A Novel Approach to Encode Two-Way Epistatic Interactions Between Single Nucleotide Polymorphisms

TL;DR

This paper introduces a new method for encoding gene-gene interactions in genetic risk models, preserving individual SNP information while capturing interactions, which enhances interpretability and model performance.

Contribution

The paper proposes three methods to encode epistatic interactions that retain SNP information, improving upon simple interaction models and facilitating better genetic risk assessment.

Findings

Interaction-preserving methods outperform simple interaction models.

Raw SNP genotypes are sufficient for complex ML models.

Explicit interaction encoding aids interpretability of genetic pathways.

Abstract

Modelling gene-gene epistatic interactions when computing genetic risk scores is not a well-explored subfield of genetics and could have potential to improve risk stratification in practice. Though applications of machine learning (ML) show promise as an avenue of improvement for current genetic risk assesments, they frequently suffer from the problem of two many features and to little data. We propose a method that when combined with ML allows information from individual genetic contributors to be preserved while incorporating information on their interactions in a single feature. This allows second-order analysis, while simultaneously increasing the number of input features to ML models as little as possible. We presented three methods that can be utilized to account for genetic interactions. We found that interaction methods that preserved information from the constituent SNPs performed significantly better than the simplest interaction method. Since the currently available ML methods are able to account for complex interactions, utilizing raw SNP genotypes alone is sufficient because the simplest model outperforms all the interaction methods Given that understanding and accounting for epistatic interactions is one of the most promising avenues for increasing explained variability in heritable disease, this work represents a first step toward an algorithmic interaction method that preserves the information in each component. This is relevant not only because of potential improvements in model quality, but also because explicit interaction terms allow a human readable interpretation of potential interaction pathways within the disease.