Genetic risk scores and family history as predictors of schizophrenia in Nordic registers

Abstract

Family history is a long-standing and readily obtainable risk factor for schizophrenia. Low cost genotyping technologies have enabled large genetic studies of schizophrenia, and the results suggest the utility of genetic risk scores (GRS, direct assessments of inherited common variant risk). Few studies have evaluated family history and GRS simultaneously to ask whether one can explain away the other. Using harmonized genomic and national register data from Denmark, Estonia, Norway, and Sweden, we confirmed that family history of schizophrenia and GRS for schizophrenia and bipolar disorder were risk factors for schizophrenia. In a joint model, the effects of GRS for schizophrenia and bipolar disorder were essentially unchanged, and the effect of family history was attenuated but remained significant. The predictive capacity of a model including GRS and family history neared the minimum for clinical utility. Combining national register data with measured genetic risk factors represents an important investigative approach for psychotic disorders. Our findings suggest the potential clinical utility of combining GRS and family history for early prediction and diagnostic improvements.

Introduction

Schizophrenia is a debilitating psychiatric disorder that is associated with substantial health and economic burden for patients, families, and society.¹ Its genetic component is substantial with estimated heritability of 77% from twin studies² and shared environment accounts for a significant fraction of liability to schizophrenia (~10%).³ These estimates are broadly consistent with national pedigree studies from Denmark and Sweden.^4,5 These results underscore the importance of both genes and environment in the etiology of schizophrenia.^6–8

A family history of schizophrenia in first-degree relatives is a strong and robust risk factor for schizophrenia,⁷ with relative risks estimates ranging from 7.0–9.3.⁷ However, given its low lifetime risk (<1%),⁴ the absolute risks of a positive family history are low and the vast majority of schizophrenia cases (>95%)⁹ do not have an immediate family history of schizophrenia.¹⁰

Knowledge of the genetic basis of schizophrenia has advanced markedly in the past decade.¹¹ A substantial fraction of the heritability of schizophrenia is due to numerous common genetic variants each carrying small changes in risk.¹² One implication of these discoveries is the capacity to estimate individual genetic risk scores (GRS),¹³ which, for the first time, quantify genetic risks for an individual.¹⁴ Case-control differences in schizophrenia GRS have been widely replicated.^15,16 GRS are a continuous measure of genetic liability and have been suggested to be a more powerful and more direct assessment of liability compared to a dichotomous family history measure.¹⁴ Unlike family history, GRS does not suffer from information loss due to recall bias, reduced fertility, or incomplete knowledge of family history.

A major distinction between family history and GRS is the extent of information captured. Family history captures the overt results of shared genetics (common and rare genetic variants) and shared environmental risks among family members. GRS are comprised of common but probably non-causal genetic markers. Indeed, there may be low correlations between these two measures,^17,18 which could be due to the small proportion of variation explained by current GRS for most traits. One study found that schizophrenia cases with a positive family history had higher GRS compared to family history negative cases.¹⁹ However, there has been little research on the joint contribution of family history and GRS to liability for schizophrenia. One study found that the effect of family history was mediated through GRS.¹⁵

In this study, we aimed to investigate the interplay of the conventional measure of family history and GRS. The Nordic countries have multiple advantages for population-based research. These countries have universal health care and a long tradition of collecting high-quality and comprehensive health care data (i.e., diagnostic data from hospital admissions and specialist outpatient treatment contacts). This makes it possible to recruit representative samples, unbiased by sociodemographic factors or insurance plans. These factors are important for generating large samples and facilitating psychiatric genetics research. The current study builds on these advantages by combining large samples across the Nordic countries. We expand upon prior research by combining data across four Nordic countries, incorporating systematically captured family history, and the use of GRS for multiple traits. We sought to determine if the effects of family history could be reduced or eradicated by adding GRS based on direct measures obtained from the genomes of cases and controls. In other words, can a direct measure of genetic liability like GRS replace cruder family history measures used in epidemiological and clinical studies?

Methods

Samples.

