59, 335–345 (2001)
Copyright © 2001 by the Society of Toxicology
TOXICOLOGICAL SCIENCES
Prediction of Eye Irritation from Organic Chemicals Using
Membrane-Interaction QSAR Analysis
Amit Kulkarni,* A. J. Hopfinger,* ,1 Rosemarie Osborne,† Leon H. Bruner,‡ and Edward D. Thompson†
*Laboratory of Molecular Modeling and Design (M/C 781), College of Pharmacy, The University of Illinois at Chicago, 833 South Wood Street, Chicago,
Illinois 60612–7231; †The Procter & Gamble Company, Miami Valley Laboratories, P.O. Box 538707, Cincinnati, Ohio 45253– 8707; and
‡Procter & Gamble Technical Centre, Ltd., Lovett House, Lovett Road, Stains, Middlesex TW18 3AZ, England
Received May 30. 2000; accepted October 13, 2000
Eye irritation potency of a compound or mixture has traditionally been evaluated using the Draize rabbit-eye test (Draize et al.,
1944). In order to aid predictions of eye irritation and to explore
possible corresponding mechanisms of eye irritation, a methodology termed “membrane–interaction QSAR analysis” (MI-QSAR)
has been developed (Kulkarni and Hopfinger 1999). A set of
Draize eye-irritation data established by the European Center for
Ecotoxicology and Toxicology of Chemicals (ECETOC) (Bagley et
al., 1992) was used as a structurally diverse training set in an
MI-QSAR analysis. Significant QSAR models were constructed
based primarily upon aqueous solvation-free energy of the solute
and the strength of solute binding to a model phospholipid
(DMPC) monolayer. The results demonstrate that inclusion of
parameters to model membrane interactions of potentially irritating chemicals provides significantly better predictions of eye irritation for structurally diverse compounds than does modeling
based solely on physiochemical properties of chemicals. The specific MI-QSAR models reported here are, in fact, close to the upper
limit in both significance and robustness that can be expected for
the variability inherent to the eye-irritation scores of the ECETOC
training set. The MI-QSAR models can be used with high reliability to classify compounds of low- and high-predicted eye irritation
scores. Thus, the models offer the opportunity to reduce animal
testing for compounds predicted to fall into these two extreme
eye-irritation score sets. The MI-QSAR paradigm may also be
applicable to other toxicological endpoints, such as skin irritation,
where interactions with cellular membranes are likely.
Key Words: eye irritation; quantitative structure-activity relationships (QSAR); membrane models; rabbits; animal alternatives.
Eye irritation traditionally has been evaluated using the
Draize in vivo rabbit eye-irritation test (Draize et al., 1944). By
this method, a 0.1-ml (or weight equivalent) sample of test
material is placed into the lower conjunctival cul-de-sac of
albino rabbit eyes; responses of the cornea, and conjunctiva
and iris are graded at standard times, generally from 1 to 35
days after dosing. The tissue grades are combined into a
1
To whom correspondence should be addressed. Fax: 312.413.3479.
E-mail: [email protected].
weighted score; the highest average score across test animals
on the various grading days is termed the maximum average
score (MAS). Recent work indicates that the extent (area and
depth) of injury produced in the cornea is the principle factor
determining acute responses and their eventual repair in that
tissue (Jester et al., 1998). However, mechanisms of eye irritation are not yet understood on a biochemical level (Bruner et
al., 1998).
The in vivo rabbit eye-irritation test has frequently been
criticized on animal welfare grounds (Rowan, 1984). Many
laboratories have been working to develop in vitro alternatives
to this test (Balls et al., 1999; Brantom et al., 1997). At the
present time, the in vitro alternatives may have a role as
screens or adjuncts to the Draize rabbit-eye test, but none are
sufficiently well validated to replace the test completely (Balls
et al., 1999). International agencies have proposed and adopted
step-wise approaches for eye-irritation assessments with the
goal of reducing the need for animal eye-irritation tests (OECD
1996). Although structure-activity and structure-property analyses are recommended as early steps in the assessment process,
a systematic approach for these analyses has not yet been
widely accepted. The current work is directed toward this need.
Quantitative structure-activity relationship (QSAR) analysis
provides a tool to relate the magnitude of a particular property,
such as an eye-irritation score, to one or more physicochemical
and/or structural parameters of a molecule. Hence, QSAR
analysis can be used to estimate eye irritation. Traditional
QSAR methods are normally limited in application to series of
chemical analogs for which the dependent property (eye irritation) is derived from a set of intramolecular descriptors based
upon an assumed common mechanism of action. However,
eye-irritation assessments are normally sought for structurally
diverse compounds. Thus, QSAR analysis is relatively limited
in utility in applications that estimate eye irritation for diverse
classes of chemicals.
The European Center for Ecotoxicology and Toxicology of
Chemicals (ECETOC) established a “standard” data set for
chemicals whose Draize rabbit eye-irritation scores have been
measured according to OECD Guideline 405 (1987). The
335
336
KULKARNI ET AL.
ECETOC data set has come to be used as a standard in the
evaluation of in vitro and QSAR methods to estimate eye
irritation. A history of the applications of QSAR and molecular
modeling to eye irritation in general, and the ECETOC data set
in particular, has been given (Kulkarni and Hopfinger, 1999).
Several QSAR, data clustering, and molecular modeling studies have been performed using the ECETOC data set. However, all of these studies only employed intramolecular physicochemical properties of the compounds of the training set as
correlation descriptors (Barratt, 1995). These previous studies
were based on the then prevalent views on the application of
QSAR and modeling methods to preclinical drug discovery. It
has been generally assumed that predicting eye irritation is
methodology-equivalent to designing an active pharmaceutical
agent. None of the previous studies were successful in developing a significant statistical QSAR model spanning all the
compounds of the ECETOC data set, because this data set is
composed of structurally diverse chemicals.
In principle, progress might be made in the QSAR analysis
of any chemically diverse data set, including the ECETOC
eye-irritation data set, if the “receptor” linked to the eyeirritation response is known and included in constructing
QSAR models. This receptor-based approach to molecular
design has been successfully used in building high-affinity
ligands and is generally called structure-based design (Kubinyi, 1993). In the case of eye irritation, uptake and diffusion of
an irritant into the keratocytes of the corneal epithelium may be
a significant event. That is, each test molecule placed in the eye
must diffuse through the cell membrane of the keratocytes
comprising the outer 7 or so layers of the corneal epithelium of
the eye. We have thus hypothesized that interactions of test
molecules with cell membranes are at least partly, responsible
for eye irritation. Moreover, the phospholipid-rich regions of a
membrane bilayer of the cell might comprise the “general
receptor” for eye irritation.
