High-quality gene/disease embedding in a multi-relational heterogeneous graph after a joint matrix/tensor...

Area of Science:

• Bioinformatics
• Computational Biology
• Network Science

Background:

• Node embedding is crucial for analyzing biological networks.
• Existing methods struggle with sparse multi-relational data between genes and diseases.
• A multi-relational heterogeneous graph format is proposed to address this.

Purpose of the Study:

• To develop a novel data integration algorithm for high-quality gene and disease embeddings.
• To fully capture information from a multi-relational heterogeneous biological network.
• To improve downstream applications by enhancing semantic embedding.

Main Methods:

• Constructed a large-scale gene-disease network with 163,024 nodes and 25,265,607 edges.
• Proposed the Joint Decomposition of Heterogeneous Matrix and Tensor (JDHMT) model.
• Integrated heterogeneous data including genes, diseases, chemicals, mutations, pathways, and phenotypes.

Main Results:

• JDHMT model achieved high-quality gene and disease embeddings.
• Intrinsic and extrinsic evaluations demonstrated superior performance over eleven SOTA methods.
• The model excelled in interpretable data clustering and link prediction tasks.

Conclusions:

• The JDHMT model offers a powerful approach for integrating and embedding complex biological networks.
• The developed gene-disease network and embeddings facilitate further research in bioinformatics.
• The study provides valuable resources for the scientific community.