Prediction of Skin Irritation from Organic Chemicals Using Membrane-Interaction QSAR Analysis

Kiran Kodithala*, A. J. Hopfinger*,1, Edward D. Thompson{dagger} and Michael K. Robinson{dagger}

* Laboratory of Molecular Modeling and Design, M/C-781, College of Pharmacy, University of Illinois at Chicago, 833 South Wood Street, Chicago, Illinois 60612–7231; and {dagger} The Procter & Gamble Company, Miami Valley Research Laboratories, P.O. Box 538707, Cincinnati, Ohio 45253–8707

Received October 2, 2001; accepted December 14, 2001


    ABSTRACT
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 
Membrane-interaction (MI) quantitative structure activity relationship (QSAR) analysis was carried out for a training set of 22 hydroxy organic compounds for which the Draize skin irritation scores, PII, had been determined. Significant MI-QSAR models were constructed in which skin irritation potency is predicted to increase with (1) increasing effective concentration of the compound available for uptake into phospholipid-rich regions of a cellular membrane, (2) increasing binding of the compound to the phospholipid-rich regions of a cellular membrane, and (3) the chemical reactivity of the compound as reflected by the highest occupied molecular orbital (HOMO) and/or lowest unoccupied molecular orbital (LUMO) of the molecule. Overall, the MI-QSAR models constructed for skin irritation are very similar, with respect to the types of descriptors, to those found for eye irritation. In turn, the skin irritation MI-QSAR models suggest a similar molecular mechanism of action to that postulated for eye irritation from MI-QSAR analysis. Significant and predictive QSAR models cannot be constructed unless test compound-membrane interaction descriptors are computed and used to build the QSAR models.

Key Words: QSAR; membrane interaction; skin irritation.


    INTRODUCTION
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 
The classic in vivo skin irritation assay widely used to obtain the skin irritation potential of a chemical compound is the Draize test, which was developed in the 1940s as part of a protocol to evaluate compounds for use as chemical warfare agents (Draize et al., 1944Go). In a Draize skin irritation study the test article is applied to the shaved skins of 3–6 rabbits and covered for 4 or 24 h. The test article is then wiped off to permit observation for signs of skin irritation for a maximum of 72 h. An empirical formula is provided by a regulatory agency to calculate the primary skin irritation score from this data for use in labeling. Though most regulations require a 24-h exposure period, 4-h exposure is adequate and permitted with certain products.

The emergence of animal protection advocacy has driven a search for alternatives to in vivo skin irritation studies. By the end of the 1980`s laws were in place worldwide, which put legal and moral obligations on all concerned to seek to replace, reduce, and/or refine laboratory animal experimentation (Balls, 1999). The alternative approaches to in vivo testing for skin irritation fall into two general classes, in vitro skin irritation screens and computational approaches. One class of in vitro studies utilizes dead human/animal skin for the evaluation of skin properties. Though in vitro studies do not have the advantages of working with living human skin, these studies are easier to perform and have achieved considerable success in predicting the skin irritation potential of individual chemical compounds and mixtures. In vitro tests have been developed using human or animal skin, three-dimensional skin "equivalent" culture systems derived from human skin cells, or noncellular "biobarrier" systems. Human skin can be difficult to obtain in sufficient quantities. Though animal skin and the skin equivalent cultures are more permeable to chemicals than human (Bronaugh et al., 1985Go), these systems have been more amenable to in vitro methods development as denoted by their recent successful validation (Fentum et al., 1998Go; Liebsch et al., 2000Go).

The principal computational approach to predicting skin irritation has been to use quantitative structure activity relationship (QSAR) methodology. The QSAR methodology is based on the evaluation of physicochemical properties of chemical compounds and trying to relate these properties to their biological activities. QSAR analyses have achieved considerable success with the prediction of biological activities of molecules when their mechanism of action is known. However, the major limitation of applying QSAR methods to estimating skin irritation is the same limitation that plagues the applicability of the QSAR paradigm to toxicological problems in general. QSAR methods are effectively limited to dealing with analog structure-activity training sets while most structure-toxicity data sets consist of structurally diverse compounds.

Recently, we have developed a methodology that combines structure-based design techniques with the QSAR paradigm (Kulkarni and Hopfinger, 1999Go; Kulkarni et al., 2001Go). The resultant methodology permits structurally diverse training sets to be investigated. The operational key to this approach is to assume that the phospholipid-rich regions of cellular membranes constitute the effective "receptor" for the compounds of the training set. The significance and utility of this approach rest upon the validity of the compound-membrane interaction being a crucial event in the expression of the toxicity. Because of the importance of the membrane in this form of QSAR, we have termed this formalism membrane-interaction (MI)-QSAR analysis.

MI-QSAR analysis has been successfully employed (Kulkarni and Hopfinger, 1999Go; Kulkarni et al., 2001Go) in the prediction of eye irritation potential of chemical compounds. In this article the results are presented from performing a MI-QSAR analysis on a training set of 22 diverse hydroxy alcohol compounds for which Draize skin irritation potencies had been previously determined (RTK web site, 2001). Part of this study includes the comparison of MI-QSAR skin irritation predictions to corresponding predictions made using "traditional" intramolecular 2D-QSAR analyses. This comparison and evaluation demonstrates that MI-QSAR analysis provides more information to estimating skin irritation potency than is realized from traditional intramolecular 2D-QSAR approaches.