In order to test this hypothesis, we simulated the uptake and
interaction of each of the ECETOC (solute) molecules with a
model phospholipid membrane, as a part of our QSAR analysis
of the ECETOC eye-irritation data set. In these simulations, the
estimated membrane-solute interaction properties from the molecular simulations are added to the intramolecular physicochemical property descriptors to provide an extended set of
trial descriptors for building eye-irritation QSAR models. This
overall methodology is called membrane-interaction QSAR
(MI-QSAR) analysis. The results indicate that MI-QSAR is a
promising approach for development of predictive models for
eye irritation.
MATERIALS AND METHODS
The difference between traditional QSAR methods and MI-QSAR analysis
is that the latter modeling considers not only physicochemical parameters
associated with the solute molecule, but also with the physicochemical parameters of the receptor (cell membranes in this case) and parameters that describe
the physicochemical interactions between the two. Accordingly, it is necessary
to characterize physicochemical parameters associated with the solute molecule, the proposed receptor, and the interaction between the two. It is also
necessary to define the in vivo endpoint to be predicted in terms relevant for the
analysis. These steps are described in detail below.
Molar eye-irritation scores. The dependent variable used in the MI-QSAR
analysis was the molar-adjusted eye score (MES), as calculated from Draize
rabbit eye-irritation test MAS values. This adjustment of the standard Draize
score was made since activities used in QSAR studies are normally expressed
as molar concentrations producing a fixed response. Thus, the MES was
determined as follows: The molarity of the each solute solution tested in vivo
was calculated using molarity ⫽ (density ⫻ 1000)/relative molecular mass of
the test chemical. Molar-adjusted eye scores were then calculated as the Draize
MAS values divided by the molarity of the solution. Table 1 contains the MES
values for the compounds of the ECETOC training set. Ionizable molecules
were not included in the training set because it is not clear if they are neutral
or charged when at, or in, the membrane. Both the neutral and charged forms
of these molecules could be considered in MI-QSAR analysis, but that was not
done in this application.
Building solute molecules and a DMPC monolayer. The 3-dimensional
structures of the solute molecules of the ECETOC training set (see Table 1)
were built using the Chemlab-II molecular modeling package (Pearlstein,
1988). A single dimyristoylphosphatidylcholine (DMPC) molecule was built
using HyperChem (1998) from available crystal structure data (Hauser et al.,
1981). The AM1 Hamiltonian in Mopac 6.0 (Mopac, 1990) was used for the
estimation of partial atomic charges on all molecules.
The phospholipid DMPC was selected as the model phospholipid in this
study and was used to build a membrane monolayer which serves as the
receptor for the eye-irritation response in the MI-QSAR analysis. The structure
of a DMPC molecule is shown in Figure 1. An assembly of 16 DMPC
molecules (4*4*1) in (x,y,z) directions, respectively, was used as the model
membrane monolayer. The size of the monolayer simulation system was
selected based on the work done by van der Ploeg and Berendsen (1982).
Additional information regarding construction of the model monolayer used in
this MI-QSAR analysis is given in (Kulkarni and Hopfinger, 1999).
We equilibrated the monolayer structure by performing molecular dynamic
simulations (MDS) at 311 degrees Kelvin (°K) for 50 pico s (ps). In order to
prevent unfavorable van der Waals interactions between a solute molecule and
the membrane DMPC molecules, one of the “center” DMPC molecules was
removed from the equilibrated monolayer and a test solute molecule inserted
in the space created by the missing DMPC molecule. Each of the test solute
molecules of the ECETOC data set was inserted at 3 different positions
(depths) in the DMPC monolayer with the most polar group of the solute
molecule “facing” toward the head group region of the monolayer. Three
corresponding MDS models were generated for each solute molecule with
regard to the trial positions of the solute molecule in the monolayer. The 3 trial
positions were, (1) solute molecule in the head group region, (2) solute
molecule between the head-group region and the aliphatic chains, and (3)
solute molecule in the tail region of the aliphatic chains.
The lowest energy geometry of the solute molecule in the monolayer was
sought using each of the 3 trial solute positions. The 3 different initial MDS
positions of styrene (one of the test solute molecules) are shown in Figure 2a
to illustrate this modeling procedure. The energetically most favorable geometry of styrene in the model DMPC monolayer is shown in Figure 2b.
Molecular dynamic simulations (MDS). MDS were carried out using the
Molsim package with an extended MM2 force field (Doherty, 1994). A simulation
temperature of 311°K was selected, since it is body temperature and is also above
the DMPC phase transition temperature. Temperature was held constant in the
MDS by coupling the system to an external fixed-temperature bath (Berendsen et
al., 1984). The trajectory step size was 0.001 ps over a total simulation time of 20
ps for each test compound. Two-dimensional periodic boundary conditions (PBC)
corresponding to the “surface plane” of the monolayer were employed (a ⫽ 32 A,
b ⫽ 32A, c ⫽ 80 A, and ␥ ⫽ 96.0°) for the DMPC molecules of the monolayer
model, but not the test solute molecule. By using periodic boundary condition it is
possible to simulate an infinite system. Also by using PBC simulations can be
337
PREDICTING EYE IRRITATION FROM ORGANIC CHEMICALS
TABLE 1
The ECETOC Draize Eye Irritation Data Set Along with the MES and Descriptor Values
of the MI-QSAR Models (Equations 3 and 4) of Each Solute Molecule
Solutes
Hydrocarbons:
3 Methyl hexane
2 Methyl pentane
Methylcyclopentane
