Genetic association studies (henceforth GAS, used in the plural) aim to identify genes or genetic variants (GVs) that are systematically associated with variation on the behavioural level (Balding Reference Balding2006; Hirschhorn & Daly Reference Hirschhorn and Daly2005). In its exploratory form, GAS test the associations between 300.000-500.000 GVs typed across the entire genome and a phenotype of interest, which can be continuous or dichotomous, such as affection-status. For instance, consider a GV with two alleles: A and B. If allele A is more frequently observed in cases than in controls, then this allele A may be associated with an increased risk for the disease under study. Where successful, GAS are a first step towards revealing functional relations between the genome and behaviour. GAS have successfully localized GVs for medical conditions such as Crohn's disease (Zhang et al. Reference Zhang, Qiu, Chen, Zu, Mao, Zhu, Zheng, Wang, Lei and Shi2009) and type-I diabetes (Barrett et al. Reference Barrett, Clayton, Concannon, Akolkar, Cooper, Erlich, Julier, Morahan, Nerup, Nierras, Plagnol, Pociot, Schuilenburg, Smyth, Stevens, Todd, Walker and Rich2009), but, despite high family-based heritability estimates, have been less successful for psychiatric traits. Disappointing results have been attributed to statistical problems (e.g., low power), genetic complexity (e.g., gene-gene interaction), and genetic heterogeneity (e.g., the genetic etiology of disease may vary across families). Incorrect characterisation of the phenotypic model, however, may also render GAS inefficient.
In the search for GVs associated with psychiatric disorders, symptoms are usually considered manifestations of an underlying latent trait, which causes the variance in the symptoms, and the observed relations between the symptoms. The ultimate aim of GAS is to identify the GVs that cause individual differences in the latent trait, as under this model the relations between GVs, on the one hand, and individual symptoms, on the other, are mediated by the latent trait. A common operationalization of a latent trait in GAS is the sum score calculated across all items or symptoms of a diagnostic test. The sum score can subsequently be subjected to dichotomization following diagnostic cut-off criteria. This score then features as the dependent variable in GAS, and GVs are used to predict the (dichotomized) sum score. Although the (dichotomized) sum score may be a crude approximation of the latent trait, the operationalization is consistent with the latent trait model.
In the network model proposed by Cramer et al., however, the observed relations between symptoms are not attributable to a common latent cause but result from the direct causal relations among the symptoms themselves. In this model, all symptoms could in principle have a unique genetic etiology, so the sum score operationalization could potentially be counterproductive in the search for GVs. One may ask, however, whether the high twin-based and family-based heritability estimates (h
2) for psychiatric disorders are consistent with Cramer et al.'s network view.
Under the network model, considerable h
2 estimates remain feasible for two reasons. First, calculations of h
2 are based on within-family comparisons. Researchers test, for example, whether monozygotic twins, who share 100% of their genetic material, are phenotypically more alike than dizygotic twins, who share on average 50% of their genetic material (Falconer Reference Falconer1989). If variance in symptoms is genetic in origin, then the genetic basis of the disorder is likely to be shared by members of the same family. Even if all symptoms have a different genetic etiology, and this etiology differs across families (i.e., genetic heterogeneity), family-based estimations of h
2 can be considerable because they are based on comparisons within families. Second, the direct causal interrelations between symptoms, such as described in the network model, will over time induce genetic correlations between symptoms that are initially genetically unrelated, that is, do not functionally share any genetic basis (as shown by van der Maas et al. [Reference Van der Maas, Dolan, Grasman, Wicherts, Huizenga and Raijmakers2006] in the context of intelligence). That is, with the passing of time, a common genetic factor may evolve from the phenotypic interactions between genetically unrelated symptoms, and the h
2 of the sum score may increase.
