Multigenerational machine learning-based genomic prediction for dermo resistance in eastern oyster Crassostrea virginica

Abstract

Dermo disease caused by the protist Perkinsus marinus poses a major threat to Eastern oyster aquaculture. We previously conducted genomic selection for dermo resistance and found increased effectiveness over phenotypic selection. Here, we report improved genomic predictions using combined data from three successive generations. We evaluated nine different genomic selection models, including three supervised machine learning architectures: gradient boosting, logistic regression, and random forest. Combining data across multiple generations did not, by itself, substantially improve the accuracy of most models, but the increased sample size provided genotyping confidence of loci with rare alleles. Correlation accuracy of all models significantly increased with the inclusion of low-frequency variants and strong-effect markers identified through a genome-wide association study. Gradient boosting machine learning models outperformed other genomic selection models across all training scenarios, suggesting enhanced capacity to learn generalizable genomic signals associated with dermo resistance. The best gradient boosting model achieved a peak correlation accuracy of 0.410, a substantial improvement over the previous peak accuracy of 0.274 from traditional models. Together, our results highlight the potential of machine learning for genomic selection and the significance of rare variants in determining dermo resistance.