Structure learning for gene regulatory networks.

Area of Science:

• Computational Biology
• Systems Biology
• Bioinformatics

Background:

• Biological network inference is challenging with high-dimensional 'omics' data due to limited sample sizes (the 'small n, large p' problem).
• Biological networks possess inherent properties like sparsity and modularity, which can be leveraged for improved inference.
• Existing methods struggle with the high dimensionality and low sample size typical of 'omics' datasets.

Purpose of the Study:

• To develop a computational framework (SHINE) for robust biological network structure learning from high-dimensional 'omics' data.
• To address the 'small n, large p' challenge by incorporating data-driven structural constraints and shared learning.
• To enable efficient learning of multiple Markov networks at large p/n ratios previously unfeasible.

Main Methods:

• SHINE (Structure Learning for Hierarchical Networks) framework utilizes data-driven structural constraints.
• Employs a shared learning paradigm for efficient inference of multiple Markov networks.
• Applied to Pan-Cancer 'omics' data across 23 tumor types and subtype-specific breast cancer data.

Main Results:

• Learned tumor-specific networks exhibit expected biological graph properties.
• SHINE successfully recaptured known biological interactions and findings from existing literature.
• Identified key genes and biological processes critical for tumor maintenance and survival in breast cancer subtypes.

Conclusions:

• SHINE provides a powerful and feasible approach for inferring biological network structures from high-dimensional 'omics' data.
• The framework effectively overcomes the 'small n, large p' limitation.
• Application to cancer data reveals novel insights into disease mechanisms and identifies potential therapeutic targets.