An interpretable integrated machine learning framework for genomic selection

Although machine learning (ML) methods have shown growing promise for genomic selection (GS), several key challenges hinder their widespread application. In this study, we conducted a comprehensive analysis comparing the performance of various ML models, along with investigations into parameter optimization, dimensionality reduction, feature selection, and the “black box” problem. We also proposed an efficient and interpretable framework, NTLS (NuSVR + TPE + LightGBM + SHAP). In the prediction of Yorkshire pig populations, NTLS outperformed the genomic best linear unbiased prediction (GBLUP) model, achieving improvements in predictive accuracy of 5.1%, 3.4%, and 1.3% for days to 100 kg (DAYS), back fat at 100 kg (BF), and number of piglets born alive (NBA), respectively. Moreover, we introduced the NuSVR model, which achieved the highest accuracy among nine compared algorithms. Our findings further highlight the importance of interpretable learning in GS and provide a detailed multi-level application of the SHAP algorithm.