Extended atom-based and bond-based group contribution descriptor and its application to melting point prediction of energetic compounds

17817 compounds were collected from the Bradley open melting point data set, including eight elements: C, H, O, N, F, S, Cl, Br, and I. An extended atom-based and bond-based group contribution descriptor was suggested to represent these compounds, which consists of a one-dimensional descriptor based on the Molecular formula, a two-dimensional group contribution descriptor based on atoms and bonds, and a structural feature descriptor. Random forest (RF), Partial Least Squares (PLS), and Deep Learning (DL) methods were used to establish models to predict melting points, and the constructed models were evaluated by correlation coefficient (R), mean absolute error (MAE) and root-mean-square error (RMSE). Among them, the best results were obtained using the model constructed by Random forest: the results of out-of-bag (OOB) cross-validation of the training set are R = 0.8977/MAE = 29.57 °C/RMSE = 40.34 °C; the predicted results of the test set are R = 0.8982/MAE = 29.68 °C/RMSE = 40.63 °C. Compared with the results obtained using the subset of this data set in a literature, the results in this study are better than the corresponding results in the literature. The established model was also used to predict an external data set consisting of 74 compounds retrieved from another literature, and the obtained results are R = 0.8946 °C/MAE = 24.51 °C/RMSE = 34.19 °C, which were significantly better than the corresponding results in the literature. If the descriptor suggested in this study is combined with RDKit descriptor that contains charge and Electronegativity information and so on, better results were achieved: the results of OOB cross-validation of the training set are R = 0.9013/MAE = 29.25 °C/RMSE = 39.76 °C; the results of the test set are R = 0.9017/MAE = 29.34 °C/RMSE = 40.07 °C.