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 v
arying functionality and should therefore be applicable to a diverse range
of systems. We include cross-validation by simultaneously training 10 diffe
rent networks, each with different training and test sets. The predicted bo
iling point is given by the mean of the 10 results, and the individual erro
r of each compound is related to the standard deviation of these prediction
s. For our best model we find that the standard deviation of the training e
rror is 16.5 K for 6000 compounds and the correlation coefficient (R-2) bet
ween our prediction and experiment is 0.96. We also examine the effect of d
ifferent conformations and tautomerism on our calculated results. Large dev
iations between our predictions and experiment can generally be explained b
y experimental errors or problems with the semiempirical methods.