Acs JPCB 7b02914

You might also like

You are on page 1of 10

This is an open access article published under an ACS AuthorChoice License, which permits

copying and redistribution of the article or any adaptations for non-commercial purposes.

Article

pubs.acs.org/JPCB

Predicting a Drug’s Membrane Permeability: A Computational Model


Validated With in Vitro Permeability Assay Data
Brian J. Bennion,†,# Nicholas A. Be,†,# M. Windy McNerney,†,‡ Victoria Lao,† Emma M. Carlson,§
Carlos A. Valdez,∥ Michael A. Malfatti,† Heather A. Enright,† Tuan H. Nguyen,⊥ Felice C. Lightstone,†
and Timothy S. Carpenter*,†

Biosciences and Biotechnology Division, Lawrence Livermore National Laboratory, Livermore, California 94550, United States

War Related Illness and Injury Study Center, Veterans Affairs, Palo Alto, California 94304, United States
§
U.S. Naval Academy, Annapolis, Maryland 21402, United States
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.


Nuclear and Chemical Sciences Division, Lawrence Livermore National Laboratory, Livermore, California 94550, United States

Global Security Directorate, Lawrence Livermore National Laboratory, Livermore, California 94550, United States
*
S Supporting Information
Downloaded via 157.47.108.80 on December 1, 2022 at 04:48:53 (UTC).

ABSTRACT: Membrane permeability is a key property to consider


during the drug design process, and particularly vital when dealing with
small molecules that have intracellular targets as their efficacy highly
depends on their ability to cross the membrane. In this work, we
describe the use of umbrella sampling molecular dynamics (MD)
computational modeling to comprehensively assess the passive
permeability profile of a range of compounds through a lipid bilayer.
The model was initially calibrated through in vitro validation studies
employing a parallel artificial membrane permeability assay (PAMPA).
The model was subsequently evaluated for its quantitative prediction of
permeability profiles for a series of custom synthesized and closely
related compounds. The results exhibited substantially improved
agreement with the PAMPA data, relative to alternative existing
methods. Our work introduces a computational model that underwent progressive molding and fine-tuning as a result of its
synergistic collaboration with numerous in vitro PAMPA permeability assays. The presented computational model introduces
itself as a useful, predictive tool for permeability prediction.

■ INTRODUCTION
Most drugs need to pass through at least one cellular
chemical properties of small molecules for the process of
membrane binding and diffusion are lipophilicity, molecular
membrane to reach their intended target. Although tight weight, and measures of molecular polarity.3 Thus, the
binding of a drug molecule to its intended target is important development of a successful drug involves a fine balance
for potency, poor membrane permeability often translates into among all these properties in a molecular scaffold, which is no
poor or nonexistent in vivo efficacy. Thus, a detailed trivial matter.
understanding of the partitioning of a given species in the Membrane permeability is a key metric in the drug design
membrane is vitally important from a pharmacokinetics and pipeline. A drug intended for an intracellular target, but with
rational drug design standpoint. In eukaryotic systems, two poor permeability will have low efficacy. As a result of these
possible transport modes are available for a molecule to pass characteristics several well-characterized in vitro and in silico
through a membrane: active and passive.1 Active transport permeability prediction methods have been developed. These
involves a transport protein that uses energy (e.g., ATP methods are relatively high throughput, and are important in
hydrolysis) to shuttle a molecule across a membrane. In the early stages of the drug discovery process with some of the
contrast, passive transport, which is the most common mode of most common and relatively simple in vitro methods being the
drug passage through membranes, involves diffusion of a parallel artificial membrane permeability assay (PAMPA), the
molecule through the membrane with no outside assistance or immobilized artificial membrane (IAM) technique, and
energy input. The rate of passive diffusion across a membrane is immobilized liposome chromatography (ILC). The PAMPA
proportional to the partition coefficient of the compound in vitro technique was developed in 1998 by Kansy et al.4 and
between the membrane (lipophilic environment) and the
external medium (aqueous milieu), the diffusion coefficient of Received: March 27, 2017
the compound through the membrane, and the compound’s Revised: April 28, 2017
concentration gradient across the membrane.2 Important Published: April 28, 2017

© 2017 American Chemical Society 5228 DOI: 10.1021/acs.jpcb.7b02914


J. Phys. Chem. B 2017, 121, 5228−5237
The Journal of Physical Chemistry B Article

Figure 1. Control compounds used in the calibration data set.

originally established to rapidly predict passive permeability interactions that a compound experiences during passage
through the gastrointestinal tract, but has since been adapted through the membrane. The interactions between the
for use in other systems such as the blood brain barrier (BBB),5 compound and the membrane are assumed to be governed
for which it demonstrated a good prediction of BBB by the chemical structure and properties of the compound,
penetration.5,6 Briefly, the technique involves a donor and an which are termed descriptors. As such, a mathematical model of
acceptor compartment separated by a filter supporting a liquid biological permeability is optimized based on a combination of
artificial membrane. The artificial membrane can be composed a variety of descriptors for the small molecule compound. Thus,
of a variety of phospholipid mixtures. The compound to be a change in structure can result in a change in permeability.
tested is placed in the donor compartment and is allowed to QSPR models have been developed to model and predict,
permeate between the donor and the acceptor compartments among others, oral bioavailability,14 intestinal absorption,15
through the artificial membrane. As the assay can be performed Caco-2 permeability,16 and blood-brain barrier penetration.17
in 96-well plates, high throughput screening of drug candidates The degree of success of these models is extremely dependent
is possible. Alternatively, IAMs mimic the phospholipid on the compounds in the training set. Thus, the applicability of
environment of the cellular plasma membrane by anchoring the approach can be limited, and transferability can be a major
synthetic lipid analogues to silica particles in monolayer density. issue.18 Even successful QSPR models can be limited because
These particles are then used as the packing material for a high- they do not provide any insight into the mechanisms that
performance liquid chromatography column.7−10 The IAM govern the permeability.19
chromatographic retention factors are used to generate For prediction of passive membrane permeability, the most
predictions of membrane permeability. These systems have critical parameter in QSPR studies is lipophilicity.20 Lip-
shown reasonable results for prediction of permeability, despite ophilicity is a crucial factor governing passive membrane
the retention time in the column not reflecting actual passage partitioning,21 and calculated lipophilicity metrics are often
across the membrane.11,12 The ILC method is used to study utilized to predict drug absorption.22 The parameter that
solute-membrane interactions, where liposomes are non- determines the lipophilicity of a molecule is LogP (the partition
covalently immobilized to gel beads. The liposomes can have coefficient of the molecule between an aqueous and lipophilic
varied compositions, including numerous different phospholi- phase, usually water and octanol). LogP is a crucial factor
pids. Lipids extracted from human red cells have also been used governing passive membrane partitioning; an increase in LogP
in ILC assays.13 enhances permeability.21 While the partition coefficient is used
An alternative approach to assess membrane permeability is to calculate properties such as membrane permeability and
to employ computational methods. Most in silico prediction water solubility, it also has importance in the prediction of
approaches use quantitative structure-permeability relationship biological activities, ADME, and toxicological end points. Thus,
(QSPR) models. These predictive, quantitative models applied reliable prediction of LogP is of massive importance in the drug
in drug design are methods that utilize statistical relationships design process, as it is important to know lipophilic properties
extracted from experimental permeability measurements of a compound before synthesis.23 There are several methods
combined with the physiochemical properties of a set of used to calculate the LogP parameter for a initial, rapid
training compounds. Thus, in QSPR studies, the permeability prediction of lipophilicity and, thus, permeability. Atomic-based
of the compound is an outcome that is the result of the various LogP predictions take each atom’s contribution to the LogP of
5229 DOI: 10.1021/acs.jpcb.7b02914
J. Phys. Chem. B 2017, 121, 5228−5237
The Journal of Physical Chemistry B Article

