Octanol/water partition coefficient, log P
There are two log P models in ADMET Predictor™: S+logP and MlogP. The S+logP model is based on artificial neural network ensembles (ANNE) constructed by our automatic model builder ADMET Modeler™ from almost 13,000 example compounds selected from the "StarList" of ion-corrected experimental logP values (Hansch, C. et al, 1995). The MlogP model, which is based on the published work of Moriguchi, et al (Moriguchi et al; 1992 and 1994), was the sole log P predictor in very early versions of ADMET Predictor and continues to be retained for comparison.
The figure below shows the correlation between the observed logP and the predicted S+logP values.
ADMET Predictor 2D S+logP Model Validation
In all of the independent model comparisons performed on public as well as industrial data sets, S+logP has consistently outperformed other predictive logP models available up to date, see (Dearden et al; 2003), (Tetko, et al; 2007) and (Mannhold, et al; 2008) for details.
Octanol/water distribution coefficient, log D
The octanol-water distribution coeffcient, log D, is determined by the ratio of analytical concentrations in octanol and water phases, respectively (Pagliara; 1997). As such, it includes all protonation states of a given compound and does depend on pH. Consequently, it is much more difficult to model log D, than log P. Its complete pH profile can be derived from pH-dependent distribution of protonation species in both solvents, as expressed by the equation below:
where the pH dependence of the model is contained in the pH-dependent fractions ionized, f(0) and f(i), which can be obtained from knowledge of the aqueous pKa's of the molecule. The sum goes over all ionized states of the molecule. The d(i) factors contain the differences between log P of the neutral species and log P(i), the log P of a given ionization state. Often, the approximation is made that the d(i) are the same for all anionic states of a molecule and can be set to a single constant, structure-independent, value d(-). Similarly, all cationic states are set to a single constant value d(+). Such assumptions were made in earlier versions of our S+logD model. The present S+logD model is quite unique in the sense that all the d(i) factors are now structure-dependent and calculated by separate artificial neural network ensembles (ANNE) from molecular descriptors. Our pKa model delivers pKa values necessary to calculate fractions ionized and S+logP provides the log P values. Thus, the structure-dependent S+logD model vastly outperforms our earlier version with constant d(i) factors. It has been trained on almost 9,000 non-ion-corrected log D values obtained from BioByte's Masterfile as well as additional measurements extracted from the literature.
ADMET Predictor 2D S+logD Model Validation
Hansch, C. et al, "Exploring QSAR: Hydrophobic, Electronic, and Steric Constraints." ACS Publications (1995).
Moriguchi I, Hirono S, Liu Q, Nakagome I, Matsushita Y. "Simple method of calculating Octanol/Water Partition Coefficient." Chem Pharm Bull. 1992; 40:127-130.
Dearden JC, Netzeva TI, Bibby R. "A Comparison of Commercially Available Software for the Prediction of Partition Coefficient." in "EuroQSAR 2002: Designing Drugs and Crop Protectants: Processes, Problems and Solutions", Bournemouth, UK, Blackwell Publishing Ltd. (2003)
Tetko IV and Poda GI. "Prediction of Log P with Property-Based Methods." in "Molecular Drug Properties: Measurement and Prediction." ed. Mannhold R. Weinheim, Germany, Wiley-VCH: 381-406. (2007)
Mannhold R, Poda GI, Ostermann C, Tetko IV. "Calculation of Molecular Lipophilicity: State of the Art and Comparison of Methods." J Pharm Sci. 2008; 98(3):861-893.
Pagliara A, Carrupt P-A, Caron G, Gaillard P, Testa B. "Lipophilicity Profiles of Ampholytes." Chem Rev. 1997; 97:3385-3340.