Applying traditional and machine learning-based GWAS approaches for marker-trait identification in wheat.

Area of Science:

• Plant genetics
• Genomics
• Computational biology

Background:

• Complex traits in wheat are governed by polygenic and interactive genomic architectures.
• Traditional genome-wide association studies (GWAS) face challenges in resolving these complex genetic patterns.
• Machine learning (ML) offers advanced methods to detect non-linear effects and improve predictive accuracy for marker-trait associations (MTAs).

Purpose of the Study:

• To evaluate and compare the performance of traditional GWAS tools and ML approaches for marker-trait identification in winter wheat.
• To assess the ability of ML models to capture non-linear genetic effects and identify novel MTAs.
• To determine the utility of ML as a complementary method to traditional GWAS for complex trait analysis in plant breeding.

Main Methods:

• Evaluated traditional GWAS tools (GAPIT, GCTA, GEMMA, sommer, TASSEL) and ML models (Elastic Net, XGBoost, Random Forest, TSLRF) using a winter wheat dataset.
• Assessed GWAS tools based on computational efficiency, model performance, and MTA consistency.
• Analyzed ML models using feature importance metrics and functional annotation of selected markers.

Main Results:

• Traditional GWAS tools showed variability in runtime and MTA detection despite using mixed linear models.
• ML models successfully identified MTAs previously detected by traditional methods.
• ML approaches uncovered novel markers, suggesting the detection of non-linear or epistatic genetic effects.

Conclusions:

• Machine learning effectively complements traditional GWAS for marker-trait identification in wheat.
• ML expands the scope of detectable genetic signals by analyzing effects beyond additivity.
• These findings provide a practical approach for analyzing complex traits and support marker-assisted breeding strategies in wheat.