Presslich, 1986). After additional manual checks, artefactfree epochs were included and baseline corrected before averaging.
The averaged waveforms to correctly detected targets were then filtered using 0.03 or 0.05 Hz high-pass and 30 or 45 Hz low-pass filters. The peak amplitude and latency of the P300 were measured at electrode location PZ (parietal midline), within the range of 250-550 ms post-stimulus.

| Polygenic score analysis
Following the method described in Purcell et al. (2009), polygenic risk profile scores were calculated separately for schizophrenia and bipolar disorder. Summary data from the most recent Psychiatric Genomics Consortium genome-wide association studies for schizophrenia (PGC2) (Ripke et al., 2014) and bipolar disorder (Sklar et al., 2011) were used. In both cases, we used data from the Psychiatric Genomics Consortium that did not overlap with the sample used in the current study. For schizophrenia polygenic scores, the discovery sample included 31,658 cases and 42,022 controls (Ripke et al., 2014), and for bipolar disorder, the discovery sample included 7,481 cases and 9,250 controls (Sklar et al., 2011).
Polygenic scores for each individual were calculated using PLINK (Purcell et al., 2007), from the number of risk alleles carried for each selected SNP (i.e., 0, 1, or 2), weighted by the log(OR) provided by the Psychiatric Genomics Consortium, and averaged across all SNPs. SNPs were selected from the Psychiatric Genomics Consortium's panel using six different significance thresholds (p T < 5 × 10 −08 , 0.001, 0.05, 0.1, 0.5, 1), hence including an increasing number of SNPs the more liberal the threshold (see Supplementary Materials for the number of SNPs included at each threshold).

| Statistical analyses
Linear regression analyses were performed to test whether schizophrenia and/or bipolar disorder polygenic scores influence endophenotypes for psychosis. These included the P300 event related potential (amplitude and latency), lateral ventricular volume, and measures of cognition (digit span, block design, and the Rey Auditory Verbal Learning Task-RAVLT immediate and delayed recall).
Endophenotype measures were standardized for each site separately (using the overall means and standard deviations within each site) to control for differences between the centers. Covariates included in all analyses were clinical group (patient, relative, or control), study site, the first three population structure principal components, age, and gender. Because the sample included related individuals, robust standard errors were used to account for effects of clustering within families. This specifies that the standard errors allow for intragroup correlation; that is, the observations are independent between families (clusters) but not necessarily within families. The change in R 2 between a model only including the covariates and a model including covariates plus the polygenic score is reported, which represents the additional proportion of the variance explained by the polygenic risk score.
Linear regression analyses were performed for each endophenotype using the entire sample-patients with psychosis, unaffected relatives of probands, and controls-examining the associations with polygenic score at the different significance thresholds of the  (Perneger, 1998;Rothman, 1990;Savitz & Olshan, 1995
relatives p = 1.1 × 10 −16 ), with patients having the highest scores, followed by relatives and lastly controls (see Figure 1, left panel).
The polygenic score for schizophrenia predicted scores on the block design task at the SNP p-value threshold of p T < 0.05, with 0.
relatives p = 2.8 × 10 −3 ), with patients having the highest scores, followed by relatives and lastly controls (see Figure 1, right panel).
The polygenic score for bipolar disorder explained 0.17% of the variance in block design (at p T < 5 × 10 −8 ), although this did not reach significance after correction for multiple testing (p = 0.02). The proportions of variances explained by the bipolar disorder polygenic score were all <0.2% (and mostly below 0.1%), and none of the other associations were significant after correcting for multiple testing.
These results are shown in Figure 2b (for full results see Supplementary   Materials).

