Boosting probabilistic graphical model inference by incorporating prior knowledge from multiple sources.

Area of Science:

• Computational Biology
• Systems Biology
• Bioinformatics

Background:

• Inferring biological regulatory networks is crucial for understanding biological systems.
• Experimental data noise and limited sample sizes hinder accurate network reconstruction.
• Integrating prior biological knowledge can improve network inference but is challenging due to data heterogeneity.

Purpose of the Study:

• To develop computational methods for integrating diverse biological information sources into a consensus structure for improved graphical model inference.
• To enhance the accuracy of regulatory network reconstruction in the presence of noisy and limited experimental data.

Main Methods:

• Proposed two novel computational methods: Latent Factor Model (LFM) and Noisy-OR.
• LFM reconstructs a hidden variable representing common correlations among external information sources using a Bayesian approach.
• Noisy-OR probabilistically identifies the strongest support for interactions across multiple information sources.

Main Results:

• Both LFM and Noisy-OR significantly improved Bayesian Network reconstruction accuracy compared to methods without prior knowledge and other competing approaches.
• Demonstrated effectiveness on KEGG signaling pathways, breast cancer gene expression data, and yeast heat shock response data.
• The framework successfully integrated diverse sources like pathway databases, Gene Ontology (GO) terms, and protein domain data.

Conclusions:

• The proposed methods offer a flexible and effective framework for integrating heterogeneous biological information to enhance regulatory network inference.
• These computational approaches address the limitations of noisy experimental data and small sample sizes, leading to more reliable biological insights.
• The framework's adaptability allows for the integration of new information sources as they become available.