Fraction-based Linear Extrapolation (FLEX) Method for Predicting Human Pharmacokinetic Clearance: Advanced Allometric Scaling Method and Machine Learning Approach

Accurate prediction of human pharmacokinetic (PK) parameters, particularly clearance (CL), is critical in early-stage drug development. Single Species Scaling (SSS) using rat PK data, with or without unbound plasma fraction (fu,plasma), is commonly used; however, its predictive accuracy is often compromised for compounds with excessive plasma-protein binding. Prior studies have suggested that compounds with fu,plasma < 0.02 in either species are prone to larger prediction errors; however, no systematic approach has addressed this limitation. In this study, we developed a new prediction method, Fraction-based Linear EXtrapolation SSS (FLEX-SSS fu Rat), which dynamically switches between SSS fu Rat and SSS Rat formulas based on an optimized fu threshold. To maximize prediction accuracy, the threshold and associated scaling coefficient were determined using a training set of 200 compounds. Additionally, we constructed a random forest machine-learning model using structural descriptors and validated all models with an independent external dataset of 62 compounds. FLEX-SSS fu Rat outperformed conventional SSS fu Rat, and the consensus model combining FLEX and RF showed the most balanced performance: 41.9% of predictions were within the 2-fold error, only 16.1% exceeded the 5-fold error, and the geometric mean fold error was 2.7. This is the first report to systematically validate the SSS fu Rat method using an independently constructed external dataset. By optimizing fu thresholds and integrating machine learning, our approach enabled more reliable human CL prediction, thereby supporting better-informed decision-making in first-in-human dose selection.