An Improved Structure−Property Model for Predicting Melting-Point Temperatures

Abstract

Physical properties and thermodynamic data are essential inputs to all computer-aided molecular design applications. Basic properties, such as the melting-point temperature, are essential for developing custom chemicals with desired thermophysical behavior. Currently, accurate correlations for the melting-point temperature are limited, including recent attempts to use quantitative structure-property relationships (QSPR). The lack of a comprehensive melting-point model can be attributed to (a) the sensitivity of the melting point to subtle variations in molecular structure and (b) the inability of existing molecular descriptors to account satisfactorily for all factors that influence the melting-point behavior. In this work, we present a new QSPR model for predicting the melting-point temperature of a diverse organic dataset. The model benefits from the inclusion of novel nonlinear descriptors developed through the use of robust genetic algorithms (GAs) and neural networks. Three new descriptors were developed to account for the nonlinear effect of molecular weight, hydrogen bonding, and molecular symmetry on the melting-point temperature. The resultant QSPR model is capable of modeling melting-point behavior with a root-mean-square error (RMSE) of 12.6 K, an average absolute percent deviation (%AAD) of 4.7 in absolute temperature, and a correlation coefficient of R 2 ) 0.95. This work differs from other literature efforts in that (a) an extensive dataset of over 1250 molecules is used for model development and (b) descriptor selection is performed using nonlinear algorithms. The results suggest that often-ignored structural descriptors such as molecular weight and symmetry have major roles in determining the melting point.

Screening of Chemical Penetration Enhancers for Transdermal Drug Delivery Using Electrical Resistance of Skin

“…The multi-well resistance chambers were equipped to perform the experiments at conditions identical to permeation experiments. First, experiments were performed using CPEs reported in the literature (14) and then extended to seven new potential CPEs which were identified by a virtual design algorithm (15)(16)(17). Herein, we show a significant agreement exists between the resistance technique and the standard permeation experiments; thus, we confirm the efficacy of the resistance technique for screening potential CPEs.…”

Prediction of Melting Point for Drug-like Compounds Using Principal Component-Genetic Algorithm-Artificial Neural Network

“…6 The solubility of a compound can be regarded as a partitioning of the compound between its crystal lattice and the solvent. If the forces holding the molecule in the crystal are high, then the solubility will be low.…”

“…4,5 Thus, it would be helpful to be able to estimate the melting point of a compound from its chemical structure. 6,7 Prediction methods for melting point, mainly can be categorized as property-property relationship (PPR), group contribution, and quantitative structure-property relationship (QSPR). 8,9 Comprehensive reviews of the subject reveal that many studies involved hydrocarbons and homologous compounds.…”

Prediction of normal melting point of pure substances by a reference series method

“…1 Consequently, property prediction techniques are significantly less reliable when applied to solid properties compared to their reliability in predicting liquid and gas phase properties. [1][2][3][4][5][6] The most widely used methods for predicting T m are the "group contribution (GC)" methods. 7 Some of the GC methods have been already introduced into commercial software packages for predicting T m on a routine basis (e.g., the Dorthmund Data Bank, DDBST, 2011 release, http:// www.ddbst.de, and CRANIUM, Molecular Knowledge Systems, http://www.molecularknowledge.com/).…”

“…For example, Marrero and Gani 11 introduced a three level GC method, Godavarthy et al 2 developed a quantitative structure-property relationship (QSPR), which uses 16 molecular descriptors in a nonlinear model whose parameters were determined using a neural network, and Lazzus 5 suggested the use of a neural network and a particle swarm algorithm to better represent the nonlinear relationship between the contribution of the various groups to T m . These methods reduced somewhat the average prediction errors.…”