TKOA-GPNN: a model framework for genotype-to-phenotype prediction in Duroc pigs

Genomic prediction has been established as a powerful tool in pig breeding, which typically entails the analysis of large-scale genomic datasets. Neural network models, renowned for their robust pattern recognition capabilities, have been successfully applied to extract complex genetic patterns from such high-dimensional genomic data.

Because growth traits are critical economic indicators in pig breeding, their accurate prediction is essential for genetic improvement. However, the underlying genetic architecture of these traits is highly complex, regulated by numerous genetic loci. Given that not all loci exert significant effects on growth traits and traditional neural networks face challenges in capturing genome-wide genetic interactions, there is an urgent need for a more efficient and accurate prediction framework.

To address this challenge, we propose a novel model, TKOA-GPNN, which comprises two components: a two-stage genomic feature selection module and a genomic prediction neural network. First, symmetric uncertainty correlation analysis is employed to reduce the feature space. Subsequently, a kepler optimization algorithm integrated with tent mapping is applied to select trait-associated loci. The selected features are then fed into the neural network (GPNN) to conduct genomic prediction.

Experimental results demonstrate that TKOA-GPNN outperforms 14 existing models with superior accuracy and efficiency. Tests based on offspring data shows favorable performance with genotyping array data, while further validation using whole-genome sequencing data confirmed the robustness of the proposed model. These results suggest that TKOA-GPNN is applicable to genomic data of different types and quality levels.

Overall, TKOA-GPNN improves both prediction accuracy and computational efficiency, providing an effective solution for intelligent breeding and genetic improvement in pigs breeding.