The AgriCoolTools project aims to develop new predictive digital tools to improve plant and animal varietal selection processes in order to accelerate the transition of agricultural systems. The ambition is to propose new approaches to predict multivariate or longitudinal traits of interest in heterogeneous populations from high-dimensional explanatory variables, such as genetic markers or low-cost intermediate phenotypes such as near-infrared spectra.
These approaches will combine advanced mechanistic modeling, statistics and machine learning methods, while providing theoretical guarantees on predictions. The interdisciplinary consortium is made up of researchers with strong complementary expertise in modeling, statistics and machine learning, as well as in biology, physiology, eco-physiology and quantitative animal and plant genetics.
The objectives are to:
develop methods to take into account biological information such as complex dependence structures between the traits of interest to be predicted and between highdimensional explanatory variables in dynamic mixed-effects models;
propose predictive approaches for multivariate traits of interest and for a time of interest based on variable selection methods in nonlinear and joint mixed effects models by providing theoretical guarantees for predictions;
improve the quality of inference and prediction by hybridizing generative mechanistic models of traits of interest and machine learning approaches trained on additional synthetic data. For each of the objectives, we will focus on the theoretical guarantees of the methods and their numerical feasibility. The tools developed will be implemented to analyze the data sets of the project partners.