A Self-Supervised Framework for Learning Biological Entities Representation by Fusing Class Information.

Area of Science:

• Bioinformatics
• Computational Biology
• Artificial Intelligence

Background:

• Ontologies are crucial for biological data annotation, integration, and analysis.
• Existing representation learning methods often overlook entity class information in ontologies.
• There is a need for methods that effectively leverage class information for enhanced bio-entity representation.

Purpose of the Study:

• To propose a unified framework, ERCI, for learning bio-entity embeddings by fusing class information.
• To develop a pluggable framework that can be integrated with existing knowledge graph embedding models.
• To improve the performance of downstream biological applications through enhanced entity representations.

Main Methods:

• ERCI jointly optimizes knowledge graph embedding and self-supervised learning.
• It fuses entity class information into the embedding generation process.
• The framework is designed to be compatible with various knowledge graph embedding models.

Main Results:

• ERCI demonstrated superior performance in predicting protein-protein interactions and gene-disease associations.
• Evaluations were conducted on standard datasets and three simulated long-tail datasets.
• The framework effectively leverages class information for improved bio-entity embeddings.

Conclusions:

• ERCI provides a powerful and flexible approach for learning bio-entity representations.
• Integrating class information significantly enhances the utility of embeddings in biological applications.
• ERCI outperforms state-of-the-art methods in key biological prediction tasks.