Automated interpretable artificial intelligence genomic prediction with AIGP

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Predicting phenotypes from genomic mutations is a significant challenge in genetics.
• Traditional methods like GBLUP and BayesR have limitations in capturing complex genetic effects, including epistasis.
• Machine learning (ML) offers a powerful alternative but often lacks interpretability.

Purpose of the Study:

• To evaluate various ML models against traditional methods for genomic prediction across diverse agricultural species.
• To identify key factors influencing genomic prediction performance, such as trait architecture and feature selection.
• To develop an interpretable and accessible toolkit for ML-driven genomic prediction research.

Main Methods:

• Evaluated 12 ML models and two traditional methods (GBLUP, BayesR) on real and simulated datasets.
• Assessed the impact of trait genetic architecture and feature selection on predictive performance.
• Employed Shapley Additive Explanations (SHAP) for quantifying SNP effects and developed the Artificial Intelligence Genomic Prediction (AIGP) toolkit.

Main Results:

• Boosting algorithms demonstrated superior performance among the evaluated ML methods.
• Trait genetic architecture and feature selection were identified as primary determinants of prediction accuracy.
• Accounting for gene interaction effects and optimizing hyperparameters significantly improved prediction accuracy.

Conclusions:

• ML, especially boosting algorithms, holds significant promise for advancing genomic prediction.
• Explainable ML approaches, integration of prior biological information, and parameter optimization are crucial for maximizing ML's potential.
• The AIGP toolkit facilitates automated model optimization and interpretability, enhancing the accessibility of ML for genomic selection.