Auxiliary Discrminator Sequence Generative Adversarial Networks (ADSeqGAN) for Few Sample Molecule Generation.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence in Chemistry

Background:

• Generative models face challenges with limited data in drug discovery.
• Scarcity of molecular datasets for specific targets hinders development (e.g., nucleic acid binders, CNS drugs).

Purpose of the Study:

• Introduce Auxiliary Discriminator Sequence Generative Adversarial Networks (ADSeqGAN) for molecular generation.
• Enhance molecular generation quality and class specificity in small-sample datasets.
• Provide a versatile framework for molecular design in data-scarce environments.

Main Methods:

• ADSeqGAN integrates an auxiliary random forest classifier into the GAN framework.
• Utilizes a pretrained generator and Wasserstein distance for stability and diversity.
• Evaluated across nucleic acid binders, CNS drugs, and CB1 ligand design.

Main Results:

• ADSeqGAN demonstrates superior generation of nucleic acid binders compared to baselines.
• Oversampling with ADSeqGAN improves CNS drug generation yields.
• Generated novel, druglike CB1 ligands with 32.8% predicted actives exceeding library hit rates.

Conclusions:

• ADSeqGAN is an effective framework for molecular design in data-scarce settings.
• The method shows significant improvements in generating specific and high-quality molecules.
• Demonstrated success in applications including nucleic acid binders, CNS drugs, and CB1 ligands.