1,9-Decadiene
Dodecane
1,5-hexadiene
cis-Cyclooctene
1,5 Dimethylcyclo-octadiene
Aromatics:
4-Bromophentole
2,4-Difluoronitrobenzene
3-Ethyltoluene
4-Fluoroaniline
Xylene
Toluene
Styrene
1-Methylpropylbenzene
1,3-Disopropylbenzene
Ketones:
Methyl amyl ketone
Methyl isobutyl ketone
Methyl ethyl ketone
Acetone
Alcohols:
n-Butanol
Isobutanol
Isopropanol
Propylene glycol*
2-Ethyl-1-hexanol
Glycerol*
Hexanol
Butyl cellsolve
Cyclohexanol
Acetates:
Ethyl acetate
Methyl acetate
Methyl trimethyl acetate
Ethyl trimethyl acetate
Cellosolve acetate
n-Butyl acetate
Ethyl-2-methylaceto-acetate
Acids:
2,2Dimethylbutanoic acid
F(H2O)
F(OCT)
E (chg ⫹ vdw)
E (vdw)
E (chg)
6.57
6.58
6.18
5.00
6.33
4.70
4.62
3.86
2.75
2.54
1.36
1.64
3.62
0.83
1.00
1.20
–2.60
–2.08
–2.80
–5.26
–5.52
–3.18
–4.48
–5.70
–17.31
–16.22
–7.00
–14.36
–24.48
–9.04
–14.70
–4.73
–4.09
–4.07
3.69
–1.20
–0.83
1.45
0.26
0.97
–13.22
–12.15
–10.69
–13.16
–23.65
–10.49
–14.96
–5.70
0.19
0.40
0.32
6.62
1.10
0.96
0.77
0.31
0.38
2.90
1.04
3.56
3.53
3.58
3.71
2.73
3.92
3.91
–4.20
5.20
–0.60
–13.68
–0.85
–0.93
–1.53
–0.26
0.09
–10.07
7.40
–5.54
–15.75
–5.06
–4.46
–5.39
–5.78
–6.66
–14.56
–9.20
–15.65
–4.04
–12.83
–4.58
–9.65
–5.36
–7.80
–8.92
–4.00
–0.64
6.50
–0.70
0.80
–0.30
–0.36
–7.48
–5.64
–5.20
–15.01
–10.54
–12.13
–5.38
–9.35
–5.00
–0.32
2.26
0.59
4.48
4.83
3.67
3.96
6.67
3.66
1.75
1.80
–7.30
–5.50
–0.73
0.11
–8.68
1.35
–9.47
–3.92
–0.67
–2.06
3.65
–0.51
6.74
–0.60
–13.12
–3.41
–7.41
–1.46
5.47
6.44
2.34
0.10
7.82
0.12
8.13
8.99
8.29
6.53
7.00
6.87
6.81
6.51
6.70
6.36
6.51
6.34
–7.45
–7.19
–7.51
–20.35
–6.49
–26.39
–7.04
–11.01
–8.08
–9.00
–8.68
–8.16
–15.80
–10.76
–23.44
–10.04
–11.85
–10.44
–18.37
–6.67
–1.69
–15.04
–9.16
–8.17
–9.04
–18.30
–10.74
–9.73
–0.26
3.94
–7.13
–3.23
0.49
1.45
0.15
0.76
–8.64
–6.41
–5.63
–7.91
–5.93
–8.66
–10.49
–18.45
–11.5
1.47
3.14
0.36
0.63
2.03
0.99
2.55
4.28
4.18
4.67
4.76
4.21
4.29
3.34
–2.82
–3.02
–2.16
–1.96
–6.83
–2.42
–2.92
–0.98
–0.46
–1.70
–2.22
–3.83
–2.02
–0.03
6.59
–15.79
–3.31
–14.31
–11.95
–27.29
–5.59
17.76
–6.82
1.57
–6.75
–3.49
–10.74
–6.77
–11.17
–8.97
–4.88
–7.56
–8.46
–16.55
1.18
5.59
4.66
–5.46
–7.72
–3.84
–2.69
–1.15
MES
LUMO
0.10
0.26
0.41
0.37
0.45
0.55
0.43
0.44
*Outliers.
performed on relatively smaller system in such a way as if the system experience
forces in bulk fluid. The angle ␥ is the angle an extended DMPC molecule makes
with the “planar surface” of the monolayer. Only a single solute molecule was
explicitly considered in each MDS.
An initial MDS on the model membrane, without a solute molecule present,
was carried out to allow for structural relaxation and distribution of the kinetic
energy over the monolayer. Each of the solute molecules was placed at each of
the 3 different positions in the monolayer, as described above, oriented with
the most polar portion of the solute toward the head-group region. The overall
simulation scheme is shown in Figure 3, and additional details of the membrane-solute MDS can be found in (Kulkarni and Hopfinger 1999).
Calculation of descriptors. Both intramolecular physicochemical properties and features of the solute molecules, and intermolecular solute-membrane
interaction properties were calculated. “Properties” and “features” will both be
referred to as descriptors from this point forward, as they constitute the trial
pool of independent variables used to build the QSAR models.
Only 2 of the descriptors from Table 2, HOMO and LUMO, were found to
338
KULKARNI ET AL.
FIG. 1. The chemical structure of a DMPC molecule having arbitrary
atom number assignments.
exhibit any type of correlation relationship to MES over the population of
QSAR models sampled in this study.
Some of the intermolecular solute-membrane interaction descriptors were
extracted directly from the MDS trajectories and are listed in Table 3. These
intermolecular descriptors were calculated using the most stable (lowest total
potential energy) solute-membrane geometry realized from MDS sampling of
the 3 initial positions (see Fig. 2a) for each of the solutes. Figure 4 shows a plot
of the total potential energy vs. MDS trajectory, from which the most energetically favorable position of styrene is identified. It can also be seen in Figure
4 that the membrane-styrene system attains energy equilibrium after about 6 ps
of MDS. That is, the energy vs. time plot is, on average, “flat” after 6 ps. The
most energetically favorable position for styrene is in the middle region of the
monolayer (see Fig. 2b) where the energy is minimum after the equilibration
phase. The intermolecular descriptors extracted from MDS trajectories are
given in Table 3. Details regarding the methods and algorithms used to
compute these descriptors can be found in (Kulkarni and Hopfinger, 1999).
F(H 2O) was, by far, the most significant MI-QSAR correlation descriptor.
Aqueous solvation free energies were calculated using a group additive version
of the hydration shell model (Hopfinger, 1973).
Construction and testing of MI-QSAR models. MI-QSAR models were
constructed using the genetic function approximation, GFA (Rogers and Hopfinger, 1994), which is a multi-dimensional optimization method based on the
genetic algorithm paradigm. The GFA algorithm is coded in the program
WOLF (Rogers, 1994). Statistical significance in the optimization of a QSAR
model using WOLF is based on Friedman’s lack of fit (LOF) measure (Friedman, 1988). The LOF measure is designed to resist overfitting, which is a
problem often encountered in constructing statistical models. Since the number
of descriptors available in MI-QSAR analysis normally exceeds the number of
observations (test compounds), the ability to prevent overfitting using GFA is
critical to the successful construction of a statistically significant MI-QSAR
model.
FIG. 2. (a) A “side” view of a styrene molecule inserted at 3 different
positions in the DMPC model monolayer prior to the start of each MDS used
in the MI-QSAR analysis; (b) the lowest-energy geometry of a DMPC-styrene
complex found in the MDS.
