Berfin Gülave (1), Luke van Boven (1), Helle W van den Maagdenberg (1), Elizabeth CM de Lange (1), JG Coen van Hasselt (1)
(1) Division of systems pharmacology and pharmacy, Leiden Academic Centre for Drug Research, Leiden University, the Netherlands
Introduction/Objective: The rate and extent of blood-brain barrier (BBB) transport are critical parameters for central nervous system (CNS) drug distribution. The extent of BBB transport is captured by the partition coefficient of unbound drug concentrations at the BBB (Kp,uu,BBB), which is defined as the ratio of the unbound drug concentrations in brain extracellular fluid (brainECF) over plasma at steady state[1]. This parameter is an essential drug-specific parameter required for generating predictions for CNS drug disposition using pharmacokinetic (PBPK) modelling approaches, including the LeiCNS-PK3.0 model developed by our group[2]. Kp,uu,BBB values are typically generated through in vitro or in vivo assays, but are often lacking for both the majority of existing drugs as well as novel investigational compounds.
Quantitative structure-property relationship (QSPR) models can be used to predict Kp,uu,BBB data for novel compounds based on their chemical structural features, by training machine learning models on existing data for Kp,uu,BBB. Many of such QSPR models have been proposed for prediction of the partition coefficient for total concentrations, based on brain homogenate data[3-5], but lack pharmacological relevance as it is the unbound concentration in the brainECF which is responsible for pharmacological effects. Furthermore, to truly specify BBB transport, Kp,uu,BBB values that are based on brainECF concentrations are needed. Currently, only one such QSPR model has been published based on a relatively small number of compounds[6] .
The aim of this study was to develop a QSPR model to predict the Kp,uu,BBB based on a larger set of compounds, testing multiple machine learning algorithms, and to demonstrate how the developed QSPR model can be integrated into the LeiCNS-PK3.0 CNS PBPK model.
Methods: Rat Kp,uu,BBB values were obtained for 106 compounds from historical in house vivo experiments and literature. For all compounds, 2D and 3D physico-chemical and structural properties were calculated using the Molecular Operating Environment (MOE) software[7] , which allows computation of compound properties while accounting for the physiological pH value in the brain microvasculature. The resulting dataset was split randomly in a training dataset with 80% of compounds (n=86) and a test dataset with the remaining 20% of compounds (n=20). We confirmed that the test and training datasets had a similar distribution within the chemical space, through principal component analysis.
Multiple machine learning algorithms including random forest, support vector machines, K-nearest neighbours, and (sparse-) partial least square were trained to build regression models. The best performing model was selected based on the prediction error (R2) based on 10-fold cross validation. To quantify predictive performance, the root mean squared error (RMSE) and true R2 are estimated. The predictive performance was further evaluated by assessing the predictive performance of the 20% of compounds that were reserved as test dataset.
Results: From all evaluated machine learning algorithms, random forest model showed the best fit. The performance of this random forest model showed that the true R2 of the training and test set were 0.936 and 0.515, respectively, while the RMSE for training and test sets was 0.232 and 0.544, respectively and 73% of all the predictions were within 2-fold error. The most important predictors identified for the Kp,uu,BBB values were the ionization potential, the highest occupied molecular orbital energy, and the sum of hydrogen bond acceptor strengths. For these predictors, negative relationships were identified, indicating that a higher value of the descriptor is associated with a lower Kp,uu,BBB value. Finally, the random forest QSPR model can be directly integrated into the LeiCNS-PK3.0 PBPK model whereby Kp,uu,BBB values can be predicted and used to generate predictions of CNS drug disposition profiles, without the need or in vitro or in vivo studies.
Conclusions: The developed QSPR model showed adequate predictive performance in prediction of Kp,uu,BBB values. The model can be further applied to support the drug discovery and development of novel investigational drugs targeting the CNS, in particular when integrated with the LeiCNS-PK3.0 PBPK model.
References:
[1] Hammarlund-Udenaes M et al. Clin Pharm Res (2008) 25, 1737-150
[2] Saleh MAA et al. J PKPD (2021) 48, 725–741
[3] Faramarzi S et al. Front Pharmacol (2022) 13, 4486
[4] Loryan I et al. Mol Pharm (2015) 12, 520-532
[5] Fridén M et al. J Med Chem (2009) 52, 6233-6243
[6] Gupta M et al. ACS Chem Neurosci (2020) 11, 205-224
[7] Molecular Operating Environment (MOE), 2022.02
Reference: PAGE 32 (2024) Abstr 10995 [www.page-meeting.org/?abstract=10995]
Poster: Methodology – AI/Machine Learning