Covariate-aware genomic prediction of blood metabolite profiles using multi-task neural networks
Predictive models of quantitative traits can combine genetic and covariate information, but overall predictive performance alone does not reveal whether differences between models arise from covariate, genetic or joint covariate-genetic effects. Circulating metabolites provide a high-dimensional set of clinically relevant quantitative traits in which these effects can be examined systematically. Although their genetic determinants are well characterised through genome-wide association studies, marginal associations do not establish how accurately metabolomic profiles can be predicted or whether nonlinear models can improve prediction by capturing complex and potentially interactive structure. Here, we developed a multi-task neural network (NN) for simultaneously predicting metabolomic profiles with a three-stage architecture separating covariate, genetic and joint contributions. In comparative analyses, the multi-task NN demonstrated the strongest mean performance across metabolites (R2=0.219), followed by the single-task NN (R2=0.211), elastic net (R2=0.207), and an activation-free multi-task model (R2=0.191). Decomposition analyses indicated that gains were mainly driven by nonlinear covariate modelling, consistent with improved prediction after incorporating nonlinear age effects into a linear model. Together, these analyses provide a framework for identifying sources of predictive differences between different models, which may be applicable to other collections of quantitative traits.