the compound and combine them additively.24 This method is °C in an overnight basis in ethanol. In some instances, the
in essence a look-up table per atom and suited to smaller, less products precipitate out of the solution (purity >99% after
complex molecules. More advanced “hybrid” LogP methods use ethanol washes during filtration) while in some cases flash
the LogP contribution of each atom, as well as a contribution column chromatography (3:7 → 7:3 EtOAc:hexanes) is needed
from the neighboring atoms, and combine them with correction for their purification. Purity of the compounds was assessed by
1
factors. In a fragment-based approach, data determined H NMR (DMSO-d6) and in all cases found to be >98%.
experimentally from different chemical moieties, or chemical Molecular Dynamics Setup. The simulation setup and
“fragments” are modeled, and these contributions are again protocol was the same as our previous published inves-
added up with correction factors. The motivation is that tigation.36 In brief, each simulation contained a single copy of
atomic-based approaches do not always properly capture the compound harmonically restrained at specified locations
certain molecular subtleties. Thus, fragment-based methods along the z-axis in a solvated DOPC bilayer system. The
tend to be better for complex, larger molecules. However, the complete system contained 5124 water molecules, 72 DOPC
assumption is the molecule contains fragments that are similar molecules, and one small compound. A typical system
to those from which the model was constructed. contained a total of ∼19 300 atoms. This system is comparable
One of the most powerful (though computer intensive) in to similar studies. We acknowledge that larger membrane
silico techniques to simulate the molecular process of diffusion systems may be required to investigate the permeability of
at the atomic level is molecular dynamics (MD).25 Coupling larger molecules. The simulations were run using GROMACS
MD with free energy techniques provides a powerful tool to 4.5.5,37 with the Berger et al. force field used for the DOPC
study membrane permeability in detail.26 While computational molecules38 as adapted by the Tieleman group (http://wcm.
models can often correlate well with experiment, knowledge- ucalgary.ca/tieleman/downloads), the GROMOS 53A6 force
based QSPR methods are often less accurate (but inexpensive) field used for the small compounds,39 and the SPC model used
compared to MD methods.27 Although human membranes are to represent the water,40 with the simulation parameters the
a complicated mixture of lipids and proteins, with a judicious same as Carpenter et al.36 The topologies for the small
choice of lipid, a single bilayer can provide a good first molecule compounds were generated using the Automated
approximation of the physicochemical properties of a cellular Topology Builder server (https://atb.uq.edu.au).41−43 The
membrane. Indeed, phosphatidylcholine lipids are the major topology builder combines a knowledge-based approach with
phospholipid within cellular membranes.28 Previous studies quantum mechanical calculations to produce parameters that
have investigated the permeation of small compounds (and a are compatible and consistent with a given force field. During
limited number of drug-like molecules) through homogeneous the initial stage, the molecule was optimized at the HF/STO-
lipid bilayers.29−35 Thus, with increasing computing power, 3G level of theory. The molecule was then reoptimized at the
these methods will become an increasingly valuable tool for B3LYP/6-31G* level of theory44−46 in implicit water. The
parameter prediction. charges were initially estimated by fitting the electrostatic
The work described herein demonstrates robust consistency potential using a Kollman−Singh scheme.47 The Hessian
between the in vitro PAMPA approach and the predictive matrix was then calculated and used to estimate the force
capacity of MD calculations. Our predictive MD approach is constants for both the bonds and angles. All compounds were
calibrated using the equivalent PAMPA experimental measure- modeled in their neutral forms and other physiologically
ments. The semiquantitative predictive power of the method- relevant charged states. The ChemAxon/Chemicalize server48
ology is then rigorously tested using a custom synthesized set of was used to calculate the pKa values for the ionizable atoms of
structurally related compounds (compounds LLNL1− each compound, and thus determine the most prevalent charge
LLNL18), demonstrating the potential utility of these methods species at physiological pH of 7.4. The parameters for the
for permeability assessment of novel chemical entities.


calibration data set are freely available from the Automated
Topology Builder server.
EXPERIMENTAL AND THEORETICAL METHODS Free Energy Calculations. The potential of mean force
In this study, two sets of compounds were characterized, the (PMF) free-energy profiles for the partitioning of compounds
calibration data set and the prediction data set. The calibration was calculated using umbrella sampling simulations. A single
data set was selected to cover a wide range of experimentally harmonic restraint with a force constant of 1000 kJ mol−1 nm−2
known permeability values and is comprised of the following was applied to the z-axis (normal to the bilayer) distance
compounds: progesterone, chlorpromazine, promazine, atro- between the center of mass of the compound and the center of
pine, diazepam, theophylline, 2-PAM, Hi-6, and MMB4 (Figure mass of the DOPC bilayer. One hundred separate simulations
1). The calibration compounds were commercially obtained were performed, with the compound harmonically restrained in
(Sigma-Aldrich). The prediction data set (compounds increments of 0.1 nm along the z-axis direction. These 100
LLNL1−LLNL18) is a structurally related set of compounds simulations completely span the 10 nm distance from bulk
that were designed and synthesized in order to provide a set of water, through the entire membrane and out into bulk water
compounds with a narrow range of permeabilities with which to again. Each simulation window was initially run for ∼50 ns, for
rigorously test our methodology and predictive capabilities. a total of ∼5 μs of simulation time for a complete set. As the
General Synthetic Procedure. The prediction set of position of a molecule within the membrane can shift its pKa
compounds used in this study was synthesized via an amide value, most compounds were simulated in both their neutral
forming reaction between an amine and an ethyl ester species. forms and most physiologically relevant charge species,
The library of generated materials can be assembled from an resulting in ∼10 μs of simulation data for each compound. A
ethyl ester and a series of amines that serve as the point total of 27 different compounds were simulated for a combined
structural diversity origin. Since the carboxylic acid component ∼250 μs of simulation time. On the basis of our previous
(ethyl ester) is not an activated species, the amine was used in experience, the initial 20 ns of each simulation was discarded as
slight excess to it (1.2 equiv) and the mixture was heated to 70 equilibration time and as such only the final ∼30 ns of each
5230 DOI: 10.1021/acs.jpcb.7b02914
J. Phys. Chem. B 2017, 121, 5228−5237
The Journal of Physical Chemistry B Article

