Mixed Model Approach for Genotypic Imputation

In the post-genomic era, new genotyping platforms such as Genotyping-by-Sequencing and SNP-chips provide large and inexpensive whole genome sequencing. The implementation of such technology in breeding programs is necessary for maximizing genetic gain through genomic selection. Accurate imputation of missing values and correction of genotyping errors become essential for filling up genotyping gaps, combining different arrays and robust genome-wide analyses such as association mapping and genomic prediction. Due to its versatility, the functionality of multivariate mixed models can be extended to imputations. Setting the marker of interest (y_j) and its flanking markers (y_i and y_k) as response variable (Y=y_i,y_j,y_k) and the genetic effect as random effect, the multivariate mixed model for imputations is expressed as Y_nx3=X_nx3μ_1x3+Z_nxnγ_1x3+ε_nx3, with var(γ)=Aσ²_γ3x3 and var(ε)=Iσ²_ε3x3. The mixed model was solved using EM-REML, implemented in REMLF90. The model is looped across the genome and fitted values of y_j are alocted into the new genotypic matrix. Genotypes and phenotypes used in this study are from the SoyNAM project (http://soynam.org), a soybean nested association panel with 40 families comprising 5596 recombinant inbred lines genotyped with a 5k SNP-chip. A subset of 4246 marker containing 1% of missing data was selected for this study. Two datasets with 10 and 20% missing data were generated from the original data by randomly assigning missing point. The objective of this study is to introduce the mixed model approach of imputation and compare its performance with common imputation methods. Imputation with mixed model has shown an increase in yield heritability.