Molecular substructure tree generative model for de novo drug design

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Deep learning accelerates drug discovery by analyzing molecular and protein features.
• Generating novel molecules with desired properties is challenging due to chemical constraints.
• Existing methods may not fully capture the complexity of molecular structures.

Purpose of the Study:

• To develop a deep learning model for generating valid and novel molecules.
• To create a molecular optimization model to enhance chemical properties.
• To address the challenges of chemical rule adherence in de novo molecule generation.

Main Methods:

• Proposed a Molecular Substructure Tree Generative Model based on Variational Auto-Encoder (VAE).
• Employed an autoregressive generative model as a decoder for molecule generation from a latent space.
• Developed a molecular optimization model utilizing the CycleGAN architecture.

Main Results:

• The generative model successfully produced valid and novel molecular structures.
• The optimization model demonstrated significant improvements in molecular properties.
• Experiments confirmed the model's ability to navigate chemical rules during generation.

Conclusions:

• The Molecular Substructure Tree Generative Model offers a robust approach for de novo molecule design.
• Deep learning, particularly VAEs and GANs, can effectively generate and optimize molecules for drug discovery.
• This work contributes to accelerating the identification of new drug candidates with tailored properties.