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

• DOI: 10.1145/3640900.3640911
• Publication Date: 2024-03-01T17:04:59Z
• Authors: Jianyang Wang; Bin Ju; Xiaoliang Qian; Minchao Ye; Wanli Huo
• Keywords: Autoencoder; Molecular graph

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