Switch on amine substrate reactivity towards hexaazaisowurtzitane cage: Insights from a tailored machine learning model

The efficient synthesis of novel hexaazaisowurtzitane cage compounds has remained a formidable challenge for years due to the complicated reaction mechanism and the uncertainty of amine substrate selection. Here, we developed a tailored machine learning model to predict the reactivity of amine substrates towards hexaazaisowurtzitane cage based on high-throughput quantum mechanical calculations of 3428 property parameters of 118 amine substrates. The customized model was developed through an appropriately weighted fusion of advanced universal models, achieving comprehensive predictive capability with an accuracy of 91.4 %, an F1 score of 89.1 %, and a recall of 91.4 %. Further, the customized model exhibits a narrow interquartile range of accuracy, surpassing universal models by 30.6–54.4 % and demonstrating robustness across various data splits. The data-driven analysis identified that electronic and geometric features are the dominant regulating factors of amine’s reactivity. Further, physics-driven insights revealed that a low electron-density environment near the nitrogen in the amine group is a key for switching on the reactivity of the amine substrates, which can be characterized by a sufficiently high NMR signal around 225.7 ppm with a narrow fluctuation of 2.6 ppm. Based on the revealed guiding factors and regulating mechanism, we selected 27 commercially available amine substrates for reactivity assessment and recommended 5 candidates with a probability exceeding 90 % for synthesis trials. This work pioneers machine learning and high-throughput quantum mechanical computationally assisted prediction of substrate selection for the rational synthesis of hexaazaisowurtzitane cages.