Machine learning-based unified models for predicting drug clearance from pharmacokinetic animal and study design variables

Abstract

Clearance (CL) is a primary pharmacokinetic (PK) parameter crucial to determine how quickly a drug is eliminated from the body, which guides the appropriate dosing interval to maintain a consistent concentration in blood. Given the importance of CL, this study aimed to use machine learning (ML) techniques to predict CL values by identifying patterns and relationships within an extracted dataset of PK variables from published articles. Variables evaluated in the extracted dataset included drug, dose, animal species, and route of administration. Nine distinct ML models were then applied to analyze the CL data, incorporating both imbalanced and balanced data generated through resampling methods. Since the CL data used in this study is a collection of all CL values (true CL and CL/F) extracted from scientific articles, the collected CL variable for both IV and non-IV administration routes are referred to as hybrid ML CL. To analyze the effect of ML models in predicting the CL values, we used the hybrid ML CL dataset for six different subsets of data including one solely from the intravenous route of administration. Linear regression, multi-layer perceptron, and random forest models consistently had the highest efficiency in predicting CL values, with an R 2 score > 0.87. However, R 2 increased to > 0.95 when analyzing only ungulates or small ruminants, and > 0.92 for the companion animal group. This study has the potential to help researchers employ computational, mathematical, and ML models to predict and estimate CL values and changes in CL values based on variables. This study focuses on evaluating the feasibility of predicting drug CL in situations where direct CL data are not available. Rather than addressing drug development processes, the research examines whether study design variables can serve as input parameters for a proposed cross-species extrapolation tool aimed specifically at predicting existing drug CL values.