Bayesian optimisation for yield in high-dimensional trait-space identifies crop ideotypes in Oil Seed Rape

Abstract

The improvement of crop yield has long been a major breeding target and is increasingly becoming a goal in many areas of plant research. Yield has been shown to be a complex trait, depending on multiple genes, plant architecture and plant-environment interactions. This complexity is frequently reduced by focussing on contributing factors to yield (yield traits). However, a quantitative understanding of the interplay between yield traits, and the effect of these relationships on yield is largely unexplored. Consequently, the extent to which crop varieties achieve their optimal morphology in a given environment and how this impacts on seed yield is unknown. Here we use causal inference to model the hierarchically structured effects of 27 macro and micro yield traits on each other over the course of plant development, and on seed yield in Spring and Winter oilseed rape plants. We perform Bayesian optimisation on the modelled yield potential, identifying the morphology of ideotype plants which are expected to be higher yielding than the existing varieties in the studied panels. We find that existing Spring varieties occupy the optimal regions of trait-space, but that potentially high yielding strategies are unexplored in extant Winter varieties. In addition to concrete recommendations for varietal improvement in oilseed rape, this work provides a novel, general methodological framework for the study of crop breeding as an optimisation problem.