MedGAN: optimized generative adversarial network with graph convolutional networks for novel molecule design.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence

Background:

• The demand for novel medicines necessitates advanced drug discovery tools.
• Generative AI offers a promising approach to accelerate the identification of new molecular structures.

Purpose of the Study:

• To optimize and fine-tune MedGAN, a deep learning model for generating novel quinoline-scaffold molecules.
• To evaluate drug-likeness properties of generated molecules, including pharmacokinetics, toxicity, and synthetic accessibility.

Main Methods:

• Utilized Wasserstein Generative Adversarial Networks and Graph Convolutional Networks for molecular generation.
• Performed hyperparameter adjustments and rigorous evaluations of generated molecules.
• Focused on generating molecules with specific scaffolds, such as quinolines.

Main Results:

• The optimized MedGAN model generated 25% valid molecules, with 92% being quinolines.
• Generated molecules exhibited high novelty (93%) and uniqueness (95%), preserving chirality and atom charge.
• Successfully generated 4831 novel quinoline molecules with favorable drug-like properties.

Conclusions:

• Deep learning models like MedGAN can effectively generate novel, drug-like molecules.
• Model performance is influenced by factors such as activation functions, optimizers, and molecular structure.
• This work enhances the application of AI in computational drug design and discovery.