MGPT: A Multi-task Graph Prompt Learning Framework for Drug Discovery.

Area of Science:

• Biomedical research
• Computational drug discovery
• Graph representation learning

Background:

• Accurate prediction of drug associations (drug-target, side effects, disease relationships) is vital for precision medicine.
• Graph representation learning methods are increasingly used for drug association studies.
• Challenges exist in applying graph pre-training to drug development, particularly for multi-task and few-shot learning.

Purpose of the Study:

• To propose a unified Multi-task Graph PrompT (MGPT) learning model.
• To provide generalizable and robust graph representations for few-shot drug association prediction.
• To address challenges in multi-task learning and limited data scenarios in drug development.

Main Methods:

• Constructed a heterogeneous graph network with entity pairs as nodes.
• Employed self-supervised contrastive learning of sub-graphs during pre-training.
• Utilized learnable functional prompts embedded with task-specific knowledge for downstream tasks.

Main Results:

• Demonstrated robust performance across various drug association tasks.
• Showcased seamless task switching capabilities.
• Outperformed competitive approaches, particularly in few-shot learning scenarios.

Conclusions:

• MGPT offers a robust solution for multi-task learning in drug development.
• The model effectively addresses the challenges of limited data in drug association prediction.
• MGPT provides generalizable graph representations crucial for advancing precision medicine.