Yorgos M. Psarellis 1, Nikhil Pillai 1, Saroj Dhakal 1, Xuelian Jia 2, Donato Teutonico 3, Panteleimon Mavroudis 1
1 Quantitative Pharmacology-Pharmacometrics, Sanofi (Cambridge, USA), 2 Center for Biomedical Informatics and Genomics, Tulane University (New Orleans, USA), 3 Quantitative Pharmacology-Pharmacometrics, Sanofi (Vitry-sur-Seine, France)
Objectives
Reliably assessing pharmacokinetics (PK) as early as possible in the R&D pipeline can potentially reduce animal studies, accelerate timelines to the clinic, and ultimately lead to molecules with higher probability of success. In this work, we employ Machine Leaning (ML) to learn PK- properties as well as full PK profiles of small molecules from structural information during intravenous (IV) administration. We demonstrate ML-enabled PK predictions for both preclinical species (rats) [1,2] and humans [3], highlighting their translational significance. We also explore how different ML strategies can be employed depending on the problem at hand: fit-for-purpose black-box models, hybrid approaches (ML + physiologically based PK – PBPK) for mechanistic insights, and Bayesian-ML for data-scarce settings.
Methods
Regarding preclinical PK predictions from molecular structure, in-house rat PK data were curated resulting in Clearance- (CL) and Volume-of-distribution (Vdss) assessments for 2928 small molecules, along with 911 concentration vs time profiles. For human PK predictions, a large dataset of physicochemical (PC) and PK properties was curated from public datasets, while concentration vs time profiles for approximately 773 compounds were digitized from literature. For consistency, the final PK datasets contain PK after single-dose IV administration normalized at 1mg/kg. Molecular structures described as SMILES strings were mapped to fingerprints or molecular descriptors [3] and used as input features in ML algorithms to predict CL, Vdss and concentration vs time profiles. In parallel, hybrid methods were explored, using PBPK models with ML predicted inputs. Lastly, Bayesian ML methods suitable for data-scarce settings were explored and tested in a small subset of the rat dataset (61 compounds).
Results
Trained ML models were able to qualitatively and quantitatively capture both preclinical (rat) and human PK profiles based on compounds’ molecular structure. Rat CL,Vdss prediction was achieved with 40% average relative error which is comparable to expected inter-animal variability. Regarding rats’ concentration vs time assessment, the trained ML algorithm predicts 72% of the compounds within 2-fold error and is able to capture qualitative PK characteristics (e.g. mono-,bi-exponential distribution phases). Reasonable exposure predictions were obtained also in the clinical setting: at least 55% of compounds within a 2-fold error, for both AUC and Cmax, as well as for the corresponding PK profiles. These results are compared with corresponding hybrid approaches (combining ML and PBPK) and adapted via Bayesian ML tο low-data regimes. Uncertainty quantification and explainability of ML frameworks is also explored, aligning ML performance at testing with the intuition of PK experts.
Conclusions
ML frameworks were successfully designed and optimized for the prediction of in-vivo PK (CL,Vdss, concentration vs time profiles) in rats and humans. Reliable PK predictions utilizing molecular structure alone, have the potential of great impact in PK-informed drug design, data-driven candidate ranking & selection, and first-in-human dose projections. Such approaches are aligned with modern pharmaceutical R&D timelines as well as with regulatory recommendations. Even though in this work we restrict our models to single-dose (1mg/kg) IV administration of small molecules, we discuss how our approaches can be adapted for different modalities, multiple dose levels and administration routes. Even in the specific case we are examining, a variety of ML-driven PK prediction approaches are available: black-box predictions for large datasets, hybrid predictions with ML/PBPK model, and Bayesian methods in low-data settings. These strategies reflect realistic scenarios a quantitative pharmacologist can encounter in preclinical and translational R&D and demonstrate the flexibility of data-driven frameworks to complement typical pharmacometrics approaches.
Funding: This work was funded by Sanofi US
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
[3] Jia, X., Teutonico, D., Dhakal, S., Psarellis, Y.M., Abos, A., Zhu, H., Mavroudis, P.D., Pillai, N., Journal of Medicinal Chemistry 2025 68 (7), 7737-7750, hhtps://10.1021/acs.jmedchem.5c00340
Reference: PAGE 34 (2026) Abstr 12306 [www.page-meeting.org/?abstract=12306]
Poster: Methodology – AI/Machine Learning