Exploration on learning molecular docking with deep learning models.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Artificial intelligence in pharmacology

Background:

• Molecular docking-based virtual screening (VS) identifies potential drug candidates from vast molecular libraries.
• Traditional VS methods are computationally intensive, often wasting resources on molecules with low target affinity.
• Efficient screening is crucial for reducing experimental testing and accelerating drug discovery.

Purpose of the Study:

• To develop and evaluate a deep learning-powered virtual screening (VS) approach for rapid and efficient screening of ultra-large compound libraries.
• To significantly reduce computational costs associated with molecular docking.
• To obtain diverse potential active compounds with high accuracy.

Main Methods:

• Implemented DL-DockVS, a novel approach combining deep learning models (regression and classification) with established docking programs.
• Trained deep learning models to learn step-by-step outcomes of pipelined docking programs.
• Utilized a self-built dataset of approximately 1.9 million molecules for verification.

Main Results:

• DL-DockVS successfully filtered out compounds with poor docking scores while retaining those with high potential across 10 DUD-E protein targets.
• Achieved excellent results in recall rate, active compound enrichment factor, and runtime speed.
• Demonstrated the approach's effectiveness on a large, diverse dataset.

Conclusions:

• DL-DockVS is a practical, effective, and transferable strategy for screening ultra-large compound libraries in the big data era.
• The approach significantly improves runtime efficiency while maintaining a high success rate.
• Enables researchers to build predictive models for specific targets, facilitating future virtual screening without re-computation.