A smoothing factor of 1.8 and 100,000 crossover operations were used to
optimize the MI-QSAR models using WOLF. Optimization of a QSAR model
was considered to be realized when descriptor usage became constant and
FIG. 3.
details.
The schedule of performing a membrane-solute MDS. See text for
PREDICTING EYE IRRITATION FROM ORGANIC CHEMICALS
339
TABLE 2
Intramolecular Solute Descriptors Considered
in the Trial QSAR Descriptor Set
Intramolecular descriptors
Source
Kappa-2-AM (topological descriptors)
HOMO (highest occupied molecular orbital energy)
LUMO (lowest occupied molecular orbital energy)
Dipole moment
Molecular volume
SA (molecular surface area)
Density
Molecular weight
Molecular refractivity
Number of hydrogen bond acceptors
Number of hydrogen bond donors
Number of rotatable bonds
Jurs-Stanton CSPA (charged partial surface-area) descriptors
Kappa descriptors (topological descriptors)
Radius of gyration
PM (principle moment of inertia)
a
a
a
b
c
c
a
b
a
a
a
a
a
a
a
c
Note. The sources are indicated as follows: a, computed using Cerius2 (MSI,
1997); b, calculated using, MOPAC 6.0 (Mopac, 1990); and c, calculated using
Chemlab II (Pearlstein, 1988).
independent of increasing crossover operations. A crossover operation is the
“birth” of a child model from its parent models. Both partial least-squares
regression (PLS) and multi-dimensional linear regression (MLR) can be used
in WOLF to establish functional data fits. MLR was used in this MI-QSAR
eye-irritation study.
In order to test and validate the MI-QSAR models, the dependent variable,
MES, was randomly “scrambled” (Waterbeemd, 1995) with respect to the set
of independent variables (descriptor set) of the compounds, to see if meaningful correlations (QSARs) could be found among the scrambled data sets. The
absence of any significant correlation for each of the scrambled data sets is
taken as evidence of the significance of the MI-QSAR models with respect to
the original, non-scrambled data set. Covariance among the descriptors in the
optimized MI-QSAR models was evaluated by constructing the linear crosscorrelation matrix of the descriptors, and by comparing relative descriptor
FIG. 4. Energy vs. simulation time plot for the 3 different initial geometries
sampled for styrene in the DMPC model monolayer as shown in Figure 2.
usage in the crossover optimization process of the GFA analysis. Figure 5
describes the flow chart for MI-QSAR analysis.
RESULTS
In our previous study (Kulkarni and Hopfinger, 1999), we
found one major outlier (propylene glycol) in constructing an
MI-QSAR model for a subset of the ECETOC data set. Nevertheless, an attempt was made in this study to build a MIQSAR model for all the compounds in the ECETOC data set,
including the outliers of the earlier subset study. The two best
MI-QSAR models for all 38 molecules of the ECETOC data
set, including the outliers from the previous study, using only
linear representations of the independent variable terms are,
MES ⫽ 1.77 – 0.18 * F(H2O)
n ⫽ 38 ; r 2 ⫽ 0.15; xv – r 2 ⫽ – 0.18; LSE ⫽ 6.40
(1)
TABLE 3
Intermolecular Solute-Membrane Interaction Descriptors Considered in the Trial QSAR
Symbol
E (total)
E inter (total)
E (chg)
E (vdw)
E (chg ⫹ vdw)
F(H 2O)
F(OCT)
Log P
Description of the descriptor
Source
Average total interaction energy of the solute and membrane (kcal/mole)
Interaction energy between the solute and the membrane at the total system minimum potential energy
(sum of electrostatic, H-bonding, and vdw energies) (kcal/mole)
Electrostatic interaction energy between the solute and the membrane at total system minimum
potential energy (kcal/mole)
Van der Waals interaction energy between the solute and the membrane at total system minimum
potential energy (kcal/mole)
Electrostatic plus van der Waals interaction energy between the solute and the membrane at the total
system minimum potential energy (kcal/mole)
The aqueous solvation free energy computed using a hydration shell model (Hopfinger, 1973)
The 1- octanol solvation free energy computed using a hydration shell model (Hopfinger, 1973)
Logarithm of the 1-octanyl/water partition coefficient
a
a
a
a
a
b
b
b
Note. The sources are indicated as follows: a, computed directly from MDS energy files using molsim simulation package, Molsim (Doherty, 1994); and b,
calculated using Chemlab–II (Pearlstein 1988).
340
KULKARNI ET AL.
FIG. 6.
A plot of the MES versus F(H 2O).
MES ⫽ 1.76(0.904) – 兵0.06(0.028) * (F(H2O)
– [– 6.5]) 2其 ⫹ 兵0.05(0.021)
⫻ (F(OCT) – [– 4.5]) 2 – 0.01(0.008)
⫻ (E(vdw) – [3.9]) 2 – 0.13(0.006) * E(chg)其
⫹ {0.05(0.031) * LUMO 2}
FIG. 5. Flow chart for developing predictive MI-QSAR models for Draize
rabbit eye irritation test.
MES ⫽ 1.74 – 0.004 * E(chg) – 0.18 * F(H2O)
n ⫽ 38 ; r 2 ⫽ 0.15; xv – r 2 ⫽ – 0.20; LSE ⫽ 6.30
(2)
F(H 2O) is the aqueous solvation-free energy of the test solute
molecule, as defined earlier, and E(chg) is the electrostatic
solute-membrane interaction energy, see Table 3. The number
of compounds used to construct the correlation equation is
given by n, r 2 is the coefficient of determination, xv-r 2 is the
leave-one-out-cross-validation coefficient and LSE is the leastsquare error of fit.
A QSAR model is usually considered significant if it has a
coefficient of determination (r 2) greater than 0.7. The two
MI-QSAR models, Equations 1 and 2, have r 2 values less than
0.2. Thus, these MI-QSAR models are not significant. Figure 6
shows a plot of MES versus F(H 2O) for the ECETOC compounds. From an inspection of this plot, it is clear that it is not
possible to build a good linear MI-QSAR model for predicting
eye-irritation potential using F(H 2O) as a descriptor. The MES
vs. F(H 2O) relationship expressed in Figure 6 is parabolic. This
apparent non-linear relationship prompted the exploration of
MI-QSAR models using a parabolic dependence of MES on
F(H 2O). Still, it is noted that the number of compounds with
highly negative F(H 2O) values is only two. One of these two
compounds, propylene glycol [F(H 2O) ⫽ –16.9 kcal/mole], as
already noted, was found as an outlier in the previous MIQSAR study. The top non-linear MI-QSAR model for all 38
solute molecules in the ECETOC data set is,
n ⫽ 38 ; r 2 ⫽ 0.78; xv – r 2 ⫽ 0.73; LSE ⫽ 1.52; F ⫽ 32.4 (3)
The constant within each square term in Equation 3 defines the
value of the corresponding descriptor that yields its maximum,
or minimum, contribution to MES. The values in parenthesis
following the regression coefficients are the 95% confidence
limits. The Fisher F-statistic, F, is also reported for Equation 3.
