V-Dock: Fast Generation of Novel Drug-like Molecules Using Machine-Learning-Based Docking Score and Molecular...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Existing drug design methods often require extensive experimental data, limiting their application for novel targets.
• Protein-ligand docking is computationally intensive, hindering broad chemical space exploration during molecular generation.

Purpose of the Study:

• To develop a computational workflow for designing novel drug-like molecules by integrating molecular property optimization and protein-ligand docking.
• To accelerate the molecular generation process by using machine learning to predict docking energy from SMILES strings.

Main Methods:

• A machine learning model was trained to predict docking energy using only SMILES strings.
• This model was integrated with the MolFinder global molecular property optimization approach.
• The combined approach, named V-dock, was used to generate novel molecules with high predicted docking scores and desired properties.

Main Results:

• Accurate prediction of docking scores was achieved using only SMILES strings.
• V-dock efficiently generated novel molecules with high predicted docking scores, target protein similarity, and desirable drug-like properties (e.g., QED).
• The predicted docking scores showed good correlation with actual docking scores, validating the approach.

Conclusions:

• V-dock offers an efficient computational strategy for de novo drug design.
• The method overcomes data limitations in traditional approaches by leveraging machine learning for rapid docking score prediction.
• This accelerates the discovery of novel drug candidates with optimized properties and high binding affinity.