Prediction of boiling points and critical temperatures of industrially important organic compounds from molecular structure

Neural Network Based Quantitative Structural Property Relations (QSPRs) for Predicting Boiling Points of Aliphatic Hydrocarbons

“…In this latter case, the resulting classification was weighted by nearly 80% of the classification weight attributable to the number of carbon atoms and about 20% to the position of double bonds. Boiling point predictions from the Fuzzy ARTMAP and back-propagation composite models were also compared with the models of Hall and Story 22 and Egolf et al, 21 for 49 hydrocarbons (alkanes and alkenes) common to these studies, revealing average absolute errors of 0.24% (0.85 K), 1.8% (5.6 K), 1.7% (4.9 K), and 2.6% (7.9 K), respectively, and maximum absolute errors of 0.91% (3.4 K), 7.1% (18.8 K), 12% (27.1 K), and 14% (31.5 K), respectively (Table 8 and Figure 8). Although the Hall and Story 22 model had a slightly lower average absolute boiling point error than the present back-propagation composite model, it did not differentiate between cis and trans isomers of alkenes (i.e., equal boiling points were predicted for such isomers).…”

“…25,26 As the literature reveals, a major challenge in neural network/QSPR development has been to establish a reliable and practical set of molecular descriptors. 8,11,21,22 As a consequence, most recent studies have explored the development of QSPRs for commonly available physicochemical parameters (e.g., boiling point, heat capacity, density, refractive index) for selected organic compound classes for which accurate and rich data sets are available. 8,17,[19][20][21][22] The use of boiling point data to test the applicability of various molecular descriptors has been particularly popular given the availability of data for large sets of organic chemical classes.…”

“…QSPRs for boiling points have been proposed by a number of investigators based on backpropagation neural networks. 8,17,[20][21][22] For example, Gakh et al 19 developed a composite neural network/QSPR boiling points model, based on 134 noncyclic alkanes ranging from 6 to 10 carbon atoms. Their model, which included five additional physicochemical properties, was based on six Weiner-type structural graph invariants representing the number of pathways for carbon lengths ranging from three to eight atoms and also included the number of carbon atoms.…”

“…A neural network-QSPR, for boiling points of a diverse class of organics, was developed by Egolf et al 21 using a data set of 298 organic compounds that included about 45 alkanes and 50 alkenes as well as halogens, alcohols, esters, ketones, carboxcyclic acids, and amines. The approach identified the following descriptors: three charged partial surface area structural (CPSA) descriptors, 32 a composite CPSA descriptor, number of oxygens, Wiener number, 4 the summation of all unique paths, or molecular ID index, 33 and charge on carbonyl or cyano carbons.…”

The Present Utility and Future Potential for Medicinal Chemistry of QSAR / QSPR with Whole Molecule Descriptors

“…Grigoras [132] furans, tetrahydrofurans the authors concluded that due to structural differences between nitrogen heterocycles and sulfur and oxygen heterocycles, various connectivity, electronic, constitutional and CPSA descriptors cannot adequately encode enough information for a combined set of heterocycles Stanton et al [134] furans, tetrahydrofurans, thiophenes, pyrans 299 MLR, NN both methods had the same quality of prediction for the training set Egolf and Jurs [135], Egolf et al [136] pyridines 572 for pyridines, in the case of the cross-validation set, the NNs outperformed conventional QSPR; descriptors that reflect hydrogen bonding and dipoledipole interactions improved the predictive models for the pyridines data set diverse organic compounds 298 for this set the back-propagation NN combination resulted in 1K improvement over the MLR alkanes 150 NN 10:7:1 architecture; the performance was slightly better in comparison with the MLR methods…”