Dynamic Bayesian Network Learning to Infer Sparse Models From Time Series Gene Expression Data.

Area of Science:

• Systems Biology
• Bioinformatics
• Computational Biology

Background:

• Inferring gene regulatory networks (GRNs) from high-dimensional, sparse biological data is a significant challenge in systems biology.
• Bayesian networks (BNs) and dynamic Bayesian networks (DBNs) are common tools for GRN inference, but traditional methods often yield numerous false positive edges.

Purpose of the Study:

• To develop novel BN scoring functions that improve the accuracy of GRN inference by minimizing spurious edges.
• To enhance the structure learning process for DBNs applied to gene expression data.

Main Methods:

• Proposed two new BN scoring functions extending the Bayesian Information Criterion (BIC) with additional penalty terms.
• Utilized these scoring functions in conjunction with DBN structure search algorithms.
• Evaluated methods on autoregressive and DREAM4 benchmarks, as well as three real time-series gene expression datasets.

Main Results:

• The novel BN scoring functions significantly reduced spurious edges compared to the standard BIC score.
• Improved precision in learned GRN graphs was observed across benchmark datasets.
• Demonstrated effectiveness in learning sparse graphs from high-dimensional time-series gene expression data.

Conclusions:

• The developed BN scoring functions offer a more accurate approach to inferring gene regulatory networks.
• These methods provide a valuable tool for systems biology research, enabling more reliable GRN reconstruction.
• The open-source R package facilitates the application of these advanced algorithms in biological research.