The fact that the h
2 of a sum score calculated across genetically unrelated symptoms can be substantial has important implications for GAS. Specifically, the high h
2 of sum scores does not mean that this operationalization is useful in the search for the actual GVs of interest. The distinction between the network model and the latent trait model is thus essential in GAS: If the network model is the true model, then the common way geneticists operationalize their phenotypic information may be counterproductive. Under the network model, GAS on the individual symptoms, or even on the relations between the symptoms, as Cramer et al. suggest, make more sense. Of course, such an approach comes with its own challenges, such as the exacerbation of the problem of multiple testing (e.g., 10 symptoms may produce 1 sum score, but 10 symptom-level ones, and 45 relational tests).
At present, the network perspective of major depressive disorder (MDD) and generalized anxiety disorder (GAD) is quite speculative. A recent study by Lux and Kendler (in press), however, lends some credence to the network view of MDD. These authors showed that nine criteria for MDD (e.g., depressed mood, fatigue) displayed markedly different relations to factors such as risk for future episodes, risk of depression in the co-twin, and patterns of comorbidity. These results suggest considerable heterogeneity within the MDD syndrome, and are inconsistent with the idea that, within disorders, all symptoms are interchangeable indicators of a single causal latent trait. Although these results do not prove the network model, Lux and Kendler's findings are hard to reconcile with the latent trait model.
While GAS may benefit from a network approach to phenotypes, genetic studies may help to validate the network approach. Network models and latent trait models may yield quite similar phenotypic covariance structures (van der Maas et al. Reference Van der Maas, Dolan, Grasman, Wicherts, Huizenga and Raijmakers2006). Behaviour genetics provides several research designs that are informative about causality, such as direction of causation models, the co-twin control design, the children-of-twins design, and Mendelian randomisation. In additional, possible genetic influence on the interrelationship among symptoms can be addressed using moderated genetic covariance structure modeling (Purcell Reference Purcell2002).
Like the latent trait model, the network perspective is a theoretical model, which still requires validation. If true, however, it constitutes an interesting alternative viewpoint and could have far-reaching consequences for GAS.
You have
Access
Genetic association studies (henceforth GAS, used in the plural) aim to identify genes or genetic variants (GVs) that are systematically associated with variation on the behavioural level (Balding Reference Balding2006; Hirschhorn & Daly Reference Hirschhorn and Daly2005). In its exploratory form, GAS test the associations between 300.000-500.000 GVs typed across the entire genome and a phenotype of interest, which can be continuous or dichotomous, such as affection-status. For instance, consider a GV with two alleles: A and B. If allele A is more frequently observed in cases than in controls, then this allele A may be associated with an increased risk for the disease under study. Where successful, GAS are a first step towards revealing functional relations between the genome and behaviour. GAS have successfully localized GVs for medical conditions such as Crohn's disease (Zhang et al. Reference Zhang, Qiu, Chen, Zu, Mao, Zhu, Zheng, Wang, Lei and Shi2009) and type-I diabetes (Barrett et al. Reference Barrett, Clayton, Concannon, Akolkar, Cooper, Erlich, Julier, Morahan, Nerup, Nierras, Plagnol, Pociot, Schuilenburg, Smyth, Stevens, Todd, Walker and Rich2009), but, despite high family-based heritability estimates, have been less successful for psychiatric traits. Disappointing results have been attributed to statistical problems (e.g., low power), genetic complexity (e.g., gene-gene interaction), and genetic heterogeneity (e.g., the genetic etiology of disease may vary across families). Incorrect characterisation of the phenotypic model, however, may also render GAS inefficient.
In the search for GVs associated with psychiatric disorders, symptoms are usually considered manifestations of an underlying latent trait, which causes the variance in the symptoms, and the observed relations between the symptoms. The ultimate aim of GAS is to identify the GVs that cause individual differences in the latent trait, as under this model the relations between GVs, on the one hand, and individual symptoms, on the other, are mediated by the latent trait. A common operationalization of a latent trait in GAS is the sum score calculated across all items or symptoms of a diagnostic test. The sum score can subsequently be subjected to dichotomization following diagnostic cut-off criteria. This score then features as the dependent variable in GAS, and GVs are used to predict the (dichotomized) sum score. Although the (dichotomized) sum score may be a crude approximation of the latent trait, the operationalization is consistent with the latent trait model.
