AI-AugETM: An AI-Augmented Exposure-Toxicity Joint Modeling Framework for Personalized Dose Optimization in Early-Phase Clinical Trials
Yachen Wang · 2025
In the current early stage of new drug development, due to the limited sample size and tight trial resources, there is an urgent need for a method that can accurately predict the concentration-toxicity response and reasonably derive the dosesafety boundary under the condition of a small population, to accelerate the design of personalized treatment pathways. In this work, we propose an AI-enhanced combined concentrationtoxicity model (AI-AugETM), which is extended on basis of the existing pharmacokinetic-toxicity combination model, and integrates deep learning technology and uncertainty inference methods, which is suitable for the modeling and evaluation of new dose pathways. Initially, the model is based on multi-source input data, and uses multi-task time series models to jointly model toxicity probability and time-varying evolution of toxicity grades, and learn the concentration-toxicity response curve under multidose pathways. Based on the results of the response curve, Bayesian uncertainty estimation and Shapley value interpretation mechanism are used to derive the effective exposure range and safety upper bound at the individual level, which can be used to assist in formulating the optimal dose range and safe upper dose limit, ensuring efficacy and controlling toxicity risk. The results show that AI-AugETM significantly improves the accuracy of toxicity prediction in pharmacokinetic toxicity combination data, and effectively narrows exposure decision boundary within 95% confidence interval.