CGBayesNets: conditional Gaussian Bayesian network learning and inference with mixed discrete and continuous data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Bayesian Networks (BN) are widely used in bioinformatics for predictive modeling.
• Existing BN software struggles with mixed discrete and continuous variables common in genomics.
• Discretizing continuous variables leads to information loss; lack of inference routines hinders prediction.

Purpose of the Study:

• Introduce CGBayesNets, a MATLAB package for clinical phenotype prediction from mixed data.
• Address limitations of existing BN software in handling multimodal genomic data.
• Provide a robust tool for prediction using conditional Gaussian Bayesian networks (CGBNs).

Main Methods:

• Implement Bayesian likelihood and inference algorithms for CGBNs.
• Offer four network learning algorithms balancing computational cost and network likelihood.
• Include functions for cross-validation, bootstrapping, and AUC manipulation for model verification.

Main Results:

• Demonstrate successful prediction of clinical phenotypes from mixed discrete and continuous genomic data.
• Showcase applications in predicting wood properties, classifying leukemia subtypes, and forecasting ICU mortality.
• Provide example analyses on public metabolomic and gene expression datasets.

Conclusions:

• CGBayesNets effectively handles mixed data types for predictive modeling in genomics.
• The package facilitates robust prediction of clinical outcomes from complex biological data.
• CGBayesNets is available as open-source MATLAB code, promoting wider research application.