Bivariate Causal Discovery and Its Applications to Gene Expression and Imaging Data Analysis

Area of Science:

• Genomics
• Epigenetics
• Biomedical Data Analysis
• Causal Inference

Background:

• Current genetic and imaging research heavily relies on statistical association, limiting understanding of complex phenotype etiology.
• Traditional causal inference methods require at least three variables, unsuitable for bivariate analyses in quantitative genetics (e.g., QTL, eQTL) and genomic-imaging studies.
• Discovering causal mechanisms is crucial for advancing genomic science and its practical applications.

Purpose of the Study:

• To introduce and evaluate bivariate causal discovery methods for continuous variables, specifically focusing on the additive noise model (ANM).
• To demonstrate the ANM's capability in distinguishing cause from effect in omics and imaging data.
• To explore the ANM's application in gene regulatory network construction and trait-imaging data analysis.

Main Methods:

• Utilizing the independence of cause and mechanism (ICM) principle for causal inference.
• Employing algorithmic information theory and the additive noise model (ANM) for bivariate causal discovery.
• Conducting large-scale simulations to assess ANM feasibility and applying ANM to gene regulatory networks and trait-imaging data.

Main Results:

• Simulations demonstrate the feasibility and performance of ANMs for bivariate causal discovery with continuous variables.
• ANM application to gene regulatory networks showcases its utility in inferring causal relationships.
• Analysis of trait-imaging data illustrates ANM's ability to identify causation and association scenarios.

Conclusions:

• The additive noise model (ANM) offers a robust approach for bivariate causal discovery in genomic and imaging data analysis.
• Transitioning from association to causation is vital for understanding complex phenotypes, and ANMs provide a promising tool.
• ANMs are a valuable method for distinguishing cause from effect between two continuous variables from observational data.