Nikhil Pillai: No financial relationships to disclose
Objectives: Accurate prediction of a new compound's pharmacokinetic (PK) profile in humans is pivotal for the success of drug discovery programs. Traditional methods such as allometric scaling and mechanistic modeling rely on parameters derived from in-vitro experimentation or in-vivo animal testing, which are not only time-consuming and resource-intensive, but also involve significant ethical concerns due to use of animals. In this study we developed a hybrid and a hierarchical Machine Learning (ML) framework to predict human PK dynamics from molecular structure.
Methods: In this study, we curated a large dataset of small molecules’ physicochemical (PC) and PK properties in humans. These properties were extracted from public datasets, while plasma concentration versus time profiles for approximately 800 compounds were digitized from published literature [1]. Utilizing these datasets, we initially developed a hybrid modeling framework that combines ML-predicted PC/PK parameters with physiologically based pharmacokinetic (PBPK) modeling to predict PK profiles in humans and compare them with observed PK dynamics. Additionally, a hierarchical ML framework was developed, using ML-predicted properties to directly predict human PK profiles.
Results: Both frameworks achieve reasonable exposure predictions for most of the drug compounds in the test dataset of 106 drugs, with the hierarchical ML modeling framework performing better than the hybrid approach. The hierarchical ML model, trained on diverse and comprehensive data, generalizes well across different drugs. The hierarchical ML framework predicted approximately 60% and 90% of compounds within a 2-fold and 5-fold error, respectively, for both AUC and Cmax, as calculated from the predicted PK profile curve. Meanwhile, the hybrid framework predicted about 40% and 80% within a 2-fold and 5-fold error, respectively, for both AUC and Cmax. Our work also highlights the benefits of considering multiple distribution models in the hybrid approach, which provides variability in addition to a single PK profile estimate. This work extends our group’s previous efforts by applying two novel modeling frameworks for human PK profile predictions, further advancing early molecular screening and design in the drug discovery process [2].
Conclusions: This effort aims to enable earlier PK profile prediction in the drug development process and ultimately assist in prioritizing compounds for future evaluation and may help in providing an alternate approach to animal testing. With the ML models presented here as a starting point, we expect that more advanced ML algorithms can push the predictive performance even higher.
Citations: [1] Lombardo, F et al. Drug Metab Dispos 2018, 46, (11), 1466-1477. [2] Pillai, N. et al. Clinical and Translational Science 2024, 17, (5), e13824.
Keywords: Machine Learning, Physiologically based Pharmacokinetic modeling, Human pharmacokinetic prediction
