To solve the problem of multiple factors limiting the prediction of organic reactions in drug synthesis, machine learning is used to achieve more accurate and efficient reaction prediction and catalyst screening. Firstly, using computers as tools and combining multidisciplinary knowledge, theoretical simulations and other methods are used to guide and assist in the design, discovery, and synthesis of new drug molecules. These methods are mainly divided into three categories based on drug small molecule structure, receptor, and molecular dynamics. Protein structure determination and docking methods are also introduced. Secondly, an improved graph convolutional network model is proposed to address the existing model issues. Firstly, the reactants are transformed into feature matrices, and the improved graph convolutional network and attention mechanism are used to predict candidate reaction centers. Then, candidate products are generated by enumerating chemical constraints. Finally, the improved graph convolutional differential network is used to evaluate and rank the candidate products. The experimental results show that the accuracy of drug synthesis prediction: on the USPTO test set, the proposed method model outperforms the WLDN model in any number of template matches, and the advantage is more pronounced when the number of template matches is low. In terms of the accuracy index of reaction product prediction, the method proposed in this paper has the smallest parameter scale, but it is significantly better than other models in Top-1, Top-2, Top-3, and Top-5 indicators. In terms of protein mechanism analysis, compounds 1 and 2 form multiple hydrogen bonds with the target, and compound 2 has a higher binding energy than compound 1, but the ICso also increases, indicating that changing the substituents on the branched chain connected to the benzene ring in the structure of fluoroquinolone drugs can affect drug activity. The above results indicate that computer-aided drug synthesis design and prediction methods based on graph convolutional neural networks are effective, and the improved graph convolutional network model performs well in drug synthesis prediction, providing new ideas and methods for drug synthesis reaction prediction.
