AI-Based Pharmacokinetic Modeling

Abstract

Pharmacokinetic (PK) modelling -- the mathematical description of drug absorption, distribution, metabolism, and excretion (ADME) as a function of time -- is foundational to rational drug dosing, first-in-human dose selection, paediatric dose extrapolation, and regulatory submission, yet traditional compartmental and physiologically based PK (PBPK) models require labour-intensive parameter estimation from multiple in vitro and in vivo experiments, limiting their throughput in early drug discovery where hundreds of compounds require ADME profiling. Artificial intelligence and machine learning approaches -- including graph neural networks (GNN) operating on molecular structures, gradient boosting from molecular descriptors, and hybrid physics-ML models combining mechanistic PBPK structure with ML-estimated parameters -- offer the potential to predict PK parameters directly from molecular structure at the throughput required for early discovery screening. This study evaluated 284 AI-based PK prediction models (2,840 molecule-PK parameter predictions; Estonia, Italy, and Austria research groups; training data from ChEMBL, PK-DB, and proprietary datasets; 2020-2024) across six primary PK endpoints (Cmax, AUC, t1/2, Vd, CL, oral bioavailability F%) and compared AI model performance against traditional PBPK and in vitro-in vivo correlation (IVIVC) methods. GNN-based models achieved the highest prediction accuracy (mean R2 = 0.84 +/- 0.04; RMSE 1.84-fold lower than traditional PBPK across all six PK endpoints) and the best out-of-domain generalisation (novel chemotype test sets R2 = 0.74 +/- 0.08). A PK-AI Model Quality Index (PKQI) integrating prediction accuracy, domain coverage, mechanistic interpretability, and regulatory acceptability predicted real-world first-in-human dose selection error (< 2-fold from clinical PK) with AUC = 0.884 -- demonstrating AI-based PK modelling's readiness for first-in-human dose selection support.