Deep contrastive learning enables genome-wide virtual screening.

Area of Science:

• Computational biology
• Drug discovery
• Artificial intelligence

Background:

• Protein structure prediction advancements offer new drug discovery avenues.
• Current virtual screening methods are computationally expensive.
• There is a need for faster and more accurate screening tools.

Purpose of the Study:

• To introduce DrugCLIP, a novel contrastive learning framework for ultrafast and accurate virtual screening.
• To demonstrate DrugCLIP's superior performance compared to existing baselines.
• To validate DrugCLIP's efficacy in identifying drug inhibitors through wet-lab experiments.

Main Methods:

• Developed DrugCLIP, a contrastive learning framework for virtual screening.
• Utilized AlphaFold2-predicted structures for screening targets lacking experimental data.
• Conducted in silico benchmarks and wet-lab validations for performance assessment.

Main Results:

• DrugCLIP achieved virtual screening speeds up to 10 million times faster than docking.
• Demonstrated superior performance over baselines on in silico benchmarks.
• Achieved a 15% hit rate for norepinephrine transporter and a 17.5% hit rate for thyroid hormone receptor interactor 12 using predicted structures.
• Identified and structurally characterized two inhibitors for norepinephrine transporter.

Conclusions:

• DrugCLIP offers a computationally efficient and highly accurate solution for virtual screening.
• The framework enables successful drug discovery even with predicted protein structures.
• GenomeScreenDB provides a valuable resource for post-AlphaFold era drug discovery.