A two-stage GAN-based instrumental variable method for causal analysis of omics data.

Area of Science:

• Genetics and Bioinformatics
• Causal Inference
• Machine Learning

Background:

• Identifying causal genes for complex diseases is challenging.
• Mendelian randomization (MR) uses genetic variants to infer causality but faces bias from violated assumptions and nonlinearities.
• Existing MR methods struggle with complex omics data and unobserved confounding factors.

Purpose of the Study:

• To develop a robust, distribution-free deep learning framework for MR analysis.
• To address violations of instrumental variable assumptions and nonlinear exposure-outcome relationships in MR.
• To enable causal inference with complex, multi-omics data.

Main Methods:

• A two-stage deep learning framework utilizing Generative Adversarial Networks (GAN) and deep functional neural networks.
• Stage 1: GAN estimates conditional gene expression distribution given genetic variants (IVs).
• Stage 2: Deep functional networks model nonlinear causal relationships between gene expression and disease outcomes.

Main Results:

• The proposed GAN-based instrumental variable (GAN-IV) method shows superior performance over traditional and deep learning-based MR methods in simulations.
• GAN-IV effectively captures complex, nonlinear causal effects and handles diverse omics data types.
• Real-data application on the ROSMAP dataset confirms GAN-IV's ability to model gene expression-disease phenotype relationships.

Conclusions:

• GAN-IV offers a powerful, distribution-free tool for causal inference in complex omics data.
• The framework accounts for unobserved pleiotropy and linkage disequilibrium, improving causal effect estimation.
• This approach advances the accurate identification of disease-associated genes and their etiological roles.