The observed versus predicted MES values, using Equation
3, are shown in Figure 7, and given as part of Table 6. By
convention, if a predicted MES value is less than zero, that is,
outside the lower base line defined by the test, the MES value
is set to zero in Table 6 and Figure 7. F(H 2O) and F(OCT) are
determined for each test compound using a group-additive
model analogous to computing Log P from ␲ constants using
FIG. 7. Predicted vs. experimental molar-adjusted eye scores as a function
of arbitrary solute number for Equation 3. Predicted MES less than zero have
been set equal to zero.
341
PREDICTING EYE IRRITATION FROM ORGANIC CHEMICALS
TABLE 4
The Linear Cross-Correlation Matrix for the Reported MES and the MI-QSAR Descriptors Used in Equations 3 and 4
MES
LUMO
F(H2O)
E (chg ⫹ vdw)
F(OCT)
E (vdw)
E (chg)
MES
LUMO
F(H2O)
F(OCT)
E (chg ⫹ vdw)
E (vdw)
E (chg)
1.00
0.36
–0.39
0.15
–0.48
0.13
0.02
1.00
–0.43
–0.23
–0.51
0.07
–0.29
1.00
0.12
0.66
–0.03
–0.10
1.00
–0.04
0.13
1.00
0.11
0.62
0.61
1.00
–0.15
1.00
Chemlab-II (Pearlstein, 1988). The AM1 form of Mopac 6.0 in
the Cerius2 (MSI, 1997) was employed to estimate the LUMO
values of each molecule in the training set. MDS of the DMPC
model membrane monolayer with a single test molecule was
used to compute E(vdw) and E(chg).
A model with a correlation coefficient squared (r 2) of 0.78 is
often considered not particularly good. A low r 2 value can be
due to 2 causes: (1) variability (uncertainty) in the dependent
variable measures of the training set and/or (2) a limited and/or
poor selection of the independent variables for constructing the
model. In the case of the ECETOC training set, and eye
irritation in general, there is a considerable variability in the
eye-irritation scoring measures that are the dependent variables. Thus, the low r 2 values from the eye-irritation MI-QSAR
models come predominantly from “noise” in the eye-irritation
scoring measures, and not necessarily from a poor selection of
descriptors. This assertion has been tested for the ECETOC
data set. The raw, multiple eye-irritation scores for each of the
compounds in the ECETOC data set are available (Bagley et
al., 1992). An analysis of the raw data permits the computation
of a mean value (MV) for each compound, which is the MES
value used in the MI-QSAR analysis and given in Table 1. It is
also possible to compute the standard deviation, SD, from the
mean for the raw scores of each compound. Using the MV and
SD of each compound, and assuming a random distribution of
scoring about the MV, it is possible to perform simulationscoring experiments (Hopfiner, 1999). In turn, it is then possible to determine the average correlation coefficient of any
simulated eye-irritation scoring set for all the compounds in the
data set to the set of MV scores. By repeating the simulation
eye-irritation scoring experiments, an average correlation coefficient of the fit to the MV scores can be made. This average
correlation coefficient can be considered the upper value that is
possible from a QSAR analysis. For the ECETOC data set, this
average correlation coefficient is 0.79. Thus, the r 2 of Equation
3 is quite reasonable when taken relative to the estimated
intrinsic noise in the dependent-variable measure.
Previous attempts to construct QSAR models for the ECETOC data set are discussed in the first paper reporting the
MI-QSAR method (Kulkarni and Hopfinger, 1999). Significant
QSAR models (r 2 ⬎ 0.7), only employing intramolecular test
molecule descriptors, could be built for small analog subsets of
the ECETOC data set. However, all QSAR models reported for
the composite ECETOC data set have r 2 in the range of 0.1 to
0.3. Thus, Equation 3 is judged to be quite significant on the
basis of its superiority in r 2 value as compared to other reported
QSAR models.
An inspection of Figure 6 reveals that there are only two
solute molecules (propylene glycol and glycerol) that are nonirritants, even though they have extremely negative F(H 2O)
values. If these two solute molecules are removed from the
training set, good linear MI-QSAR models can be constructed.
Thus, propylene glycol and glycerol were removed from the
training set and linear MI-QSAR models were constructed. The
top linear MI-QSAR model identified in the GFA analysis is,
MES ⫽ – 0.66(0.488) – 0.01(0.042) * E(chg ⫹ vdw)
– 0.48(0.045) * F(H2O) ⫹ 0.39(0.024) * LUMO
n ⫽ 36; r 2 ⫽ 0.71; xv – r 2 ⫽ 0.65; LSE ⫽ 1.15; F ⫽ 38.9 (4)
The values for the descriptors used in MI-QSAR models,
Equations 3 and 4, for the ECETOC data set are given as part
of Table 1. The linear cross-correlation matrix for the descriptors in Equations 3 and 4 and the MES is in Table 4. The
cross-correlation matrix for the residuals of fit of the MI-QSAR
models (Equations 3 and 4) is given in Table 5.
The scrambling experiments to ascertain the validity of
Equations 3 and 4 lead to models with r 2 ⬍ 0.12. These low r 2
values for the scrambling experiments suggest that both Equations 3 and 4 are significant MI-QSAR models and not the
result of chance correlations.
TABLE 5
The Cross-Correlation Matrix for the Residuals of Fit
of the MI-QSAR Models, Equations 3 and 4
Equation 3
Equation 4
Equation 3
Equation 4
1
0.34
1
342
KULKARNI ET AL.
DISCUSSION
Significant MES MI-QSAR models were obtained for the
structurally diverse compounds (solutes) of the ECETOC eyeirritation database by combining intramolecular physicochemical properties of the test solute compounds with corresponding
intermolecular solute-membrane and solute-solvent interaction
properties. Solute-membrane interaction (binding) energies
[E(vdw), E(chg ⫹ vdw) and E(chg)] comprise one set of
intermolecular interaction properties of the MI-QSAR models.
F(H 2O) and F(OCT), the aqueous and 1-octanol solvation free
energies of the solute, respectively, contribute a second set of
intermolecular properties, although they are estimated by a
scheme solely based on the chemical structure of the solute.
