Multivariable G-E interplay in the prediction of educational achievement

Abstract

Polygenic scores are increasingly powerful predictors of educational achievement. It is unclear, however, how sets of polygenic scores, which partly capture environmental effects, perform jointly with sets of environmental measures, which are themselves heritable, in prediction models of educational achievement. Here, for the first time, we systematically investigate gene-environment correlation (rGE) and interaction (GxE) in the joint analysis of multiple genome-wide polygenic scores (GPS) and multiple environmental measures as they predict tested educational achievement (EA). We predict EA in a representative sample of 7,026 16-year-olds, with 20 GPS for psychiatric, cognitive and anthropometric traits, and 13 environments (including life events, home environment, and SES) measured earlier in life. Environmental and GPS predictors were modelled, separately and jointly, in penalized regression models with out-of-sample comparisons of prediction accuracy, considering the implications that their interplay had on model performance. Jointly modelling multiple GPS and environmental factors significantly improved prediction of EA, with cognitive-related GPS adding unique independent information beyond SES, home environment and life events. We found evidence for rGE underlying variation in EA (rGE = .38; 95% CIs = .30, .45). We estimated that 40% (95% CIs = 31%, 50%) of the polygenic scores effects on EA were mediated by environmental effects, and in turn that 18% (95% CIs = 12%, 25%) of environmental effects were accounted for by the polygenic model, indicating genetic confounding. Lastly, we did not find evidence that GxE effects significantly contributed to multivariable prediction. Our multivariable polygenic and environmental prediction model suggests widespread rGE and unsystematic GxE contributions to EA in adolescence.

Fig 1. Multivariable prediction of educational achievement.

Panel A = repeated 10-fold cross validation in training set, for the environmental (E), multi-polygenic score (G), joint (G+E), and interaction (G*E) prediction models. Panel B = Hold-out set prediction of EA for best models obtained via repeated cross validation in training set. Error bars are 95% bootstrapped confidence intervals. Panel C = G+E model used in hold-out set prediction. Figure shows variables selected via repeated cross-validation in the training set, and relative importance. Panel D = Comparison of prediction accuracy for models tested as bootstrapped R² difference between nested models in the hold-out set. Distributions represent independent (non-mediated) genetic effects (G+E−E), environmental effects (G+E−G), and G*E effects (G*E–G+E). Note. PGS = polygenic scores, ENV = Environmental measures. ASD = Autism Spectrum Disorder, BIP = Bipolar Disorder, BMI = Body Mass Index, EA3 = Educational Attainment, IQ3 = Intelligence, OCD = Obsessive Compulsive Disorder, PTSD = Post-Traumatic Stress Disorder, SCZ = Schizophrenia.

Fig 2. Relative contributions of model selected variables for the G+E model in the prediction of educational achievement.

Figure shows partial regression coefficients, and 95% CIs around estimates. Naive = partial regression coefficients from multiple regression of selected variables in Training set; Hold-out = partial regression coefficients of selected variables in the hold-out set; Conditional = partial regression coefficients of training set for selected variables estimated with a conditional probability from a truncated distribution (see methods section). Note. ASD = Autism Spectrum Disorder, ADHD = Attention-Deficit Hyperactivity Disorder, BIP = Bipolar Disorder, EA3 = Educational Attainment, IQ3 = Intelligence, MDD = Major Depressive Disorder, SWB = Subjective Well-Being, OCD = Obsessive Compulsive Disorder, PTSD = Post-Traumatic Stress Disorder, Risk PC1 = First Principal Component of Risky behaviors, SCZ = Schizophrenia.

Fig 3. Interaction network of glinternet model.

Note. Edges width represent interactions weights. E = Environmental measure, G = Genome-wide polygenic score. Polygenic scores acronyms: ASD = Autism Spectrum Disorder, ADHD = Attention-Deficit Hyperactivity Disorder, BIP = Bipolar Disorder, EA3 = Educational Attainment, IQ3 = Intelligence, Income = household income, MDD = Major Depressive Disorder, SWB = Subjective Well-Being, OCD = Obsessive Compulsive Disorder, PTSD = Post-Traumatic Stress Disorder, Risk PC1 = First Principal Component of Risky behaviors, SCZ = Schizophrenia.

Fig 4

Panel A = schematic representation of mediation analysis; βC = effect of a predictor X on an outcome Y; βa = effect of X on a mediator (M); βb = effect of M on Y after adjusting for X; βC’ = effect of X on Y after adjusting for M. Panel B = Directed acyclic graph (DAG) showing E_ea mediated effects of G_ea on EA in the hold out-set; β_ge = causal path between G_ea and E_ea equivalent to r_G,E; β_eEA = direct independent E_ea effects on EA; β_gEA = total G_ea effects on EA. Panel C = DAG showing G_ea mediated effects on EA (genetic confounding, see methods and discussion); β_eg = causal path between E_ea and G_ea equivalent to r_G,E; β_gEA = direct independent G_ea effects on EA; β_eEA = total E_ea effects on EA. Note. Blue paths represent G model effects, yellow paths represent E model effects.