Genomic selection is a powerful tool for enhancing genetic gain in wheat breeding. However, prediction accuracy (PA), which is directly proportional to genetic gain in genomic-based breeding programs, remains low for complex traits such as wheat grain yield (GY). This study aimed to increase PA for GY in different selection environments (SEs) by leveraging CIMMYT's historical dataset.

Ten years of GY data observed in six different SEs for elite breeding lines were used. The Genomic Best Linear Unbiased Prediction (GBLUP) model assessed PA using five years as the test population (from 2018-2019 to 2022-2023) and all available years back up to 2013-2014 as the training population. Generally, we observed that as the number of training years increased, PA tended to improve or stabilize.

For instance, by testing the year 2022-2023, the SE BLHT (late heat stress) showed a notable increase in PA, reaching 0.23 when using five years back for training compared to 0.11 with only one year back, then PA was found stable by training from five to nine years back (from 0.23 to 0.26). Similarly, in B2IR (intermediate drought) SE exhibited an improvement from 0.12 to 0.21 using one and four years back, subsequently no significative improvements were observed by training from four to nine years back (from 0.21 to 0.24).

Conversely, some traits such as F5IR (flat optimal environment) did not show a significant increase, with PA fluctuating around 0.09-0.14 regardless of the training years used. This indicates that while increasing the number of training years generally enhances PA for GY, the degree of improvement can vary significantly across different SEs.

Our findings suggest that leveraging an extended historical dataset generally enhances genomic prediction accuracy of complex traits like wheat grain yield. These insights provide valuable guidance for optimizing genomic selection strategies in wheat breeding, ultimately contributing to the development of high-yielding, resilient wheat varieties.