We evaluated data from Denmark, Estonia, Norway, and Sweden,¹⁶ all of which have universal public health care systems. Case/control status and family history of schizophrenia and other psychiatric disorders were extracted from national health registers that systematically capture ICD diagnoses from inpatient admissions and outpatient specialist treatment contacts. As in previous studies,^11,16,20 the case definition included individuals with schizophrenia or schizoaffective disorder. For family history, the first-degree relatives of cases and controls were identified via linkage with multi-generational registers.

Danish cases were ascertained through psychiatric departments in the Copenhagen area and were of Danish parentage. Controls of Danish parentage were collected from the Danish Blood Donor Corps in the same catchment area. All subjects gave written informed consent and the Danish Data Protection Agency and the ethics committees of Denmark approved the human subjects protocol. The Estonian cohort comes from the population-based Estonian Genome Project.²¹ The project was conducted according to the Estonian Gene Research Act and all participants provided informed consent. Participation was voluntary and no gender balance was controlled for. In total, 52,000 individuals aged 18 years or older participated. General practitioners and physicians in hospitals recruited participants but schizophrenia was diagnosed by psychiatrists according to ICD-10 criteria. Controls were drawn from a larger pool of genotyped biobank samples by matching on sex, age, and genetic ancestry. All controls were population-based and have not been sampled for any specific disease. Norwegian cases of European ancestry, born in Norway, were recruited from psychiatric hospitals in the Oslo region and diagnosed using the SCID.²² Healthy controls were randomly selected from statistical records of persons from the same catchment area as cases. Controls were required to have no family history of any psychiatric disorder. All participants provided written informed consent and the human subjects protocol was approved by the Norwegian Scientific-Ethical Committee and the Norwegian Data Protection Agency. Cases and controls from the Swedish Schizophrenia Study¹¹ were identified via the Swedish Hospital Discharge Register which captures all public and private inpatient hospitalizations. Case inclusion criteria included ≥2 hospitalizations with a discharge diagnosis of schizophrenia or schizoaffective disorder and age ≥18 years. Case exclusion criteria included hospital register diagnosis of any medical or psychiatric disorder mitigating a confident diagnosis of schizophrenia as determined by expert review. The validity of this case definition of schizophrenia was strongly supported by clinical, epidemiological, genetic epidemiological and genetic evidence.¹¹ Controls were selected at random from Swedish population registers. Control inclusion criteria included never being hospitalized for schizophrenia or bipolar disorder and age ≥18 years. All procedures were approved by ethical committees and all subjects provided written informed consent (or legal guardian consent and subject assent).

Genetic risk scores (GRS) were estimated for all target Nordic subjects using best practices (Supplementary Figure 1).^13,16 We derived GRS for schizophrenia (SCZ), bipolar disorder (BIP), major depressive disorder (MDD), autism spectrum disorder (ASD), body mass index (BMI), and educational attainment (EDU). These traits all have significant genetic correlations with schizophrenia.^23–25 We obtained GWAS results for SCZ, BIP, ASD, and MDD from the PGC (URLs), and reran meta-analysis of case-control data from European samples but excluding Nordic sites. The results of meta-analyses from these non-Nordic samples were considered as discovery sets to select SNPs and to weight the risk alleles during the calculation of GRS. We found limited evidence for sample overlap with our target sets (Supplementary Table 1). GWA summary statistics for BMI were obtained from GIANT.²⁶ GWA summary statistics for the total number of years of education (a proxy for general intelligence) from SSGAC were used to derive GRS (after excluding Swedish studies).²⁴ Meta-analyses were performed using METAL.²⁷ The genotype data in the Nordic target sets were processed using the PGC Ricopilli pipeline for QC and imputation. Cryptic relatedness was checked in the Ricopilli pipeline, where one in each pair of related individuals ($\widehat{\pi }$ > 0.2) was excluded. Variants were excluded for the following reasons: not biallelic, lacking dbSNP rsID, strand-ambiguous, MAF<0.05 or >0.95, or poor imputation quality (R² or INFO < 0.8). To select a relatively independent set of SNPs for the calculation of GRS, we ran LD-clumping (r² <0.1 in 1Mb window) on the overlapping SNPs using 1000 Genomes Project European samples as LD reference. Supplementary Table 2 has details on the number of SNPs used for calculation of GRS.

