Sparse prediction informed by genetic annotations using the logit normal prior for Bayesian regression tree...

Area of Science:

• Genetics
• Bioinformatics
• Statistical Genetics

Background:

• Predicting complex diseases using high-dimensional genetic variants like single nucleotide polymorphisms (SNPs) is crucial for research and clinical applications.
• Challenges include weak signals, high dimensionality, and linkage disequilibrium among SNPs, complicating model building.
• Functional annotations offer potential to improve SNP-based prediction models.

Purpose of the Study:

• To propose an alternative prior for Bayesian additive regression trees (BART) to enhance prediction of complex diseases and traits.
• To develop a framework that incorporates functional annotations and accounts for SNP correlations.
• To address limitations of default BART priors in sparse genetic data scenarios.

Main Methods:

• Proposed an alternative prior for BART based on the logit normal distribution.
• Developed a framework to integrate functional annotations and prior information on SNP correlations.
• Conducted simulation studies and a genome-wide prediction analysis using Alzheimer's Disease Neuroimaging Initiative data.

Main Results:

• The proposed logit normal prior enhances BART's ability to handle sparse genetic data.
• The method effectively models informative functional annotations and between-SNP correlations.
• Demonstrated improved prediction performance in simulation and real-world Alzheimer's disease data.

Conclusions:

• The novel BART prior offers an adaptive and effective approach for genetic prediction, particularly in high-dimensional and sparse settings.
• Incorporating functional annotations and SNP correlations via this prior improves prediction accuracy.
• This method has significant implications for identifying causal genes and advancing precision medicine.