Cognitive intermediate phenotype and genetic risk for psychosis

Highlights

• General cognitive ability and working memory are robust intermediate phenotypes.

• Molecular genetic analyses advanced heritability and genetic overlap estimation.

• Use of cognitive intermediate phenotypes in genomic studies is gaining momentum.

• Intermediate phenotypes may aid functional characterization of psychosis risk genes.

Intermediate phenotypes (IPs) are defined as measurable liability traits underlying complex phenotypes, posited to be more genetically tractable than the phenotypes themselves. Here we review evidence for cognition as an IP of psychosis, and highlight topical advances in the literature: first, heritability estimation of cognitive abilities using genomewide complex-trait analysis; second, evidence that cognition lies upstream to schizophrenia liability; third, use of polygenic risk scores rather than single genetic variants to examine genetic overlap between cognitive IPs and schizophrenia; and fourth, use of cognitive IPs for schizophrenia risk gene discovery and functional characterization. We end with future directions in using cognitive IPs to study genetic risk of psychosis, including methodological refinements and shifting research focus from identifying IPs to using them.

Introduction

Behavioral genetic research employing family, twin and adoption designs have consistently shown that psychotic symptoms and disorders are under substantial genetic influence, suggesting that the genome contains considerable relevant information about their underlying pathology [1, 2]. However, the first few decades of genetic linkage and candidate gene-based association studies of schizophrenia spectrum disorders were disproportionately slow to bear fruit. One explanation for this limited success might be the heterogeneity and subjectivity of the phenotypes used in early studies (i.e., clinical diagnoses). Accordingly, some researchers proposed replacing the use of qualitative diagnoses with quantifiable liability traits underlying psychiatric disorders (e.g., laboratory-based physiological or cognitive deficits) as the phenotypes in molecular genetic studies [3]. Such liability traits, termed ‘intermediate phenotypes’ (IPs) or ‘endophenotypes’, were posited to lie along the causative pathway between phenotype (i.e., clinical symptoms) and genotype (i.e., risk genetic variations). Hence, they were assumed to be more directly and strongly influenced by susceptibility genes of psychosis than symptom manifestations. By decomposing a complex phenotype like psychosis into circumscribed and more genetically tractable traits, it was hoped that the neurobiology and genetic architecture of psychosis could be clarified.

IPs can be biochemical, endocrinological, neurophysiological, neuroanatomical, or cognitive in nature, as long as they satisfy established criteria of being associated with the illness in the population; heritable; primarily state-independent (i.e., not fluctuating with illness course); cosegregating with the illness within families; and found in unaffected family members at a higher rate than in the general population [3]. Cognition has received widespread support as a putative IP for psychosis. The purpose of this review is to examine the robustness of general and specific cognitive abilities as IPs of psychosis, and to summarize recent attempts and limitations in using cognitive IPs to shed light on the genetic architecture of psychosis.

The flourishing field of cognitive IPs in psychosis owes much to early studies of cognition in unaffected siblings and co-twins of patients with schizophrenia in the 1990s and early 2000s. These seminal studies not only shed light on the familial basis of cognitive impairments in schizophrenia (e.g., working memory, memory, attention, and executive function) [4, 5, 6, 7], but also formed the paradigmatic basis for subsequent studies. The past decade has seen a proliferation of putative cognitive IPs for psychosis. Alterations in general (e.g., intelligence) as well as specific neurocognitive abilities (e.g., working memory, verbal memory and attention) have been suggested as IPs. More recently, some IPs have been combined to improve statistical power. Here, we provide an update on the extent to which these candidate IPs fit established criteria for psychosis.

General cognitive ability (GCA) is the apex in hierarchical models of cognitive abilities [8], and encompasses specific cognitive domains such as attention, memory and executive function. GCA impairments in schizophrenia are widely reported [9, 10], and research continues to provide support for GCA as an IP for psychosis. The association of GCA with psychosis in the general population was validated by a large epidemiological study [11], while state independence of GCA was robustly demonstrated in clinical ultrahigh-risk (UHR) populations [12, 13, 14], as well as individuals in active and remitted states of first episode psychosis (FEP) [15, 16]. Although one study reported a widening gap in GCA between individuals with FEP and controls in longitudinal studies, findings suggested that it could be ascribed to patients’ failure to gain from practice effects relative to controls, rather than GCA deterioration over the course of illness [17].

