Representing core gene expression activity relationships using the latent structure implicit in Bayesian networks.

Area of Science:

• Bioinformatics
• Systems Biology
• Computational Biology

Background:

• Gene regulatory networks (GRNs) and co-expression networks are vital in biomedical research.
• These networks are often overly complex, resembling 'hairballs,' hindering interpretation and biological insight.

Purpose of the Study:

• To develop a simplified network model for summarizing core gene expression relationships.
• To improve the interpretability of biological networks for biomedical studies.

Main Methods:

• Proposed a Bayesian network approach, termed LatentDAG, to model gene expression activities.
• Compared LatentDAG's performance against conventional co-expression and ChIP-seq networks.
• Integrated LatentDAG with graph neural networks for downstream biological tasks.

Main Results:

• LatentDAG significantly reduces network complexity (by two orders of magnitude) compared to existing methods.
• The LatentDAG model reveals clearer gene clusters and module separation.
• Demonstrated LatentDAG's ability to bridge transcriptional regulatory networks with other biological networks (e.g., RNA-binding protein interactions).
• LatentDAG, when combined with graph neural networks, outperformed other networks in predicting gene conservation and gene clustering.

Conclusions:

• LatentDAG offers a substantially simpler and more interpretable representation of gene expression relationships.
• This novel network approach enhances the utility of biological networks for various computational biology tasks.
• The LatentDAG framework provides a powerful tool for dissecting complex gene regulatory mechanisms.