Predicting pK_a Values of Substituted Phenols from Atomic Charges: Comparison of Different Quantum Mechanical Methods and Charge Distribution Schemes

Abstract

The acid dissociation (ionization) constant pK(a) is one of the fundamental properties of organic molecules. We have evaluated different computational strategies and models to predict the pK(a) values of substituted phenols using partial atomic charges. Partial atomic charges for 124 phenol molecules were calculated using 83 approaches containing seven theory levels (MP2, HF, B3LYP, BLYP, BP86, AM1, and PM3), three basis sets (6-31G*, 6-311G, STO-3G), and five population analyses (MPA, NPA, Hirshfeld, MK, and Löwdin). The correlations between pK(a) and various atomic charge descriptors were examined, and the best descriptors were selected for preparing the quantitative structure-property relationship (QSPR) models. One QSPR model was created for each of the 83 approaches to charge calculation, and then the accuracy of all these models was analyzed and compared. The pK(a)s predicted by most of the models correlate strongly with experimental pK(a) values. For example, more than 25% of the models have correlation coefficients (R²) greater than 0.95 and root-mean-square errors smaller than 0.49. All seven examined theory levels are applicable for pK(a) prediction from charges. The best results were obtained for the MP2 and HF level of theory. The most suitable basis set was found to be 6-31G*. The 6-311G basis set provided slightly weaker correlations, and unexpectedly also, the STO-3G basis set is applicable for the QSPR modeling of pK(a). The Mulliken, natural, and Löwdin population analyses provide accurate models for all tested theory levels and basis sets. The results provided by the Hirshfeld population analysis were also acceptable, but the QSPR models based on MK charges show only weak correlations.

Quantum mechanics approaches to drug research in the era of structural chemogenomics

“…This new charge assignment is expected to be more accurate, especially for complexes with significant charge transfer or polarization. Among several QM methods for calculation of partial atomic charges, 66-68 the one most relevant to protein-ligand binding calculations is Restrained Electrostatic Potential (RESP 68 ). However, due to the computational complexity of QM charge calculations, much effort in the recent years was devoted to development of faster schemes closely approximating RESP and other QM changes.…”

pKa prediction for acidic phosphorus‐containing compounds using multiple linear regression with computational descriptors

“…[14,15] The prediction accuracies vary depending on the method or substrate; the R 2 values range from less than 0.85 to higher than 0.99, and the root-mean-square errors vary from higher than 0.75 pK a units to lower than 0.40 units. [13][14][15][16]20] Harding et al reviewed the descriptors related to pK a . [3,14,15] The results reported in the literature show that multiple linear regression (MLR) generally gives better results than simple linear regression, for both empirical and quantum QSPR approaches.…”

How Does the Methodology of 3D Structure Preparation Influence the Quality of pKa Prediction?

“…The most appropriate charge calculation schemes for these purposes seem to be MPA (Mulliken population analysis), NPA (Natural population analysis) and AIM (atoms in molecules). 8,15,17 Semiempirical QM charges have also been employed in QSPR models for p K a prediction (e.g., AM1, PM3 or PM6 theory levels in combination with MPA). 11,14,17–19 A major drawback of the QM charges is the computational effort required for the calculation of the wave function.…”

“…8,15,17 Semiempirical QM charges have also been employed in QSPR models for p K a prediction (e.g., AM1, PM3 or PM6 theory levels in combination with MPA). 11,14,17–19 A major drawback of the QM charges is the computational effort required for the calculation of the wave function. For this reason, the computational complexity of obtaining QM charges is at least θ ( B 4 ), where B is the number of basis functions.…”

“…In previous works focused on p K a prediction via QSPR, 8,17,19,21,22 3D structures were mainly obtained from the DTP NCI database 23 (which uses CORINA to generate the 3D structures) or directly designed by CORINA. 24 But there are other tools and databases which are often used as sources of 3D structures.…”