Prediction of melting point of indols based on using molecular structure artificial neural networks

By applying the method of quantitative structure property relationship, was determined a model to predict the temperature of the melting point of indoles from their molecular structure (n = 86). Using computer programs such as Gaussian 98 and PCDM, a series of electronic, topological and geometric descriptors were calculated. For the preparation of the prediction was used a multiple linear regressions model between molecular descriptors and the temperature of the melting points of indols present in the database. Giving as a result a coefficient of determination (R-2) and standard error of estimate (EEA) of 0.73 and 27.42 degrees C respectively. Through a neural network backpropagation (5: 4: 1) model of multiple linear regression was optimized, including possible non-linear relationships between molecular structure and temperature of the melting point of the indols, obtaining better results in predicting melting point for the training group (R-2 = 0.9978) and the validation group (R-2 = 0.9987). The mean square error (MSE) associated with the training group and validation for the network model was 0.006 and 0.006 respectively.