SEMgraph

SEMgraph provides functions for estimating and analyzing directed acyclic graphs (DAGs) from data, implementing methods for structure learning and causal inference as well as tools for manipulating and visualizing graphs.

Key Features:

• Structure Learning: Implements SEMdag(), a two-step order-based search approach for fast learning of high-dimensional linear structural equation models (SEMs).
• Node Ordering Methods: Supports knowledge-based (KB) ordering using prior knowledge and bottom-up (BU) data-driven ordering methods.
• Topological Layer Ordering: Enables knowledge-based approach based on topological layers for denser graph structures.
• Graph Manipulation and Visualization: Provides tools for manipulating and visualizing directed acyclic graphs.
• Flexible Algorithm Definition: Allows users to define distinct structure learning algorithms.
• High-Dimensional Data Handling: Addresses high-dimensional issues with reduced computing load.

Scientific Applications:

• Disease Prediction: Recovers graph structures with good disease prediction performance evaluated by random forest (RF) supervised learning algorithm.
• ALS Analysis: Applied to Amyotrophic Lateral Sclerosis (ALS) expression data for causal structure learning.
• Breast Cancer Analysis: Applied to Breast cancer (BRCA) RNA-seq data for causal structure learning.
• COVID-19 Analysis: Applied to Coronavirus disease (COVID-19) expression data for causal structure learning.
• Myocardial Infarction Analysis: Applied to ST-elevation myocardial infarction (STEMI) expression data for causal structure learning.
• Causal Discovery Benchmarking: Evaluated against ARGES, GES, PC, LiNGAM, CAM, and NOTEARS causal discovery techniques.

Methodology:

SEMdag() employs a two-stage order-based search assuming a linear structural equation model with equal variance error terms. The method uses either knowledge-based (KB) ordering from prior knowledge or bottom-up (BU) data-driven ordering to learn high-dimensional DAGs from observed expression or RNA-seq data, with evaluation through training and testing datasets and random forest supervised learning.