Evaluation of machine learning models for predicting breeding values of key performance traits in Iranian crossbred (Thoroughbred × Turkmen) racehorses
Related Experiment Video
Updated: Oct 19, 2026
Induction and Evaluation of Inbreeding Crosses Using the Ant, Vollenhovia Emeryi
Published on: October 5, 2018
M Taned1, M B Zandi1, M P Eskandari Nasab2
1Department of Animal Science, Faculty of Agriculture, University of Zanjan, Zanjan, Iran; Equine Research Group, Faculty of Agriculture, University of Zanjan, Zanjan, Iran.
Background:
Accurate prediction of breeding values (BVs) is fundamental to genetic improvement in racehorses. Although single-trait mixed models (STM), typically implemented using BLUP-based procedures, remain the standard framework for genetic evaluation, machine learning (ML) methods offer a flexible, data-driven approach for learning complex relationships among phenotypic, environmental, and pedigree-derived information. These capabilities have encouraged researchers to explore the application of ML methods for predicting breeding values.
Objectives:
This study compared the performance of 13 ML algorithms for predicting estimated breeding values (EBVs) of Race Completion Time (RCT), Rank at the End of Competition (REC), and Average Speed (AS) in Iranian crossbred racehorses.
Methods:
A total of 55,070 race records from 5,296 horses were analyzed. Reference EBVs were estimated using single-trait mixed models (STM) fitted by maximum likelihood in Wombat. Thirteen ML algorithms, including linear regression models, tree-based ensembles, boosting methods, and artificial neural networks (ANNs), were trained using 75% of the data and evaluated on the remaining 25%. Predictive performance was assessed using accuracy, correlation, mean absolute error (MAE), mean squared error (MSE), bias, and dispersion.
Results:
All ML models outperformed the conventional STM across the three traits. ANN consistently achieved the highest predictive performance, with accuracies of 0.63, 0.72, and 0.56 and correlations of 75%, 86%, and 71% for RCT, REC, and AS, respectively. XGBoost, LightGBM, Gradient Boosting and Random Forest also showed strong predictive ability, whereas linear models and distance-based methods exhibited comparatively lower performance. Prediction bias remained close to zero, and dispersion coefficients were generally close to unity across ML models.
Conclusions:
Machine learning, particularly ANN and boosting-based algorithms, substantially improved EBV prediction compared with the conventional STM. These findings demonstrate the potential of ML as a complementary tool for accelerating routine genetic evaluation and supporting more reliable selection decisions in horse breeding programs.