Prediction of boiling points of organic heterocyclic compounds using regression and neural network techniques

High quality models which relate structural descriptors to normal boiling points have been developed for large, diverse groups of heterocyclic compounds using both linear regression and neural network techniques. Parallel experiments were designed to compare the performance of these complementary modeling techniques on two different data sets. A formerly studied data set comprised of 299 tetrahydrofuran (THF), thiophene, furan, and pyran compounds was reexamined using neural networks. In addition, a new data set of 572 pyridine compounds was investigated to increase our understanding of the nitrogen-containing heterocycles. First, several new descriptors were developed to explore chemical principles which govern the boiling point process. In particular, descriptors that reflect hydrogen bonding and dipole-dipole interactions proved especially useful for improving the predictive models in the pyridine regression work. With each data set, neural networks were trained to predict boiling points with close to experimental accuracy using the back-propagation learning algorithm. Results from these boiling point investigations show that once the key structural features are indentified through traditional regression techniques, neural networks generally provide access to superior predictive equations. On the basis of this information, further studies were initiated to explore using neural networks as a tool to upgrade structural feature selection. Results from this phase of the study demonstrate that this methodology can be used to identify the most informationally rich descriptors.