SEMdag: Fast learning of Directed Acyclic Graphs via node or layer ordering.

Area of Science:

• Causal inference and graphical models
• Bioinformatics and computational biology
• Statistical genetics

Background:

• Directed Acyclic Graphs (DAGs) are crucial for representing causal relationships between variables.
• Learning DAG structures from observational data is a significant challenge in various scientific fields.
• Existing methods for causal discovery often struggle with high-dimensional datasets.

Purpose of the Study:

• To introduce SEMdag(), a novel two-step approach for learning high-dimensional linear Structural Equation Models (SEMs).
• To evaluate the performance of SEMdag() in recovering plausible DAGs and predicting disease outcomes.
• To compare SEMdag() against established causal discovery techniques using real-world biological data.

Main Methods:

• SEMdag() employs a two-stage order-based search, utilizing either prior knowledge (Knowledge-based, KB) or a data-driven (Bottom-up, BU) strategy.
• The method assumes linear SEMs with equal variance error terms.
• Performance was assessed using RNA-seq data from four diseases (ALS, BRCA, COVID-19, STEMI) and evaluated with Random Forest (RF) for disease prediction.

Main Results:

• SEMdag() successfully recovered graph structures that demonstrated strong disease prediction performance.
• The Bottom-up (BU) approach showed better results for sparse initial graphs, while both BU and Knowledge-based (KB) performed well on denser graphs.
• The KB approach, particularly when leveraging topological layers, achieved the highest predictive scores.

Conclusions:

• SEMdag() offers a computationally efficient and flexible tool for high-dimensional causal structure learning.
• The method provides superior disease prediction capabilities compared to existing causal discovery techniques.
• SEMdag() is implemented in the R package SEMgraph, facilitating its accessibility and application in research.