DAGBagM

DAGBagM learns directed acyclic graphs (DAGs) from datasets containing both continuous and binary variables to infer gene and protein regulatory networks and identify causal biomarkers in proteogenomic studies.

Key Features:

• Joint modeling of continuous and binary variables: Employs distinct statistical models for continuous and binary data allowing both variable types to serve as parents or children in the learned DAG.
• Bootstrap aggregation strategy: Integrates bootstrap aggregating to improve estimation accuracy, reduce false positives, and aggregate structure across resamples.
• Prior knowledge incorporation: Provides a framework to incorporate prior edges into network reconstruction during aggregation.
• Comparative performance: Demonstrates superior edge detection between continuous and binary nodes compared with score-based hill climbing (HC), constraint-based PC-algorithm (PC-alg), and max-min hill climbing (MMHC).
• Optimized HC implementation: Implements a hill-climbing (HC) algorithm reported to be faster than the HC implementation in R packages such as bnlearn.

Scientific Applications:

• Regulatory network inference: Infers gene and protein regulatory networks from mixed-type datasets for downstream causal analysis.
• Proteogenomic biomarker discovery: Applied to proteomics datasets, including ovarian cancer studies, to identify potential prognostic protein biomarkers associated with clinical outcomes.

Methodology:

Uses distinct statistical models for continuous and binary variables, integrates bootstrap aggregation for structure averaging and false-positive control, allows incorporation of prior edges, and employs a score-based hill-climbing (HC) algorithm with an optimized implementation.