Machine Learning Meets Pharmacokinetics: A Comparative Analysis of Predictive Models for Plasma Concentration-Time...

Area of Science:

• Computational chemistry
• Pharmacokinetics
• Drug discovery

Background:

• Accurate prediction of pharmacokinetic (PK) profiles from molecular structures is crucial for efficient drug discovery.
• Existing methods often require extensive experimental data, delaying the process.

Purpose of the Study:

• To systematically compare five distinct computational frameworks for predicting rat plasma concentration-time profiles directly from molecular structures.
• To evaluate the performance of machine learning (ML) and physics-informed neural networks (PINNs) in PK prediction.

Main Methods:

• Five methodologies were evaluated: NCA-ML, PBPK-ML, CMT-ML, CMT-PINN, and PURE-ML.
• Models were trained and validated on a consistent dataset using a standardized evaluation framework.
• Performance was assessed using R²-log, Spearman correlation, and percentage of predictions within twofold error.

Main Results:

• The CMT-PINN approach demonstrated the highest predictive performance (R²-log: 0.854, Spearman: 0.933), closely followed by PURE-ML (R²-log: 0.789, Spearman: 0.896).
• CMT-PINN and PURE-ML achieved 65.9% and 61.0% prediction accuracy within twofold error, respectively.
• Models trained directly on concentration-time data outperformed those using derived PK parameters, especially with limited data.

Conclusions:

• Predicting PK behavior from molecular structures before synthesis is feasible.
• Computational approaches like CMT-PINN and PURE-ML enable informed compound selection early in drug discovery, reducing costs and timelines.
• These methods have the potential to reduce reliance on animal studies and accelerate pharmaceutical development.