Predicting First-in-Human Pharmacokinetics: Comparative Evaluation of Standard PBPK, High-Throughput PBPK, and Machine Learning.

Accurate prediction of human pharmacokinetics is essential for selecting safe starting doses in first-in-human (FiH) trials. This study compares the predictive accuracy of three methodologies including standard physiologically based pharmacokinetic (PBPK) modeling incorporating animal verification, high-throughput (HT) PBPK based on in vitro inputs, and machine learning (ML). We evaluated their performance on 40 diverse small molecules that entered clinical development at Roche between 2003 and 2024. Standard PBPK predictions were assessed prospectively from original reports, while HT-PBPK and ML predictions were obtained retrospectively. Predicted area under the curve (AUCinf) and maximum plasma concentration (Cmax) for oral administration were compared to observed clinical data. While all three approaches demonstrated a useful level of accuracy with at least 49% of predicted AUCinf and Cmax values within 2-fold, standard PBPK was the most precise method with 65 and 63% of AUCinf and Cmax within that margin, respectively. This higher precision likely results from the integration of expert judgment and animal in vivo data for model refinement, whereas HT-PBPK and ML provide valuable animal-free, high-throughput alternatives. Ensemble models leveraging two or three models were explored, which generally improved accuracy in comparison to single models. This highlights the complementary nature of mechanistic and ML-based models. Currently, standard PBPK is still the preferred method for FiH dose selection for which highest accuracy and mechanistic insight is sought. However, although less accurate, we believe that HT-PBPK and ML are viable animal-free alternatives with value for early human dose prediction, reducing costs and cycle times.