LUMO, the lowest unoccupied molecular orbital energy, is the
only true intramolecular solute descriptor of Equations 3 and 4.
Aqueous solubility has long been qualitatively identified as
influencing the toxic spectrum of a compound. However, there
has not been a general computational tool to compute aqueous
solubility, or free energy of aqueous solvation, until recently.
Hence, the work reported here might be one of the first predictive toxicity studies employing a measure of solute aqueous
solubility. The aqueous solvation free energy [FH 2O] is
roughly proportional to the aqueous solubility of the molecule
(Kulkarni and Hopfinger, 1999). Increasingly negative F(H 2O)
values correspond to increasing aqueous solubility of a solute.
Similarly, increasingly negative F(OCT) values correspond to
increasing organic solubility of a solute in a nonpolar (1octanol) medium. In Equation 4, it is seen that aqueous solvation-free energy is negatively correlated with MES. This relationship suggests that water-soluble compounds have a greater
propensity to be eye irritants than hydrophobic compounds.
However MES and F(H 2O) are parabolically related in Equation 3. Compounds that have very negative F(H 2O) values have
low MES values. Thus, compounds that have very high aqueous solubility are non-irritants to the eye.
The solute-membrane interaction energy descriptors in
Equations 3 and 4 are also negatively correlated with the MES.
Thus, as the “binding energy” of a solute molecule to the
phospholipid regions of a membrane increases (a more negative descriptor value), its irritation potency is predicted to
increase. F(OCT) in Equation 3 is conceptually viewed as a
psuedo-solute-membrane interaction descriptor, which aids in
incorporating all the ECETOC compounds into a single significant MI-QSAR model.
LUMO, which appears in both Equations 3 and 4, measures
the electrophilicity of a molecule which, in turn, is interpreted
as a measure of molecular reactivity and stability. As LUMO
increases (relative to other molecules), the molecule is more
stable and less reactive. MES is predicted to increase as LUMO
increases in both Equations 3 and 4, although the relationship
is parabolic in Equation 3, but linear in Equation 4. The linear
versus parabolic dependence of MES on LUMO is a consequence of regression fitting for slightly different data sets with
Equation 3, based on all 38 ECETOC compounds, and Equation 4, derived by removing 2 outlier compounds. LUMO is
also associated with the ability of a compound to produce
color, that is, to act as a dye in solution. Hence, LUMO might
also reflect color changes observed in eye-irritation scoring not
necessarily related to irritation.
Combining the interpretations of the aqueous solvation and
solute-membrane interaction energy descriptors in Equations 3
and 4 leads to the following points. If a solute molecule is
water-soluble, it possesses some polar moieties. These polar
groups can also have favorable binding interactions with the
phospholipid regions of a membrane, probably involving the
phospholipid head groups. Polar alcohols are known to disturb
phospholipid membrane structure (McKarns et al., 1997),
which is consistent with this picture. The eye-irritation MIQSAR models given by Equations 3 and 4 suggest that the
eye-irritation potency of a solute molecule, as measured by the
Draize test, is highly dependent on the aqueous solubility of the
solute.
The solute-membrane interaction energy terms in Equations
3 and 4 suggest that eye-irritation potency increases with
increasing binding of the solute to the phospholipid regions of
the membrane. A straightforward interpretation of this type of
descriptor term is that disruption of membrane structure, and
likely function, resulting from strong interactions between solutes and phospholipids promotes eye irritation.
A mechanistic generalization of eye irritation can be made
from the discussion above and Equations 3-4. The F(H 2O)
descriptor reflects the number solute molecules available on the
aqueous/saline surface of the eye that could potentially disrupt
membrane structure. That is, F(H 2O) is a solute concentration
measure. The intermolecular membrane-solute interaction energy descriptors provide measures of the intrinsic membrane
disrupting potencies of each of the individual solute molecules.
MES is thus controlled by an effective solute concentration
coupled to the intrinsic membrane disruption propensity of the
solute. This mechanistic interpretation of the MI-QSARs models is similar to the model of Abraham and coworkers (Abraham et al., 1998) in terms of an effective solute concentration.
Their model identifies the significance of transferring the solute
from its application state (pure organic liquid or solid dispersed
into aqueous solution) to “an organic biophase” (the biological
structure of the eye). In other words, the concentration of the
solute in the organic biophase is crucial to eye-irritation potency.
All attempts to build a good linear MI-QSAR model for the
entire ECETOC data set resulted in models having 2 major
outliers; propylene glycol and glycerol. Both of these alcohols
have extremely negative F(H 2O) values. Thus, according to
Equation 4, which incorporates a linear dependence between
MES and F(H 2O), these 2 solutes should be highly irritating to
the eye. But experimental data shows that they are non-irritating. One plausible explanation for this apparent dichotomy is
that, for a solute molecule to partition significantly between the
343
PREDICTING EYE IRRITATION FROM ORGANIC CHEMICALS
TABLE 6
The Observed and Predicted, Using Equation 3, MES and MAS Scores for the ECETOC Data Set
MES
Compounds
Hydrocarbons:
3 Methyl hexane
2 Methyl pentane
Methylcyclopentane
1,9-Decadiene
Dodecane
1,5-hexadiene
cis-Cyclooctene
1,5-Dimethylcyclooctadiene
Aromatics:
4-Bromophentole
2,4-Difluoronitrobenzene
3-Ethyltoluene
4-Fluoroaniline
Xylene
Toluene
Styrene
1-Methylpropylbenzene
1,3-Disopropylbenzene
Ketones:
Methyl amyl ketone
Methyl isobutyl ketone
Methyl ethyl ketone
Acetone
Alcohols:
n-Butanol
Isobutanol
Isopropanol
Propylene glycol
2-Ethyl-1-hexanol
Glycerol
Hexanol
Butyl cellsolve
Cyclohexanol
Acetates:
Ethyl acetate
Methyl acetate
Methyl trimethyl acetate
Ethyl trimethyl acetate
Cellosolve acetate
n-Butyl acetate
Ethyl-2-methyl-acetoacetate
Acids:
2,2Dimethylbutanoic acid
MAS
Observed
Predicted
Residual
Observed
Predicted
Residual
0.10
0.26
0.41
0.37
0.45
0.55
0.43
0.44
0
0
1.39
0.33
0.32
0.93
1.15
0
0.10
0.26
–0.98
0.04
0.13
–0.38
–0.72
0.44
0.67
2.00
3.67
2.00
2.00
4.67
3.33
2.83
0
0
12.40
1.77
1.40
7.93
8.89
0
0.67
2.00
–8.73
0.23
0.60
–3.26
–5.56
2.83
0.19
0.40
0.32
6.62
1.10
0.96
0.77
0.31
0.38
1.94
0.75
2.00
7.24
1.75
1.10
1.64
0.61
0
–1.75
–0.35
–1.68
–0.62
–0.65
–0.14
–0.87
–0.30
0.38
1.33
3.67
2.33
69.83
9.00
9.00
6.75
2.00
2.00
13.55
6.88
14.58
76.39
14.34
10.30
14.39
3.95
0
–12.22
–3.21
–12.25
–6.56
–5.34
–1.30
–7.64
–1.95
2.00
2.26
0.59
4.48
4.83
0.73
0
5.65
4.04
1.53
0.59
–1.17
0.79
16.25
4.75
50.00
65.75
5.28
0
63.10
55.05
10.97
4.75
–13.10
10.70
5.47
6.44
2.34
0.10
7.82
0.12
8.13
8.99
8.29
4.12
5.62
5.42
0
5.95
0
6.62
7.68
6.80
1.35
0.82
–3.08
0.10
1.87
0.12
1.51
1.31
1.49
60.75
60.25
30.50
1.33
50.00
1.67
64.75
68.67
79.75
37.46
52.56
70.64
0
38.05
0
52.75
58.66
65.43
23.29
7.69
–40.14
1.33
11.95
1.67
12.00
10.01
14.32
1.47
3.14
0.36
0.63
2.03
0.99
2.55
1.25
2.29
2.63
1.27
2.98
2.00
0.80
0.22
0.85
–2.27
–0.64
–0.95
–1.01
1.75
15.00
39.50
2.67
4.17
15.00
7.50
18.00
12.74
28.81
19.51
8.40
22.00
8.63
5.66
2.26
10.69
–16.84
–4.23
–7.00
–1.13
12.34
5.59
2.83
2.76
44.67
22.60
22.07
Note. If an actual predicted MES value was less than zero, both the MES and MAS predicted values were set to zero in the table.
