General melting point prediction based on a diverse compound data set and artificial neural networks

We report the development of a robust and general model for the prediction of melting points. It is based on a diverse data set of 4173 compounds and employs a large number of 2D and 3D descriptors to capture molecular physicochemical and other graph-based properties. Dimensionality reduction is performed by principal component analysis, while a fully connected feed-forward back-propagation artificial neural network is employed for model generation. The melting point is a fundamental physicochemical property of a molecule that is controlled by both single-molecule properties and intermolecular interactions due to packing in the solid state. Thus, it is difficult to predict, and previously only melting point models for clearly defined and smaller compound sets have been developed. Here we derive the first general model that covers a comparatively large and relevant part of organic chemical space. The final model is based on 2D descriptors, which are found to contain more relevant information than the 3D descriptors calculated. Internal random validation of the model achieves a correlation coefficient of R(2) = 0.661 with an average absolute error of 37.6 degrees C. The model is internally consistent with a correlation coefficient of the test set of Q(2) = 0.658 (average absolute error 38.2 degrees C) and a correlation coefficient of the internal validation set of Q(2) = 0.645 (average absolute error 39.8 degrees C). Additional validation was performed on an external drug data set consisting of 277 compounds. On this external data set a correlation coefficient of Q(2) = 0.662 (average absolute error 32.6 degrees C) was achieved, showing ability of the model to generalize. Compared to an earlier model for the prediction of melting points of druglike compounds our model exhibits slightly improved performance, despite the much larger chemical space covered. The remaining model error is due to molecular properties that are not captured using single-molecule based descriptors, namely both inter- and intramolecular interactions and crystal packing, for which examples of and reasons for outliers are given.