Learning transcriptional regulatory networks from high throughput gene expression data using continuous three-way...

Area of Science:

• Systems Biology
• Computational Biology
• Statistical Learning

Background:

• Probability-based statistical learning methods like mutual information and Bayesian networks are used for reverse engineering biological relationships.
• Previous methods face challenges with continuous variables, complex three-way interactions, and differentiating causal from confounding relationships.

Purpose of the Study:

• Introduce MI3, a novel statistical learning strategy to address limitations in existing biological network inference tools.
• Improve the realistic representation of underlying biological systems by enhancing mechanistic relationship detection.

Main Methods:

• Developed MI3, a new statistical learning strategy.
• Tested MI3 using synthetic and experimental biological data, including a microarray dataset for MYC transcription factor regulatory network inference.
• Compared MI3 performance against control methods like Bayesian networks and classical mutual information.

Main Results:

• MI3 achieved high sensitivity and precision (0.77/0.83 absolute, 0.99 relative) on synthetic data, significantly outperforming control methods.
• Inferred regulatory networks using MI3 were distinct from those of control methods and effectively differentiated causal from confounding models.
• MI3 identified key MYC cofactors and regulatory mechanisms, supported by existing literature, revealing repeated use of limited regulatory mechanisms.

Conclusions:

• MI3 demonstrates superior performance compared to frequently used control methods for inferring mechanistic relationships.
• MI3 offers a powerful approach for understanding complex biological and other systems.
• The MI3 method is available as an R package ('mi3') for broader accessibility.