aqueous phase and an organic phase, a proper balance between
its aqueous and organic phase solubility is required. If a solute
molecule is highly soluble in the aqueous phase, it won’t enter
the organic phase and vice versa. Propylene glycol and glycerol are highly soluble in water. They prefer to stay in the
aqueous phase and not enter phospholipid regions of the membrane. Hence, the net solute concentration available to disrupt
the membrane structure is extremely low. If this hypothesis is
true, then MES may indeed have an approximate parabolic
relationship to F(H 2O), as found in Equation 3, for all the
ECETOC compounds. From an inspection of Figure 6, the
vertex of the “parabola” corresponds to a F(H2O) value of
about –11.3 kcal/mole. This value of F(H 2O) maximizes eye
irritation (MES). The vertex value for F(H 2O) in Equation 3 is
– 6.5 kcal/mole. The difference in the vertex value in Figure 6
and Equation 3 is mainly due to the influence of the other
344
KULKARNI ET AL.
descriptor terms in the MI-QSAR model given by Equation 3.
A parabolic fit of MES to only F(H 2O) gives a F(H 2O) vertex
value of –11.3 kcal/mole.
Using the vertex value of F(H 2O) as a reference, if the
F(H 2O) values increase, the solute molecules becomes less
water soluble and the net number of solute molecules available
for transfer into the membrane decreases. If F(H 2O) becomes
more negative, the solute molecules are highly water-soluble
and do not transfer into the membrane. In both these cases, the
number of solute molecules realized within the membrane
decreases and, correspondingly, so does eye irritation (MES).
It is important to point out 2 biochemical factors not considered in the MI-QSAR formalism. First, possible interactions
of a solute with membrane proteins are not considered. If this
class of interactions is important to the expression of eye
irritation of a solute, MI-QSAR analysis is not applicable and
will fail to provide a meaningful prediction of an MES. Based
on the consistently accurate estimation of eye irritation from
the MI-QSAR models (e.g., Fig. 7), however, it does not
appear that direct protein interactions play any substantial role
for these chemicals.
Secondly, at the current stage of development of MI-QSAR
analysis, cellular membrane specificity, in terms of specific
phospholipids, has not been considered. The MI-QSAR models
are based solely on DMPC monolayer “receptor” models.
However, there is no reason that other phospholipid membrane
models cannot be considered in a MI-QSAR analysis. A library
of membrane “receptor” models could be constructed and
employed in extended MI-QSAR investigations to determine
model sensitivity to membrane composition and structure.
Such multiple phospholipid membrane modeling could examine tissue-specific membrane lipid compositions such as in the
cornea and conjunctiva, and adjunct structures such as tear
film.
Figure 7 and Table 6 report the predicted vs. experimental
MES values using Equation 3. It is clear that the predicted
MES values track closely to those actually determined by
Draize eye-irritation tests. These results indicate that inclusion
of solute-membrane interaction properties in the MI-QSAR
analysis provide a better prediction (and description) of eye
irritation across chemical classes than can be obtained on the
basis of the QSAR analysis of the physicochemical parameters
of the test chemicals alone, e.g., Barratt (1995). It is proposed
that a MI-QSAR approach be used to develop “standard”
QSAR analyses for eye irritation, which would then be incorporated into risk assessment processes for eye irritation, such
as the process proposed by the OECD (1998).
Table 6 also contains the observed MAS values, and the
predicted MAS values based on the predicted MES values from
Equation 3, and the estimated solute densities determined as
described in the Materials and Methods section. When the
MES value is predicted to be less than zero, a value of zero is
assigned for both the predicted MES and MAS scores. Where
there are large differences between observed and predicted
MAS values, the corresponding observed MAS values are
often large, that is, the compounds are highly irritating. The
major source of the “magnification” of difference between
predicted and observed MAS values, relative to MES values,
resides in the estimations of the solute densities. A small
change in the density for propylene glycol by 0.2 units results
in a change of the predicted MAS value by 12 units. Thus, a
small change in estimated solute density often results in a
magnified change in the predicted MAS value. Work is currently underway in our laboratory to find a suitable approach to
estimating the “effective” density of a test solute molecule.
Overall, the most significant feature of this study is the
successful treatment of a representative set of structurally
diverse compounds from the ECETOC eye-irritation training
set by including interactions of these compounds with membrane models. This approach, membrane-interaction (MI)QSAR analysis, may be a breakthrough method to reduce
animal testing in several areas of toxicology. In addition to eye
irritation, other areas that may involve membrane interactions
in their biochemical mechanisms, and therefore, could be
meaningfully investigated using MI-QSAR analysis include
skin sensitization and irritation, aquatic toxicity, drug-membrane receptor interactions, and general modeling of bioavailability. Additional work and applications of MI-QSAR analysis continue in our laboratory with the hope of learning more
about both the reliability and general utility of the method.
