Yorgos Psarellis, PhD: No relevant disclosure to display
Nikhil Pillai: No financial relationships to disclose
Objectives: Reliably predicting exposure prior to in-vivo studies can potentially accelerate drug discovery timelines and save a significant amount of preclinical resources [1]. The objective of this work is to use Machine Learning (ML) to predict concentration-time curves from physicochemical/ in vitro properties with quantifiable uncertainty. We focus on intravenous administration of small molecules in rats, and develop a Bayesian ML framework that can handle small datasets, well-suited for the preclinical space.
Methods: A dataset of 61 small molecules was curated from internal databases, collecting PK curves, physicochemical properties (molecular weight, pKa’s, fraction unbound, lipophilicity) and in vitro clearance. After data cleaning and postprocessing, a Gaussian Process Regression (GPR) [2] scheme was optimized and trained to predict whole PK curves for each compound. Additionally, Physiologically-Based Pharmacokinetic (PBPK) simulations [3] were obtained for the same compounds (using the inputs in the dataset) as an extra “in-silico” source of information. This resulted in a hybrid mechanistic-ML (or multifidelity), framework, where PBPK simulation outputs are calibrated based on the physicochemical/ in vitro inputs.
Results: The GPR trained on 80% of the dataset was able to capture well qualitative characteristics of the PK curves, and quantify the uncertainty of the predictions. Interestingly, the proposed ML framework is able to capture non-standard PK responses, such as multiple phases and complex transitions, owing to the non-parametric nature of the GPR framework. The predicted values were within 2-fold error of the observed PK (MAPE: 114%, MdAPE: 36%) Similar performance was observed for the hybrid (mechanistic-ML) framework (MAPE: 100%, MdAPE: 31%) . The reported prediction errors align well with the expected uncertainty in preclinical PK studies, resulting mainly from interanimal variability. Furthermore, the GPR framework is able to smoothly interpolate between sparse PK datapoints “filling in the gaps” while accounting for uncertainty.
Conclusions: The proposed Bayesian ML framework shows great promise in the preclinical space, as it can (i) handle noisy data, (ii) adapt to small datasets, (iii) quantify the uncertainty of the predictions. By fusing physicochemical, in vitro and in silico (PBPK) information, GPR can predict small molecule PK curves in rats after IV administration and capture important in vivo information qualitatively and quantitatively. Interestingly, PBPK simulations alone fail to capture important PK information, and therefore the ML framework proposed here presents as an alternative. This work can be considered as the “Proof-of-Concept”, for application to data-scarce modalities (such as mAbs) or the possibility of training ML “experts” for certain classes of molecules within the same modality.
Citations: [1] Pillai N., Abos A., Teutonico D., Mavroudis P.D., Machine learning framework to predict pharmacokinetic profile of small molecule drugs based on chemical structure. Clin Transl Sci. 2024; 17:e13824. https://doi:10.1111/cts.13824
[2] Rasmussen C.E., Williams C.K.I., Gaussian Processes for Machine Learning, the MIT Press, 2006, ISBN026218253X, www.GaussianProcess.org/gpml
[3] Willmann S., Lippert J., Sevestre M., Solodenko, J. Fois F., Schmitt W., PK-Sim®: a physiologically based pharmacokinetic ‘whole-body’ model, BIOSILICO (Vol. 1, Issue 4, pp. 121–124). Elsevier BV, 2003 https://doi.org/10.1016/s1478-5382(03)02342-4
