Semi-supervised Hierarchical Drug Embedding in Hyperbolic Space.

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Accurate drug representations are crucial for drug repositioning and predicting side effects.
• Existing methods struggle to incorporate drug hierarchy knowledge or place novel molecules within such hierarchies.

Purpose of the Study:

• To develop a semi-supervised drug embedding method that leverages both chemical structures and drug hierarchy information.
• To enable accurate prediction of drug properties, new uses, and side effects for both existing and novel molecules.

Main Methods:

• Utilized a Variational Auto-Encoder (VAE) framework to encode molecular chemical structures.
• Incorporated supervised information from an expert-crafted drug hierarchy to guide embedding in hyperbolic space.
• Combined unsupervised learning of chemical grammar with supervised hierarchical relations.

Main Results:

• The learned drug embedding accurately reproduces chemical structures and recapitulates hierarchical drug relationships.
• The method successfully infers pharmacological properties of novel molecules by identifying similar drugs in the embedding space.
• Demonstrated superior performance in predicting new drug uses and discovering new side effects compared to existing methods.

Conclusions:

• The developed semi-supervised drug embedding effectively integrates chemical and hierarchical information for enhanced drug representation.
• This approach offers a powerful tool for drug discovery, repositioning, and understanding drug properties and interactions.