In the network model proposed by Cramer et al., however, the observed relations between symptoms are not attributable to a common latent cause but result from the direct causal relations among the symptoms themselves. In this model, all symptoms could in principle have a unique genetic etiology, so the sum score operationalization could potentially be counterproductive in the search for GVs. One may ask, however, whether the high twin-based and family-based heritability estimates (h 2) for psychiatric disorders are consistent with Cramer et al.'s network view.
Under the network model, considerable h 2 estimates remain feasible for two reasons. First, calculations of h 2 are based on within-family comparisons. Researchers test, for example, whether monozygotic twins, who share 100% of their genetic material, are phenotypically more alike than dizygotic twins, who share on average 50% of their genetic material (Falconer Reference Falconer1989). If variance in symptoms is genetic in origin, then the genetic basis of the disorder is likely to be shared by members of the same family. Even if all symptoms have a different genetic etiology, and this etiology differs across families (i.e., genetic heterogeneity), family-based estimations of h 2 can be considerable because they are based on comparisons within families. Second, the direct causal interrelations between symptoms, such as described in the network model, will over time induce genetic correlations between symptoms that are initially genetically unrelated, that is, do not functionally share any genetic basis (as shown by van der Maas et al. [Reference Van der Maas, Dolan, Grasman, Wicherts, Huizenga and Raijmakers2006] in the context of intelligence). That is, with the passing of time, a common genetic factor may evolve from the phenotypic interactions between genetically unrelated symptoms, and the h 2 of the sum score may increase.
The fact that the h 2 of a sum score calculated across genetically unrelated symptoms can be substantial has important implications for GAS. Specifically, the high h 2 of sum scores does not mean that this operationalization is useful in the search for the actual GVs of interest. The distinction between the network model and the latent trait model is thus essential in GAS: If the network model is the true model, then the common way geneticists operationalize their phenotypic information may be counterproductive. Under the network model, GAS on the individual symptoms, or even on the relations between the symptoms, as Cramer et al. suggest, make more sense. Of course, such an approach comes with its own challenges, such as the exacerbation of the problem of multiple testing (e.g., 10 symptoms may produce 1 sum score, but 10 symptom-level ones, and 45 relational tests).
At present, the network perspective of major depressive disorder (MDD) and generalized anxiety disorder (GAD) is quite speculative. A recent study by Lux and Kendler (in press), however, lends some credence to the network view of MDD. These authors showed that nine criteria for MDD (e.g., depressed mood, fatigue) displayed markedly different relations to factors such as risk for future episodes, risk of depression in the co-twin, and patterns of comorbidity. These results suggest considerable heterogeneity within the MDD syndrome, and are inconsistent with the idea that, within disorders, all symptoms are interchangeable indicators of a single causal latent trait. Although these results do not prove the network model, Lux and Kendler's findings are hard to reconcile with the latent trait model.
While GAS may benefit from a network approach to phenotypes, genetic studies may help to validate the network approach. Network models and latent trait models may yield quite similar phenotypic covariance structures (van der Maas et al. Reference Van der Maas, Dolan, Grasman, Wicherts, Huizenga and Raijmakers2006). Behaviour genetics provides several research designs that are informative about causality, such as direction of causation models, the co-twin control design, the children-of-twins design, and Mendelian randomisation. In additional, possible genetic influence on the interrelationship among symptoms can be addressed using moderated genetic covariance structure modeling (Purcell Reference Purcell2002).
Like the latent trait model, the network perspective is a theoretical model, which still requires validation. If true, however, it constitutes an interesting alternative viewpoint and could have far-reaching consequences for GAS.