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

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning for chemistry

Background:

• Generative models struggle with small datasets common in drug discovery.
• Scarcity of data for specific targets like CNS drugs and nucleic acid binders hinders development.

Purpose of the Study:

• Introduce Auxiliary Discriminator Sequence Generative Adversarial Networks (ADSeqGAN) for molecular generation.
• Improve quality and class specificity of generated molecules in data-scarce scenarios.

Main Methods:

• Integrated an auxiliary random forest classifier into a GAN framework.
• Utilized a pretrained generator and Wasserstein distance for stability and diversity.
• Evaluated ADSeqGAN on nucleic acid binders, CNS drugs, and CB1 ligands.

Main Results:

• ADSeqGAN demonstrated superior generation of nucleic acid binders compared to baseline models.
• Oversampling with ADSeqGAN improved CNS drug generation yields.
• Generated novel, druglike CB1 ligands with a high predicted active rate (32.8%).

Conclusions:

• ADSeqGAN is a versatile framework for molecular design in data-limited settings.
• The method shows significant promise for developing targeted therapeutics, including CNS drugs and specific ligand classes.