Generalization power of threshold Boolean networks.

Area of Science:

• Computational Biology
• Systems Biology
• Network Science

Background:

• Threshold Boolean networks model gene regulation and social dynamics.
• Inferring these networks requires learning parameters from configuration data.
• Complete state transition matrices are often unavailable in practice.

Purpose of the Study:

• Investigate the generalization power of threshold Boolean networks.
• Assess network inference accuracy with reduced or degraded training data.
• Evaluate the preservation of original system fixed points.

Main Methods:

• Empirical experiments on networks of varying sizes and connectivities.
• Utilized the perceptron learning algorithm for network training.
• Examined degraded data scenarios and fixed point preservation.

Main Results:

• Larger networks require less data for accurate inference (e.g., 9-node networks needed 46% data vs. 5-node networks needing 62.5%).
• Higher node indegree correlates positively with increased data requirements for inference.
• Around 40% of data is generally sufficient to retain system fixed points.

Conclusions:

• Network size is inversely related to data requirements for inference.
• Node connectivity impacts the amount of data needed for accurate threshold Boolean network reconstruction.
• Sufficient data exists to preserve key dynamical properties like fixed points.