This study may be emblematic of the progress made in the
quantitative prediction of toxicological endpoints. The predictive toxicity models developed for eye irritation appear to be
sufficiently robust to be used to reduce animal testing by
eliminating compounds of predicted high and low eye-irritation
scores from the pool of animal test compounds. However, other
areas of toxicology, such as chemical carcinogenicity, remain
less tractable to computational prediction methods. Still, the
promise of success seems within reach, and in fact, a competition is now being held to see which computational
methods work best in predicting chemical carcinogenicity (see
The Predictive Toxicology Challenge, http://www.methods.
informatik.uni-freiburg.de/).
ACKNOWLEDGMENTS
AJH and AK are pleased to acknowledge the financial support of The
Procter & Gamble Company. Resources of the Laboratory of Molecular
Modeling and Design at UIC, and of The Chem21 Group, Inc. were used in
performing this work.
REFERENCES
Abraham, M. H., Kumarsingh, R., Cometto-Muniz, J. E., and Cain, W. S.
(1998). Quantitative structure-activity relationship (QSAR) for a Draize
eye-irritation database. Toxicol. in Vitro 12, 201–207.
Bagley, D. M., Botham, P. A., Gardner, J. R., Holland, G., Kreiling, R., Lewis,
R. W., Stringer, D. A., and Walker, A. P. (1992). Eye irritation: Reference
chemicals data bank. Toxicol. in Vitro 6, 487– 491.
PREDICTING EYE IRRITATION FROM ORGANIC CHEMICALS
Balls, M., Berg, N., Bruner, L. H., Curren, R. D., de Silva, O., Earl, L. K.,
Esdaile, D. J., Fentem, J. H., Liebsch, M., Ohno, Y., Prinsen, M. K.,
Spielmann, H., and Worth, A. P. (1999). Eye irritation testing: The way
forward. ATLA 27, 53–77.
Barratt, M. D. (1995). The role of structure activity relationships and expert
systems in alternative strategies for the determination of skin sensitization,
skin corrosivity, and eye irritation. ATLA 23, 111–122.
Berendsen, H. J. C., Postma, J. P. M., Gunsteren, W. F. V., DiNola, A., and
Haak, J. R. (1984). Molecular dynamics with coupling with to an external
bath. J. Chem. Phys. 81, 3684 –3690.
Brantom, P. G., Bruner, L. H., de Silva, O., Easty, D. L., and Pape, W. (1997).
A summary report of the COLIPA international validation study on alternatives to the Draize rabbit-eye-irritation test. Toxicol. in Vitro 11, 141–179.
Bruner, L. H., de Silva, O., Earl, L. K., Easty, D. L., Pape, W., and Spielmann,
H. (1998). Report on the COLIPA Workshop on Mechanisms of Eye
Irritation. Altern. Lab. Anim. 26, 811– 820.
Doherty, D. C. (1994). Molsim User⬘s Guide, Ver. 3.0. Chicago.
Draize, J. H., Woodard, G., and Calvery, H. O. (1944). Methods for the study
of irritation and toxicity of substances applied to the skin and mucous
membranes. J. Pharmacol. Exp. Ther. 82, 377–390.
Friedman, J. (1988). Multivariate Adaptive Regression Spline. Technical Report No. 102. Stanford University, Stanford, CA.
Hauser, H., Pascher, I., Pearson, R. H., and Sundell, S. (1981). Preferred
conformation and molecular packing of phosphatidylethanolamine and
phosphatidylcholine. Biochim. Biophys. Acta 650, 21–51.
Hopfinger, A. J. (1973). Conformational Properties of Macromolecules. Academic Press, New York and London.
Hopfinger, A. J. (1999). 4D-QSAR User⬘s Manual, Release 2.0. The Chem21
Group, Chicago.
HyperChem (1998). HyperChem, Release 4.5 for MS Windows. Hypercube
Inc., Waterloo, Ontario.
Jester, J. V., Li, J. F., Petroll, W. M., Parker, R. D., Cavanaugh, H. D., Carr,
G. J., Smith, B., and Maurer, J. K. (1998). Area and depth of surfactantinduced corneal injury correlates with cell death. Invest. Ophthalmol. Vis.
Sci. 39, 922–936.
345
Kubinyi, H. (1993). 3D QSAR in Drug Design Theory Methods and Applications. ESCOM Science Publishers B.V., Germany.
Kulkarni, A. S., and Hopfinger, A. J. (1999). Membrane-interaction QSAR
analysis: Application to the estimation of eye irritation by organic compounds. Pharm. Res. 16, 1244 –1252.
McKarns, S. C., Hansch, C., Caldwell, W. S., Morgan, W. T., Moore, S. K.,
and Doolittle, D. J. (1997). Correlation between hydrophobicity of shortchain aliphatic alcohols and their ability to alter plasma membrane integrity.
Fundam. Appl. Toxicol. 36, 62–70.
Mopac (1990). Mopac, Ver. 6.0, Release Notes. Frank J. Seiler Research Laboratory, United States Air Force Academy, Colorado Springs,
CO.
MSI (1997). Cerius2, Molecular Simulations Users Guide, Ver. 3.0. MSI, San
Diego.
OECD (1987). Acute Eye Irritation/Corrosion OECD Guideline for Testing of
Chemicals 405. OECD, Paris.
OECD (1996). Final Report of the OECD Workshop on Harmonization and
Validation of Acceptance Criteria for Toxicological Test Methods. OECD,
Paris.
OECD (1998). Revised Proposal for the Harmonization of Hazard Classification Based on Eye Irritation/Corrosion. OECD, Paris.
Pearlstein, R. A. (1988). Chemlab-II Users Guide. Chemlab, Chicago.
Ploeg, P. v. d., and Berendsen, H. J. C. (1982). Molecular dynamic simulation
of a bilayer membrane. J. Chem. Phys. 76, 3271–3276.
Rogers, D. (1994). WOLF 6.2 GFA Program, Molecular Simulations, San
Diego.
Rogers, D., and Hopfinger, A. J. (1994). Applications of genetic function
approximation to quantitative structure-activity relationships and quantitative structure-property relationships. J. Chem. Inf. Comput. Sci. 34, 854 –
866.
Rowan, A. N. (1984). The future of animals in research and training: The
search for alternatives. Toxicology 4, 508 –516.
Waterbeemd, H. (1995). Methods in Molecular Design. VCH Publishers, New
York.