Machine Learning and Statistical Models for Genomic Selection of Phytophthora Fruit Rot Resistance in Watermelon

Breeding for Phytophthora fruit rot (PFR) resistance caused by the devastating pathogen Phytophthora capsici remains challenging in watermelon mainly due to phenotyping constraints. To overcome these bottlenecks, we developed and validated genomic selection models using an interspecific recombinant inbred line (RIL, F₁₁) population derived from USVL531-MDR (resistant, Citrullus mucosospermus) × USVL677-PMS (susceptible, Citrullus lanatus) and a segregating F₂ population (USVL531-MDR × 'Calhoun Grey'). Mature fruits from both populations were artificially inoculated under controlled conditions and phenotyped for lesion diameter, visible pathogen growth diameter, and sporulation intensity. Fourteen genomic selection models comprising ten parametric approaches (Bayesian and regression-based), one semi-parametric model (RKHS), and three machine learning algorithms (Random Forest, SVM, XGBoost) were evaluated for their performance. Parametric models achieved high predictive accuracy in the RIL population (r = 0.700-0.718 in 10-fold cross-validation; r = 0.694-0.713 in 5-fold cross-validation), while the F₂ population showed moderate accuracies, with GBLUP performing well under 10-fold cross-validation (r = 0.476) and Bayesian ridge regression under 5-fold cross-validation (r = 0.433). Coincidence analysis demonstrated that genomic prediction effectively captured top-performing RILs, with 10% and 20% selection intensities showing strong overlap with phenotypic selection and Bayesian models achieving up to 85% coincidence. These results establish a solid framework for application of genomic selection in watermelon PFR resistance breeding programs.