Optimizing genomic prediction with transfer learning under a ridge regression framework

Genomic selection (GS) is a predictive plant and animal methodology that allows the selection of plants and animals based on predictions without the need to measure the phenotype. However, its practical application requires challenging prediction accuracy due to the noise observations collected in experiments in these areas. Many strategies and approaches have been proposed to improve the prediction accuracy of this methodology. This paper explores the use of transfer learning in the context of GS. Transfer learning with (1) ridge regression (RR) (Transfer RR) and (2) analytic RR (ARR) (Transfer ARR) were applied from cultivars in the proxy environment to predict those cultivars in the goal environments. Also, we compared the performance of models RR and ARR without transfer learning. We used 11 real multi-environment datasets (wheat and rice) and evaluated them in terms of Pearson's correlation (Cor) and normalized root mean square error (NRMSE). Our study shows empirical evidence that the Transfer RR or Transfer ARR approaches significantly enhanced predictive performance. Across the datasets, Transfer RR (or Transfer ARR) method improved Cor by 22.962% and NRMSE by 5.757%, in comparison to models RR and ARR. These results underscore the potential of Transfer RR (or Transfer ARR) when enhancing predictive accuracy in this context.