Neural Network Prediction Model of Cocrystal Melting Temperature Based on Molecular Descriptors and Graphs

Abstract

Among the physical properties characterizing cocrystals, melting temperature is one of the primary properties. Its prediction has been done by researchers, but in the known prediction models, there are some limitations. In order to adapt to the requirements of data development further and improve the quality of prediction, two prediction models of cocrystal melting temperature, the artificial neural network (ANN(I)) and the message passing neural network (MPNN), are proposed in this paper with molecular descriptors and molecular graphs as the inputs. As for ANN(I), molecular descriptors are applied to the prediction of cocrystal melting temperature for the first time, and the importance of features is evaluated by SHapley Additive explanation. For dataset 1 established in this paper, the prediction accuracy of ANN(I) is 95.75% on the training set and 95.67% on the test set. The feature evaluation of ANN(I) reveals that GATS2v, AMR, BCUTc-1h, MLFER_BH, BCUTp-1l, and ATSC1c play major roles in predicting the melting temperature. The ANN(I) model based on molecular descriptors could explain the relationship between the molecular structure and melting temperature to some extent, but its prediction accuracy is relatively poor. Thus, the message passing neural network (MPNN), a deep learning model based on molecular graphs, is proposed. For the test sets of two datasets used in this paper, the prediction accuracy of MPNN is 99.84 and 99.63%, respectively. Also, in the data pre-processing, the MPNN extracts the structural features of molecules through molecular graphs, which could save time and workload a lot. The excellent performance suggests that the MPNN model might be a more suitable tool for the prediction of cocrystal properties in the future.