Original article
In silico prediction of brain and CSF permeation of small molecules using PLS regression models

https://doi.org/10.1016/j.ejmech.2007.11.011Get rights and content

Abstract

Computational partial least square (PLS) regression models were developed, which can be applied to predict central nervous system (CNS) penetration of drug-like organic molecules. For modeling, a dataset of 77 structurally diverse compounds was used with reported steady-state rat brain to plasma ratios (BPR). Information on steady-state cerebrospinal fluid distribution (CSF to plasma ratio or CSFPR) was available for 37 of these compounds. The molecules were from different chemical series and included bases, acids, zwitterions and neutral molecules. They were CNS active and were therefore assumed to penetrate the blood–brain barrier and/or the blood–liquor barrier. Using these PLS models, the dataset could be described accurately (r2 = 0.78, StErrorEst = 0.30 and r2 = 0.75, StErrorEst = 0.28 for BPR and CSFPR, respectively). Molecular descriptors used for the prediction of passive membrane transport were lipophilicity, polar and hydrophobic surface areas as well as structural parameters and net charge at physiological pH. There was no apparent correlation between experimental brain and CSF exposure. Consequently, different PLS models and guiding rules were developed and discussed for the prediction of BPR or CSFPR. The present models provide a cost-effective and efficient strategy to guide synthetic efforts in medicinal chemistry at an early stage of the drug discovery and development process.

Introduction

There is an increasing interest in predicting the process of passive translocation of drugs from the blood stream to the brain, in particular for pharmaceutical companies focusing on the development of drugs that act on targets in the central nervous system (CNS). The pharmacological activity of such CNS medicines not only depends on receptor affinity but also on the achieved compound concentration in brain. In many instances, however, access of chemicals to the brain is restricted at the level of the brain capillary endothelial wall that forms the blood–brain barrier (BBB). Direct measurements of BBB permeability or brain uptake of drugs is difficult and time-consuming and requires sophisticated in vitro experimental systems or animal experiments. Some examples of used techniques include in vitro models of the BBB [1], [2], in vivo pharmacokinetic and tissue distribution studies [3], [4] or in situ brain perfusion and capillary depletion experiments [5], [6]. In a clinical setting, access to brain tissue is not possible. It has therefore been suggested to use cerebrospinal fluid (CSF) as a surrogate marker for drug concentrations in brain tissue since this CNS compartment is accessible by lumbar puncture in human or ventricular puncture in experimental animals [7]. All mentioned techniques can be applied to detailed mechanistic studies with selected test compounds, however, their use for routine drug screening is not possible due to their limited throughput. This situation has created an interest in predictive in silico permeability models, which can be used to analyze compound libraries in an industrial setting and to guide medicinal chemists in drug discovery and development [8], [9], [10], [11].

It was the aim of the present study to develop and validate a computational blood–brain barrier permeation model, which can be used to predict the extent of passive uptake of drug-like organic molecules into brain tissue as well as CSF. We could thereby make use of a proprietary in-house database, which contains brain permeation data for 77 drug-like molecules and CSF exposure data for a subset of 37 molecules. All compounds were CNS active, derived from several structurally unrelated chemical series and were assumed not to be substrates of drug transporting proteins such as P-glycoprotein [12]. It was not possible to describe the dataset with existing brain permeation models, such as the quantitative structure–activity relationship (QSAR) model by Clark [13]. This as well as our interest in CNS permeation of drugs led to the development of a new and improved partial least square (PLS) regression model.

Section snippets

In vivo dataset

The proprietary in-house dataset in this study consists of CNS active compounds. These drug-like compounds are from different chemical classes covering a broad range of physicochemical and structural properties (Table 1). In the beginning, 91 compounds with reported brain to plasma ratios (BPR) were analyzed and a subset of 77 high-quality data was selected for this study (Table 2). Exclusion criteria were, for example, a poor aqueous solubility of the tested compounds (below 1 μg/mL),

Results and discussion

The distribution of molecules between the blood compartment and the central nervous system (CNS) is a complex process, which is the consequence of active and passive transport across two cellular barriers in the CNS. These are the blood–brain barrier (BBB), located at the endothelial cells of the brain vasculature, and the blood–liquor barrier (choroid plexus) [23]. One of the distinguishing properties of these barriers is the presence of high-resistance tight junctions, which seal the

Acknowledgements

We thank Dr. Sonia Poli and Dr. Philippe Coassolo for their advice and support. For the determination of distribution coefficients and solubility values we would like to thank Pia Warga, Virginie Micallef and Isabelle Parrilla.

References (33)

  • D.D. Shen et al.

    Adv. Drug Deliv. Rev.

    (2004)
  • D.E. Clark

    Drug Discov. Today

    (2003)
  • M. Lobell et al.

    J. Pharm. Sci.

    (2003)
  • D.E. Clark

    J. Pharm. Sci.

    (1999)
  • Y.L. Huang et al.

    Brain Res. Bull.

    (1996)
  • K. Takacs-Novak et al.

    J. Pharm. Biomed. Anal.

    (1996)
  • K. Takacs-Novak et al.

    J. Pharm. Biomed. Anal.

    (1994)
  • W.M. Meylan et al.

    J. Pharm. Sci.

    (1995)
  • W.M. Pardridge

    Adv. Drug Deliv. Rev.

    (1995)
  • R.P. Austin et al.

    J. Pharm. Sci.

    (1995)
  • M.A. Deli et al.

    Cell. Mol. Neurobiol.

    (2005)
  • M. Török et al.

    Exp. Brain Res.

    (2003)
  • G.Z. Zheng et al.

    Bioorg. Med. Chem. Lett.

    (2006)
  • T. Ooie et al.

    J. Pharmacol. Exp. Ther.

    (1997)
  • D. Triguero et al.

    J. Neurochem.

    (1990)
  • A. Cerletti et al.

    J. Drug Target.

    (2000)
  • Cited by (43)

    • Modeling ADMET properties based on biomimetic chromatographic data

      2023, Cheminformatics, QSAR and Machine Learning Applications for Novel Drug Development
    • In vitro blood brain barrier models: An overview

      2022, Journal of Controlled Release
      Citation Excerpt :

      As a result, it is crucial to the predictive power of estimations that the dataset selection process is correct. Passive diffusion as the major route of transport over the BBB is assumed when in silico predictions omit other BBB transport pathways, such as nanoparticle- or carrier-mediated transport, receptor-mediated transport and active efflux or influx transport [130]. Advanced strategies have been established to enhance the prediction capacity of these computer models.

    • In vitro models and systems for evaluating the dynamics of drug delivery to the healthy and diseased brain

      2018, Journal of Controlled Release
      Citation Excerpt :

      This method takes the advantage of currently available data on molecules with known logBB and subsequently models the correlation between logBB and molecular properties of the target molecules. Different molecular properties, known as molecular descriptors, such as molecular volume, solvent accessible surface area (SASA), polarity, hydrogen bond acidity and hydrogen bond basicity, polarizability, lipophilicity, solvation energy and octanol/water partition coefficient have been used in QSAR modeling [325,327–339]. Training of the model with a dataset is the basis of the QSAR technique, therefore both the quality and quantity of the training dataset are important.

    View all citing articles on Scopus
    View full text