Prediction\nof Multi-Pharmacokinetics Property in Multi-Species:\nBayesian Neural Network Stacking Model with Uncertainty

Pharmacokinetic (PK) properties of a drug are vital attributes\ninfluencing its therapeutic effectiveness, playing an important role\nin the drug development process. Focusing on the difficult task of\npredicting PK parameters, we compiled an extensive data set comprising\nparameters across multiple species. Building upon this groundwork,\nwe introduced the PKStack ensemble model to predict PK parameters\nacross diverse species. PKStack integrates a variety of base models\nand includes uncertainty in its predictions. We also manually collected\nPK data from animals as an external test set. We predicted a total\nof 45 tasks for nine PK parameters in five species, and in general,\nthe prediction accuracy was better for intravenous injections, including\nparameters such as human <i>V</i><sub>d</sub> (R<sup>2</sup> = 0.72, RMSE = 0.31), human CL (R<sup>2</sup> = 0.52, RMSE = 0.32),\nand others. In addition to predictive accuracy, we also considered\nthe interpretability of the results and the definition of the model’s\napplication domain. Based on the findings, our model has great potential\nfor practical applications in drug discovery.