A Machine Learning Framework to Improve Rat Clearance Predictions and Inform Physiologically Based Pharmacokinetic...

Area of Science:

• Pharmacokinetics and Drug Metabolism
• Computational Chemistry and Molecular Modeling
• Machine Learning in Drug Discovery

Background:

• Physiologically based pharmacokinetic (PBPK) models are crucial for predicting drug efficacy and safety.
• Accurate estimation of unbound intrinsic clearance (CLint,u) is a key challenge in PBPK modeling for drug discovery.
• Current in vitro-to-in vivo extrapolation methods can be complex and require extensive experimental data.

Purpose of the Study:

• To develop and compare machine learning (ML) strategies for predicting rat CLint,u for PBPK applications.
• To evaluate the performance of ML-based predictions against the standard in vitro bottom-up approach.
• To assess the utility of ML models in improving the prediction of in vivo pharmacokinetic parameters.

Main Methods:

• Collected in vivo and in vitro data for 2639 proprietary compounds.
• Developed three ML-based strategies to predict CLint,u, including back-calculation from in vivo data and bias prediction.
• Compared ML approaches with the standard in vitro bottom-up method using temporal cross-validation.

Main Results:

• ML model trained on back-calculated CLint,u achieved an absolute average fold error (AAFE) of 3.1, outperforming the bottom-up approach (AAFE 3.6-16).
• ML model incorporating bias prediction improved AAFE from 16 to 2.9 and log Pearson r^2 from 0.1 to 0.29 compared to bottom-up.
• ML approaches offer advantages such as reduced need for experimental in vitro data and circumvention of certain scaling corrections.

Conclusions:

• Machine learning strategies provide a powerful and accurate alternative for predicting CLint,u in PBPK modeling.
• ML-based prediction of CLint,u can enhance existing drug discovery workflows and improve in vivo pharmacokinetic predictions.
• These computational approaches streamline the drug development process by providing reliable clearance estimates with potentially less experimental burden.