Fragment-based deep molecular generation using hierarchical chemical graph representation and multi-resolution graph...

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Graph neural networks

Background:

• Graph generative models are increasingly used for molecular structure construction.
• Fragment-based strategies offer a promising approach for de novo molecular design.
• Existing methods may benefit from enhanced structural diversity and novel moiety generation.

Purpose of the Study:

• To develop a novel graph generative model for molecular structures.
• To enhance the generation of diverse and novel chemical moieties.
• To improve molecular generation by incorporating fragment clustering.

Main Methods:

• Decomposition of molecules into chemical fragments.
• Grouping fragments into clusters based on local structural environments.
• Implementation of a multi-resolution graph variational autoencoder (MRGVAE).
• Hierarchical learning and decoding of molecular structures in a fine-to-coarse manner.

Main Results:

• The MRGVAE model demonstrated competitive performance in molecular evaluation metrics.
• The fragment cluster layer improved the generation of novel chemical moieties.
• Enhanced structural diversity was observed in the generated molecules.
• The model shows potential for creating chemical structures absent in the training set.

Conclusions:

• The proposed multi-resolution graph approach effectively generates diverse molecular structures.
• Incorporating fragment clusters enhances the novelty and diversity of generated molecules.
• This framework provides a foundation for integrating chemical domain knowledge into generative models for drug discovery.