P05-07
Unbound Fraction Optimized Method for Predicting Human Pharmacokinetic Clearance: Advanced Allometric Scaling Method and Machine Learning Approach
Yuki UMEMORI *1, Koichi HANDA2, Saki YOSHIMURA1, Michiharu KAGEYAMA1
1Translational Science, Discovery DMPK, Axcelead Tokyo West Partners
2Discovery Science, Drug Discovery Chemistry, Axcelead Tokyo West Partners
( * E-mail: yuki.umemori@axcelead-twp.com )
Accurate prediction of human pharmacokinetic (PK) parameters, particularly human clearance (CL), to estimate human dosing in the drug discovery phase is crucial for enhancing the success rate of drug development. Among the various methods, Single Species Scaling (SSS) using rat PK data and the unbound fraction in plasma (fu) from both human and rat, known as SSS fu rat, has been widely employed. However, the datasets used in the SSS fu rat are limited to about 200 compounds, and leaving the accuracy with external datasets unclear; allometric scaling models have been built without consideration of separated dataset traditionally. Recent advances have also employed machine learning models to predict human CL, leveraging structural information.
In this study, we prepared 200 training and 62 external test compounds that was obtained experimentally. Using these data we investigated the conventional SSS rat; a new method, the Unbound Fraction Optimized SSS (UFO SSS) fu rat, which predicts using the SSS rat method for compounds with low fu and a newly calculated allometric equation for remaining compounds; a Random Forest (RF) machine learning model using ECFP4; a consensus model of UFO SSS and RF.
We first analyzed the training dataset of 200 compounds and found that compounds with an fu value of less than 0.03 in either human or rat exhibited poor predictive accuracy using the SSS fu rat. This value (0.03) was used as the threshold value of fu in UFO SSS rat. To investigate our approach, we compared the predictive performance of each model with 62 external test datasets. For SSS fu rat, UFO SSS fu rat, RF, and the consensus model, the percentages of compounds within 2-fold error were 37.1%, 41.9%, 40.3%, and 41.9%; the percentages of compounds exceeding 5-fold error were 29.0%, 21.0%, 21.0%, and 16.1%; the Geometric Mean Fold Error were 5.5. 4.8, 2.6, and 2.6, respectively. Hence, we conclude that the consensus model achieved the best overall performance, demonstrating its superior predictive accuracy.
We can stress that these models were validated by the external dataset of 62 compounds, presenting a novel approach that enhances the accuracy of human clearance predictions in the drug discovery phase. By combining an allometric method optimized with unbound fraction data and machine learning techniques, our approach provides a more reliable prediction method for human pharmacokinetics, ultimately contributing to the success of drug development.