Feature-Selected Gradient Boosting Outperforms Parametric Baselines for Genomic Prediction of Mastitis Susceptibility

Author

• Gavin Lynch, California Polytechnic State University, San Luis ObispoFollow

Available at: https://digitalcommons.calpoly.edu/theses/3425

Date of Award

8-2026

Degree Name

MS in Computer Science

Department/Program

Computer Science

College

College of Engineering

Advisor

Sumona Mukhopadhyay

Advisor Department

Computer Science

Advisor College

College of Engineering

Abstract

Predicting complex disease traits, such as mastitis, from high-dimensional genomic data is a significant challenge in modern dairy cattle breeding. Traditional genomic selection models, including mixed linear models and Bayesian approaches, are industry standards but may not fully capture complex, non-linear interactions within the genome. This study investigates the potential of machine learning models, particularly Transformer and tree-based architectures, to improve prediction accuracy. Using a covariate-residualized feature-selection pipeline, we compare Transformer, Random Forest, and LightGBM models against GBLUP and BayesR in 1,860 first-lactation cows (93 clinical mastitis cases) genotyped at 624,300 SNPs. Results indicate that LightGBM outperforms traditional genomic prediction methods on ranking metrics, achieving the highest Average Precision (0.130) against GBLUP (0.097) and BayesR (0.088), though both parametric baselines sit near a covariates-only floor (0.085) that uses breed and herd-year alone, indicating that little of their accuracy is marker-derived.

Two extensions build on the same model comparison pipeline. Incorporating Ensembl Variant Effect Predictor (VEP) functional-consequence annotations, as weights in feature selection and within in a Bayesian model (BayesRC), decreased (-0.03 AP for LightGBM at 1000 SNPs) rather than improved performance. The second extension replaced the binary phenotype with the first-lactation somatic cell score (996 cows), a continuous, more densely recorded udder-health trait. This likewise failed to increase predictive performance; every genomic model fell short of a covariates-only baseline. However, these findings identify feature-selected gradient boosting as the most performant of the architectures tested for genomic prediction of complex disease traits in dairy cattle, offering a promising avenue for future research.