Bayesian blind source separation for data with network structure.

Area of Science:

• Systems Biology
• Bioinformatics
• Computational Biology

Background:

• Biological regulatory systems generate complex multivariate signaling data.
• Prior knowledge of network structures aids in interpreting this data.
• Existing methods struggle with the complexity and missing values in signaling data.

Purpose of the Study:

• To develop a novel algorithm, emGrade, for analyzing multivariate signaling data within known regulatory networks.
• To improve the estimation of source signals and model parameters using a Bayesian network framework.
• To enhance the interpretability and flexibility of blind source separation (BSS) for biological network analysis.

Main Methods:

• Utilized a blind source separation (BSS) model adapted for multivariate signaling data.
• Incorporated Bayesian networks to capture the structure of source signals.
• Developed the emGrade algorithm employing expectation-maximization for parameter and signal estimation.
• Assumed stationary signals to manage parameter space complexity.

Main Results:

• emGrade demonstrated improved estimation performance compared to other BSS algorithms for network data.
• The Bayesian approach effectively handled repeated and missing observation values.
• Model selection criteria allowed for determining the number of source signals and evaluating network structures.
• Simulations confirmed the recovery of source signals based on graph structure and data dimensionality.

Conclusions:

• emGrade offers a statistically interpretable and flexible method for analyzing biological regulatory networks.
• The algorithm enhances signal extraction from complex multivariate data, even with missing values.
• emGrade provides a robust framework for inferring network properties and source signals in systems biology.