Comparison of Statistical Methods and Reliability of Genomic Prediction in Coffea canephora Populations

Simulation and empirical results have shown that genomic predictions present sufficient accuracy to help increase success in breeding programs. Although many crops have benefited from this novel approach, studies in the Coffea genus are still in their infancy. Until now, there have been no studies of how predictive models work across populations and environments or, even, their performance for different complex traits. Considering that predictive models are based on biological and statistical assumptions, it is expected that their performance varies depending on the true underlying genetic architecture of the phenotype. We used real data from two experimental populations of Coffea canephora, evaluated in two environments (sites) and SNPs identified by Genotyping-by-Sequencing (GBS) to investigate the genotype-phenotype relationship. We considered Bayesian models, with different prior distributions for the marker effects, and regularized linear regression models. We assessed predictive abilities using a Replicated Training-Testing evaluation, with 30 repetitions, and different metrics to compare the model performances. In addition, we investigated SNP effects to learn about underlying biology related to genomic regions affecting the phenotype and their interactions. For the three traits evaluated, there were minimal differences in predictive accuracy among models. A slight advantage of Bayesian methods was observed, although more computation was required. Predictions within-population, on average, were more accurate than between populations. Biological insights revealed genetic variants with specific signals within populations and environments. Consequently, these results have great potential to reshape traditional breeding programs, including genomic predictions for improved breeding strategies.