simulation was used for all subsequent analyses. The weighted profiles that were distributed around the calculated data
histogram analysis method,49 as implemented within GRO- according to the normal distribution of that data.
MACS, was used to calculate the PMF for each compound, and Resampling of the PMF curves was calculated using the
all free-energy profiles were normalized to zero in bulk water. bootstrapping method as implemented within the g_wham52
Diffusion and Permeability. The membrane permeability analysis program of GROMACS. Again, 50 separate PMF
of a small molecule was calculated using the same methodology profiles were generated, from which the mean and standard
as previously published.36 Briefly, information from the PMF deviations calculated. These standard deviations of the PMF
1D energy landscape was combined with diffusion coefficients were used as an additional quality control metric. Any PMF
within the membrane to calculate the membrane permeability profile that exhibited a standard deviation of >0.5 kcal mol−1
rates. The position-specific diffusion coefficients were calculated was subjected to extended simulation and supplementary
from the MD simulation data using the method of Hummer.50 sampling in order to reduce this deviation. The PMF histogram
For each umbrella sampling simulation, the compound overlap was also checked to ensure correct sampling.
positional variance and autocovariance as a function of lag In order to determine the error of the Peff, the data in the 50
time were calculated. These calculations were repeated for a diffusion profiles were combined with the data in the 50 PMF
number of subsamples of each trajectory. The resulting profiles using eq 3 to produce 50 similar, but different, Peffs.
autocovariance curves decay roughly exponentially with The standard deviation for this set of 50 Peffs was taken as the
increasing lag time. The characteristic time, τ, of each error.
autocovariance decay was estimated by making a least-squares The Parallel Artificial Membrane Permeability Assay
fit to the log of the autocovariance data. The diffusion (PAMPA). PAMPA was applied to screen compounds for
coefficient for each subsample was then calculated as: passive diffusion. This study employed the Gentest Precoated
PAMPA Plate System (Corning Discovery Labware). The
var(Z)
D(⟨Z⟩) = system is composed of a 96-well plate/insert system. Two fluid-
τZ (1) filled compartments (donor well and receiver well) are
separated by a polyvinylidene fluoride (PVDF) filter plate
where ⟨Z⟩ is the average of the z-axis position of the compound
precoated with a phospholipid-oil-phospholipid trilayer primar-
center of mass during the simulation and var. (z) is the variance
ily consisting of DOPC phospholipids.53 Compounds of
of the compound center of mass (the autocovariance at “lag
interest were dissolved in Hanks Balanced Salt Solution
zero”), and the average over the subsamples was calculated. The
(HBSS) at a concentration of 100 μM. Solutions were added
standard deviation of the diffusion coefficient over the
subsamples was also calculated for use in a sensitivity analysis, to donor wells at a volume of 0.3 mL. The filter plate,
described below. The effective membrane permeability, Peff, was containing acceptor wells filled with 0.2 mL blank HBSS, was
calculated from the effective resistivity, Reff, (as derived by placed on top of the donor plate. The entire system was
Marrink and Berendsen51), incubated for 5 h at 25 °C. Following incubation, solution was
removed from donor and acceptor wells. Compounds were
1 Z then reserved for analysis via ultra performance liquid
Peff
= R eff = ∫0 R (z ) d z
(2) chromatography (UPLC). Certain compounds exhibited very
low traversal to the acceptor well, such that material could not
where R(z) is the resistivity of every “slice” of the membrane at be detected via UPLC upon initial analysis. Such compounds
position “z”, were concentrated from donor and acceptor solution fractions
by extraction into methanol, followed by vacuum concentration
e β(ΔG(z)) and resuspension in the minimal amount of appropriate UPLC
R (z ) =
D(z) (3) solvent.
UPLC Conditions. Compounds in solution were quantified
where β is the inverse of the Boltzmann constant times the using the Acquity ultra performance liquid chromatography
temperature, and ΔG(z) is the free energy from the PMF (UPLC) system (Waters) with the Acquity BEH C18 column
calculations. The integral is over the width of the membrane. (1.7 μm × 2.1 mm × 50 mm) at a pressure of 7500 PSI. Here,
Error Estimation. As a metric for confidence in our results, 10 μL were injected for each sample. Detection was based on
the error of Peff was determined. In order to calculate this retention time and UV absorbance. Specific detection protocols
overall error of Peff, we first had to separately resample both the were determined, where necessary for each compound
PMF and diffusion profiles. The sensitivity analysis used a (Supporting Information).
random resampling of the position-specific diffusion coef- Permeability Calculations. The concentration data was
ficients. As described previously, for each compound, position- used to calculate effective permeability (cm/s) as follows:
specific diffusion coefficients were calculated as well as the
standard deviation of each position-specific coefficient. These −ln[1 − CA(t )/Ceq]
calculations were based on independent subsamples of the Pe =
trajectory from each umbrella-sampling simulation. Since the A × (1/VD + 1/VA) × t (4)
mean and variation of each calculated coefficient is an
independent calculation, the correlation of the variations is where A is the filter insert area (0.3 cm2), VD is the volume of
zero. The resampled diffusion coefficients were thus based on the donor well (0.3 mL), VA is the volume of the acceptor/
independent draws from a set of normal distributions, where receiver well (0.2 mL), CA(t) is the concentration in the
each such distribution has a mean and standard deviation acceptor well at time t, and t = incubation time (18 000 s). Ceq
corresponding to the calculated mean and standard deviation of was calculated according to the following equation:
each position-specific diffusion coefficient. We used 50 random
Ceq = [CD(t ) × VD + CA(t ) × VA]/(VD + VA) (5)
draws for each coefficient to produce a set of 50 diffusion
5231 DOI: 10.1021/acs.jpcb.7b02914
J. Phys. Chem. B 2017, 121, 5228−5237
The Journal of Physical Chemistry B Article

