Learning to Generate and Predict New Conditional Small Organic Molecules

Organic molecules are essential in various fields, including drug discovery, pharmaceuticals, materials science, and agrochemicals. However, synthesizing new organic molecules in laboratories is both costly and time-intensive. This paper introduces a novel approach for generating and predicting new conditional small organic molecules. The approach utilizes a Generative Adversarial Network (GAN), specifically a Wasserstein GAN (WGAN), to generate new molecular structures, and a Vision Transformer (ViT) to classify these generated molecules. To ensure the organic compounds generated are chemically feasible, the WGAN is conditioned with specific parameters, such as bond types and elemental composition. A generated dataset, named GenOrgo, includes 400 new molecules categorized into four classes: Homocycles, Heterocycles, Oxygen- and Nitrogen-containing Acyclics, and Hydrocarbons. The ViT is then evaluated on the GenOrgo dataset, demonstrating its effectiveness in accurately predicting the categories of generated organic molecules. The code and the dataset are available at github.com/RaghadAllehyani/WGAN.