Towards explainable interaction prediction: Embedding biological hierarchies into hyperbolic interaction space

Area of Science:

• Computational Biology
• Pharmacology
• Machine Learning

Background:

• Traditional drug development is slow and expensive.
• Computational methods accelerate drug discovery but often lack explainability.
• Machine learning models for drug-target interaction prediction need improved interpretability.

Purpose of the Study:

• To develop more interpretable drug-target interaction prediction models.
• To integrate biological hierarchies (drug and protein) into a joint-embedding latent space.
• To compare Euclidean and hyperbolic embedding spaces for interaction prediction.

Main Methods:

• Developed a similarity-based prediction model in a latent space aligned with biological hierarchies.
• Integrated drug and protein hierarchies using embedding regularization.
• Conducted a comparative analysis of Euclidean versus hyperbolic embeddings.
• Utilized dimensionality reduction for latent space analysis and visualization.

Main Results:

• Hierarchy regularization enhances model interpretability without reducing predictive performance.
• Hyperbolic embeddings, with regularization, improve the quality of embedded biological hierarchies.
• The proposed approach provides visual insights into the model's latent space.

Conclusions:

• The interpretable hyperbolic latent space facilitates informed drug discovery applications.
• Integrating hierarchies and using hyperbolic geometry improves drug-target interaction prediction transparency.
• The method is compatible with existing explainable AI solutions for further transparency.