Gene-environment interaction analysis via deep learning.

Area of Science:

• Genetics
• Computational Biology
• Biostatistics

Background:

• Gene-environment (G-E) interactions are crucial for understanding complex diseases.
• Current G-E interaction analysis primarily relies on regression techniques.
• Deep learning offers potential advantages in flexibility and prediction accuracy for biological data.

Purpose of the Study:

• To develop a novel deep learning-based method for G-E interaction analysis.
• To address the gap in deep learning applications for G-E interaction studies.
• To create a method that handles unspecified nonlinear effects and respects variable selection hierarchies.

Main Methods:

• Developed a deep neural network approach combined with penalization for G-E interaction analysis.
• The method performs simultaneous model estimation and selection of main genetic effects and G-E interactions.
• Ensures the "main effects, interactions" variable selection hierarchy is respected.

Main Results:

• Simulations demonstrate superior prediction and feature selection performance compared to existing methods.
• The approach effectively identifies important genetic effects and their interactions with environmental factors.
• Practical utility is confirmed through analysis of lung adenocarcinoma and skin cutaneous melanoma survival data.

Conclusions:

• The proposed deep learning method offers a powerful advancement for G-E interaction analysis.
• This approach can improve the understanding of complex diseases by leveraging modern computational techniques.
• The study provides a valuable new tool for genetic and omics research.