Prediction of In Vivo Pharmacokinetic Parameters and Time-Exposure Curves in Rats Using Machine Learning from the...

Area of Science:

• Pharmacology
• Computational Chemistry
• Drug Discovery

Background:

• Animal pharmacokinetic (PK) data and in vitro systems are crucial for predicting drug elimination and disposition.
• Accurate prediction of in vivo drug properties aids in selecting better drug candidates and reducing animal testing.
• Understanding drug clearance and disposition in animals is key for extrapolation to human studies.

Purpose of the Study:

• To develop machine learning models for predicting rat in vivo PK parameters and concentration-time profiles.
• To utilize molecular chemical structure and in vitro data for PK predictions.
• To compare traditional machine learning and deep learning approaches for PK modeling.

Main Methods:

• Generated machine learning models using internal in vivo rat PK data for over 3000 diverse compounds.
• Employed molecular chemical structure and measured/predicted in vitro parameters as model inputs.
• Evaluated traditional machine learning algorithms and deep learning methods, including graph convolutional neural networks.

Main Results:

• The best models achieved R² = 0.63 for clearance and R² = 0.55 for oral bioavailability.
• Models accurately predicted key PK parameters based on chemical structure and in vitro data.
• Demonstrated the efficacy of deep learning approaches, specifically graph convolutional neural networks.

Conclusions:

• Developed a fast and cost-efficient method for predicting drug PK profiles.
• Enables prediction of virtual compounds at the design stage to guide molecule optimization.
• Facilitates prioritization of compounds for in vivo assays, reducing experimental burden.