| Additional analyses
Given the fact that we are including both controls and relatives that are Hence, research suggests that there is a genetic overlap between cognitive performance and schizophrenia susceptibility (Lencz et al., 2014;Toulopoulou et al., 2010Toulopoulou et al., , 2015, and our finding with the block design task is in line with this. This provides support for the notion that this measure of spatial visualization is an endophenotype for schizophrenia, and that genetic risk variants are shared between the two traits. However, there was no association between measures of working and verbal memory and this polygenic score, and furthermore, no associations reached significance for bipolar disorder polygenic score after correction for multiple testing. This could be due to a lack of power, as these genetic effects are likely to be subtle as discussed below. It is nevertheless interesting that for bipolar disorder, the association with block design approached significance at the most stringent threshold; this is the genome-wide significant threshold and included only four SNPs that have been associated with bipolar disorder. These did not overlap with SNPs included in the schizophrenia score at this threshold, thus indicating a potential genetic overlap between block design impairment and bipolar disorder risk. Cognition has also been put forward as a possible endophenotype for bipolar disorder (Gkintoni, Pallis, Bitsios, & Giakoumaki, 2017;Miskowiak et al., 2017;Trotta, Murray, & MacCabe, 2015), although the evidence is much more limited than for schizophrenia and further research is required.
We did not find an association between polygenic risk scores for schizophrenia and bipolar disorder and the P300 event related potential. Similarly to this, studies by Hall et al. (2015) in a sample of 392 patients with schizophrenia and controls, and Liu et al. (2017) including a community-based sample of over 4,000 individuals, both failed to show associations between the P300 and polygenic score for schizophrenia. Nevertheless, research has suggested that the P300 has a significant genetic component. Abnormalities in unaffected firstdegree relatives of patients have been identified (Schulze et al., 2008;Thaker, 2008), its heritability is around 60% (Hall et al., 2006;van Beijsterveldt & van Baal, 2002), and about 27% of variance in P300 amplitude can be accounted for by common genetic variation . Furthermore, a significant genetic overlap of about 34% between the P300 amplitude and bipolar disorder has been observed (Hall, Rijsdijk, Kalidindi, et al., 2007). It is possible that the overlap in common variants involved in both psychosis and the P300 is small, suggesting subtle effect sizes that are difficult to detect with very large samples required.
As for the influence of polygenic scores on measures of brain volumes, Terwisscha van Scheltinga, Bakker, van Haren, Derks, Buizer-Voskamp, Boos, et al. (2013) and Papiol et al. (2014) both looked at total brain, white and gray matter volumes, and associations with Ventricular volume has a genetic basis with heritability estimates ranging from 30% to 70% (Carmelli, Swan, DeCarli, & Reed, 2002;Kremen et al., 2010Kremen et al., , 2012Peper et al., 2009;Schmitt et al., 2007). McDonald et al. (2002) found increased volumes in unaffected relatives of individuals with schizophrenia in families with more than one affected member, but not in relatives from families with only a single known case. In a meta-analysis of 1,065 unaffected relatives of patients with and 1,100 healthy controls, Boos, et al. (2007) did not find an overall effect in relatives, which is consistent with group comparisons in our sample. Enlargement of cerebral ventricles remains the best replicated biomarker in schizophrenia and bipolar disorder.
The samples investigating unaffected relatives including our own are of modest size and probably have limited power to detect anatomical changes, which we would expect to be much milder than those observed among patients. Of course, the ventricular enlargement described in psychosis might also be due to illness progression, or to the effects of treatment with antipsychotic medication over time.
Nevertheless, it is striking that the variance explained by the schizophrenia scores is larger for ventricular volume than for any other endophenotype we examined, with p-values approaching significance, and the question remains whether with larger samples one might see an association.
Although research has shown that there is a genetic component contributing to variability in the biomarkers investigated here, these are all complex (multifactorial and heterogeneous) phenotypes, and environmental factors play important roles too. Furthermore, all endophenotypes are likely to have complex genetic influences, including a substantial polygenicity (de Geus, 2014;Geschwind & Flint, 2015;Munafò & Flint, 2014;Rees, O'Donovan, & Owen, 2015), and only a subset of SNPs associated with psychosis will also be related to particular endophenotypes, and vice versa, suggesting that effect sizes for the associations of overlapping genetic factors will be small (Lencz et al., 2014). This has indeed been found for the phenotypes investigated so far, with the amount of variance explained by polygenic scores mostly below 1% Lencz et al., 2014;McIntosh et al., 2013;Papiol et al., 2014;Terwisscha van Scheltinga, Bakker, van Haren, Derks, Buizer-Voskamp, Boos, et al., 2013;Terwisscha van Scheltinga, Bakker, van Haren, Derks, Buizer-Voskamp, Cahn, et al., 2013;Van der Auwera et al., 2015;Whalley et al., 2012Whalley et al., , 2013Whalley et al., , 2015.  Where a bar appears missing this is because the variance explained is too low to display given the scale used in the figure. The lowest p-value for each endophenotype is displayed above the corresponding bar; the p-value in bold shows a significant finding. RAVLT, Rey auditory verbal learning task; imm, immediate recall; del, delayed recall [Color figure can be viewed at wileyonlinelibrary.com] that genotyping of all samples was done at the same laboratory using the same platform, and that all genetic analyses and quality control were completed in a unified way. Furthermore, it is precisely through this multi-center effort that we were able to achieve a very large sample, a key strength of this study. As the Psychiatric Genomics Consortium's work shows, large international collaborations are essential in genetic studies of common diseases and traits (Lee et al., 2013;Ripke et al., 2014;Sklar et al., 2011;Smoller et al., 2013).
