Novel molecule design with POWGAN, a policy-optimized Wasserstein generative adversarial network.

Area of Science:

• Artificial Intelligence in Chemistry
• Generative Adversarial Networks (GANs)
• Drug Discovery and Development

Background:

• Existing reinforcement-guided generative adversarial networks (GANs) face challenges in producing non-fragmented, property-oriented molecules at scale.
• A trade-off often exists between optimizing molecular properties and maintaining sample diversity in generative models.
• Current GANs struggle to balance molecule connectivity, novelty, and drug-likeness simultaneously.

Purpose of the Study:

• To introduce Policy-Optimised Wasserstein GAN (POWGAN), an adaptive reward-scaling strategy for improved molecule generation.
• To enhance the capabilities of generative models for producing valid, novel, and property-oriented molecules at scale.
• To overcome limitations in current GANs regarding molecular fragmentation and property optimization without compromising diversity.

Main Methods:

• Developed POWGAN, a graph-based generator incorporating dynamically scaled rewards into adversarial training.
• Integrated POWGAN into the MedGAN architecture, creating R-MedGAN, to target graph connectivity (non-fragmentation).
• Evaluated model performance on quinoline, indole, and imidazole scaffolds, assessing connectivity, novelty, uniqueness, and drug-likeness properties (SAS, LogP).

Main Results:

• R-MedGAN achieved 1.00 fully connected quinoline-like molecules, a significant improvement over the baseline (0.62), while maintaining high novelty (0.93) and uniqueness (0.95).
• The model generated over 12,000 novel quinoline-like molecules, populating previously unexplored regions of chemical space.
• Drug-likeness properties improved substantially: Synthetic Accessibility Scores (SAS) increased from 8% to 65%, and lipophilicity (LogP) from 17% to 45%.

Conclusions:

• POWGAN and R-MedGAN effectively address limitations in molecule-generating GANs by enabling simultaneous optimization of molecular topology and property.
• The adaptive reward-scaling strategy enhances generative throughput, preserves diversity, and improves drug-likeness, demonstrating generalizability across different molecular scaffolds.
• This work presents a robust, scalable platform for high-throughput, goal-directed chemical exploration, advancing the state-of-the-art in AI-driven drug discovery.