A Quantum Mechanical/Neural Net Model for Boiling Points with Error Estimation

Abstract

We present QSPR models for normal boiling points employing a neural network approach and descriptors calculated using semiempirical MO theory (AM1 and PM3). These models are based on a data set of 6000 compounds with widely varying functionality and should therefore be applicable to a diverse range of systems. We include cross-validation by simultaneously training 10 different networks, each with different training and test sets. The predicted boiling point is given by the mean of the 10 results, and the individual error of each compound is related to the standard deviation of these predictions. For our best model we find that the standard deviation of the training error is 16.5 K for 6000 compounds and the correlation coefficient (R2) between our prediction and experiment is 0.96. We also examine the effect of different conformations and tautomerism on our calculated results. Large deviations between our predictions and experiment can generally be explained by experimental errors or problems with the semiempirical methods.

Toward better QSAR/QSPR modeling: simultaneous outlier detection and variable selection using distribution of model features

“…A good QSAR model not only has good summary statistics such as RMSECV and Q 2 , but also yields high stability of predictions for each sample. Several studies have indicated that the high stability of predictions of models correlates with the accuracy of predictions [63][64][65][66][67]. Thus, the standard deviation of prediction errors for each sample can be used as an additional metric characterizing the Fig.…”

QSAR and QSPR based solely on surface properties?

“…Murray and Politzer [7] introduced a series of descriptors based on the statistics of the MEPvalues at the triangulation points on a calculated molecular surface as powerful descriptors for physical properties. We [5,[8][9][10][11][12][13][14][15] have used slight modifications of these descriptors in a series of QSPR models. Thus, Coulomb interactions have been treated systematically using surface-based descriptors.…”

“…This approach is generally justified because the quality of the experimental data does not allow resolution of conformational effects and because even quite significant conformational changes involving intramolecular hydrogen bonds only change, for instance, the predicted boiling point by about as much as the uncertainty in the predicted value. [5] However, even 3D-descriptors usually have a strong element of the molecular topology (atom or group counts etc.). These elements are often regarded as being essential because they allow chemists to interpret the results in terms of the chemical structure and modifications that may improve the activity or a given physical property.…”

Artificial Neural Network‐Based Quantitative Structural Property Relationship for Predicting Boiling Points of Refrigerants

“…The absolute mean errors of the 8-12-1 feed-forward model were 27.7 K for the training set (n ¼ 1168) and 20.8 K for the test set (n ¼ 153). The corresponding errors in case of the Fuzzy ARTMAP model were 2.0 and 13.5 K. The most general boiling point model has been developed by Clark and his co-workers [18]. They used a data set of 6629 compounds with very diverse functionality, containing elements H, B, C, N, O, F, Al, Si, P, S, Cl, Zn, Ge, Br, Sn, I, and Hg.…”