Comparing Scientific Machine Learning With Population Pharmacokinetic and Classical Machine Learning Approaches for...

Area of Science:

• Pharmacokinetics and Machine Learning
• Drug Development and Personalized Medicine

Background:

• Classical machine learning (ML) models for individualizing drug dosages lack pharmacokinetic (PK) interpretability.
• Population PK (PopPK) models struggle with complex covariate relationships.

Purpose of the Study:

• Compare classical ML, PopPK, and a novel scientific ML (MMPK-SciML) framework for drug plasma concentration prediction.
• Evaluate MMPK-SciML's ability to estimate PopPK parameters and inter-individual variability (IIV).
• Assess MMPK-SciML performance using fluorouracil (5FU) and sunitinib datasets.

Main Methods:

• Developed and applied the MMPK-SciML framework utilizing multimodal patient covariate data.
• Compared MMPK-SciML predictions against classical ML and established PopPK models.
• Utilized datasets for intravenously administered 5FU (541 concentrations) and orally administered sunitinib (302 concentrations).

Main Results:

• Classical ML models failed to adequately describe drug concentration data.
• MMPK-SciML achieved accurate drug plasma concentration predictions for test patients.
• MMPK-SciML demonstrated superior accuracy over PopPK for 5FU; comparable accuracy for sunitinib.

Conclusions:

• MMPK-SciML shows significant promise as an alternative to traditional PopPK modeling.
• The framework effectively estimates PopPK parameters and IIV without assuming covariate relationships.
• Further investigation of MMPK-SciML is warranted, particularly with sufficient training data.