Using Deep Learning Models as a Genetic Architecture for the Simulation of Breeding Schemes

Area of Science:

• Quantitative genetics
• Bioinformatics
• Machine learning in animal breeding

Background:

• Classical quantitative genetic simulation models often fail to replicate real-life observations due to rapid depletion of genetic variance.
• Current models struggle to capture complex biological interactions underlying trait development, limiting their predictive accuracy.
• Deep learning (DL) shows potential in modeling intricate genetic interactions, suggesting improved simulation capabilities.

Purpose of the Study:

• Introduce and evaluate novel DL-based genetic simulation models.
• Compare the performance of DL models against classical models in retaining additive genetic variance.
• Assess the impact of DL model complexity on genetic variance preservation under selection.

Main Methods:

• Simulated a full-sib pig breeding scheme using real haplotypes.
• Applied directional truncation selection over 20 generations.
• Compared genetic variance retention across classical models (A, ADAA, ADAAADDD) and DL models (DL_simple, DL_medium, DL_complex).

Main Results:

• Classical models retained 55-64% of initial additive genetic variance.
• DL_simple model lost all additive variance.
• DL_medium retained 92-98%, and DL_complex increased additive variance by 296-314%.

Conclusions:

• DL-based models can significantly improve the retention of additive genetic variance in simulations.
• Model architectural complexity is key; more complex DL models better capture epistatic interactions.
• DL models offer a promising approach for more realistic long-term genetic simulations, highlighting the role of non-additive effects.