Gene-gene interaction detection with deep learning

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Current genetic interaction detection methods are limited, focusing only on top SNPs and simple multiplicative relationships.
• The impact of complex genetic interactions on observed phenotypes remains largely uncharacterized.
• Discovering complex gene-gene interactions is crucial for understanding disease etiology.

Purpose of the Study:

• To develop and validate an open-source framework for enhanced detection of complex genetic interactions.
• To move beyond simple SNP-SNP interactions and multiplicative models in genetic association studies.
• To improve the power and accuracy of identifying gene-gene interactions influencing phenotypes.

Main Methods:

• Utilized a neural network to model the relationship between all SNPs within selected genes and phenotypes.
• Employed Shapley scores between hidden nodes (gene representations) to quantify complex interactions.
• Developed a novel permutation procedure for assessing the statistical significance of neural network-derived interactions.

Main Results:

• The proposed framework demonstrated superior performance in detecting complex interactions on simulated data compared to existing methods.
• In a UK Biobank cholesterol study, nine significant genetic interactions were identified.
• These nine interactions showed replication in an independent FINRISK dataset, confirming their biological relevance.

Conclusions:

• The novel framework significantly enhances the power to detect complex genetic interactions.
• This approach provides a more comprehensive understanding of the genetic architecture of complex traits.
• The findings have implications for genetic association studies and personalized medicine.