OBJECTIVE: Genomic best linear unbiased prediction (GBLUP) is a key method in genomic prediction, relying on the construction of a genomic relationship matrix (G-matrix). Although various methods for G-matrix construction have been proposed, the performance of these methods across different species has not been thoroughly compared. METHODS: This study systematically evaluated the performance of six genomic relationship matrix (G-matrix) construction methods in improving the prediction accuracy of GBLUP models across four species: pigs, bulls, wheat, and mice. The methodological framework included: (1) an initial unscaled matrix
(2) five scaled methods utilizing allele frequency centralization. The scaled methods comprised: (a) three variance-weighted approaches using allele frequencies fixed at 0.5 (G05), observed frequencies (GOF), or average minor allele frequencies (GMF)
(b) two centralized methods with weighting by either the trace of the numerator matrix (GN) or reciprocals of each locus's expected variance (GD). RESULTS: The GD matrix demonstrated significant prediction accuracy improvements for pig traits. Conversely, most scaled G-matrices showed minimal effects on mice, wheat, and bull, even with underperforming unscaled baselines in prediction accuracy compared to the original unscaled matrix. The learning curve for bull data showed the choice of G-matrix had minimal impact on prediction accuracy when the reference population size and genetic marker density reached a certain threshold. DISCUSSION: The study concluded that the optimal G-matrix construction method varies across species, with population structure being a key factor. These findings highlight the importance of species-specific optimization in genomic prediction and suggest that the influence of G-matrix construction diminishes in large-scale, high-density genomic datasets.