5. Conclusions

In this paper, we developed a novel chemical reaction superiority evaluation index based on the information from the reaction database using deep learning (DL) methods. For the modeling of RSscore, a new type of chemical reaction graph descriptor with atom mapping relationship was constructed, which provides sufficient chemical reaction information and significantly improved the model classification ability. Label smoothing was also employed to reduce the overfitting of the model and enable differentiation of reaction superiority within the same categories. Contrastive learning pre-training and supervised learning fine-tuning method were used to improve the generality of the model and the accuracy of classification.
The effect of different message passing methods was investigated and the AIR residual on the GNN model is used to generate the RSscore. It proves that the GINE message passing methods combined with AIR residual demonstrate the best outcomes on classification. Additionally, the data distribution of the entire dataset is analyzed and visualized using UMAP dimensionality reduction, which showed that the developed model effectively distinguishes reaction superiority and generates a robust evaluation metric. The effectiveness of this RSscore provides a crucial evaluation index for the computer-aided synthesis planning.