Predicting Pharmacokinetics in Rats Using Machine Learning: A Comparative Study Between Empirical, Compartmental, and...

Area of Science:

• Pharmacokinetics
• Drug Discovery
• Computational Chemistry

Background:

• Drug development requires high potency and favorable pharmacokinetic (PK) properties for sustained efficacy.
• In vivo PK studies are crucial for dose estimation in preclinical and clinical settings.
• Predicting ADME properties with machine learning (ML) is established, but PK profile prediction is emerging.

Purpose of the Study:

• To systematically compare different approaches for predicting PK profiles in rats.
• To evaluate ML integration with empirical or mechanistic PK models for pre-synthesis prediction.
• To assess the accuracy of various PK prediction methods using internal preclinical data.

Main Methods:

• Comparison of four PK profile prediction approaches: NCA-based, pure ML, compartmental modeling, and physiologically based pharmacokinetic (PBPK) modeling.
• Utilized internal preclinical data for over 1000 small molecules.
• Evaluated prediction accuracy using geometric mean fold errors for plasma concentration-time profiles.

Main Results:

• Pure ML, compartmental, and PBPK modeling approaches showed comparable accuracy in PK profile prediction.
• These three methods outperformed standard non-compartmental analysis (NCA)-based predictions.
• Accurate prediction of PK profiles was achieved for a large dataset of small molecules.

Conclusions:

• ML-integrated PK modeling significantly improves the ability to predict drug behavior prior to synthesis.
• This approach enhances the prioritization of drug candidates with desirable pharmacokinetic properties.
• Facilitates more efficient drug discovery and development pipelines.