We calculated GRS in the target sets as the sum of the SNP dosages weighted by the effect from the discovery set (logistic regression ß for the case/control psychiatric disorders or multiple regression ß for BMI and EDU) across all SNPs under a specified P-value threshold. All ß values were recoded to be positive. GRS were calculated for eight P-value cutoffs (P_T): ≤5×10⁻⁸, ≤1×10⁻⁵, ≤1×10⁻³, ≤0.01, ≤0.05, ≤0.1, ≤0.5, and ≤1. We used PLINK (v1.9) for the calculation of GRS.²⁸

Statistical methods.

We used logistic regression to test the association of family history and different GRS on schizophrenia case/control status. Family history was dichotomized as negative (i.e., no first-degree relative with schizophrenia or schizoaffective disorder) or positive (i.e., at least one first-degree relative with the condition).

The associations between the GRS of four psychiatric disorders with schizophrenia were first tested within each Nordic dataset. We included 20 ancestry principal components (PCs). The variance in schizophrenia (Nagelkerke’s²⁹ pseudo-R²) explained by GRS was calculated as the difference in R² from the full model including the GRS and the baseline model with PCs only (Supplementary Figure 2). Assuming a life-time risk of 1%, we reported variance explained on the liability scale adjusted for case-control ascertainment.^16,30 GRS were standardized within each target set prior to the main analyses to account for variation in numbers of SNPs used for GRS calculation across studies. We used the GRS from the SNP set that was selected from each of the discovery sets at the P-value cut-off of 0.05, given that variance in schizophrenia explained by most of GRS plateaued at this cut-off (Supplementary Figure 2). We included the GRS of all four psychiatric disorders due to the known moderate to strong genetic overlap among these disorders²³ (Supplementary Table 3). We included GRS for BMI and EDU given their genetic correlations with schizophrenia.²⁵ GRS were included as continuous variables. To aid the comparison of ORs from GRS and from family history, we also divided the GRS of SCZ into three subgroups: low genetic risk (GRS in the 1^st quintile), intermediate genetic risk (2^nd–4^th quintile), and high genetic risk (5^th quintile).

The main analyses included: (a) fitting separate logistic regression models for family history (adjusting for study: SCZ diagnosis ~ family history + study indicator) and for the GRS of four psychiatric disorders and of BMI and EDU (adjusting for study and PCs: SCZ diagnosis ~ GRS_SCZ + GRS_BIP + GRS_MDD + GRS_ASD + GRS_BMI + GRS_EDU + study indicator + PCs); and (b) joint estimation of family history and all GRS (adjusting for the studies and PCs: SCZ diagnosis ~ family history + GRS_SCZ + GRS_BIP + GRS_MDD + GRS_ASD + GRS_BMI + GRS_EDU + study indicator + PCs). We also tested the significance of the interaction term of family history and the GRS of SCZ in addition to the main effects. All analyses were performed using R (v3.3.2).

Results

This study included 5,959 schizophrenia cases and 8,717 controls from four Nordic countries. Table 1 shows sample characteristics for individual studies. The prevalence of a positive family history was approximately 1% in controls (except for Norwegian controls ascertained for the absence of family history) and 10–14% in cases (Table 1). Schizophrenia cases had more first-degree relatives with bipolar disorder or other mental disorders than controls (Table 1). Consistent with epidemiological estimates⁹, a positive family history was associated with an OR of 9.70 (95% CI=7.99–11.79, Table 2), which explained 3.5% of variance on the liability scale of schizophrenia calculated from a logistic regression model and adjusted for case-control ascertainment.

Table 1.

