A new method for exploring gene-gene and gene-environment interactions in GWAS with tree ensemble methods and SHAP...

Area of Science:

• Genetics and Bioinformatics
• Computational Biology
• Statistical Genomics

Background:

• Identifying gene-gene and gene-environment interactions in genome-wide association studies (GWAS) is complex due to the vast number of potential combinations and unknown interaction types.
• Traditional parametric regression models are limited to detecting pre-specified interactions.
• Nonparametric tree ensemble models can detect unspecified interactions but have historically posed interpretation challenges, though recent explainability methods offer solutions.

Purpose of the Study:

• To develop and validate a tree ensemble- and SHAP-based methodology for discovering and interpreting gene-gene and gene-environment interactions within large-scale biobank datasets.
• To address the limitations of parametric models in detecting novel genetic interactions.
• To leverage advanced model explainability techniques for robust interpretation of complex genetic interactions.

Main Methods:

• Proposed a novel method combining tree ensemble models with SHapley Additive exPlanations (SHAP) for interaction identification and interpretation.
• Utilized independent cross-validation runs to implicitly scan the entire genome for interactions.
• Applied the method to UK Biobank data, focusing on obesity as the phenotype.

Main Results:

• Successfully identified potential gene-gene and gene-environment interactions in large-scale biobank data.
• Results align with existing research, pinpointing single nucleotide polymorphisms (SNPs) previously associated with obesity.
• Demonstrated effective interpretation and visualization of identified interaction candidates.

Conclusions:

• The developed method successfully identifies interaction candidates that may be missed by conventional parametric regression models.
• Further research is warranted to rigorously assess the uncertainty associated with the identified interaction candidates.
• The methodology is suitable for application in large-scale biobanks characterized by high-dimensional genetic data.