EXPLORING THE REPRODUCIBILITY OF PROBABILISTIC CAUSAL MOLECULAR NETWORK MODELS

Area of Science:

• Biomedical research
• Life sciences
• Computational biology

Background:

• Network reconstruction algorithms, including Bayesian networks (BNs), are vital for analyzing complex biological data.
• BNs offer a framework for inferring causal relationships and integrating prior knowledge.
• The reproducibility of BNs in biological research remains underexplored.

Purpose of the Study:

• To investigate the criteria for generating reproducible Bayesian networks (BNs) in transcription-based regulatory networks.
• To assess the impact of sample size on BN reproducibility using independent biological datasets.

Main Methods:

• Utilized whole blood (GTEx Consortium) and liver (STARNET) datasets.
• Constructed BNs on data subsampled at varying levels.
• Compared subsampled networks with those from complete datasets.
• Validated findings using simulated networks.

Main Results:

• Reproducibility of edges in BNs is highly dependent on sample size.
• Identification of highly connected key driver nodes in BNs demonstrates high confidence across different sample sizes.
• Biological research using BNs requires careful consideration of reproducibility.

Conclusions:

• Systematic evaluation of BN reproducibility is necessary for reliable biological hypothesis generation.
• While edge reproducibility is sensitive to sample size, core network structures (key drivers) are more robust.
• Future studies should prioritize and report BN reproducibility metrics.