Explaining protein-protein interactions with knowledge graph-based semantic similarity.

Area of Science:

• Biomedical informatics
• Artificial intelligence in science
• Machine learning for drug discovery

Background:

• Machine learning (ML) and artificial intelligence (AI) are increasingly used in biomedical applications like protein-protein interaction prediction.
• Explainable AI (XAI) is crucial for scientific discovery, enabling understanding of ML mechanisms and data bias.
• Knowledge graphs (KGs) represent domain knowledge but are often explored using non-explainable embeddings.

Purpose of the Study:

• To develop an explainable method for representing entities in knowledge graphs for biomedical applications.
• To enhance the interpretability and predictive performance of machine learning models in complex biological domains.
• To provide an alternative to non-explainable knowledge graph embeddings.

Main Methods:

• Proposed KGsim2vec, a novel method for generating explainable vector representations using aspect-oriented semantic similarity in knowledge graphs.
• Utilized various machine learning models (decision trees, genetic programming, random forest, eXtreme gradient boosting) to predict entity relations.
• Computed similarity across multiple semantic aspects within the knowledge graph.

Main Results:

• Considering multiple semantic aspects in entity similarity representation improved both explainability and predictive performance.
• KGsim2vec outperformed traditional black-box methods like knowledge graph embeddings and graph neural networks.
• The developed models were capable of capturing biological phenomena and revealing data biases.

Conclusions:

• KGsim2vec offers a more explainable and effective approach for biomedical applications compared to current embedding-based methods.
• The method enhances scientific discovery by providing interpretable insights into biological relationships and data characteristics.
• This work advances the integration of explainable AI with knowledge graphs for robust biomedical data analysis.