Critical evaluation of the theory and practice of feed-forward neural networks for genomic prediction

Genomic prediction (GP) has catalyzed increased rates of genetic gain in animal and plant breeding. Recently, deep learning (DL) has been explored to increase GP accuracy by incorporating diverse data types and learning complex, non-linear patterns in datasets. However, DL consistently fails to significantly improve prediction accuracy over gold standard genomic BLUP (gBLUP) models. In this study, we first review the theory behind neural networks and reproducing kernel Hilbert spaces (RKHS) regression to contextualize 3 claimed benefits of DL over linear models: incorporation of diverse data types, avoidance of feature engineering, and universal approximation behavior. We also propose a taxonomy of prediction problems so that model comparisons do not confound differences in the predictive skill of different model classes with differences in the input data. Second, we leverage a maize multi-environment trial dataset to train DL and RKHS models that implicitly capture non-linear patterns between genomic, soil, weather, and management inputs and grain yield. The results demonstrate that feature engineering using principal components of SNPs generally degrades prediction accuracy across model classes. Furthermore, DL models persistently fail to outperform RKHS models across prediction problems. Finally, we evaluate the theoretical critiques with the empirical results, confirming the theoretical arguments. Nevertheless, a small portion of the possible DL model space has been explored, leaving open the possibility of DL making significant contributions to GP problems through additional aspects not considered here. We conclude by suggesting several avenues for further theoretical and practical research, including the resolution of several disciplinary differences.