CausNet

CausNet discovers optimal Bayesian Networks from high-dimensional datasets to infer causal relationships among variables.

Key Features:

• Dynamic Programming Algorithm: Employs dynamic programming with dimensionality reduction and parent set identification to enable learning Bayesian Networks in large-variable datasets.
• Generational Orderings Based Search: Explores the network space using generational orderings-based search to consider possible parent sets without exhaustive enumeration.
• Support for Diverse Data Types: Supports continuous and categorical predictors and continuous, binary, and survival outcomes.
• Scalability: Handles datasets with thousands of variables and identified an optimal network in 3.4 minutes for 513 genes with a survival outcome on standard personal computing hardware.
• Performance Superiority: Outperformed three state-of-the-art algorithms in comparative simulations for optimal Bayesian Network identification.
• Customizable Parameters: Allows specification of correlation thresholds, false discovery rate (FDR) cutoffs, and maximum node in-degree.
• Scoring Options: Provides Bayesian Information Criterion (BIC) and Bayesian Gaussian equivalent (Bge) scoring options.
• Application Versatility: Applicable to high-dimensional fields such as genomics and has been applied to an ovarian cancer gene expression dataset.

Scientific Applications:

• Gene regulatory network inference: Infers gene regulatory networks from high-dimensional expression data.
• Disease mechanism elucidation: Identifies potential causal relationships underlying disease mechanisms.
• Biomarker identification: Supports identification of key genetic markers associated with clinical outcomes, including survival.
• Survival outcome modeling: Integrates survival outcomes into network discovery from molecular data.
• High-dimensional causal inference: Enables causal inference in any domain that requires analysis of high-dimensional datasets.

Methodology:

Uses dynamic programming with dimensionality reduction and parent-set identification combined with generational-orderings-based search and scoring by BIC or Bge, with configurable correlation thresholds, false discovery rate cutoffs, and maximum node in-degree.