Grant: Discovery Grants Program - Individual

Genomic prediction (GP) uses computational methods to select varieties with desired phenotypes based on whole-genome markers from a population. GP is an important tool in crop development, allowing breeders to select varieties of a crop with desirable phenotypes. GP excels when applied to traits with complex genetic architectures, such as yield. GP research increasingly uses deep learning (DL) to enhance predictive accuracy. Key issues have emerged in crop GP due to the availability of large genomic datasets capable of use in DL, as well as the addition of growing season information for plants grown in diverse environmental locations. This program of research has a long-term goal to improve DL predictions, aid breeding programs, and provide fundamental insight into genomic regions related to important agronomic traits.  Rich time series field-level weather data is an important component of predicting phenotypes in crop GP. My results show that DL models that incorporate environmental data can significantly improve phenotype prediction even in previously unobserved environments. This data aids in modelling Genotype Environment Interaction (GEI). More complex DL models can be used with these data sets, which improves prediction and reduces overfitting. The use of weather data also allows DL models to predict performance in novel or future environmental conditions. Suitable representation of weather data is essential for successful DL prediction. An objective of this proposal is to develop DL models that are suited for time series data, such as attention-based models. Using DL to select genomic markers that are relevant in particular environments will allow insight into GEI. Domain knowledge on the effect of environmental conditions at different stages of growth will develop models that transform weather data into features that are centered on plant growth stage. Additionally, unstructured data can also be used to improve GP prediction, including field management data. Feature selection in multi-environment crop GP will also be considered. Our research has also demonstrated that performing feature selection with the aim of generating a universal set of markers for all environments may not be suitable. This research objective will develop new feature selection methods that allow integration of environmental conditions to improve prediction, by retaining genomic markers that are critical under different growing conditions. This proposal will develop techniques for predicting future growing seasons based on historical records. The DL models that predict future growing seasons will rely on the representations of weather data produced in this research: the models for future conditions will generate data in a format known to allow GEI to be predicted. This DL research will be used to allow crop breeders to make better decisions, accelerate breeding programs, reduce long-term costs, and aid in the development of crops for future climate conditions.