edge2vec: Representation learning using edge semantics for biomedical knowledge discovery.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Representation learning offers advanced tools for knowledge graph mining.
• Existing methods often overlook the complexity of heterogeneous biomedical graphs.
• Biomedical data requires capturing rich semantics of nodes and edges (e.g., genes, drugs, diseases).

Purpose of the Study:

• To develop a novel representation learning model, edge2vec, that explicitly considers edge semantics in heterogeneous graphs.
• To address the challenge of analyzing complex biomedical knowledge domains.

Main Methods:

• Proposed the edge2vec model, which learns graph representations by considering edge semantics.
• Trained an edge-type transition matrix using Expectation-Maximization.
• Employed stochastic gradient descent for node embedding learning on heterogeneous graphs.

Main Results:

• edge2vec was validated on three biomedical tasks: entity classification, compound-gene bioactivity prediction, and information retrieval.
• The model significantly outperformed state-of-the-art methods by incorporating edge-types into node embedding.
• Demonstrated superior performance across all evaluated biomedical domain tasks.

Conclusions:

• The proposed edge2vec method adds significant value to existing graph analytical methodologies.
• edge2vec is applicable to real-world biomedical knowledge discovery challenges.
• Highlighting the importance of edge semantics for effective biomedical graph analysis.