Novel Molecules Generation Using Graph Generative Adversarial Networks.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence in drug discovery

Background:

• De novo molecular generation is crucial but challenging in drug discovery.
• Existing graph-based models face issues like mode collapse and limited control.
• Need for improved generative models for valid, diverse, and property-aware molecules.

Purpose of the Study:

• To introduce NovMolG-GAN, a novel graph-based generative adversarial framework.
• To address limitations of existing generative models in molecular design.
• To enable stable training for generating valid, novel, diverse, and drug-like molecules.

Main Methods:

• Utilized a graph-based generative adversarial framework (NovMolG-GAN).
• Integrated graph attention mechanisms, a learnable reward network, and mode-seeking regularization.
• Employed proximal policy optimization-based reinforcement learning for stable training.

Main Results:

• Achieved high molecular validity (99.6%), novelty (99.4%), and uniqueness (88.5%) on ChEMBL-35.
• Demonstrated competitive quantitative drug-likeness scores.
• Showcased controllability in conditional generation, steering towards synthesis-feasible molecules.

Conclusions:

• NovMolG-GAN offers a stable and effective approach for de novo molecular generation.
• The framework facilitates goal-directed molecular design with enhanced control.
• Results highlight NovMolG-GAN's potential as a flexible foundation for drug discovery.