Docking Score ML: Target-Specific Machine Learning Models Improving Docking-Based Virtual Screening in 155 Targets.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in cheminformatics

Background:

• Molecular docking methods struggle with accurate energy prediction due to limitations in scoring functions for complex protein-ligand interactions.
• This inaccuracy leads to biases and errors in virtual screening and target identification for drug development.

Purpose of the Study:

• To introduce "Docking Score ML", an advanced scoring function designed to overcome the limitations of conventional molecular docking.
• To enhance the accuracy and efficiency of virtual screening and target prediction in drug discovery.

Main Methods:

• Developed "Docking Score ML" using over 200,000 docked complexes from 155 cancer targets.
• Utilized bioactivity data from ChEMBL, fine-tuned with supervised and deep learning techniques.
• Employed feature fusion, integrating Graph Convolutional Network (GCN) with multiple docking functions for enhanced prediction.

Main Results:

• Extensive validation using datasets like DUDE, DUD-AD, and LIT-PCBA demonstrated superior performance.
• Multitarget analysis, including drugs like sunitinib, confirmed the methodology's effectiveness.
• The fusion of GCN with docking functions significantly outperformed traditional approaches in accuracy.

Conclusions:

• "Docking Score ML" offers a substantial improvement over conventional methods for molecular docking.
• The tool provides an efficient and accurate solution for virtual screening and reverse docking in drug discovery.
• This AI-driven approach holds promise for accelerating the identification of potential drug candidates.