According to the comparison of the model evaluation results presented inFig. 4(a) , the backbone Model with GINE aggregation method and AIR residual shows the best performance among the several directly trained ones, where the accuracy, F1-score and ROC-AUC were 0.897, 0.898 and 0.951, respectively. It indicates that the backbone model, equipped with the GINE layers and the AIR residual, is effective in extracting more robust features for distinguishing reactions superiority. The testing set ROC curves of different models are shown in Fig. 4(b) , which indicate the generalization ability and prediction accuracy of the model pre-trained using contrastive learning method are significantly better than those of the direct training ones. The accuracy, F1-score and ROC-AUC of the model with contrastive learning pre-training are 0.903, 0.903 and 0.965, respectively, which shows that the pre-training via contrastive learning method can effectively improve the generalization ability of the model.

To visualize the effect of feature extraction by the pre-trained backbone model and the corresponding data space, the features extracted from the reactions are projected into 2 dimensions using Uniform Manifold Approximation and Projection (UMAP)⁴⁸ which is plotted in Fig. 4(c) . By analyzing the distributions of superior and inferior reaction data points, a clear distinction can be observed between the main distributions of superior and inferior reactions, indicating that the pre-trained with contrastive learning model performs better in distinguishing the reaction superiority and providing suitable advice for reaction selections.