Fig. 4: Model evaluation comparison and visualization of the backbone model extraction effect

3.2 Case studies for reaction evaluation

Here, three types of reaction examples are used to illustrate the effectiveness of RSscore. (Fig. 5 ) First, the synthetic reactions of the rivaroxaban intermediate (4-phenylmorpholin-3-one) is employed as an example to evaluate the RSscore for reactions with same class. (See Fig. 5(a) , Fig. 5(b) ) The substitution reactions depicted in Fig. 5(a) and 5(b) allow for a straightforward analysis of reaction superiority through the leaving group effect. Except for the same leaving part, the basicity of the chloride leaving group in Fig. 5(a) is lower than the basicity of the ethyl ester group in Fig. 5(b) . It indicates that the corresponding anion of the Fig. 5(a) reaction leaving group is more stable, leading to milder reaction conditions in the Fig. 5(a) reaction. According to the experiment data from the Reaxys database, the Fig. 5(a) reaction demonstrates the same reaction temperature, a higher reaction yield and a shorter reaction time compared to the Fig. 5(b) reaction. It validates the leaving group effect and indicates that the Fig. 5(a) reaction exhibits higher superiority. In accordance of the evaluation metric proposed in this paper, the RSscore achieved 0.7939 in Fig. 5(a) and 0.7651 in Fig. 5(b) . It indicates that the Fig. 5(a) reaction exhibits a higher reaction superiority, aligning with the results obtained from the experimental and mechanistic analyses. This demonstrates that the RSscore can detect the impact of electronic effects, and effectively distinguishes the superiority differences among reactions within the same class.
Synthetic reactions of Pomalidomide intermediate (4-nitrothalidomide) using different reaction agents are conducted to determine whether the RSscore can detect the impact of the reaction agent. The experimental data shows that the reaction usingFig. 5(c) agents exhibit lower reaction time, higher reaction yield and milder reaction temperature condition compared to the reaction using Fig. 5(d) agents. After completing the reaction atom mapping, the RSscore for reactions using Fig. 5(c) andFig. 5(d) agents are calculated as 0.7518 and 0.4812, respectively. This demonstrates that the RSscore are consistent with experimental data and provide a better reflection of the reaction agents influence on chemical reactions.
Finally, as an exploration of the assessment effect of the RSscore in different reactions, the synthesis reactions Olaparib intermediate (4-(4-fluoro-3-(piperazine-1-carbonyl)benzyl)phthalazin-1(2H)-one) are used as an example for comparison. Although the final product and the reacted sites for the reactions in Fig. 5(e) and Fig.5(f) are the same, the reaction type are completely different, with Fig. 5(e) being a hydrolysis reaction and Fig. 5(f) being a condensation reaction. Based on the comparison of the experimental data between the reactions in Fig. 5(e) andFig. 5(f) , the reaction in Fig. 5(e) is superior in terms of reaction time, yield, and temperature condition. The calculated RSscore for Fig. 5(e) is 0.7836 and for Fig. 5(f) is 0.6399. The magnitude relationship of the RSscore supports the superiority of the reaction in Fig. 5(e) , which aligns with the experimental data.