Boiling Point and Critical Temperature of a Heterogeneous Data Set:  QSAR with Atom Type Electrotopological State Indices Using Artificial Neural Networks

Abstract

Two sets of heterogeneous organic compounds were analyzed with artificial neural networks using atom type electrotopological state indices. The first set contains the boiling point for 298 compounds; 30 were placed in a testing set. The neural network model used atom type E-state indices for the 19 atom types present in the data set; the actual network used for prediction had a 19:5:1 architecture. This model produced a mean absolute error (MAE) of 3.93 K for the overall set, 3.86 for the training set, and 4.57 for the test set. The average relative percent error for 10 runs is 0.94% for the whole data set and 1.12% for the test set. The second set contains critical temperatures for 165 compounds; 18 were placed in the testing set. The neural network possessed a 19:4:1 architecture and produced an MAE of 4.52 K for the whole set, 4.39 K for the training set, and 5.59 K for the test set. The average relative percent error for 5 runs is 0.77% for the whole data set and 0.95% for the test set.