Transfer learning-enabled density prediction model for energetic molecule screening

Identifying candidates with high crystal density using Quantitative Structure-Performance Relationship (QSPR) models is a primary task in screening energetic molecules. However, due to the scarcity of high-density training data, QSPR models are often inaccurate in the high-density region. Therefore, screening novel energetic molecules with existing QSPR models carries the risk of obtaining false candidates and missing true ones. We propose a transfer learning scheme to build QSPR models for the density of energetic molecules that present high accuracy for both the high-density and the full density ranges. Parameter transfer and multi-step instance transfer are employed to train the models, with an independent test set composed entirely of high-density and nitro-containing molecules used to evaluate their accuracy. Our findings indicate that tuning the packing coefficient distribution of the dataset is more effective than tuning the crystal density distribution for improving the accuracy and generalization ability of the models. The final target model is trained on a dataset that has undergone two rounds of packing coefficient distribution tuning, achieving MAE values of 0.029 g·cm -3 for the high-density region and 0.028 g·cm -3 for the full density range of nitro-containing molecules, respectively. The target model, coupled with existing prediction methods for other important properties, is applied in a virtual screening of novel energetic molecules. Sixteen candidates with high detonation energy and good chemical stability are identified, and two of them has been experimentally validated. The workflow of this project is as follows: (a) Data collection from the Cambridge Structural Database (CSD) and Reaxys, followed by data curation per predefined rules to establish a dataset for QSPR model training and molecular skeletons for virtual screening molecule generation; (b) Training density prediction models via D-MPNN; (c) Generating candidate molecules through exhaustive docking of organic fragments; (d) Molecule screening after labeling key properties. • A transfer learning scheme is proposed for addressing the imbalanced regression problem occurred in training the density prediction models of energetic molecules. • Tuning datasets according to packing coefficient distribution is more effective than crystal density distribution for improving the accuracy and generalization ability of the models. • The target model is utilized to accurately screen out 15 high-density candidates of energetic molecules, and 2 of them has already been confirmed experimentally.