1 Outlier detection in the statistical accuracy test of a pKa prediction Milan Meloun1*, Sylva Bordovská1 and Karel Kupka2 1Department of Analytical Chemistry, Faculty of Chemical Technology, Pardubice University, CZ-532 10 Pardubice, Czech Republic, 2TriloByte Statistical Software, s.r.o., Jiráskova 21, CZ-530 02 Pardubice, Czech Republic email:
[email protected], *Corresponding author Abstract: The regression diagnostics algorithm REGDIA in S-Plus is introduced to examine the accuracy of pKa predicted with four programs: PALLAS, MARVIN, PERRIN and SYBYL. On basis of a statistical analysis of residuals (classical, normalized, standardized, jackknife, predicted and recursive), outlier diagnostics are proposed. Residual analysis of the ADSTAT program is based on examining goodness-of-fit via graphical and/or numerical diagnostics of 15 exploratory data analysis plots, such as bar plots, box-and-whisker plots, dot plots, midsum plots, symmetry plots, kurtosis plots, differential quantile plots, quantile-box plots, frequency polygons, histograms, quantile plots, quantile- quantile plots, rankit plots, scatter plots, and autocorrelation plots. Outliers in pKa relate to molecules which are poorly characterized by the considered pKa program. The statistical detection of outliers can fail because of masking and swamping effects. Of the seven most efficient diagnostic plots (the Williams graph, Graph of predicted residuals, Pregibon graph, Gray L-R graph, Index graph of Atkinson measure, Index graph of diagonal elements of the hat matrix and Rankit Q-Q graph of jackknife residuals), the Williams graph was selected to give the most reliable detection of outliers. The 2 2 six statistical characteristics, Fexp, R , R P, MEP, AIC, and s in pKa units, successfully examine the specimen of 25 acids and bases of a Perrin´s data set classifying four pKa prediction algorithms.