where CD(t) is the concentration in donor well at time t, and


the other values are the same as previously defined.
LogP Calculations. CLogP. CLogP is a proprietary method
that is owned by BioByte/Pomona College. This method is a
fragment-based approach, whereby experimentally determined
fragments are modeled using QSPR (rather than per atom).
The CLogP program54−56 has become the benchmark, or gold-
standard, method for LogP prediction in the last 35 years and is
the most extensively tested fragment-based LogP calculator
used in drug design.
miLogP. The Molinspiration method for LogP prediction
(miLogP) is also based on group/fragment contributions. The
fragment contributions were obtained by fitting a training set of
>12 000 mostly drug-like molecules with experimental LogP
data. The authors state that miLogP is calculated using the Figure 2. PMF free energy curves for the passage of the control
methodology developed by Molinspiration as a sum of compounds from water (x-axis >3 nm) to the center of the bilayer (x-
fragment-based contributions and correction factors.57,58 axis <1 nm). As the membranes are symmetrical, the absolute
Chemicalize LogP. The LogP calculations (termed “(c)- distances of the compound from the bilayer center are used generate
LogP”) implemented from the Chemaxon/Chemicalize serv- these curves. Generally, a negative free energy at the bilayer center
er48 are based on a pool of predefined fragments. This pool of correlates with a more permeable compound.
fragments is based on the data in Viswanadhan et al.,59 with a
few minor adaptations (such as redefining atomic types to
accommodate electron delocalization and contributions of ionic
forms, or considering the potential for hydrogen bonding to
form a six membered ring between suitable donor/acceptor
atoms).
MOE SLogP. Chemical Computing Group’s MOE60 uses a
hybrid LogP prediction method termed “SLogP”. This method
is based on an atomic contribution model.61 The SLogP
training set to calibrate the model was ∼7000 structures.

■ RESULTS
Our computational methodology is first calibrated against the
well-established PAMPA protocol using a set of diverse, highly
studied compounds. The predictive power of the approach is Figure 3. Comparison between the experimentally measured Peff
then tested with a synthesized set of structurally similar (using PAMPA, x-axis) and the computationally predicted Peff (using
compounds and compared against the equivalent predictive the PMF, y-axis). The plots are divided into regions of differing
power of readily available LogP calculators. permeability. A red region is defined as impermeable, an orange region
Calibration. We are focused on improving the predictive is defined as having a low permeability, a yellow region is defined as
capability of our membrane permeability model based on our having a medium permeability, and a green region is defined as having
previous work.36 In order to do this, we continue to run a high permeability.
simulations and carry out experimental assay validation on our
data set of “calibration” compounds (Figure 1). This set of is extremely good (R2 = 0.97). Using these measured
compounds is comprised of molecules that fit a specific set of experimental permeability rates, as well as permeability data
criteria: (1) a significant amount of previous experimental work from literature, we are able to divide our data into four groups;
is available to compare against our results, and (2) the high permeability (green), medium permeability (yellow), low
compounds must also cover a wide range of permeability values permeability (orange), and impermeable (red). The boundaries
to evaluate the performance of our model across this range. between these regions are defined by using the PeffPAMPA
The potential of mean force (PMF) free energy curves for midpoints between different compounds of known permeability
these compounds is generated from the umbrella sampling MD categories (identified from literature). The boundary between
simulations (Figure 2), and in keeping with previous results,36 it the “impermeable” and “low permeability” groups is chosen
shows that (as a general rule) if a compound has a higher free such that the size of the “low permeability” region is uniformly
energy in the middle of the bilayer, it will be less permeable. distributed about the PeffPAMPA value for the “low permeability”
The PMF values for the calibration compounds (along with control compound (theophylline). These cutoff values of the
diffusion data also extracted from the MD simulations) are used experimentally measured PeffPAMPA are determined such that
to determine Peff (effective permeability) of the compounds impermeable compounds have a log PeffPAMPA of ←6.14, low
(termed “PeffPMF”). These PeffPMF values are then compared to permeability compounds have a log PeffPAMPA of > −6.14 and <
the equivalent P eff measured from running the same −5.66, medium permeability compounds have a log PeffPAMPA of
compounds through in vitro PAMPA assays (termed > −5.66 and < −5.33, and high permeability compounds have a
“PeffPAMPA”), which also measure passive permeability (Figure log PeffPAMPA of > −5.33. By fitting against the linear regression
3). line, we can use these in vitro cutoff values to define the
The linear correlation between our predicted P effPMF equivalent cutoff values for our predicted log PeffPMF (Figure 3
permeability rates and those experimentally measured PeffPAMPA and Table 1). Thus, the cutoff values for our predictions are
5232 DOI: 10.1021/acs.jpcb.7b02914
J. Phys. Chem. B 2017, 121, 5228−5237
The Journal of Physical Chemistry B Article