Advances in heritability estimation have not only confirmed sizeable genetic contributions to GCA, but also revealed novel clues about its genetic architecture. Heritability is traditionally estimated by comparing observed and expected phenotypic resemblance in related samples, such as twins and siblings or parents and offspring. Recent twin studies have employed larger samples or meta-analysis to increase statistical power in heritability estimation [18, 19, 20]. A new method for estimating genetic influence has furthermore made it possible to estimate the heritability of GCA in unrelated individuals [19, 21••, 22••]. This method, called ‘genomewide complex-trait analysis (GWCT)’, generates heritability estimates from genotyped DNA markers (e.g., single nucleotide polymorphisms; SNPs), which are typically available in genomewide association analyses (GWAS) [23]. Since the twin method estimates take into account both additive and non-additive genetic effects, while GWCT estimates include only additive genetic effects tagged by SNPs, one could deduce the genetic architecture of GCA by comparing heritability estimates generated from these two methods [19, 23]. Based on a study that directly compared twin-based estimates (0.46) and SNP-based estimates (0.35) in the same sample [19], it was projected that more than half of the heritability of GCA could eventually be found in well-powered GWAS, while the remainder might be attributed to other genetic variants, non-additive effects, and environmental factors [24].

With regards to familial association and cosegregation, recent studies and meta-analysis of unaffected relatives of individuals with schizophrenia showed that relatives performed more poorly than controls [25, 26]. This suggested that GCA deficits vary quantitatively with degree of genetic risk for psychosis, as would be expected if there is substantial overlap in genes contributing to GCA and psychosis. There is also evidence for cosegregation of GCA and psychosis within families [27, 28]. In summary, recent studies lent support to GCA as a putative IP based on fulfillment of established criteria.

Although a range of specific cognitive deficits has been posited as candidate IPs, here we focus on impairments in working memory (WM), verbal memory (VM) and sustained attention, as they have received relatively more attention in the recent literature (Figure 1). These are good candidate IPs because they are well-replicated impairments in subclinical and clinical psychosis [29, 30, 31, 32•, 33, 34, 35, 36], have a relatively established biological basis [37•, 38], and are easy to measure, even in large samples and within neuroimaging scanners.

A plethora of recent evidence supported WM as being state-independent [15, 31, 39, 40], over-expressed in unaffected siblings [41], and cosegregated with psychosis within families [36, 41, 42, 43]. Evidence for VM and sustained attention is less extensive, though both have also been shown to be state-independent [44, 45, 46], and to display familial association [17, 42, 47, 48, 49] and cosegregation with psychosis [36, 47]. Although WM, VM and sustained attention have each been demonstrated to be heritable [50, 51, 52, 53], the extent to which they remain so even after accounting for GCA deficits is under-studied. Only one recent study has demonstrated that spatial WM performance remained heritable after controlling for GCA [51], supporting it as an independently useful cognitive IP.

Concerning familial association and cosegregation of specific cognitive abilities with psychosis, of particular interest are studies that combined behavioral data with neuroanatomical [54] or neuroactivation data [32•, 43, 47, 55, 56]. For WM and sustained attention, such studies showed that even when no significant difference in task performance was apparent in UHR individuals versus controls [32^•], or in unaffected relatives versus controls [36, 56], there could be differences in task-related brain activations. For VM, neither task performance on the Logical Memory subtest of the Wechsler Memory Scale (WMS) nor task-related brain activation was found to discriminate between schizophrenia patients and unaffected siblings [47]. This stood in contrast from earlier twin and family studies that showed modest genetic overlap between episodic memory (measured by verbal paired associates and Logical Memory subtests of WMS) and schizophrenia [57, 58]. More studies are needed to clarify or reconcile such inconsistencies. Taken together, WM stands out as the most promising specific cognitive IP [59].

Many of the studies cited above used individual neurocognitive tasks (e.g., N-back tasks for WM, list learning tasks for VM, Continuous Performance Test for attention), which might be vulnerable to task idiosyncrasies. A better strategy might be to combine individual cognitive tasks into composite measures, such as by factor analysis [22••, 50] or combining IPs with high shared genetic variations (e.g., IQ and WM) [37•, 60•, 61, 62]. Theoretically, such composite measures might confer more statistical power [37•, 60•]. Empirically, they have been shown to be heritable [60^•] and stable over a 10-year illness course in FEP [62]. However, the extent to which such composite measures show better penetrance than individual IPs has yet to be tested empirically.

