Generating molecules with specific chemical properties for target proteins can accelerate the drug development process and open new avenues for developing treatments for diseases with known pathogenic target proteins. However, current approaches to generate molecules with desired properties face several challenges, including prolonged generation time, complexity in learning parameters, and unqualified chemical properties. To address these issues, we proposed a structure-aware diffusion model, termed KGMG. This method incorporated the protein pocket as a constraint and integrated cutting-edge technologies such as KNN (K-Nearest Neighbors), equivariant graph neural networks, and self-attention mechanism. The core concept of KGMG was based on the 3D point cloud representation of protein pocket and its bound molecule. First, KNN was employed to construct both local and global graphs for each atom, followed by the uses of equivariant graph neural networks to iteratively update the atomic features and coordinates. Next, a self-attention mechanism was applied to fuse the updated atomic features and coordinates, forming the forward propagation process of diffusion model. Finally, through a backward denoising process, the model progressively restored the data, generating new molecules for a specific target protein. KGMG exhibited superior performance across multiple evaluation metrics.
Zeng et al. (Wed,) studied this question.