Sample characteristics

	DENMARK		Estonia		Norway		Sweden
	Case	Control	Case	Control	Case	Control	Case	Control
N	463	956	233	1153	339	401	4924	6207
Age, mean	50.7	49.9	46.2	46.4	32.7	34.8	53.9	56.3
% Male	58.3%	61.8%	26.2%	26.8%	55.8%	50.1%	59.7%	51.2%
Family history: number of first-degree relative with SCZ or schizoaffective disorder (%)
0	363 (78.4)	875 (91.5)	n.a.	n.a.	278 (82.0)	401 (100)	4001 (81.3)	6015 (96.9)
1	33 (7.1)	11 (1.2)	n.a.	n.a.	40 (11.8)	.	581 (11.8)	106 (1.7)
>1	13 (2.8)	1 (0.1)	n.a.	n.a.	3 (0.9)	.	94 (1.9)	5 (0.1)
Family history: number of first-degree relative with bipolar disorder (%)
0	389 (84.0)	880 (92.1)	n.a.	n.a.	300 (88.5)	401 (100)	4102 (83.3)	5829 (93.9)
1	20 (4.3)	7 (0.7)	n.a.	n.a.	19 (5.6)	.	490 (10.0)	272 (4.4)
>1	0 (0)	0 (0)	n.a.	n.a.	1 (0.3)	.	84 (1.7)	25 (0.4)
Family history: number of first-degree relative with other mental disorders (%)
0	234 (50.5)	626 (65.5)	n.a.	n.a.	157 (46.3)	401 (100)	2474 (50.2)	4130 (66.5)
1	132 (28.5)	222 (23.2)	n.a.	n.a.	111 (32.7)	.	1204 (24.5)	1279 (20.6)
>1	43 (9.3)	39 (4.1)	n.a.	n.a.	52 (15.3)	.	998 (20.3)	717 (11.6)

The percentages in family history measures do not add up to 1 due to missing data; no sufficient information to determine the number of first-degree relatives with psychiatric disorders in the Estonian cohort; controls in the Norwegian cohort were required to have no family history of any psychiatric disorder.

Table 2.

Separate and joint estimations of family history and GRS

	Separate estimation		Joint estimation
	OR (95% CI)	P-value	OR (95% CI)	P-value
Family history	9.70 (7.99–11.79)	3E–116	8.37 (6.85–10.23)	5E–96
GRS PT ≤ 0.05
SCZ	1.83 (1.76–1.91)	5E–191	1.82 (1.74–1.90)	2E–159
BIP	1.26 (1.22–1.31)	5E–34	1.25 (1.20–1.30)	1E–26
MDD	1.03 (0.99–1.07)	0.11	1.02 (0.98–1.06)	0.26
ASD	1.03 (1.00–1.07)	0.06	1.04 (1.00–1.08)	0.04
GRS of BMI & EDU, PT ≤ 0.05
BMI	0.97 (0.93–1.00)	0.08	0.97 (0.93–1.01)	0.17
EDU	1.04 (1.00–1.08)	0.03	1.04 (1.00–1.08)	0.08

^*Results based on the dataset excluding Norwegian study was provided in Supplementary Table 6.

The mean values of GRS for all four psychiatric disorders were higher among cases than controls within each study site (Figure 1, Supplementary Table 4). Case-control differences were most pronounced for GRS for SCZ and BIP. Across all four study sites, the variance on the liability scale explained by single GRS was 5.3%, 1.8%, 0.2% and 0.02% for SCZ, BIP, MDD and ASD. When GRS for all four psychiatric disorders were included in the model, the total variance explained was 6.0%, which was less than the sum of variances explained by individual GRS due to significant correlations among the GRS (Supplementary Table 5). An increase in SCZ GRS of one standard deviation was associated with 83% greater risk for schizophrenia (OR=1.83, 95% CI=1.76–1.91), and the corresponding figures were 26% for BIP GRS, 3% for ASD GRS, and 3% for MDD GRS (Table 2). GRS for BMI and EDU were associated with schizophrenia at borderline significance (OR=0.97 for GRS of BMI, and 1.04 for GRS of EDU).

Figure 1. Boxplots for GRS mean differences between cases/controls (P<0.05).

A) GRS of schizophrenia; B) GRS of bipolar disorder; C) GRS of major depressive disorder; D) GRS of autism spectrum disorder in the four Nordic studies (‘DENM’ Denmark, ‘ESTO’ Estonia, ‘NORW’ Norway, and ‘SWE’ Sweden). Controls in grey and cases in red.

