An improved QSPR study of standard formation enthalpies of acyclic alkanes based on artificial neural networks and genetic algorithm

A QSPR model based on artificial neural networks (ANN) was developed to study the standard formation enthalpies of 85 kinds of acyclic alkanes. The ANN was trained applying quick error back-propagation (BP) algorithm. Meanwhile, twenty-five well-known topological indices were used as structural descriptors for each alkane molecule, and they were also considered to be the potential input variables for the proposed ANN-QSPR model. Optimization of an input variable representation to the ANN-QSPR model was carried out via genetic algorithm (GA). Then, the final optimized structure representation of all the alkanes contains only 17 variables. The input variable selection strategy based on GA improved the prediction results both for training and test samples. Moreover, a novel QSPR approach based on the combination of GA and ANN to improve the prediction results in test set was also proposed, which is achieved by optimizing initial learning rate, learning momentum, the number of hidden neurons and in fact the randomly-generated values of starting weights in ANN according to GA. In the novel QSPR model, the genetic input variable selection strategy can also improve the prediction results of ANN considerably.