Evaluation of tree-based statistical learning methods for constructing genetic risk scores.

Area of Science:

• Genetics
• Statistical Learning
• Bioinformatics

Background:

• Genetic risk scores (GRS) aggregate genetic variants like single nucleotide polymorphisms (SNPs) for trait prediction.
• Current GRS methods often use generalized linear models, limiting their ability to capture complex genetic interactions or non-linear relationships.

Purpose of the Study:

• To investigate tree-based methods, including random forests and logic regression, as alternatives for constructing GRS.
• To evaluate modified versions of these tree-based methods for enhanced GRS construction.

Main Methods:

• Comparison of random forests and logic regression against elastic net for GRS construction.
• Utilized simulation studies and a real-world cohort dataset for evaluation.
• Investigated logic bagging, a modification of logic regression, for its predictive performance.

Main Results:

• Tree-based approaches, particularly logic bagging, demonstrated superior performance over elastic net in constructing GRS for binary traits.
• Logic bagging achieved high predictive power, as indicated by area under the curve and statistical power.
• Even without considering epistasis, tree-based methods often outperformed regularized regression.

Conclusions:

• Random forests and logic bagging are recommended for GRS construction, especially when potential epistasis among SNPs is suspected.
• Comprehensive joint hyperparameter optimization is crucial for developing optimal prediction models.