Graph-Based Generative Adversarial Networks for Molecular Generation with Noise Diffusion

One of the crucial objectives in modern drug design is the generation of high-quality novel molecules. Currently, numerous deep learning methods have been applied in the field of molecular generation. Modeling molecules as graph-structured data provides a unique representation that offers richer structural information than sequential data. However, unlike data such as images, conventional embedding methods struggle to capture the topological structure of graphs and distinguish between different types of nodes and edges, leading to a loss of molecular structural information. Additionally, traditional Generative Adversarial Networks(GANs) often concentrate on fitting training data, leading to a concentrated distribution that limits the diversity of generated molecules. In this paper, we propose a Diffusion-GAN framework based on molecular graphs to address these challenges. The graph autoencoder effectively embeds molecular graph data, preserving structural information. The Diffusion-GAN module progressively introduces noise through a forward diffusion chain to broaden the sampling distribution, thereby enhancing the diversity of generated samples. We conducted molecular graph generation tasks on the QM9, ZINC and MOSES datasets to demonstrate the effectiveness of this method. Our method exhibits higher validity and diversity than several classical molecular generation algorithms.