Deep neural networks for genomic prediction do not estimate marker effects

Area of Science:

• Genetics
• Bioinformatics
• Plant Breeding

Background:

• Genomic prediction is crucial for plant and animal breeding.
• Nonlinear models like deep neural networks (DNNs) are hypothesized to capture complex epistatic effects.
• However, DNNs often do not surpass classical linear methods in genomic prediction accuracy.

Purpose of the Study:

• To investigate why powerful nonlinear models, such as DNNs, do not consistently outperform linear models in genomic prediction.
• To propose and test the theory that DNNs utilize shortcut learning, focusing on overall genetic relatedness.

Main Methods:

• Utilized datasets from crop plants: lentil, wheat, and Brassica carinata.
• Evaluated DNN performance by providing only marker match locations between individuals.
• Compared prediction accuracy achieved through this simplified input against standard genomic prediction approaches.

Main Results:

• Demonstrated that DNNs achieve similar prediction accuracy when provided only with marker match locations.
• This indicates the network's predictions are based on overall genetic similarity, not specific marker interactions like epistasis.
• The network showed indifference to the actual marker values, supporting the shortcut learning hypothesis.

Conclusions:

• Deep neural networks may default to shortcut learning in genomic prediction, leveraging overall genetic relatedness.
• This reliance on general relatedness, rather than specific marker effects (including epistasis), explains their failure to outperform linear models.
• Future research should focus on mitigating shortcut learning in DNNs for improved genomic prediction.