Systematic evaluation of five classical mechanistic methods for tissue-plasma partition coefficients across diverse...

Tissue-plasma partition coefficient (K_p) is a key parameter governing drug distribution and the predictive accuracy of physiologically based pharmacokinetic models. Although several classical mechanistic methods have been developed for K_p prediction, systematic comparisons across multiple tissues and compound classes remain limited. In this study, we systematically compared 5 classical mechanistic methods using a standardized rat dataset to assess their predictive performance and applicability. A PubMed literature search was performed to compile a dataset of 781 K_p values from 119 compounds across 11 rat tissues, covering 6 compound classes. Physicochemical properties were collected from literatxure and public databases. Five mechanistic methods-Poulin and Theil, Berezhkovskiy, Rodgers and Rowland, Willmann, and Schmitt-were implemented in R software (R Foundation for Statistical Computing) according to their published algorithms. Predictive performance was evaluated using 3-fold error proportion, average fold error, absolute average fold error, and root mean square error. Overall, the Rodgers and Rowland method performed the best (68.63% within 3-fold error; absolute average fold error = 2.62, average fold error= 0.82). Across compound classes, the Rodgers and Rowland method was superior for most classes, particularly strong bases (74.24% within 3-fold error), whereas the Poulin and Theil method performed better for neutral compounds (71.15% within 3-fold error). Across tissues, predictions were more accurate for skin and heart, whereas kidney, adipose, and brain remained challenging. These findings indicate that K_p prediction performance varies substantially across different in silico methods, compound classes, and tissue types. SIGNIFICANCE STATEMENT: This study systematically evaluates 5 classical mechanistic methods for predicting tissue-plasma partition coefficients across 11 rat tissues and 6 compound classes using a standardized dataset of 781 values. Our findings identify the most accurate and robust method, reveal compound-specific and tissue-specific limitations, and provide practical guidance for physiologically based pharmacokinetic modeling, supporting improved drug distribution prediction in preclinical research and drug development.