Although common variants are thought to explain up to 30% of heritability in psychosis, genome wide association studies to date have only significantly identified about 3% of this (Fernandes et al., 2013;Lee, DeCandia, et al., 2012). More can be captured by calculating polygenic scores, although false positives will also be included (Iyegbe, Campbell, Butler, Ajnakina, & Sham, 2014;Wray et al., 2014). It is important to note that a larger discovery sample used to calculate polygenic scores is likely to include a higher proportion of true positive hits, and hence lead to enhanced performance of the polygenic scores as predictors of disease risk (Chatterjee et al., 2013;Dudbridge, 2013;Plomin, 2013;Wray et al., 2014). Compared to the discovery sample size used to calculate the schizophrenia polygenic score (including about 31,700 cases Ripke et al., 2014) the discovery sample for the bipolar disorder score was more than four times smaller-including only about 7,500 cases (Sklar et al., 2011)-and consequently this is the most compelling explanation of the lack of findings with the bipolar disorder polygenic score.
Importantly, there are highly significant differences in polygenic scores between the clinical groups, both in this sample and in previous studies (Bramon et al., 2014;Derks et al., 2012;Purcell et al., 2009;Ripke et al., 2014), indicating that this measure does capture genetic variants that differ between patients, unaffected relatives, and healthy controls. However, currently their predictive power is still low, and polygenic scores are not able to predict illness status accurately enough to be used in clinical practice. This would require very large discovery data sets, a large catalogue of genetic risk variants (potentially including both common and rare variants), and most likely the inclusion of a combination of genetic and non-genetic risk factors such as cognition, brain imaging, or family history, as well as age and gender (Chatterjee et al., 2013;Dima & Breen, 2015;Dudbridge, 2013;Iyegbe et al., 2014;McCarroll & Hyman, 2013;Wray, Yang, Goddard, & Visscher, 2010).
In future, as our understanding of the genetic architecture of psychosis improves, and as discovery samples become larger, the performance of the polygenic scores is likely to be further enhanced.
Polygenic scores could then be useful for testing hypotheses about the functional effects of risk variants, or to investigate the associations between disease risk and severity of illness, symptoms dimensions, and treatment or functional outcomes. This method could potentially be used to stratify populations into groups with shared genetic features, or to identify individuals at high-risk of developing psychosis who would benefit from early therapeutic interventions (Maier et al., 2015;Wray et al., 2014). Furthermore, using polygenic scores based on selected genetic risk variants clustering on specific functional pathways, rather than a broad selection of SNPs, could become beneficial in the investigation of the specific effects that genetic risk factors for psychosis have on brain function/structure and cognition.
In conclusion, results from this large multi-center study indicate that the combined effect of common genetic risk variants for schizophrenia is associated with spatial visualization (as measured by the block design task), providing further evidence that this measure is an endophenotype for the disorder with shared genetic risk variants.
No other associations between polygenic scores for schizophrenia or bipolar disorder and endophenotypes reached significance, possibly due to a lack of power, with larger samples needed to detect these small effects. As discovery samples get larger, and additional and better targeted genetic information is included, the performance of polygenic scores will be further enhanced. Larger association studies using these scores on deeply phenotyped samples may in future provide a promising approach to investigate the effects and mechanisms of genetic risk variants for psychosis.