The assumption that IPs lie on the causal pathway from gene to phenotype is mostly presumed or indirectly inferred through the criterion of state independence, rather than directly born out by experimentation. Through the use of reciprocal causation modeling in a twin sample, Toulopoulou and colleagues investigated the direction of causation among three variables: cognitive IPs (IQ, immediate recall and delayed recall), brain volume changes, and schizophrenia liability. Results showed that cognitive IPs lied upstream to schizophrenia liability, while brain volume changes lied downstream to schizophrenia liability [63^••]. Such results suggested that while cognition is causative to schizophrenia risk, brain volume changes might be a consequence. This is the first study to date to establish causal pathways between genes and cognition, and is an important finding in demonstrating the robustness of cognition as an IP.

A recent review has suggested that the field has been more successful at identifying IPs than actually using them in mapping genetic risk of psychiatric disorders in the past decade [64]. Here, we provide emerging evidence that paints a more hopeful picture, and highlight the potential of cognitive IPs in advancing the study of the genetic risk of psychosis.

Early promises were that IPs would accelerate gene discovery, since they were assumed to be genetically simpler and more directly influenced by psychosis risk genes [3, 65]. Hence, loci for cognitive IPs could first be identified, and then be tested for association in psychotic cohorts. However, recent studies have revealed that cognitive IPs are also polygenic (i.e., influenced by many genes of small effect sizes) and are probably no less genetically complex than psychosis [22••, 66]. Thus, early studies trying to map specific genetic variants associated with cognitive IPs using sample sizes in the order of thousands were only met with sporadic success [52, 67, 68]. Recently, however, large datasets have been made possible through meta-analysis. In a sample of more than 50,000 middle-aged and older adults aggregated from 31 GWAS cohorts, three genome-wide significant genomic regions containing 13 SNPs associated with general cognitive function were identified [22^••]. One of these regions (14q12) contained the NPAS3 (neuronal PAS domain protein 3) gene, which had been previously associated with schizophrenia. This study showed that although cognitive IPs are polygenic, sufficiently-powered GWAS could nonetheless map specific genetic variants associated with cognitive IPs, which could then be followed up in psychotic cohorts. Given that gene discovery using cognitive IPs has only just begun to gain momentum, we anticipate novel findings and replications in future studies.

Despite budding success in schizophrenia risk gene mapping, a growing number of researchers are of the opinion that the biggest contribution of IPs lies instead in mapping the neurobiological signature of schizophrenia (i.e., using cognitive IPs to aid the identification of genetically mediated neurobiological pathways of schizophrenia) [69, 70••]. The rationale is that, although GWAS has been successful in identifying some psychosis susceptibility genes [71], it cannot inform how the discovered genetic variations might impact function to increase risk for developing psychosis. By examining the impact of schizophrenia-associated loci on cognitive IPs, researchers could deduce brain functions or other biological mechanisms by which disease risk variants act to produce the complex outcome of psychosis [69, 70••].

Cognition and psychosis are both polygenic; hence, single genetic variants do not capture much phenotypic variance. To circumvent this problem, recent studies employed ‘polygenic risk scores’ (PRS), which aggregates the effects of individual risk alleles into a cumulative genetic risk measure for each phenotype, to increase statistical power [72]. Significant correlations between two phenotypic PRS can be interpreted as the degree of genetic overlap between those phenotypes (e.g., schizophrenia and an IP), thus implicating involvement of the IP in psychosis liability. In the first molecular genetic study employing the PRS strategy to examine genetic overlap between cognition and psychosis, a sample of over 5000 individuals demonstrated that schizophrenia patients possessed fewer alleles associated with good GCA and more alleles associated with poor GCA [67]. Such results suggested that poor GCA might be a deficit mechanism through which psychosis genetic risk is conferred.

Similar results have been found in attention, spatial WM and memory [73, 74, 75]. Increased schizophrenia PRS have been shown to predict WM-related frontal lobe dysregulation across cases and controls, suggesting the involvement of WM and its biological substrates in genetic risk for the psychosis continuum [72]. The cell adhesion molecule (CAM) pathway was also found to be related to poor memory and attention in patients with psychosis, suggesting that CAM might encode cognitive impairment through its known involvement in disruptions to neurones and synapse formation, ultimately leading to psychosis [74]. Future research that continues to study cognitive IPs across several levels of analyses will clarify the cascade of psychosis pathoneurobiology from genes to the proteins they encode, to cellular systems and signaling pathways, to neural system dysfunction, to cognitive dysfunction, and finally symptom manifestations in psychosis [65].