Integrating spatial modeling and genomic prediction to support early breeding of Acrocomia aculeata

Acrocomia aculeata is a promising neotropical oil palm with great potential for food, bioenergy, and biobased industries. However, breeding programs are constrained by the species' long juvenile period, the difficulty of evaluating natural populations under heterogeneous environmental conditions, and the limited genomic resources. Integrating genomic prediction with spatial information represents a promising strategy for improving the identification of superior individuals while accounting for environmental variation. This study evaluated the performance of genomic prediction models incorporating spatial effects for three fruit traits: pulp dry mass (PDM), kernel dry mass (KDM), and pulp oil content (OC). A total of 167 naturally occurring individuals distributed across four provenances were phenotyped and genotyped. Four SNP datasets were evaluated, including three derived from distinct SNP-calling strategies (de novo assembly, the Elaeis guineensis reference genome, and the A. aculeata transcriptome) and a combined dataset containing SNPs from all three individual approaches. Genomic prediction models were fitted considering additive and additive-plus-dominance genetic effects together with a spatial kernel to account for environmental heterogeneity. Predictive performance was evaluated using leave-one-out and provenance-based cross-validation schemes. Population genetic analyses revealed high within-population genetic diversity and negligible genetic differentiation among provenances (F_ST = 0.001–0.003), indicating strong genetic connectivity across the study area. The estimated variance components demonstrated substantial differences among SNP datasets and traits. For PDM, genomic additive variance ranged from 15.07 to 36.41, whereas dominance variance ranged from 0.00 to 7.50. For KDM, additive variance ranged from 0.17 to 0.51, with dominance variance remaining low (0.01–0.09). For OC, additive variance ranged from 18.71 to 42.85, whereas dominance variance ranged from 9.35 to 29.27. In contrast, the spatial kernel explained a considerable proportion of the phenotypic variation across all traits, particularly for OC, where it accounted for up to 69.32% of the explained variance, highlighting the importance of modeling spatial heterogeneity when evaluating natural populations. Across validation schemes, predictive abilities reached 0.35 for PDM, 0.43 for KDM, and 0.50 for OC, with the oil palm genome and combined SNP datasets showing the most consistent performance. Spatial interpolation of predicted genetic effects enabled the identification of superior genotypes distributed throughout the landscape rather than concentrated within specific sampling areas. These findings demonstrate that integrating spatial modeling with genomic prediction provides an effective framework for supporting early-stage breeding, germplasm characterization, and the identification of elite individuals in A. aculeata.