ACD/pK_a Prediction Accuracy vs. Experiment

A true and fair measure of the accuracy of predictive software is its ability to accurately predict diverse compound datasets. We recently took the opportunity to search the current literature for experimental pK_a data. This data was collated and used to test the prediction accuracy of the ACD/pK_a algorithm.

Experimental data was gathered for 277 compounds, providing 285 pK_a measurements for acidic and basic ionizing groups. Of these, 284 corresponding pK_a values were predicted by the ACD/pK_a algorithm. For compounds with differing published experimental values (within 0.5 pK_a) an average was used. For compounds where the difference in experimental pK_a showed a discrepancy of more than this, further study of the data was undertaken to decide on the ionization taking place and select the appropriate pK_a value. Approximated apparent pK_a predictions were used in this study since these represent best the true behavior of a compound in aqueous solution, and would therefore be most likely to reflect the findings of experimentally determined pK_a.

Figure 1: A scatter plot of experimental pKa data vs. predictions made using ACD/pKa.			Figure 2: A histogram illustrating the error range in predictions.
N =284	R2 = 0.95		Av. Error = 0.39	Std. Error = 0.76

Results of this study show that 80% of predictions (228 individual pK_a values) for this diverse set of compounds were within an error of 0.5 pK_a units; while 91% (260 of 285 pK_a values) were predicted with an error of less than 1 pK_a. No specific ionization centre or functional group was found to be predicted consistently poorly, although four of the compounds with the poorest prediction accuracy (errors of greater than 2.5 pK_a units compared to the experimental value) had ionization events at extremely acidic pH (below zero). Click here to see the table of data.

Within this dataset, we can look at a particular series of compounds to examine how well the algorithm performed for that chemical space. If we choose the benzimidazoles, of which there were a total of 13 in the full dataset, we see at a quick glance that ~77% (10 out of 13) were predicted within 0.75 pK_a; 3 are predicted poorly with an error greater than 2 pK_a. Closer examination of this class in the desktop pK_a module showed that prediction of single pK_a values for this class were far more accurate, with an average error of 0.18 pK_a (compared to 0.68 pK_a in approximated apparent values) (see Table below).

Table 1. Experimental and predicted pK_a for Benzimidazole class

It is therefore worth noting that the different types of pK_a predicted by the ACD/pK_a algorithm should be investigated especially when looking at a particular chemical space. Single pK_a values are useful in tracking protonation/deprotonation for a single ionization centre in a series of compounds to determine relative changes.

While this comparison takes into account only limited series of compounds, and is therefore not a comprehensive study of the prediction accuracy of the ACD/pK_a algorithm, it reflects the high standards our customers have come to expect from ACD/Labs PhysChem software. To study the accuracy of prediction for your chemical domain, we invite you to evaluate the software.

References

• L. Xing and R.C. Glen, J. Chem. Inf. Comput. Sci., 42: 796–805, 2002
• U.A. Chaudry and P.L.A. Popelier, J. Org. Chem., 69: 233–41, 2004
• F. Eckert and A. Klamt, J. Comput. Chem., 27: 11–19, 2006
• T. N. Brown and N. Mora-Diez, J. Phys. Chem. B, 110: 9270–79, 2006
• H. Lu, X. Chen, and C.G. Zhan, J. Phys. Chem. B, 111: 10599–605, 2007
• M. Meloun and S. Bordovská, Anal. Bioanal. Chem., 389: 1267–81, 2007