Neural network modeling of physical properties of chemical compounds

Three different models relating structural descriptors to normal boiling po ints, melting points, and refractive indexes of organic compounds have been developed using artificial neural networks. A newly elaborated set of mole cular descriptors was evaluated to determine their utility in quantitative structure-property relationship (QSPR) studies. Applying two data sets cont aining 190 amines and 393 amides, neural networks were trained to predict p hysical properties with close to experimental accuracy, using the conjugate d gradient algorithm. Obtained results have shown a high predictive ability of learned neural networks models. The fit error for the predicted propert ies values compared to experimental data is relatively small. (C) 2001 John Wiley & Sons, Inc.