Multi-task Gaussian process for imputing missing data in multi-trait and multi-environment trials

Abstract

A method based on a multi-task Gaussian process using self-measuring similarity gave increased accuracy for imputing missing phenotypic data in multi-trait and multi-environment trials. Multi-environmental trial (MET) data often encounter the problem of missing data. Accurate imputation of missing data makes subsequent analysis more effective and the results easier to understand. Moreover, accurate imputation may help to reduce the cost of phenotyping for thinned-out lines tested in METs. METs are generally performed for multiple traits that are correlated to each other. Correlation among traits can be useful information for imputation, but single-trait-based methods cannot utilize information shared by traits that are correlated. In this paper, we propose imputation methods based on a multi-task Gaussian process (MTGP) using self-measuring similarity kernels reflecting relationships among traits, genotypes, and environments. This framework allows us to use genetic correlation among multi-trait multi-environment data and also to combine MET data and marker genotype data. We compared the accuracy of three MTGP methods and iterative regularized PCA using rice MET data. Two scenarios for the generation of missing data at various missing rates were considered. The MTGP performed a better imputation accuracy than regularized PCA, especially at high missing rates. Under the 'uniform' scenario, in which missing data arise randomly, inclusion of marker genotype data in the imputation increased the imputation accuracy at high missing rates. Under the 'fiber' scenario, in which missing data arise in all traits for some combinations between genotypes and environments, the inclusion of marker genotype data decreased the imputation accuracy for most traits while increasing the accuracy in a few traits remarkably. The proposed methods will be useful for solving the missing data problem in MET data.

Comparison of F-tests for Univariate and Multivariate Mixed-Effect Models in Genome-Wide Association Mapping

“…Phenotype imputation is essential for several estimation algorithms for multivariate mixed-effect models because of transformation using eigen vectors, which is required for high computational efficiency (e.g., Zhou and Stephens, 2014; Lee and van der Werf, 2016). Phenotype imputation using multivariate information has been proposed for GWA (Dahl et al, 2016) and plant breeding (Hori et al, 2016). Although we evaluated the multivariate F -test without imputation in the present study, the performance of the multivariate and univariate F -tests based on imputed data would also depend on the magnitude of correlation between variates ( r ) as observed here because it will influence the imputation accuracy regardless of the imputation methods.…”

Identifying Sepsis Subphenotypes via Time-Aware Multi-Modal Auto-Encoder

“…3DMICE combines MICE [3] and Gaussian Process [7] to impute missing values, which integrates cross-variable and longitudinal information.…”

Mixture-based Multiple Imputation Model for Clinical Data with a Temporal Dimension

“…Due to the fact that many patients only have records with a limited number of time points, time series of inpatient clinical laboratory tests fall short of such a requirement. Hori et al [40] extend MTGP to imputation on multi-environmental trial data in a third-order tensor. However, this approach is not applicable to clinical data with a large number of patients, since the covariance matrix of observed values needs to be computed together with its inverse, which is intractable for our datasets.…”