Multivariate multi-informant genomic prediction of developmental psychopathology from childhood to emerging adulthood

Background The statistics of mental health problems in the UK portrays a dark picture: 1 in 10 children have a diagnosable mental health disorder, approximately three children in every classroom (Green, McGinnity, Meltzer, Ford, & Goodman, 2009; Murphy & Fonagy, 2012). These mental health problems may have a long shadow later in life for children (Kelly-Irving et al., 2013; Kessler, Foster, Saunders, & Stang, 1995); they also affect children's families (Bogels & Brechman-Toussaint, 2006) as well as the overall economy (Prince et al., 2007). In addition to being associated with adverse life outcomes, these mental health problems can lead to self-harm and suicide (Fombonne, Wostear, Cooper, Harrington, & Rutter, 2001). Furthermore, studies have shown that the earlier these symptoms appear, the worse the outcome tends to be (Otto et al., 2001). It is of vital importance to identify children who are at greatest risk of developing mental health problems as early as possible because earlier intervention or prevention tends to be most effective (Colizzi, Lasalvia, & Ruggeri, 2020). One way to identify children at highest risk of developing mental health problems is to study their vulnerability to any mental health problems rather than focusing on the risk of a specific disease, taking into account both genetic and environmental risk factors. Recent research has provided empirical evidence for a general factor of psychopathology (p-factor), indicating that the range of psychopathology symptoms, from internalizing to externalizing problems, can be summarized by a common latent factor that explains a large part of the variance. This p-factor, capturing the vulnerability to mental health problems, emerges taking into account dimensionality of disorders (quantitative continuum of symptoms) as well as diagnoses, persistence and co-occurrence of the range of disorders (Allegrini et al., 2020; Caspi et al., 2014; Martel et al., 2017), both during childhood and adulthood. The p-factor is predictive of multiple negative outcomes, general wellbeing and impairment throughout life (Caspi & Moffitt, 2018). The p-factor is analogous to the g-factor (general cognitive ability, intelligence) indicating that a common factor characterizes a range of cognitive abilities, although there is some specificity for verbal, non-verbal and spatial abilities (Carroll, 1993). Similarly, there is some specificity for psychopathology symptoms after accounting for the p-factor (Allegrini et al., 2020; Selzam, Coleman, Caspi, Moffitt, & Plomin, 2018). An important factor to consider is what explains the wide individual differences in p-factor. Individual differences in mental health problems are explained by both genetic and environmental factors. Twin and family studies have shown that virtually all dimensions of psychopathology are explained by substantial genetic influence, with heritability estimates ranging from 30-80% (Knopik, Neiderhiser, DeFries, & Plomin, 2017). Behavioural genetic and genomic studies have also indicated substantial genetic correlations across dimensions of psychopathology, as well as across medical diagnoses, implying that shared genetic factors contribute to the co-occurrence of the range of dimensions of psychopathological traits and diagnosed mental health disorders (Allegrini et al., 2020; Knopik et al., 2017). The common p-factor is also found to be highly heritable (50-60%) during childhood and adolescence and highly heritable across raters (Allegrini et al., 2020). Great progress has been made in identifying specific genetic markers associated with a range of mental health problems (Visscher et al., 2017). Individually these markers are not very useful because their effect sizes are very small. However, markers can be aggregated in genome-wide polygenic scores (GPS) to predict psychopathology symptoms across development (Middeldorp & Wray, 2018). The predictive ability of GPS constructed from genome-wide association (GWA) studies of a range of mental health traits has been shown to be significant, but the predictive accuracy is still modest. The goal of the proposed study is to boost the predictive power of GPS by focusing on the p factor. Multivariate prediction models that leverage joint prediction of multiple GWA summary statistics have been shown to be effective in boosting power and predictive accuracy of GPS (Allegrini et al., 2020, 2019; Grotzinger et al., 2019; Maier et al., 2018; Turley et al., 2018). There are several multivariate methods available for researchers to use (Genomic SEM, MTAG, SMTpred). The predictive accuracy for the cognitive traits has been shown to be similar between these different methods (Allegrini et al., 2019). However, to our knowledge it has not been tested empirically for mental health outcomes, that is if the different multivariate methods differ for the prediction for psychopathology symptoms (and diagnoses of mental health problems) across development. The current project will investigate how the genetics of the p-factor can best be captured using methods leveraging GWA summary statistics -- both multi-trait GWA prediction (e.g. Genomic SEM; MTAG) and multi-polygenic score models (e.g. multiple regression; prediction models using the elastic net). We will test whether the way genomic p is constructed leads to differences in prediction in both the dimensional p- factor (continuous measures of psychopathology) and clinical outcomes (diagnoses by a psychiatrist; p-factor) during childhood and early adulthood. We will systematically test the multivariate prediction of GPS (using both multi-GPS and multi-trait GWA prediction) to the p-factor from childhood to emerging adulthood (using data from self-, parent-, teacher-ratings). This project will leverage the most recent GWA summary statistics of psychiatric problems in adulthood. In addition, the project will attempt to boost the power to predict the p factor by including GPS of nonpsychiatric traits, most notably educational attainment and intelligence. It has been shown both phenotypically and genetically that educational attainment and intelligence are predictors of psychiatric problems both during childhood and in adulthood (Harden et al., 2019; Plomin & Deary, 2015). We will test the extent to which polygenic scores derived from GWA summary statistics from intelligence, educational attainment and personality-related phenotypes like neuroticism and risk-taking will increase the prediction of the p-factor. We will test the prediction in both dimensional/quantitative measures of mental health as well as clinical outcomes of mental health using multivariate genomic prediction methods building on work by Mallard and colleagues (2019), which suggests that diagnostic criteria in adulthood are only moderately related to self-reported symptoms (Mallard et al., 2019). Here we will test the prediction of p- factor in self-, parent- and teacher ratings, as well as clinical diagnoses across childhood and adulthood. Finally, we will study the specificity of psychiatric problems after accounting for genomic p factor. There are symptom-specific genomic factors, or residual variance in psychopathology after controlling for multivariate GPS predictions (genomic p-factor) (Grotzinger et al., 2019). Here we will test how this symptom-specific variance could increase prediction of specific symptoms of psychopathology (internalizing, externalizing and neurodevelopmental disorders) during childhood and emerging adulthood, over and above the genomic p factor. Genomic SEM will be employed to fit the common pathway model to the multivariate data using summary statistics of adult psychiatric GWA, testing for both common and specific factors of psychopathology. Using the summary statistics derived from Genomic SEM, we will construct GPS to test the variance explained in the phenotypic p-factor and the specific psychopathology symptoms from childhood to emerging adulthood in two representative samples from the UK and Sweden. Our project will take a systematic look at how the genetics of p can best be aggregated to predict the predisposition to psychopathology from childhood to early adulthood. This prediction could potentially be useful in order to identify children at greatest risk who would benefit most from early prevention and intervention. Additionally, the future potential impact of this work would allow using the genomic p factor in studies as a control variable. When studies explore the environmental risk factors of psychopathology, genetic p allows to partially control for genetic confounding, thus getting a cleaner measure of non-genetic, environmental risk factors of adverse mental health.