Methods and compositions for imputing or predicting genotype or phenotype

A universal method of parametrically representing genotypic or phenotypic association data from a training data set obtained from a plant population or a plant sample set to impute or predict a genotype and/or a phenotype in a non-training data obtained from a non-training plant population or a non-training plant sample data, the method comprising:

• generating a universal continuous global latent space representation by encoding discrete or continuous variables derived from a genotypic or phenotypic association training data into latent vectors through a machine learning-based global autoencoder framework, wherein the global latent space is independent of the underlying genotypic or phenotypic association;
• generating a local latent representation by encoding a subset of the discrete or continuous variables derived from the genotypic or phenotypic association training data set into latent vectors through a machine learning-based local autoencoder framework, wherein the local latent space is generated with inputs from the local autoencoder and the global autoencoder;
• decoding the global latent representation and the local latent representation by a local decoder, thereby imputing or predicting the genotype or phenotype of the non-training data by the combination of the decoded global latent representation and the local latent representation;
• and
• selecting one or more plant populations or members thereof based on the imputed or predicted genotype or phenotype.