Formalizing method choice: a data-driven selection of mechanistic models for predicting human heart drug partitioning

Abstract We developed a workflow based on a human-derived dataset, aiming to predict the human heart-to-plasma partition coefficient, $$\:{K}_{p}$$ . The workflow combines established mechanistic tissue-composition models with an AI/ML-based selector that supports compound-specific selection among mechanistic heart $$\:{K}_{p}$$ prediction methods. Three published mechanistic approaches were implemented and parameterized in a harmonized manner consistent with the described methodologies to generate baseline $$\:{K}_{p}$$ predictions. The ML-based selector was trained using physicochemical, ADME, and pharmacokinetic descriptors to recognize compound-specific differences in model performance. This hybrid strategy preserves mechanistic interpretability while using empirical pattern recognition to formalize a priori model choice across compounds. In internal cross-validation, the selector reduced extreme outliers and provided a reproducible, data-driven rule for compound-specific method selection, although it did not materially outperform the strongest standalone mechanistic baseline in numerical error. Although no individual mechanistic method performed consistently across the full chemical space, the Poulin–Theil approach showed the best overall performance on human whole-heart data. The workflow was also applied prospectively to transthyretin stabilizers, for which reliable heart $$\:{K}_{p}$$ estimates may support further mechanistic modelling of cardiac drug exposure and downstream pharmacological effects. Together, these findings suggest that combining mechanistic calculators with an AI/ML-based selector may provide a practical basis for a priori, compound-specific method selection, reducing reliance on subjective choice of a single mechanistic approach.