Cases with a positive family history had a significantly higher mean value of the GRS of SCZ when compared with those with a negative family history (mean=0.47 as compared to 0.32; P=7×10⁻⁵; Supplementary Figure 3). In the absence of a family history, the risk to schizophrenia was increased by 2.48-fold (95% CI=2.21 to 2.77) and by 5.49-fold (95% CI=4.79 to 6.30) for individuals with intermediate and high genetic risk when compared to the low GRS group.

Joint estimation of the risks of family history and all GRS yielded a lower OR associated with family history (OR=8.37, 95% CI=6.85–10.23; Table 2) as compared to the results from the separate models. The effects from all the GRS were essentially unchanged. Collinearity was not severe owing to a low correlation between GRS and family history (Spearman correlation of 0.11 among all samples). Family history and multiple GRS, together, explained 8.9% of variance on the liability scale. The interaction between family history and the GRS was not significant (P=0.11) when tested together with the main effects. Thus, in the joint model, the effect of family history remained significant despite inclusion of multiple GRS that assessed common variant liability directly.

Discussion

We leveraged the unique resources of genomic and medical register data from four Nordic countries, and provided evidence that genetic risk profile and family history are strong predictors of schizophrenia, even when considered jointly.

We first replicated previous findings that GRS of schizophrenia and related disorders were robustly associated with schizophrenia, together accounting for 6% of variation on the liability scale after adjustment of case-control ascertainment.³⁰ The variance explained was smaller than reported before³¹ because, by necessity, we had to exclude all Nordic samples from the discovery samples. This highlighted that the accuracy of GRS will improve as sample sizes amassed for discovery continue to increase. We used Nordic registers to accurately identify family history, and found that a positive family history (defined as ≥1 first-degree relatives with schizophrenia or schizoaffective disorder) was associated with a nearly 10-fold risk to schizophrenia, and subjects with a positive family history carried higher GRS burden compared to those without family history. The risk estimate was highly consistent with the estimates of relative risks in the literature.⁹ However, a comprehensive measure of family history explained around half of the variance explained by directly measured genetic risk. In accordance with simulations in a previous report,³² GRS outperforms traditional family history, and will likely improve further with larger studies now in analysis.³³

The effect of family history was attenuated but remained significant when jointly analyzed with GRS measures for schizophrenia, bipolar disorder, MDD, ASD, educational attainment, and BMI. This indicates that family history contains information not captured by these multiple GRS. Intuitively, the reduction in the effect of family history was due to shared genetic information captured by both measures. We expect that the evolving GRS with better accuracy will continue to capture a higher proportion of schizophrenia risk associated with family history, although it will be inherently capped at a level determined by the SNP heritability, i.e., the phenotypic variance attributable to common genetic variants that are tagged on the genotyping arrays. The independent effects captured by family history could be due to common or rare genetic variation not well captured by the SNP arrays used in these GWAS as well as shared environmental risk factors. On the other hand, few individuals would have a positive family history, which may limit its utility as a predictive measure.

We note that, in the absence of a family history, there is a risk gradient associated with different profiles of GRS. The combination of GRS and comprehensive family history explained 9% of variance in schizophrenia liability — better discrimination accuracy can be achieved by jointly considering GRS and family history. This has proven useful in other diseases. For prostate cancer, integrating family history, prostate cancer GRS, and protein biomarkers led to a clinically important prediction algorithm that is now being implemented in routine clinical care.³⁴ In our data, joint modelling both measures yielded an area under the receiver operating characteristic (AUC) curve of 0.73. This level of discriminative accuracy is close to the AUC threshold of 0.75 which is regarded as the minimum for clinical utility.^35,36 Future studies should evaluate the impact of additional epidemiological and genetic risk factors (including the burden of rare copy number variants and exon variation). Together with other risk factors, a reasonable predictive utility may be attainable.

In conclusion, by integrating register data and carefully curated common-variant SNP data from multiple Nordic countries, we determined that “family history” contains information that is not captured by GRS for multiple psychiatric disorders and educational attainment. By combining family history with these multiple GRS, we obtained levels of discriminative accuracy that – with replication in other studies and in samples with a lower prior probability of schizophrenia than in these case-control cohorts – could prove to be clinically useful.