Using GANs with adaptive training data to search for new molecules.

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Drug Discovery
• Machine Learning for Molecular Design

Background:

• Drug discovery requires extensive exploration of vast chemical spaces.
• Generative Adversarial Networks (GANs) are powerful tools for exploring chemical space and optimizing compounds.
• Standard GAN training can lead to mode collapse, limiting exploration of novel chemical entities.

Purpose of the Study:

• To develop a GAN training approach that promotes incremental exploration and mitigates mode collapse.
• To enhance the discovery of novel chemical compounds for drug development.
• To improve the applicability of GANs in the field of pharmaceutical research.

Main Methods:

• Integration of Genetic Algorithm concepts into GAN training.
• Utilizing valid generated samples to replace existing training data.
• Implementing both random and guided selection strategies, including recombination, for data replacement.

Main Results:

• The proposed GAN training method significantly enhances the generation of novel compounds compared to traditional approaches.
• Updates to the training dataset through sample replacement effectively limit mode collapse.
• Demonstrated a substantial increase in the number of unique compounds produced during training.

Conclusions:

• The novel GAN training strategy effectively overcomes mode collapse, a common limitation in generative models.
• This approach substantially improves the exploration of chemical space, leading to the discovery of more diverse and novel compounds.
• The findings highlight the potential of this enhanced GAN methodology to accelerate and broaden drug discovery efforts.