Abstract
The partitioning of molecules between an aqueous and an organic medium is of major interest for pharmaceutical development and the chemical industry. It characterizes the impact of substances to the environment and to humans, e.g., their accumulation in living organisms. It is usually quantified in terms of the octanol-water partition coefficient K(OW) of these substances. Although this is a clearly defined thermodynamic property, different experimental approaches exist for its estimation. Using active pharmaceutical ingredients (APIs) as examples, we demonstrate the large scatter in experimentally determined partition coefficients reported in the literature. This is especially serious for weak bases or weak acids, which account for around 95% of all APIs. In some cases, reported K(OW) values for the same substance differ by even several orders of magnitude. This is particularly worrying because this property is crucial for approval procedures of APIs and is also used as input for a whole range of estimation methods, such as machine-learning algorithms. In this work, we discuss the physical reasons for the unusually high variety of reported K(OW) values. Using physicochemical laws, it is shown that the large scatter of the data is not caused by analytical uncertainties but by the extrapolation of the experimental data to a solute concentration of zero. Based on this, we propose a new approach for evaluating experimental data on partition coefficients. This approach involves extrapolating experimentally determined distribution coefficients with respect to pH rather than concentration. We will show that this reduces the uncertainty of the experimentally obtained K(OW) values, narrowing the difference between the highest and the lowest value for the same substance of currently about 2.4 to about 0.5 logarithmic units. The new approach can be combined with any existing experimental method for concentration analysis. Moreover, the obtained data agree very well with theoretical values obtained from thermodynamic modeling explicitly considering solute ionization, thus validating the proposed approach.