Stacked kinship CNN vs. GBLUP for genomic predictions of additive and complex continuous phenotypes

Abstract

Abstract Deep learning is impacting many fields of data science with often spectacular results. However, its application to whole-genomepredictions in plant and animal science or in human biology has been rather limited, with mostly underwhelming results.While most works focus on exploring alternative network architectures, in this study we propose an innovative representation ofmarker genotype data. Different types of genomic kinship matrices are stacked together in a 3D pile from where 2D grayscaleslices are extracted and fed to a deep convolutional neural network (DNN). We tested nine scenarios of simulated phenotypeswith combinations of additivity, dominance and epistasis, and compared the DNN to GBLUP-A (Genomic BLUP computedusing only the additive kinship matrix) and GBLUP-optim (Genomic BLUP with additive, dominance, and epistasis kinshipmatrices, as needed).Results varied depending on the accuracy metric employed, with DNN performing better in terms of Root Mean Squared Error(1%-12% lower than GBLUP-A; 1%-9% lower than GBLUP-optim) but worse in terms of Pearson’s correlation (0.505 for DNNcompared to 0.672 and 0.669 of GBLUP-A and GBLUP-optim for fully additive case; 0.274 for DNN, 0.279 for GBLUP-A, and0.477 for GBLUP-optim for fully dominant case).