Evolutionary structure constrains genomic prediction accuracy more than model complexity in mango (Mangifera indica L.)

In genomic prediction, it remains unclear whether increasingly complex or ensemble models improve prediction over established linear approaches, and why prediction accuracy varies among traits. Here, we evaluated a comprehensive suite of genomic prediction models, including linear mixed models, Bayesian variable selection, kernel methods, machine learning algorithms, graph attention networks, and stacked ensembles, in mango (Mangifera indica L.). Across 5 traits, prediction accuracy converged across linear, Bayesian, kernel, and ensemble models, with only marginal gains derived from stacking and no systematic advantage of machine learning approaches. Ensemble ablation and weight analyses revealed that predictive signal was dominated by additive and smooth kernel components, while more complex learners contributed little or negatively upon performance. To explain these trait-dependent patterns in predictability, we quantified the phylogenetic signal using genome-wide marker-based trees. All traits showed a significant phylogenetic signal, with the magnitude varying widely and strongly associated with prediction accuracy (r ≈ 0.71). Traits with strong phylogenetic structure achieved the highest prediction accuracies, whereas traits with a weaker signal were consistently harder to predict, regardless of model choice. Together, these results confirm that, in mango, genomic prediction accuracy is determined more by evolutionary structure and trait architecture rather than increasing model complexity. Aligning prediction strategies with the evolutionary basis of trait variation may therefore be more effective than adopting increasingly complex models.