(W-024) Small molecule exposure prediction (intravenous (IV) and oral (PO)) using Machine learning (ML) in combination with mechanistic modeling.

Nikhil Pillai, PhD – Principal Scientist, Sanofi; Georgios Psarellis, PhD – Senior Computational Scientist, Global Modeling and Simulation, DMPK, Sanofi, Cambridge, MA, USA, Sanofi; Alexandra Abos, PhD – Data Scientist, Commercial Data Science, Sanofi, Barcelona, Spain, Sanofi; Donato Teutonico, PhD – Pharmacometrician Expert PBPK, Translational Medicine and Early Development, Sanofi, Chilly-Mazarin, France,, Sanofi; Panteleimon Mavroudis, PhD – Distinguished Scientist, DMPK Modeling & Simulation Group Head, Global Modeling and Simulation, DMPK, Sanofi, Cambridge, MA, USA, Sanofi

Objective: The goal of this proof-of concept work is to evaluate whether using machine learning (ML)-derived parameters in tandem with physiologically based pharmacokinetic (PBPK) modeling can reasonably predict plasma exposure after IV and PO administration of small molecules. This framework aims to provide early PK prediction solely based on molecular structure enabling early go, no-go decisions.

Methods: In this study, ML is used to predict pharmacokinetic (PK) and physicochemical (PC) properties that are then used as input to mechanistic models to ultimately predict rat plasma exposure for both IV [1] and PO administration. As a first step, ML models were built to predict parameters required as input to PBPK model. Different ML algorithms such as random forest [2], support vector regression (SVR) [3], XGboost [4] and message-passing neural networks (MPNN) [5] were tested for their performance. For PBPK modeling five distribution models were tested: Berezhkovskiy [6], PK-Sim standard [7], Poulin and Theil [8], Rodgers and Rowland [9] and Schmidt [10]. PBPK model was implemented in PK-Sim, part of the Open Systems Pharmacology Suite version 11.0 [11]. R version 4.2.0 was used to perform the simulations. To ensure data consistency and reduce variability caused by assay choice, we utilized historical data exclusively generated at the Sanofi Boston site. Administered doses were 1 mg/kg for IV administration and 3 mg/kg for PO administration. The test set comprised of 61 compounds, while the training data for various PC properties ranged from 250 to 500 compounds.

Results: ML predictions for PC parameters demonstrated reasonable accuracy, with less than a two-fold error (mean percentage error (MAPE) < 1) in the final model. Deep learning methods were found to be less accurate than classical ML approaches, possibly due to the limited amount of data in the training set. Both MAPE and root mean square error (RMSE) were employed to select the best model for each parameter. PBPK-based predictions utilizing ML-derived parameters were simulated and the resulting profile was compared to the observed rat PK exposure data. Generally, the PBPK model adequately described the PK profile of most compounds administered by IV or PO. Furthermore, simulations revealed significant variations in the PK profile characteristics depending on the distribution model employed.

Conclusions: Conclusively, the ML/PBPK framework presented in this study yields satisfactory IV/PO exposure predictions. Our research underscores the importance of incorporating various distribution models when predicting PK from molecular structure, alongside acknowledging the accompanying variability rather than relying solely on a single PK profile estimate. This endeavor is geared towards facilitating early PK prediction in drug development and, ultimately, aiding in the prioritization of compounds for further evaluation.

Citations:

1. Mavroudis, Panteleimon D., et al. "Application of machine learning in combination with mechanistic modeling to predict plasma exposure of small molecules." Frontiers in Systems Biology 3 (2023): 1180948


2. Liaw, A., and Wiener, M. J. R. N. (2002). Classification and regression by random Forest. MJRN 2 (3), 18–22.


3. Awad, M., and Khanna, R. (2015). “Support vector regression,” in Efficient learning machines: Theories, concepts, and applications for engineers and system designers (Berkeley, CA: Apress), 67–80.


4. Chen, Y., Jin, J. Y., Mukadam, S., Malhi, V., and Kenny, J. R. (2012). Application of IVIVE and PBPK modeling in prospective prediction of clinical pharmacokinetics: Strategy and approach during the drug discovery phase with four case studies. Biopharm. Drug Dispos. 33 (2), 85–98. doi:10.1002/bdd.1769


5. Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. (2020). “Message passing neural networks,” in Machine learning meets quantum physics. Editor K. T. Schütt (Cham: Springer International Publishing), 199–214.


6. Berezhkovskiy, L. M. (2004). Volume of distribution at steady state for a linear pharmacokinetic system with peripheral elimination. J. Pharm. Sci. 93 (6), 1628–1640.


7. Willmann, S., Lippert, J., and Schmitt, W. (2005). From physicochemistry to absorption and distribution: Predictive mechanistic modelling and computational tools. Expert Opin. Drug Metab. Toxicol. 1 (1), 159–168.


8. Poulin, P., and Theil, F. P. (2000). A priori prediction of tissue:plasma partition coefficients of drugs to facilitate the use of physiologically-based pharmacokinetic models in drug discovery. J. Pharm. Sci. 89 (1), 16–35


9. Rodgers, T., Leahy, D., and Rowland, M. (2005). Physiologically based pharmacokinetic modeling 1: Predicting the tissue distribution of moderate-to-strong bases. J. Pharm. Sci. 94 (6), 1259–1276


10. Schmitt, W. (2008). General approach for the calculation of tissue to plasma partition coefficients. Toxicol Vitro 22 (2), 457–467.


11. Open Systems Pharmacology (open-systems-pharmacology.org)