FLONE: fully Lorentz network embedding for inferring novel drug targets

Area of Science:

• Computational biology
• Machine learning
• Bioinformatics

Background:

• Graph-based machine learning is widely used for drug target prediction in drug-disease-target (DDT) networks.
• Existing methods often fail to consider disease types during prediction and cannot fully utilize the hierarchical information within DDT networks.
• This limitation can lead to suboptimal drug target prediction accuracy.

Purpose of the Study:

• To improve drug target prediction by explicitly considering disease types and utilizing the hierarchical structure of DDT networks.
• To propose a novel hyperbolic embedding-based method, FLONE, for enhanced target prediction.
• To enable FLONE to handle previously unseen drugs and targets by incorporating external domain knowledge.

Main Methods:

• Formulated drug target prediction as a knowledge graph completion task.
• Developed FLONE, a hyperbolic embedding method designed to capture hierarchical topological information in DDT networks.
• Devised hyperbolic encoders to integrate external domain knowledge for handling novel samples.

Main Results:

• FLONE demonstrated superior accuracy in drug target prediction compared to Euclidean embedding methods on two DDT networks.
• The use of hyperbolic space effectively captured the hierarchical relationships within DDT networks.
• Hyperbolic encoders enabled FLONE to generalize to unseen drugs and targets, enhancing its practical applicability.

Conclusions:

• Hyperbolic embedding is a powerful approach for modeling hierarchical DDT networks and improving drug target prediction.
• FLONE offers a more accurate and versatile solution for identifying drug targets, especially when considering specific disease contexts.
• The method's ability to handle novel entities makes it valuable for real-world drug discovery scenarios.