Table 1. Experimentally Measured and Computationally better outcome than obtaining false negatives that actually do
Predicted Permeation Rates of the Calibration Compounds possess permeability, but are predicted to be impermeable.
Comparison to LogP Predictions. The semiquantitative
permeability prediction results generated from our PMF-based
methodology are compared with the equivalent calibration and
predictions using several commonly used LogP calculation
tools (Table 3). LogP prediction tools are used in lieu of QSPR
models, which are primarily built upon LogP values. These
LogP calculation tools are the atomic/hybrid methodology of
MOE’s SLogP, and the fragment-based methods of miLogP
(from Molinspiriation), (c)LogP (from Chemicalize), and
CLogP (from Biobyte/Pomona College). The details of these
techniques are described in the Methods section. The LogP
calculation tools are subjected to the same analytic approach as
for the PeffPMF results. The various LogP calculations are first
calibrated against the PeffPAMPA measurements for the control
compounds to define LogP cutoffs between the four different
permeability categories. These defined thresholds for each of
the LogP tools are used to categorize our 18 custom
impermeable compounds, log PeffPMF < −2.05; low permeabilty compounds (SI, Figures S1 and S2).
compounds, log PeffPMF > −2.05 and <0.15; medium The LogP prediction tools performed more poorly than our
permeabilty compounds, log PeffPMF > 0.15 and <1.62; high PMF-based methodology in both the correlation against the
permeability compounds, log PeffPMF > 1.62. These values are in control compounds and the semiquantitative prediction. While
agreement with our previous studies, where the boundary the correlation coefficient for our PMF-based methodology is
between impermeable/poorly permeable and highly permeable 0.97, the range for the LogP tools is only 0.44−0.75. The LogP
is defined as a log PeffPMF of ∼0.36 By defining what the PeffPMF tools have difficulty with the unusual structures of our negative
cutoffs and thresholds are for these different permeability controls that are known to be impermeable. Furthermore, all of
categories, we are able to compute the PeffPMFs for novel the LogP tools have a significant number of false negative
compounds and use those values to predict the permeability permeability predictions, with no method able to correctly
categories for those compounds. identify more than two (of the eight) compounds that exhibit
Prediction. Using these cutoffs, we attempt to semi- any form of experimentally categorized permeability. In fact,
quantitatively predict the permeability of 18 of our synthesized two of the methods (SLogP and miLogP) fail to identify any of
set of structurally related compounds. This prediction is these compounds that have categorized permeability. Indeed,
significantly more difficult than correlating the control regardless of calibration, each of the LogP calculation tools
compounds, as the range of permeabilities of these compounds assign a higher LogP (greater permeability) to some
is much smaller. Despite this, however, we were able to compounds measured to be impermeable than those measured
correctly predict the permeability category of 78% (14 of 18) of to have a medium (or even high) permeability. In comparison,
the novel compounds (Figure 4). The four compounds that are our PMF-based methodology has no such occurrences and
correctly identifies all of the compounds with experimentally
measured categorized permeability. Indeed, even the four
impermeable compounds that our method incorrectly catego-
rize as low permeability compounds still all have lower
predicted permeability that the four compounds correctly
identified to have low permeability. Thus, a “manual
recalibration” of the impermeable/low permeability threshold
cutoffs would allow our data to be categorized such that all 18
compounds are predicted correctly. This is impossible to
achieve with any of the LogP data sets. Additionally, the
compounds miscategorized by the PMF method are only a
single cutoff (predicts to be “low permeability” but measures as
“impermeable”) removed from their correct region, while many
Figure 4. Prediction of the permeability category of our set of 18
of the LogP predictions are misplaced by two, or even three
structurally related compounds. Almost 80% of the compounds had
their permeability category correctly predicted. regions (predicts to be “impermeable” but measures as “high
permeability”). As can be expected, the fragment-based
approaches perform marginally better than the atomic/hybrid
method. CLogP is the top performing LogP tool (albeit only
assigned into the wrong permeability category are computa- slightly, with only a ∼ 60% success rate and only predicting 2/8
tionally predicted to have a “low permeability”, while permeable compounds). This is again anticipated as CLogP is
the current gold-standard tool for LogP prediction.


experimentally they are classified as impermeable (Table 2).
Thus, we have obtained a few “false positive” results. However,
we have successfully identified (and categorized) all eight of the DISCUSSION
compounds with some level of permeability (low, medium, or Assessing permeability is of critical importance to under-
high). From a drug design point of view, a false positive is a standing and predicting in vivo efficacy and bioavailability of
5233 DOI: 10.1021/acs.jpcb.7b02914
J. Phys. Chem. B 2017, 121, 5228−5237
The Journal of Physical Chemistry B Article

Table 2. Measured and Predicted Permeabilities for the Set of Structurally Related LLNL Compounds
permeability category
compound logPeffPMF log PeffPAMPA measured predicted
LLNL1 −6.52 ± 0.39 −1.17 ± 0.33 impermeable low
LLNL2 −6.53 ± 0.23 −1.69 ± 0.21 impermeable low
LLNL3 −6.12 ± 0.29 −0.16 ± 0.24 low low
LLNL4 −6.52 ± 0.21 −0.90 ± 0.29 impermeable low
LLNL5 −6.57 ± 0.17 −2.27 ± 0.30 impermeable impermeable
LLNL6 −6.73 ± 0.74 −3.90 ± 0.19 impermeable impermeable
LLNL7 −6.45 ± 0.26 −3.28 ± 0.27 impermeable impermeable
LLNL8 −6.84 ± 0.12 −0.93 ± 0.26 impermeable low
LLNL9 −6.46 ± 0.41 −3.57 ± 0.30 impermeable impermeable
LLNL10 −5.50 ± 0.20 0.27 ± 0.42 medium medium
LLNL11 −5.91 ± 0.15 −0.10 ± 0.51 low low
LLNL12 −5.80 ± 0.14 −0.73 ± 0.25 low low
LLNL13 −6.86 ± 0.41 −2.59 ± 0.65 impermeable impermeable
LLNL14 −7.05 ± 0.17 −3.82 ± 0.18 impermeable impermeable
LLNL15 −6.10 ± 0.37 −0.64 ± 1.75 low low
LLNL16 −5.83 ± 0.24 −0.64 ± 0.30 low low
LLNL17 −5.36 ± 0.16 1.46 ± 0.19 medium medium
LLNL18 −5.23 ± 0.28 1.82 ± 0.12 high high

Table 3. Comparison between the Characterized Permeability of Novel Synthesized Compounds LLNL1-18 as Predicted Using
PMF Methods versus Various LogP Prediction Tools
method calibration compound correlation (R2) LLNL1−LLNL18 compounds correct (%) false positives false negatives “permeable” compounds correct
PMF 0.97 78 4 0 8/8
SLogP 0.53 56 0 8 0/8
miLogP 0.75 56 0 8 0/8
(c)LogP 0.45 56 1 7 1/8
CLogP 0.44 61 1 6 2/8

candidate drug compounds. A range of physiological mem- the calibration compounds are structurally diverse and their
branes are relevant to such concerns. Permeability across experimental permeabilities are spread over ∼5 log units, while
gastrointestinal membranes is critical to bioavailability62 the prediction set are structurally similar and only spread over
following oral administration. Transdermal permeation must <2 log units. Calculations based on this prediction set are thus a
be evaluated for topical administration and management of skin challenging task, especially as the properties of many of the
pathologies.63 Traversal across the BBB represents a particularly compounds are so similar. Furthermore, several of the
confounding challenge, and is critical to delivery of compounds experimental permeabilities of the prediction set are measured
requiring central nervous system access for treatment of at the lower end of the sensitivity threshold of the PAMPA
infection, neurologic malignancies, and neurodegenerative methodology, resulting in higher relative uncertainty in this
disease,64,65 among other maladies. Computational prediction region due to the inherent challenge of experimentally
of permeability represents a need that has been historically measuring very low levels of diffusion precisely (Figure 4).
challenging and is often limited by the requirement for a robust Amplified relative uncertainty at the low end of the
training set, which ultimately restricts applicability of the permeability spectrum is correspondingly observed in computa-
resultant model. This study sought to address this need by tional predictions; we have previously shown that equivalent
refining and prospectively validating an MD-based model, errors in the PMF profile have a larger effect on the calculated
demonstrating effective prediction of permeability for candidate Peff when the compound has a lower permeability versus a
compounds across a physiological lipid membrane. higher permeability (if there is an energy barrier at the bilayer
The results of our study show that the developed center rather than an energy well).36 Thus, any error in
computational model can predict the PAMPA-defined perme- calculating the PMF for these poorly permeable compounds
ability category of a compound with greater accuracy than will be reflected in larger error in the Peff.
compared LogP-based methods. This is demonstrated both These results are a significant improvement in accuracy over
when examining well-characterized calibration compounds predictions made using current high throughput LogP
exhibiting a wide range of permeation capacity, as well when calculation techniques. Indeed, some of the LogP-based
prospectively studying a set of novel compounds with a more methods fail completely when attempting to categorize the
narrow dynamic range. The methodology is extremely permeability of the compounds. Depending on which LogP
successful in characterizing the nine selected calibration tool is used, the same compound could be characterized into
compounds (R2 = 0.97), and is able to correctly categorize three different regions. Furthermore, our PMF-based predicted
∼80% of our internal “prediction set” of 18 compounds. permeabilities are not dependent on training sets, but only the
As expected, the correlation with the calibration compounds physical properties recapitulated in the MD simulations. Our
is higher than with the prediction set of compounds because first-principles method is especially useful for a novel set of
5234 DOI: 10.1021/acs.jpcb.7b02914
J. Phys. Chem. B 2017, 121, 5228−5237
The Journal of Physical Chemistry B Article

