Deep neural networks for genomic prediction do not estimate marker effects. Issue 3 (1st October 2021)

• Record Type: Journal Article
• Authors: Ubbens, Jordan; Parkin, Isobel; Eynck, Christina; Stavness, Ian; Sharpe, Andrew G.
• Is Part Of: œplant genome. Volume 14:Issue 3(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-10-01
• Subjects: Plant genomes -- Periodicals; Plant genome mapping -- Periodicals; 572.862
• DOI: 10.1002/tpg2.20147
• Languages: English
• ISSNs: 1940-3372

Abstract

Abstract: Genomic prediction is a promising technology for advancing both plant and animal breeding, with many different prediction models evaluated in the literature. It has been suggested that the ability of powerful nonlinear models, such as deep neural networks, to capture complex epistatic effects between markers offers advantages for genomic prediction. However, these methods tend not to outperform classical linear methods, leaving it an open question why this capacity to model nonlinear effects does not seem to result in better predictive capability. In this work, we propose the theory that, because of a previously described principle called shortcut learning, deep neural networks tend to base their predictions on overall genetic relatedness rather than on the effects of particular markers such as epistatic effects. Using several datasets of crop plants [lentil ( Lens culinaris Medik.), wheat ( Triticum aestivum L.), and Brassica carinata A. Braun], we demonstrate the network's indifference to the values of the markers by showing that the same network, provided with only the locations of matches between markers for two individuals, is able to perform prediction to the same level of accuracy. Core Ideas: The capacity of deep neural networks does not match performance in genomic prediction. Deep neural networks are not disadvantaged when they cannot access values of the markers. Deep neural networks likely attend primarily to genetic relatedness, not marker effects.