piMGM: incorporating multi-source priors in mixed graphical models for learning disease networks

Area of Science:

• Computational biology
• Systems medicine
• Bioinformatics

Background:

• Learning probabilistic graphs from mixed data (gene expression and clinical data) is crucial for systems medicine.
• Existing methods struggle to effectively leverage imperfect pathway database information for mixed graphical model (MGM) learning.

Purpose of the Study:

• To introduce piMGM, a novel method for accurate probabilistic graph structure learning from mixed data.
• To incorporate priors from multiple experts with varying reliability into MGM learning.
• To accurately score the reliability of expert prior information, even with limited sample sizes.

Main Methods:

• Developed piMGM, a method for probabilistic graph learning over mixed data.
• Implemented a mechanism to incorporate and weight prior information from multiple experts.
• Scored expert reliability and utilized these scores to guide graph learning.

Main Results:

• piMGM accurately learns probabilistic graph structures from mixed data.
• The method reliably scores expert prior information reliability, even at low sample sizes.
• piMGM's performance is robust against unreliable priors.
• Applied piMGM to TCGA data, identifying important breast cancer pathways and improving cancer subtype classification.

Conclusions:

• piMGM offers an accurate and robust approach for probabilistic graph learning with mixed data.
• The method effectively leverages expert knowledge by assessing and incorporating reliability.
• piMGM has significant potential applications in systems medicine, particularly in identifying disease-relevant pathways and enhancing classification.