基于动态推理方法的小分子生成OA

Small molecule generation plays an increasingly vital role in scientific fields such as drug design and new materials development.In recent years,diffusion models have been widely employed for molecular structure generation tasks owing to their exceptional generative capabilities.However,existing diffusion-based methods often treat molecular properties as static conditions,failing to adequately capture the dynamic relationship between structure and properties,thus hindering the precise control of target properties.To ad-dress this challenge,we propose a dynamic inference-based conditional molecular generation model that inte-grates an edge-guided discrete diffusion mechanism with a molecular property prediction module.This ap-proach facilitates the joint modeling and dynamic co-optimization of molecular graph structures and their corre-sponding properties during the generation process.Specifically,we employed a graph neural network to esti-mate the properties of intermediate molecular graphs at each diffusion step.During the denoising phase,these estimations were jointly incorporated into the loss function along with the target properties,thereby enhancing the model′s ability to enforce structure-property consistency.Experimental results demonstrate that our method significantly outperforms existing baselines across various property control tasks.In the targeted prop-erty control setting,the proposed method reduced the Mean Absolute Error(MAE)of HOMO and μ by 21%and 18%,respectively.Furthermore,in the random conditional generation tasks,the MAE for these properties were reduced by approximately 33.3%and 31.7%,respectively.These results validate the effec-tiveness of the dynamic inference mechanism in improving property controllability and the quality of the gener-ated molecules.