BioBLP: a modular framework for learning on multimodal biomedical knowledge graphs.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Graph Representation Learning

Background:

• Biomedical knowledge graphs (KGs) represent complex relationships, with embeddings used for link prediction.
• Existing methods often overlook multimodal entity attributes (e.g., protein sequences, molecular graphs).
• Biomedical KGs feature entities with heterogeneous data modalities, posing challenges for uniform representation.

Purpose of the Study:

• To develop a framework for incorporating multimodal data into biomedical KG embeddings.
• To analyze the performance of multimodal embeddings against traditional methods.
• To address entities with missing attributes and optimize training efficiency.

Main Methods:

• Proposed a modular framework (BioBLP) for learning KG embeddings with multimodal entity attributes.
• Developed an efficient pretraining strategy to reduce training time.
• Trained and evaluated models on a large biomedical KG (~2 million triples) for link prediction and drug-protein interaction prediction.

Main Results:

• Multimodal approach showed competitive performance in standard link prediction, outperforming baselines in drug-protein interaction prediction.
• Attribute incorporation improved performance for entities with low node degrees (approx. 75% of diseases).
• Pretraining strategy significantly boosted performance and reduced training runtime; attribute encoder optimization increased costs.

Conclusions:

• BioBLP facilitates multimodal data integration in biomedical KGs.
• Attribute data integration offers benefits for specific entity subsets, particularly understudied ones.
• The findings suggest potential for enhanced scientific discovery in areas with limited data.