Abstract

Glass transition temperature (T_g) is an important inherent property of an amorphous solid material which is usually determined experimentally. In this study, the relation between T_g and melting temperature (T_m) was evaluated using a data set of 71 structurally diverse druglike compounds. Further, in silico models for prediction of T_g were developed based on calculated molecular descriptors and linear (multilinear regression, partial least-squares, principal component regression) and nonlinear (neural network, support vector regression) modeling techniques. The models based on T_m predicted T_g with an RMSE of 19.5 K for the test set. Among the five computational models developed herein the support vector regression gave the best result with RMSE of 18.7 K for the test set using only four chemical descriptors. Hence, two different models that predict T_g of drug-like molecules with high accuracy were developed. If T_m is available, a simple linear regression can be used to predict T_g. However, the results also suggest that support vector regression and calculated molecular descriptors can predict T_g with equal accuracy, already before compound synthesis.