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

Abstract

Predicting pharmacokinetic (PK) profiles from molecular structures constitutes a significant advancement in pharmaceutical research with substantial implications for expediting the entire drug discovery process. Our investigation presents a comprehensive comparative analysis of five distinct methodological frameworks for predicting rat plasma concentration-time profiles: four approaches that integrate mechanistic modeling with computational techniques, and one approach employing pure machine learning (ML) architecture without physiological constraints. As reference (1), we used a method that predicts non-compartmental analysis (NCA) parameters to predict plasma PK profiles with a one-compartmental PK model (NCA-ML). Alternative approaches include: (2) a physiologically based PK (PBPK) model utilizing ML-predicted in vitro characteristics; (3) CMT-ML, where neural networks predict compartmental PK model parameters; (4) CMT-PINN, employing physics-informed neural networks; and (5) PURE-ML, using decision trees to predict concentration values at specific timepoints based on predicted volume of distribution and clearance. Assessment using multiple performance metrics demonstrated that the CMT-PINN approach achieved superior predictivity for PK profiles. Models trained directly on concentration-time, rather than derived PK parameters, delivered markedly improved predictive performance, particularly when working with limited training datasets. Our findings confirm the viability of predicting PK behavior from molecular structures before synthesis, representing a significant advancement for subsequent design. Implementation of these computational approaches enables informed compound selection early in the discovery pipeline, concentrating valuable resources for advanced preclinical investigations on only the most promising candidates, with anticipated beneficial effects on overall development timelines.