Masked graph modeling for molecule generation.

Area of Science:

• Computational chemistry
• Machine learning for drug discovery
• Molecular graph representation

Background:

• De novo molecular design is crucial for drug discovery and materials science.
• Existing computational methods face challenges in generating valid and novel molecules.
• Graph-based representations offer a powerful framework for molecular modeling.

Purpose of the Study:

• To introduce a novel masked graph model for in-silico molecular design.
• To evaluate the model's performance on established molecular distribution benchmarks.
• To demonstrate the model's ability to generate molecules with specific properties.

Main Methods:

• Developed a masked graph neural network model.
• Trained the model by iteratively masking and replacing graph components (nodes/atoms and edges/bonds).
• Evaluated using QM9 and ChEMBL datasets with the GuacaMol benchmark.

Main Results:

• The masked graph model effectively learns molecular distributions.
• Achieved a superior trade-off between molecule validity and novelty compared to existing models.
• Outperformed graph-based methods and competed with SMILES-based approaches on distributional metrics.

Conclusions:

• The masked graph model is a promising approach for de novo molecular generation.
• The model can generate molecules with desired properties while preserving physiochemical similarity.
• This work advances the field of AI-driven molecular design for various applications.