Non-homogeneous dynamic Bayesian networks with edge-wise sequentially coupled parameters.

Area of Science:

• Computational Biology
• Network Inference
• Machine Learning

Background:

• Non-homogeneous Dynamic Bayesian Networks (NH-DBNs) are used for learning networks with time-varying parameters.
• Conventional NH-DBNs treat segments independently, while coupled models enforce similarity, potentially hindering accuracy.
• Existing methods struggle with effectively modeling dynamic network changes across segments.

Purpose of the Study:

• To develop a novel consensus NH-DBN model that balances segment-specific learning with parameter similarity.
• To infer network structure, segmentation, and segment-specific parameters adaptively.
• To improve the accuracy of dynamic network reconstruction.

Main Methods:

• Proposed a consensus NH-DBN model integrating features of uncoupled and coupled approaches.
• Developed a method to infer for each edge whether its parameters should be coupled or uncoupled across segments.
• Applied the model to synthetic, yeast, and A. thaliana gene expression data.

Main Results:

• The consensus NH-DBN model achieved higher network reconstruction accuracies compared to state-of-the-art methods on synthetic and yeast data.
• For A. thaliana gene expression data, the model inferred a plausible network topology.
• The model provided insights into light-dependent gene interactions.

Conclusions:

• The consensus NH-DBN model offers a more flexible and accurate approach to modeling dynamic biological networks.
• This method enhances the understanding of time-varying biological processes by adaptively learning network structures.
• The model provides a powerful tool for inferring complex biological interactions from time-series data.