(T0930-10-57) Predicting Pharmacokinetic Response of Small Molecules in Rats via Machine Learning

Purpose: Predicting exposure from molecular structure can potentially accelerate drug discovery timelines, and save a significant amount of preclinical resources. In this work, machine learning models are trained to predict (i) in-vivo clearance (CL), (ii) volume of distribution (Vdss), and (iii) time-dependent plasma concentration, based on molecular structure. The models presented here focus on intravenous (IV) administration of small drug molecules in rats.

Methods: Two groups of in-house pharmacokinetic (PK) data were collected and preprocessed. The first group included CL and Vdss values for 2928 small molecule drugs after IV administration in rats, and the second group the corresponding exposure profile from 911 of them. The corresponding molecular structure was parametrized either by Extended Connectivity Fingerprints (ECFP) or molecular descriptors. Then, supervised learning (gradient boosting – XGBoost) was employed to link such molecular structure descriptions to CL or Vdss. Finally, CL or Vdss were used as predictors (along with time t) to recover or predict full PK profiles (time-dependent drug concentration in plasma).

Results: Trained machine learning models were able to predict CL, Vdss for most compounds in the testing dataset within 40% relative error (see Fig. 1) – which is comparable to expected inter-animal variability.  Such performance marks a significant improvement compared to previous efforts where smaller datasets were utilized [1]. Next, full pharmacokinetic profiles were recovered/ predicted using CL, Vdss and time as input variables (Fig. 2). The trained machine learning model shows improved performance (MAPE: 75%, R2=0.67 - log scale) compared to previous efforts [2], and is able to capture qualitative PK characteristics (e.g. mono-, bi-exponential distribution phases).

Conclusion: Machine learning models were developed and optimized for the prediction of in-vivo PK profiles (clearance, volume of steady state distribution and time-dependent plasma exposure) in rats. The predictive power of such models strongly depends on the amount and diversity of molecular structures in the training dataset. Efforts to curate larger datasets and optimize machine learning models are necessary to expand predictions to different routes of administration (e.g. oral) and to different species (preclinical or even clinical). Importantly, such models can be utilized for further downstream tasks, such as efficacious dose prediction thus facilitating decision making in the early stages of drug development.

References: [1] Mavroudis, P. D., Teutonico, D., Abos, A., & Pillai, N. (2023). Application of machine learning in combination with mechanistic modeling to predict plasma exposure of small molecules. In Frontiers in Systems Biology (Vol. 3). Frontiers Media SA. https://doi.org/10.3389/fsysb.2023.1180948
[2] Pillai, N., Abos, A., Teutonico, D., & Mavroudis, P. D. (2024). Machine learning framework to predict pharmacokinetic profile of small molecule drugs based on chemical structure. In Clinical and Translational Science (Vol. 17, Issue 5). Wiley. https://doi.org/10.1111/cts.13824

Acknowledgements: The authors acknowledge Alex Marker, Henrik Cordes, Felix Jost, Yongyi Luo, Sirimas Sudsakorn. This research is funded by Sanofi.

Median Absolute Percentage Error for in vivo clearance (CL) and Volume of distribution at steady state (Vdss) using either ECFP or molecular descriptors as inputs.

Left - Parity plot for predicted PK profile (each point represents a concentration of a molecule at some point in time). Right – sample predicted and observed PK profiles for molecules in the training and testing datasets respectively.