Nonlinear spatiotemporal factors improve prediction for complex traits

Ross DeVito and Melissa Gymrek; University of California San Diego | Nature Communications | peer reviewed | 29 August 2026

Problem: genetic models often use limited linear covariate adjustment. Contribution: gradient-boosted null models incorporate location, time and nonlinear interactions; the authors report a median 7.3% R² gain with one downstream method across 16 UK Biobank phenotypes. Significance: a small workflow addition can expose environmental effects. Limitations: the metric is author-reported from a specific biobank and needs external population validation.