Multi-Task Learning for Genomic Prediction Robust to Missing Data in Multi-Environmental Trials

The analysis of multi-environmental trial (MET) data often suffers from missing data. A genomic prediction method robust to missing data is required to get accurate predictions based on MET data. Moreover, a method robust to missing data may allow us to reduce cost through thinning out lines tested in MET. Multi-task learning methods that allow simultaneous prediction for multiple traits and multiple environments are expected to be robust to missing data because they can leverage from genetic correlation among environments and traits. The robustness of multi-task learning methods, however, has not been assessed in detail. In this study, we compared the prediction accuracy of multi-task learning methods for MET data with a large missing rate. We simulated missing data under two scenarios and different missing rates. Performance of eight methods, which include multi-task learning methods as well as single-trait and single-environment methods, were assessed and compared with each other. Under the ‘uniform’ scenario, in which missing data arises randomly, three methods that model genotype by environment interaction explicitly were robust to missing data at high missing rates. The accuracy of other methods including ones that model among-trait correlation decreased significantly when missing rate increased. Under the ‘genotype’ scenario, in which missing data arises in all traits in an environment for randomly selected genotypes, the accuracy of all the method decreased significantly when the missing rate increased. The results suggest that multi-task learning models are robust to missing data when phenotypic data are available for sufficient number of genotypes.