Prediction of drug-target interactions through multi-task learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Protein-ligand binding identification is crucial for drug discovery.
• Multi-task learning (MTL) facilitates knowledge sharing with limited data per task.
• Standard MTL can lead to performance trade-offs or degradation.

Purpose of the Study:

• To develop a general MTL scheme that enhances average performance.
• To minimize individual task performance degradation in MTL.
• To improve the efficiency of drug discovery processes.

Main Methods:

• Group selection based on chemical similarity of ligand sets for target grouping.
• Joint training of similar targets within selected groups.
• Knowledge distillation with teacher annealing, using single-task models as guides.

Main Results:

• The proposed MTL scheme achieved higher average performance than single-task learning (STL) and classic MTL.
• MTL proved particularly effective for tasks with initially low performance.
• Knowledge distillation successfully mitigated individual task performance degradation.

Conclusions:

• The novel MTL approach enhances drug discovery by improving predictive performance.
• Group selection and knowledge distillation are effective strategies for robust MTL.
• This method offers a promising direction for computational drug design and development.