compounds for which there is no QSPR or experimental data under the auspices of the U.S. Department of Energy by
available, and we do not run the risk of overfitting our model to Lawrence Livermore National Laboratory under Contract DE-
a narrow spectrum of physicochemical parameters. AC52−07NA27344. LLNL-JRNL-725792.
Limitations in permeability and, ultimately, bioavailability, of
candidate drug compounds can substantially slow and derail the
development process. Ideally, computational model systems
■ ABBREVIATIONS
MD, molecular dynamics; PAMPA, parallel artificial membrane
would serve as a first pass screen for permeation. Existing permeability assay; IAM, immobilized artificial membrane; ILC,
methods such as QSPR and LogP are not computationally immobilized liposome chromatography; BBB, blood brain
intensive, but are restrictive in their input capacity, and were barrier; QSPR, quantitative structure-permeability relationship;
shown in our study to produce a large proportion of false PMF, potential of mean force
negatives, thereby running the risk of ruling out potentially
promising leads. Although MD-based simulations are more
computationally intensive than LogP calculations (a few days
■ REFERENCES
(1) Seddon, A. M.; Casey, D.; Law, R. V.; Gee, A.; Templer, R. H.;
compared to a few minutes), the result is a more accurate, Ces, O. Drug interactions with lipid membranes. Chem. Soc. Rev. 2009,
actionable model that does not discard valuable compounds. 38, 2509−19.
Further, the computational time required for the method (2) Wolak, D. J.; Thorne, R. G. Diffusion of macromolecules in the
described in our study is still less than the time required to brain: implications for drug delivery. Mol. Pharmaceutics 2013, 10,
synthesize, purify, and experimentally characterize the same 1492−504.
novel compound. (3) Banks, W. A. Characteristics of compounds that cross the blood-
From a drug design perspective, this predictive capability brain barrier. BMC Neurol. 2009, 9, S3.
(4) Kansy, M.; Senner, F.; Gubernator, K. Physicochemical high
would facilitate compound evaluation by ruling out imperme- throughput screening: parallel artificial membrane permeation assay in
able candidates with a demonstrably low false-negative rate. the description of passive absorption processes. J. Med. Chem. 1998,
The resultant data would provide a more in depth character- 41, 1007−10.
ization of the permeability of hits identified though high (5) Di, L.; Kerns, E. H.; Fan, K.; McConnell, O. J.; Carter, G. T. High
throughput virtual screening or analogues of existing lead throughput artificial membrane permeability assay for blood-brain
compounds. Application of the described predictive methods barrier. Eur. J. Med. Chem. 2003, 38, 223−32.
would facilitate faster iterative improvement, allowing for (6) Mensch, J.; Melis, A.; Mackie, C.; Verreck, G.; Brewster, M. E.;
selective retention of compounds with access to the intended Augustijns, P. Evaluation of various PAMPA models to identify the
physiological target. Such a model could represent an important most discriminating method for the prediction of BBB permeability.
component of a more comprehensive design pipeline, capable Eur. J. Pharm. Biopharm. 2010, 74, 495−502.
(7) Ong, S.; Liu, H.; Pidgeon, C. Immobilized-artificial-membrane
of evaluating leads independent of structure or novelty of the
chromatography: measurements of membrane partition coefficient and
given compound.


predicting drug membrane permeability. J. Chromatogr. A 1996, 728,
113−28.
ASSOCIATED CONTENT (8) Taillardat-Bertschinger, A.; Galland, A.; Carrupt, P. A.; Testa, B.
* Supporting Information
S Immobilized artificial membrane liquid chromatography: proposed
The Supporting Information is available free of charge on the guidelines for technical optimization of retention measurements. J.
ACS Publications website at DOI: 10.1021/acs.jpcb.7b02914. Chromatogr. A 2002, 953, 39−53.
(9) Verzele, D.; Lynen, F.; De Vrieze, M.; Wright, A. G.; Hanna-
Figures S1 and S2, the calibration plots and permeability Brown, M.; Sandra, P. Development of the first sphingomyelin
prediction plots for the LogP-based methods (PDF) biomimetic stationary phase for immobilized artificial membrane


(IAM) chromatography. Chem. Commun. (Cambridge, U. K.) 2012, 48,
AUTHOR INFORMATION 1162−1164.
(10) Yang, C. Y.; Cai, S. J.; Liu, H. L.; Pidgeon, C. Immobilized
Corresponding Author artificial membranes - Screens for drug membrane interactions. Adv.
*(T.S.C.) Biosciences and Biotechnology Division, L-367, Drug Delivery Rev. 1997, 23, 229−256.
Physical and Life Sciences Directorate, Lawrence Livermore (11) Osterberg, T.; Svensson, M.; Lundahl, P. Chromatographic
National Laboratory 7000 East Ave., Livermore, CA 94550. E- retention of drug molecules on immobilised liposomes prepared from
mail: carpenter36@llnl.gov. egg phospholipids and from chemically pure phospholipids. Eur. J.
Pharm. Sci. 2001, 12, 427−39.
ORCID (12) Stenberg, P.; Norinder, U.; Luthman, K.; Artursson, P.
Timothy S. Carpenter: 0000-0001-7848-9983 Experimental and computational screening models for the prediction
Author Contributions of intestinal drug absorption. J. Med. Chem. 2001, 44, 1927−37.
# (13) Beigi, F.; Gottschalk, I.; Lagerquist Hagglund, C. L.; Haneskog,
The manuscript was written through contributions of all
authors. All authors have given approval to the final version of L.; Brekkan, E.; Zhang, Y. X.; Osterberg, T.; Lundahl, P. Immobilized
liposome and biomembrane partitioning chromatography of drugs for
the manuscript. The authors denoted with the # label prediction of drug transport. Int. J. Pharm. 1998, 164, 129−137.
contributed equally. (14) Yoshida, F.; Topliss, J. G. QSAR model for drug human oral
Notes bioavailability. J. Med. Chem. 2000, 43, 2575−2585.
The authors declare no competing financial interest. (15) Suenderhauf, C.; Hammann, F.; Maunz, A.; Helma, C.;


