A few-shot link prediction framework to drug repurposing using multi-level attention network.

Area of Science:

• Computational Biology
• Pharmacology
• Artificial Intelligence

Background:

• Drug repurposing accelerates the discovery of new therapeutic uses for existing drugs.
• Link prediction in medical knowledge graphs models drug-disease relationships.
• Few-shot learning is crucial due to the rarity of novel drug-disease links.

Purpose of the Study:

• To develop an effective few-shot link prediction framework for drug repurposing.
• To address challenges posed by sparse data and weak interactions in medical knowledge graphs.
• To improve the accuracy and efficiency of identifying novel drug indications.

Main Methods:

• A meta-learning framework integrating a multi-level attention network (MLAN).
• Utilized gating mechanisms and graph attention networks to filter noise and highlight relevant neighborhood information.
• Employed a set transformer for robust triplet-level interaction learning and a model-agnostic meta-learning strategy for rapid optimization.

Main Results:

• The proposed MLAN-based framework demonstrated significant advantages over state-of-the-art few-shot link prediction methods.
• Achieved superior performance on specialized few-shot medical link prediction datasets (COVID19-One, BIOKG-One).
• Validated the framework's ability to capture valuable information in low-data scenarios.

Conclusions:

• The unified meta-learning framework effectively addresses challenges in few-shot drug repurposing.
• The MLAN approach provides valuable insights for predicting novel drug-disease relationships.
• This method offers a promising direction for accelerating drug discovery through advanced AI techniques.