Using multi-step proposal distribution for improved MCMC convergence in Bayesian network structure learning

Area of Science:

• Computational Biology
• Systems Biology
• Bioinformatics

Background:

• Bayesian networks model probabilistic and causal relationships in systems.
• Learning biological networks (e.g., gene regulation) is complex due to large model spaces.
• Markov Chain Monte Carlo (MCMC) methods are often slow for Bayesian network structure inference, especially with growing datasets.

Purpose of the Study:

• To enhance the convergence speed of MCMC methods in Bayesian network structure learning.
• To develop an adjustable proposal distribution for more efficient exploration of the structure space.
• To improve the inference of biological network structures, such as signaling pathways.

Main Methods:

• Implemented an adjustable proposal distribution for MCMC sampling.
• Tested the method's ability to propose a wide range of structural changes.
• Applied the method to infer network structures from phosphoprotein data.

Main Results:

• Demonstrated improved convergence rates in Bayesian network structure space.
• Showcased enhanced network structure inference capabilities.
• Successfully analyzed human primary T cell signaling network phosphoprotein data.

Conclusions:

• The proposed adjustable proposal distribution significantly accelerates MCMC convergence for Bayesian network structure learning.
• This approach offers a more efficient way to infer complex biological networks.
• The method shows promise for analyzing large-scale biological signaling data.