Second order optimization for the inference of gene regulatory pathways

Area of Science:

• Systems Biology
• Computational Biology
• Bioinformatics

Background:

• Gene regulatory pathway modeling is crucial due to increasing experimental data.
• Gradient descent methods for model training yield variable results.
• Existing methods may lack robustness in inferring optimal gene regulatory pathways.

Purpose of the Study:

• To present a new second-order learning rule for inferring optimal gene regulatory pathways.
• To develop an optimization rule independent of learning parameters, improving upon gradient descent.
• To provide a robust method for calculating maximal gene expression based on initial conditions.

Main Methods:

• A novel second-order learning rule based on Newton's method is proposed.
• Flow vectors are estimated using biomass conservation principles.
• Constraints are formulated with weighting coefficients to calculate maximal target gene expression.

Main Results:

• The proposed optimization rule is independent of the learning parameter, unlike gradient descent.
• The algorithm was benchmarked and validated on functions and literature-derived gene regulatory networks.
• The new method demonstrated superior performance compared to gradient descent learning.

Conclusions:

• The developed second-order learning rule offers a more stable and effective approach for gene regulatory pathway inference.
• Extensive comparisons validate the effectiveness of the proposed methodology against existing methods like extreme pathway analysis.