BioMedKG: multimodal contrastive representation learning in augmented BioMedical knowledge graphs.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Data Science

Background:

• Biomedical Knowledge Graphs (BKGs) are crucial for integrating diverse data to understand complex biological relationships.
• Effective link prediction in BKGs can identify novel connections, such as potential drug-disease associations.
• Existing BKGs face limitations in comprehensively integrating multimodal data.

Purpose of the Study:

• To develop a novel multimodal approach for enhancing link prediction in Biomedical Knowledge Graphs.
• To introduce PrimeKG++, an enriched BKG incorporating biological sequences and textual descriptions.
• To improve the generalizability and accuracy of link prediction, even for unseen entities.

Main Methods:

• A multimodal approach unifying embeddings from specialized Language Models (LMs) with Graph Contrastive Learning (GCL) for intra-entity relationships.
• Utilizing a Knowledge Graph Embedding (KGE) model to capture inter-entity relationships for link prediction.
• Developing PrimeKG++, an enriched knowledge graph with multimodal data (biological sequences, textual descriptions).

Main Results:

• The proposed method demonstrates strong generalizability, enabling accurate link predictions for unseen nodes.
• Experimental validation on PrimeKG++ and the DrugBank dataset shows the method's effectiveness and robustness.
• The approach successfully combines semantic and relational information into a unified representation.

Conclusions:

• The novel multimodal approach significantly enhances link prediction accuracy in Biomedical Knowledge Graphs.
• PrimeKG++ provides a robust foundation for discovering complex biomedical relationships and potential drug-target interactions.
• The developed methodology and resources are publicly available, facilitating further research in the field.