We propose a novel approach for molecule generation from small datasets, the Auxiliary Discriminant Sequence Generative Adversarial Network (ADSeqGAN). Existing generative models struggle with limited training data, particularly in the field of drug discovery, where molecular datasets for specific therapeutic targets, such as nucleic acid binders or central nervous system (CNS) drugs, are scarce. ADSeqGAN significantly improves the quality and class specificity of molecule generation by incorporating a random forest classifier as an additional discriminator into the GAN framework. This study further enhances training stability and diversity by incorporating a pretrained generator and the Wasserstein distance. We evaluated ADSeqGAN on three use cases: nucleic acid and protein target molecules, CNS drugs, and CB1 ligand design. It outperformed the baseline model in the generation of nucleic acid binders, and achieved higher yields than existing novel drug design models in the generation of CNS drugs through oversampling. In the CB1 ligand design, the predicted activity was 32.8%, as evaluated by the target-specific LRIP-SF score function, generating novel drug-like molecules that outperformed both CB1-focused and general-purpose libraries. Overall, ADSeqGAN provides a versatile framework for molecular design in data-poor scenarios, demonstrating applications in nucleic acid binding agents, central nervous system drugs, and CB1 ligands.
