Learning epistatic polygenic phenotypes with Boolean interactions

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Detecting epistatic drivers of human phenotypes is computationally challenging with traditional regression methods.
• Existing approaches struggle with higher-order interactions and large-scale genomic data.
• Multiplicative terms in regression may not accurately represent biological interactions.

Purpose of the Study:

• Introduce the epiTree pipeline for extracting higher-order genetic interactions from genomic data.
• Utilize tree-based models within the Predictability, Computability, Stability (PCS) framework.
• Develop a method to identify and assess epistatic interactions beyond pairwise relationships.

Main Methods:

• epiTree pipeline selects variants based on tissue-specific gene expression.
• Iterative random forests (iRF) identify candidate Boolean interactions (pairwise and higher-order).
• Significance tests use a stabilized likelihood ratio test and bootstrap sampling for PCS epistasis p-values.

Main Results:

• epiTree successfully predicted red hair, identifying known and novel non-linear interactions around MC1R.
• For multiple sclerosis (MS), epiTree prioritized novel interactions around HLA-DRB1, a known MS-associated variant.
• The pipeline demonstrated effectiveness in uncovering complex genetic interactions for human phenotypes.

Conclusions:

• The epiTree pipeline offers a powerful approach to detect higher-order epistatic interactions.
• It enhances prediction accuracy and reduces the search space for experimental validation.
• This method holds potential for advancing our understanding of the genetic architecture of complex diseases.