DDMG: A New Discrete Diffusion Model Developed for Molecular Graph Generation

Recent progress in diffusion generative models have reshaped the paradigm in the field of generation. Inspired by these advancements, the research community has extended the application of diffusion frameworks to molecular design. Nevertheless, current diffusion based molecular generation approaches suffer from an incomplete utilization of the structural features inherent in molecules. To surmount these obstacles, this paper propose DDMG (Discrete Diffusion for Molecular Generation), a novel molecular graph diffusion model featuring three pivotal innovations. First, this paper have engineered a mask-and-replace noise strategy. This strategy can safeguard molecular semantics during the noise injection process, thereby significantly boosting the quality of generated molecules. Second, a novel graph transformer is employed in which graph nodes are initialized with random noise. This design overcomes the limitations in expressiveness imposed by traditional graph neural networks. Third, a dynamic graph padding method is implemented, which allows for conditional generation and targeted molecular optimization by effectively handling variablesized graphs. Quantitative validation spanning multiple generative chemistry benchmarks reveals that DDMG attains state-of-the-art performance across validity, novelty, and diversity metrics, holding great promise for future applications in molecular design.