Huwyler, J. Combinatorial QSAR modeling of human intestinal
ACKNOWLEDGMENTS absorption. Mol. Pharmaceutics 2011, 8, 213−224.
(16) Ano, R.; Kimura, Y.; Shima, M.; Matsuno, R.; Ueno, T.;
We thank the Defense Threat Reduction Agency (DTRA) for Akamatsu, M. Relationships between structure and high-throughput
funding (CB14-CBS-03-1-0127). We thank D. A. Kirshner for a screening permeability of peptide derivatives and related compounds
helpful discussion and the Livermore Computing Grand with artificial membranes: application to prediction of Caco-2 cell
Challenge for the computer time. This work was performed permeability. Bioorg. Med. Chem. 2004, 12, 257−264.

5235 DOI: 10.1021/acs.jpcb.7b02914


J. Phys. Chem. B 2017, 121, 5228−5237
The Journal of Physical Chemistry B Article

(17) Liu, R. F.; Sun, H. M.; So, S. S. Development of quantitative hydration, constant pressure and constant temperature. Biophys. J.
structure-property relationship models for early ADME evaluation in 1997, 72, 2002−2013.
drug discovery. 2. Blood-brain barrier penetration. J. Chem. Inf. (39) Oostenbrink, C.; Villa, A.; Mark, A. E.; Van Gunsteren, W. F. A
Comput. Sci. 2001, 41, 1623−1632. biomolecular force field based on the free enthalpy of hydration and
(18) Stouch, T. R.; Kenyon, J. R.; Johnson, S. R.; Chen, X. Q.; solvation: The GROMOS force-field parameter sets 53A5 and 53A6. J.
Doweyko, A.; Li, Y. In silico ADME/Tox: why models fail. J. Comput.- Comput. Chem. 2004, 25, 1656−1676.
Aided Mol. Des. 2003, 17, 83−92. (40) Berendsen, H. J. C.; Postma, J. P. M.; van Gunsteren, W. F.;
(19) Swift, R. V.; Amaro, R. E. Back to the future: can physical Hermans, J. Interaction models for water in relation to protein
models of passive membrane permeability help reduce drug candidate hydration. In Intermolecular Forces; Pullman, B., Ed.; D. Reidel:
attrition and move us beyond QSPR? Chem. Biol. Drug Des. 2013, 81, Dordrecht, The Netherlands, 1981; pp 331−342.
61−71. (41) Malde, A. K.; Zuo, L.; Breeze, M.; Stroet, M.; Poger, D.; Nair, P.
(20) Liu, X.; Testa, B.; Fahr, A. Lipophilicity and Its relationship with C.; Oostenbrink, C.; Mark, A. E. An Automated force field Topology
passive drug permeation. Pharm. Res. 2011, 28, 962−977. Builder (ATB) and repository: Version 1.0. J. Chem. Theory Comput.
(21) Refsgaard, H. H. F.; Jensen, B. F.; Brockhoff, P. B.; Padkjaer, S. 2011, 7, 4026−4037.
B.; Guldbrandt, M.; Christensen, M. S. In silico prediction of (42) Canzar, S.; El-Kebir, M.; Pool, R.; Elbassioni, K.; Mark, A. E.;
membrane permeability from calculated molecular parameters. J. Med. Geerke, D. P.; Stougie, L.; Klau, G. W.; et al. Charge group
Chem. 2005, 48, 805−811. partitioning in biomolecular simulation. J. Comput. Biol. 2013, 20,
(22) Leung, S. S. F.; Mijalkovic, J.; Borrelli, K.; Jacobson, M. P. 188−198.
Testing physical models of passive membrane permeation. J. Chem. Inf. (43) Koziara, K. B.; Stroet, M.; Malde, A. K.; Mark, A. E. Testing and
Model. 2012, 52, 1621−1636. validation of the Automated Topology Builder (ATB) version 2.0:
(23) Katritzky, A. R.; Kuanar, M.; Slavov, S.; Hall, C. D.; Karelson, prediction of hydration free enthalpies. J. Comput.-Aided Mol. Des.
M.; Kahn, I.; Dobchev, D. A. Quantitative correlation of physical and 2014, 28, 221−33.
chemical properties with chemical structure: utility for prediction. (44) Becke, A. D. Density-functional thermochemistry. 3. The role of
Chem. Rev. 2010, 110, 5714−5789. exact exchange. J. Chem. Phys. 1993, 98, 5648−5652.
(24) Ghose, A. K.; Crippen, G. M. Atomic physicochemical (45) Lee, C. T.; Yang, W. T.; Parr, R. G. Development of the Colle-
parameters for three-dimensional-structure-directed quantitative struc- Salvetti correlation-energy formula into a functional of the electron-
ture-activity relationships. 2. Modeling dispersive and hydrophobic density. Phys. Rev. B: Condens. Matter Mater. Phys. 1988, 37, 785−789.
interactions. J. Chem. Inf. Model. 1987, 27, 21−35. (46) Perdew, J. P.; Wang, Y. Accurate and simple analytic
(25) Borhani, D. W.; Shaw, D. E. The future of molecular dynamics representation of the electron-gas correlation-energy. Phys. Rev. B:
simulations in drug discovery. J. Comput.-Aided Mol. Des. 2012, 26,
Condens. Matter Mater. Phys. 1992, 45, 13244−13249.
15−26. (47) Singh, U. C.; Kollman, P. A. An approach to computing
(26) Deng, Y. Q.; Roux, B. Computations of standard binding free
electrostatic charges for molecules. J. Comput. Chem. 1984, 5, 129−
energies with molecular dynamics simulations. J. Phys. Chem. B 2009,
145.
113, 2234−2246.
(48) http://www.chemicalize.org.
(27) Swift, R. V.; Amaro, R. E. Modeling the pharmacodynamics of
(49) Kumar, S.; Bouzida, D.; Swendsen, R. H.; Kollman, P. A.;
passive membrane permeability. J. Comput.-Aided Mol. Des. 2011, 25,
Rosenberg, J. M. The Weighted Histogram Analysis Method for free-
1007−1017.
energy calculations on biomolecules. 1. The method. J. Comput. Chem.
(28) Devaux, P. F. Static and dynamic lipid asymmetry in cell
membranes. Biochemistry 1991, 30, 1163−73. 1992, 13, 1011−1021.
(29) Bemporad, D.; Essex, J. W.; Luttmann, C. Permeation of small (50) Hummer, G. Position-dependent diffusion coefficients and free
molecules through a lipid bilayer: A computer simulation study. J. Phys. energies from Bayesian analysis of equilibrium and replica molecular
Chem. B 2004, 108, 4875−4884. dynamics simulations. New J. Phys. 2005, 7, 34.
(30) Boggara, M. B.; Krishnamoorti, R. Partitioning of nonsteroidal (51) Marrink, S. J.; Berendsen, H. J. C. Simulation of water transport
antiinflammatory drugs in lipid membranes: A molecular dynamics through a lipid membrane. J. Phys. Chem. 1994, 98, 4155−4168.
simulation study. Biophys. J. 2010, 98, 586−595. (52) Hub, J. S.; de Groot, B. L.; van der Spoel, D. g_wham-A free
(31) Holland, B. W.; Gray, C. G.; Tomberli, B. Calculating diffusion weighted Histogram Analysis implementation including robust error
and permeability coefficients with the oscillating forward-reverse and autocorrelation estimates. J. Chem. Theory Comput. 2010, 6,
method. Phys. Rev. E 2012, 86, 036707. 3713−3720.
(32) MacCallum, J. L.; Bennett, W. F. D.; Tieleman, D. P. (53) Chen, X.; Murawski, A.; Patel, K.; Crespi, C. L.; Balimane, P. V.
Distribution of amino acids in a lipid bilayer from computer A novel design of artificial membrane for improving the PAMPA
simulations. Biophys. J. 2008, 94, 3393−3404. model. Pharm. Res. 2008, 25, 1511−20.
(33) Marrink, S. J.; Berendsen, H. J. C. Permeation process of small (54) Leo, A. J. Calculating the hydrophobicity of chlorinated-
molecules across lipid membranes studied by molecular dynamics hydrocarbon solutes. Sci. Total Environ. 1991, 109-110, 121−130.
simulations. J. Phys. Chem. 1996, 100, 16729−16738. (55) Leo, A. J. The octanol water partition-coefficient of aromatic
(34) Orsi, M.; Essex, J. W. Permeability of drugs and hormones solutes - the effect of electronic interactions, alkyl chains, hydrogen-
through a lipid bilayer: insights from dual-resolution molecular bonds, and ortho-substitution. J. Chem. Soc., Perkin Trans. 2 1983,
dynamics. Soft Matter 2010, 6, 3797−3808. 825−838.
(35) Carpenter, T. S.; Parkin, J.; Khalid, S. The Free Energy of Small (56) Leo, A. J. Calculating Log P(oct) from structures. Chem. Rev.
Solute Permeation through the Escherichia coli Outer Membrane Has a 1993, 93, 1281−1306.
Distinctly Asymmetric Profile. J. Phys. Chem. Lett. 2016, 7, 3446−3451. (57) http://www.molinspiration.com.
(36) Carpenter, T. S.; Kirshner, D. A.; Lau, E. Y.; Wong, S. E.; (58) Ertl, P.; Rohde, B.; Selzer, P. Fast calculation of molecular polar
Nilmeier, J. P.; Lightstone, F. C. A method to predict blood-brain surface area as a sum of fragment-based contributions and its
barrier permeability of drug-like compounds using molecular dynamics application to the prediction of drug transport properties. J. Med.
simulations. Biophys. J. 2014, 107, 630−41. Chem. 2000, 43, 3714−3717.
(37) Hess, B.; Kutzner, C.; van der Spoel, D.; Lindahl, E. GROMACS (59) Viswanadhan, V. N.; Ghose, A. K.; Revankar, G. R.; Robins, R.
4: Algorithms for highly efficient, load-balanced, and scalable K. Atomic physicochemical parameters for 3 dimensional structure
molecular simulation. J. Chem. Theory Comput. 2008, 4, 435−447. directed Quantitative Structure - Activity Relationships. 4. Additional
(38) Berger, O.; Edholm, O.; Jahnig, F. Molecular dynamics parameters for hydrophobic and dispersive interactions and their
simulations of a fluid bilayer of dipalmitoylphosphatidycholine at full application for an automated superposition of certain naturally-

