Structure-Based de Novo Molecular Generator Combined with Artificial Intelligence and Docking Simulations

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Deep learning molecular generation models are crucial for drug discovery.
• Existing models often neglect the three-dimensional (3D) structure of target proteins.
• This limitation hinders effective drug design and lead optimization.

Purpose of the Study:

• To develop an advanced deep learning-based molecular generator, SBMolGen.
• To incorporate target protein 3D structure information into the molecular generation process.
• To improve the generation of novel, active drug-like molecules.

Main Methods:

• Integration of a recurrent neural network (RNN) with Monte Carlo tree search (MCTS).
• Inclusion of molecular docking simulations within the generation loop.
• Utilized SBMolGen for generating molecules against four diverse protein targets (kinases and GPCRs).

Main Results:

• Generated molecules demonstrated superior binding affinity (docking scores) compared to known active compounds.
• The model produced molecules with a broader and more diverse chemical space distribution.
• SBMolGen successfully generated novel molecules with high predicted binding activity.

Conclusions:

• SBMolGen effectively addresses limitations of previous models by incorporating target protein 3D structure.
• The generated molecules show promise for novel drug discovery with enhanced binding affinity.
• The tool provides 3D docking poses, facilitating subsequent structure-based drug design efforts.