Evaluation of Machine Learning Models for Predicting Breeding Values of Key Performance Traits in Iranian Crossbred (Thoroughbred × Turkmen) Racehorses.

Overview

• This research evaluated various machine learning algorithms for predicting breeding values of key performance traits in Iranian crossbred racehorses, comparing their accuracy against traditional genetic evaluation methods.
• The study found that machine learning, especially artificial neural networks and boosting methods, provided more accurate and reliable predictions than the conventional single-trait mixed models.

Background and Objective

• Accurate prediction of breeding values (BVs) is critical for genetic improvement in racehorses because it helps in selecting superior animals for breeding, ultimately enhancing desired traits such as speed and race performance.
• The traditional approach for genetic evaluation is the single-trait mixed model (STM), commonly implemented using BLUP (Best Linear Unbiased Prediction), which relies on pedigree and phenotype data.
• Machine learning (ML) methods offer a flexible alternative that can model complex relationships in data, including phenotypic, environmental, and pedigree-derived variables, potentially improving predictions.
• This study aimed to compare 13 different ML algorithms’ abilities to predict estimated breeding values (EBVs) for three key performance traits in Iranian Thoroughbred × Turkmen crossbred racehorses:
• Race Completion Time (RCT)
• Rank at the End of Competition (REC)
• Average Speed (AS)

Methods

• Data consisted of 55,070 race records from 5,296 individual horses, providing a large, diverse dataset for analysis.
• Reference EBVs were calculated using STM fitted by maximum likelihood via the Wombat software, which served as a baseline for comparison.
• Thirteen ML algorithms evaluated included:
• Linear regression models
• Tree-based ensemble methods (e.g., Random Forest)
• Boosting algorithms (e.g., XGBoost, LightGBM, Gradient Boosting)
• Artificial Neural Networks (ANNs)
• Distance-based and other methods
• The dataset was split into training (75%) and testing (25%) subsets to build and assess model performance objectively.
• Performance metrics used were:
• Accuracy of prediction
• Correlation between predicted and reference EBVs
• Mean Absolute Error (MAE)
• Mean Squared Error (MSE)
• Bias (difference between predicted and actual mean EBVs)
• Dispersion (consistency of prediction variance)

Results

• All machine learning models outperformed the conventional single-trait mixed model (STM) across the three performance traits.
• The Artificial Neural Network (ANN) consistently achieved the highest predictive accuracy and correlation for all traits:
• Accuracy: 0.63 (RCT), 0.72 (REC), 0.56 (AS)
• Correlation with reference EBVs: 75% (RCT), 86% (REC), 71% (AS)
• Other strong ML performers included:
• XGBoost
• LightGBM
• Gradient Boosting
• Random Forest
• Linear models and distance-based ML methods showed comparatively lower predictive performance, indicating they may not capture complex trait relationships well.
• Prediction bias across ML models remained near zero, confirming that predictions were neither systematically over- nor under-estimated.
• Dispersion coefficients were close to one for most ML models, indicating consistent variability in their predictions relative to the reference.

Conclusions and Implications

• Machine learning methods, and especially artificial neural networks and boosting-based algorithms, can substantially improve the accuracy and reliability of EBV predictions compared to traditional genetic evaluation techniques.
• The improved predictive capacity of ML models enables more informed and trustworthy selection decisions, which can accelerate genetic progress in racehorse breeding programs.
• Machine learning should be considered a complementary tool alongside traditional methods for routine genetic evaluations in animal breeding.
• Future work may include integrating environmental and genomic data to further enhance prediction models and testing these approaches in other breeds or species.