5236 DOI: 10.1021/acs.jpcb.7b02914


J. Phys. Chem. B 2017, 121, 5228−5237
The Journal of Physical Chemistry B Article

occurring nucleoside antibiotics. J. Chem. Inf. Model. 1989, 29, 163−


172.
(60) Molecular Operating Environment (MOE). 2013.08; Chemical
Computing Group Inc.: Montreal, QC, Canada, 2016.
(61) Wildman, S. A.; Crippen, G. M. Prediction of physicochemical
parameters by atomic contributions. J. Chem. Inf. Comput. Sci. 1999,
39, 868−873.
(62) Dahan, A.; Beig, A.; Lindley, D.; Miller, J. M. The solubility-
permeability interplay and oral drug formulation design: Two heads
are better than one. Adv. Drug Delivery Rev. 2016, 101, 99−107.
(63) Ferreira, M.; Silva, E.; Barreiros, L.; Segundo, M. A.; Costa
Lima, S. A.; Reis, S. Methotrexate loaded lipid nanoparticles for topical
management of skin-related diseases: Design, characterization and skin
permeation potential. Int. J. Pharm. 2016, 512, 14−21.
(64) Nair, M.; Jayant, R. D.; Kaushik, A.; Sagar, V. Getting into the
brain: Potential of nanotechnology in the management of NeuroAIDS.
Adv. Drug Delivery Rev. 2016, 103, 202−17.
(65) Sheng, R.; Tang, L.; Jiang, L.; Hong, L.; Shi, Y.; Zhou, N.; Hu, Y.
Novel 1-Phenyl-3-hydroxy-4-pyridinone derivatives as multifunctional
agents for the therapy of Alzheimer’s disease. ACS Chem. Neurosci.
2016, 7, 69−81.

5237 DOI: 10.1021/acs.jpcb.7b02914


J. Phys. Chem. B 2017, 121, 5228−5237

You might also like