Prediction of Compound Plasma Concentration-Time Profiles in Mice Using Random Forest

Area of Science:

• Computational chemistry and cheminformatics
• Pharmacokinetics and drug metabolism
• Machine learning in drug discovery

Background:

• Traditional pharmacokinetic (PK) models focus on parameters like clearance (CL) and volume of distribution (Vd).
• Predicting the full concentration-time profile offers deeper insights into drug efficacy and safety, beyond static PK parameters.
• Existing in silico models often lack the ability to capture complex concentration-time dynamics.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting plasma concentration-time profiles after intravenous (i.v.) and oral (p.o.) dosing.
• To assess the predictive accuracy of various machine learning approaches, including Random Forest (RF).
• To identify key in vitro parameters influencing absorption, distribution, metabolism, and excretion (ADME) through feature importance analysis.

Main Methods:

• Development of machine learning models using MACCS Keys descriptors and in silico/in vitro PK parameters.
• Comparative analysis of Random Forest (RF), message passing neural network, and 2-compartment models.
• Validation using 5-fold cross-validation (5-fold CV) and leave-one-project-out validation (LOPO-V).

Main Results:

• The Random Forest (RF) model demonstrated superior predictive accuracy for both i.v. and p.o. concentration-time profiles.
• Key in vitro parameters like predicted human Vd (hVd), mouse intrinsic CL, unbound fraction of mouse plasma, and Caco2 permeability were identified as crucial predictors.
• The RF model effectively predicted complex phenomena such as twin peaks, outperforming baseline compartment models.

Conclusions:

• Machine learning, particularly RF, provides a robust and accurate method for predicting drug plasma concentration-time profiles.
• The model's ability to integrate diverse data sources (chemical descriptors, in vitro data) enhances its predictive power.
• This fit-for-purpose model offers a valuable tool for accelerating lead optimization in drug discovery due to its accuracy and speed.