GenMAS: A Machine Learning-Based Modeling and Assessment Strategy for ADMET Profiling and its Application in PBPK Simulation

Evaluation of pharmacokinetic (PK) properties is crucial for the discovery of high-quality drug candidates while minimizing unnecessary biological and clinical trial costs. Compared with in vitro and in vivo studies, in silico methods offer a faster and more cost-efficient alternative. However, despite advances in computational approaches, accurate PK assessment and simulation remain challenging. Current industrial applications are often hindered by a heavy reliance on animal data, limited flexibility, and insufficient project-specific customization. In this study, we present GenMAS, a machine learning (ML)-based strategy designed for comprehensive modeling and evaluation of PK properties. GenMAS was evaluated for its performance in ADMET property prediction and its applicability to physiologically based pharmacokinetic (PBPK) simulation. This strategy synergistically integrated three functional modules: (1) ADMET property prediction module, (2) ML-informed PBPK modeling module, and (3) self-developed high-throughput ADMET-PBPK screening module. The ADMET prediction module was assessed using 10-fold cross-validation, internal test sets, and external validation sets. Cross-validation demonstrated robust performance, with the models yielding an average AUC (area under the curve) of 0.879 and an accuracy of 0.828. For independent test sets, classification accuracy ranged from 70.2% to 88.7%, while average R2 for physicochemical properties exceeded 0.76. The predictive robustness of GenMAS was further confirmed using in-house project datasets, demonstrating its utility in real-world drug discovery. Furthermore, we developed a ML-informed PBPK module and a high-throughput ADMET-PBPK screening module. In the ML-informed module, ML-predicted compound-specific parameters were integrated into a GastroPlus® rat PBPK model. For 9 intravenously administered compounds, the predicted systemic exposures fell within a two-fold error range of observed data, with a 66.7% accuracy for elimination half-life. The high-throughput module couples ADMET predictions with a minimal PBPK model to enable customizable PK screening. Notably, GenMAS can complete ADMET predictions and PK simulations for 100 compounds in approximately 12 s. GenMAS offers a scalable, structure-based workflow to support early-stage drug discovery.