Liyanage, D.S., Lee, Sukkyoung, Yang, Hyerim, Lim, Chaehyeon, Omeka, W. K.M., Sandamalika, W.M.Gayashani, Udayantha, H.M.V., Kim, Gaeun, Hanchapola, H.A.C.R., Ganeshalingam, Subothini, Jeong, Taehyug, Oh, Seong Rip, Won, Seung Hwan, Koh, Hyoung Bum, Kim, Mun Kwan, Jones, David B., Massault, Cecile, Jerry, Dean R., and Lee, Jehee (2025) Genomic prediction of survival traits in the response of olive flounder (Paralichthys olivaceus) to viral hemorrhagic septicemia virus: Comparing machine learning models and traditional approaches. Aquaculture, 595 (Part 2). 741685.

PDF (Pubished Version) - Published Version Restricted to Repository staff only

3

Abstract

Genomic prediction utilizes relationships between phenotypes and thousands of genetic markers dispersed across the genome of a species under selection to estimate an individual's breeding value. Viral hemorrhagic septicemia (VHS) is a devastating disease that causes high mortality in the olive flounder (Paralichthys olivaceus). Selection of VHS resistant individuals using traditional pedigree-based approaches is slow and ineffective; therefore, we investigated the potential of genomic selection to identify superior VHS-resistant individuals. In this study, various statistical models and algorithms, including the multilayer perceptron (MLP) and convolutional neural network (CNN), were compared for their prediction accuracy of breeding values. Other models assessed include pedigree-based best linear unbiased prediction (PBLUP), genomic best linear unbiased prediction (GBLUP), Bayesian A (BA), Bayesian B (BB), Bayesian C (BC), Bayesian Lasso (BL), Bayesian ridge regression (BRR), elastic net (EN), ridge regression (RR), and random forest (RF). These models were assessed for their ability to predict VHS resistance based on genomic data obtained from high-quality 70 K single nucleotide polymorphism (SNP) Affymetrix® Axiom® myDesign™ Genotyping Array from 865 animals. Furthermore, we investigated the population structure of the selected flounder population using a genomic relatedness matrix, PCA analysis, kinship coefficients, and multidimensional scaling. The results revealed that RF had the highest prediction accuracy for the three VHS-resistance traits (binary survival, days to death, and time of death), followed by BRR and GBLUP. PBLUP exhibits the lowest accuracy for these traits. Machine learning (ML) models, such as RF, outperformed traditional PBLUP and GBLUP, showing the greatest improvement over other Bayesian methods (BA, BB, and BC). The optimal parameters were determined for the best models, including specific marker sizes and sample size recommendations. The selected models and their parameters significantly improved the prediction of VHS resistance, demonstrating the potential of genomic selection to outperform the traditional pedigree-based methods. Our findings indicate that genomic selection can be more effective using ML models than conventional approaches in predicting VHS resistance and offers a means to enhance the genetic resistance of olive flounder in aquaculture to this disease.

Actions (Repository Staff Only)

Item Control Page