Modeling nonlinear and interaction effects of spatiotemporal and other non-genetic factors improves phenotypic prediction for complex traits.

Adjusting for non-genetic factors can improve genetic association testing and polygenic prediction, yet most studies rely on linear adjustments for a small set of covariates. Location and time covariates are typically not included in genetic studies, but they can serve as useful proxies for environmental exposures such as climate, pollution, and behavioural differences. These features especially highlight the weakness of linear covariate adjustments, which cannot capture their cyclical or nonlinear effects nor interactions between covariates. To evaluate the benefits of modeling nonlinear covariate effects, including spatiotemporal features, we adopted a null model approach in which an auxiliary nonlinear model predicts the phenotype from non-genetic covariates alone. This prediction is then included as an additional covariate in downstream analysis, allowing association studies, polygenic scores, and fine-mapping to account for nonlinear covariate effects and interactions without modifying existing tools. Using 16 continuous phenotypes from the UK Biobank, we first evaluated gradient boosted decision tree (GBDT) and neural network null models with combinations of additional time-of-day, time-of-year, and birth or home location covariates and determined how well these models could predict phenotypes from covariates alone compared to linear models. A GBDT null improved average R