Unravelling genomic and environmental contributions to nitrogen-use efficiency in tropical maize through Bayesian causal forests

Abstract

Nitrogen fertilisation is one of the largest variable input costs in maize production, and genotypes differ in how efficiently they convert added nitrogen into yield. Breeding for nitrogen-use efficiency therefore depends on knowing not only the average response of the crop to nitrogen but how that response varies across genotypes and environments. We frame this as a problem of causal inference and estimate the effect of an ideal versus low nitrogen regime on three maize traits jointly (grain yield, plant height and ear height) using a multivariate Bayesian causal forest. To our knowledge, this is among the first studies to quantify the effect of nitrogen in a genotyped breeding panel as a formally identified causal quantity, estimated from a designed contrast within the potential-outcomes framework, rather than as the genotype-environment association that genomic selection and association mapping report. The trial suits a causal reading: 614 single-cross hybrids were grown under both nitrogen levels in adjacent plots sharing weather and soil, so nitrogen was assigned by design rather than confounded with genotype or environment, and the estimated propensity score was close to 0.5. The leading principal components of the additive genomic relationship matrix were placed in the moderator set, so that both the baseline outcome and the treatment effect could vary with genomic background. Ideal nitrogen raised all three traits (grain yield +0.59, 95% credible interval 0.51 to 0.67; plant height +0.095; ear height +0.044, in adjusted-phenotype units). The heterogeneity of the effect was almost entirely environmental, with a small genotypic component and negligible genotype-by-environment interaction; the posterior credible interval of the effect was wider than the spread among genotypes, so differences in responsiveness among hybrids were only weakly resolved. Reading the two tree ensembles separately indicated that genomic background governs the yield level while climate governs the response to nitrogen, the leading genomic contribution being a single axis. A reaction-norm variance decomposition and cross-validated prediction corroborated this reading. A multivariate causal forest with genomic moderators, read through its two ensembles and reported with posterior uncertainty, is a suitable and transparent tool for studying nitrogen response in breeding material.