A zero altered Poisson random forest model for genomic-enabled prediction.

Area of Science:

• Genomics
• Statistical Genetics
• Machine Learning

Background:

• Genomic selection (GS) relies heavily on accurate statistical models for predicting breeding values.
• Count response variables in GS often exhibit excess zeros, posing challenges for conventional models.
• Machine learning approaches offer potential solutions for complex genomic data structures.

Purpose of the Study:

• To introduce and evaluate novel zero-inflated random forest models (ZAP_RF and ZAPC_RF) for genomic selection.
• To assess the performance of these models in handling count data with excess zeros.
• To compare the predictive accuracy of the proposed models against traditional methods.

Main Methods:

• Application of two zero-inflated random forest models: ZAP_RF and ZAPC_RF.
• Comparison with conventional Random Forest (RF) and Generalized Poisson Regression (GPR) models.
• Validation using two real-world genomic datasets.

Main Results:

• The proposed zero-inflated random forest models (ZAP_RF and ZAPC_RF) demonstrated superior prediction performance.
• These models effectively addressed the issue of excess zeros in count response variables.
• Outperformed conventional RF and GPR models in predictive accuracy.

Conclusions:

• Zero-inflated random forest models are highly effective for genomic selection with excess zero count data.
• The ZAP_RF and ZAPC_RF models offer a significant advancement over traditional statistical and machine learning methods in GS.
• These findings have important implications for improving the efficiency of breeding programs through enhanced genomic prediction.