    MATERIALS AND METHODS
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 
Training set and skin irritation scores.
In the Draize skin irritation test, the test material is introduced under gauze patches to intact sites on the clipped dorsum of the rabbit. Applications are normally made on 6 rabbits. The patches are secured by adhesive tape and the entire trunk of the animal is wrapped in a semi-occlusive dressing for 4 h. After 4 h the patches are removed, the test sites cleaned, and any resulting reaction graded for erythema and oedema. The reactions are also scored at 24, 48, and 72 h (Draize, 1944). Skin reactions are graded separately for erythema/eschar and edema, each on a 0–4 grading scale. For erythema/eschar: 0 = no erythema; 1 = very slight erythema, barely perceptible; 2 = well-defined erythema; 3 = moderate to severe erythema; 4 = severe erythema (beet redness) to slight eschar formation (injuries in depth). For edema: 0 = no edema; 1 = very slight edema, barely perceptible; 2 = slight edema (edges of area well defined by raising); 3 = moderate edema (raised approximately 1 mm); 4 = severe edema (raised more than 1 mm and extending beyond the area of exposure).

These grades are averaged across all animals and grading time points, and then the averages are combined to derive a primary index (PII). Test materials producing PII values less than 2 are considered mildly irritating, 2 to 5 are moderately irritating, and greater than 5 are severely irritating. Variations around this basic method have formed the regulatory classification of skin corrosion and skin irritation worldwide.

Table 1Go contains the PII values for the training set of 22 hydroxy organic compounds (RTK web site, 2001) for which MI-QSAR models were sought. Some of the multiple RTK descriptions of irritation for a compound in the training set are not consistent. For example, 2-tert-butylphenol is not rated as a severe irritant, but one report in its RTK file indicates this chemical produced necrosis. The low irritation score for 2-tert-butylphenol is based on it not producing edema. Overall, we have accepted and used the reported PII values given in the RTK files for the compounds of our training set.


View this table:
[in this window]
[in a new window]
 
TABLE 1 Training Set of 22 Hydroxy Organic Compounds Used to Construct the MI-QSAR Models
 
The PII values are derived from constant molecular weight samples of the test compounds. QSAR models are normally constructed from molar concentration measures. Molar concentration measures permit predictions to be made from a QSAR model on a molecule to molecule basis that is independent of the molecular weights of the molecules. Thus, the PII values in Table 1Go were converted to their molar concentration equivalents [PII (molar)] that are also listed in Table 1Go. Conversion to molar units from molecular weight units is described in detail by Kulkarni et al. (2001). The Draize eye irritation measures, maximum average score (MAS), and molar-adjusted eye score (MES) were also available for the 22 hydroxy organic compounds and are reported in Table 1Go.

The cross-correlation matrix of the 4 irritation measures is given in Table 2Go. For this particular training set there are high cross-correlations between PII and PII (molar) [0.89] and between MES and MAS [0.94]. Such high cross-correlations among molecular weight and molar concentration irritation measures cannot be expected for arbitrary sets of organic compounds. However, the high PII and PII (molar) cross-correlation is an important finding since the range in the PII (molar) values is only 0.00 to 1.94. The narrow range and the relatively high uncertainty in the PII (molar) scores combine to limit the applicability of constructing reliable QSAR models using the PII (molar) data. Basically, the average difference among PII (molar) measures is likely smaller than the average error in making a PII (molar) scoring. However, the PII measures, which span a range of 0.00 to 8.00, can used to construct significant MI-QSAR models. Even though these models are molecular weight dependent, the high correlation between PII and PII (molar) measures guarantees the QSAR descriptor terms, and corresponding inferred mechanism of action, should be meaningful.


View this table:
[in this window]
[in a new window]
 
TABLE 2 Cross-Correlation Matrix for the Training Set Compounds
 
The molecular weight concentration skin and eye irritation measures (PII and MAS) and molar concentration skin and eye irritation measures [PII (molar) and MES] are 0.60 and 0.79 respectively (see Table 2Go). These moderate correlations possibly suggest both similarities and differences between the skin and eye irritation mechanisms of action. These similarities and differences are discussed later in the article.

Construction of DMPC monolayer.
The selection of the 1,2-dimyristoyl-sn-glycero-3-phosphatidylcholine (DMPC) monolayer was based on the MI-QSAR studies on eye irritation (Kulkarni and Hopfinger, 1999Go; Kulkarni et al., 2001Go). The assumption has been made that both skin and eye irritation do not have specific "irritation receptors" in the membranes, but solute uptake into the phospholipid rich regions of a membrane and resulting disruption of the membrane are related to "irritation." A DMPC monolayer was selected to model a "skin cellular membrane." The single DMPC molecule was built using HyperChem software (HyperChem, 1998Go) from the available crystal data (Hauser et al., 1981Go). The aliphatic chains of DMPC molecule were assigned to the trans-planar, local minimum energy conformation. The AM1 Hamiltonian in Mopac 6.0 (Mopac, 1990Go) was used for the estimation of partial atomic charges on all molecules.

The structure of a DMPC molecule is shown in Figure 1Go. An assembly of 16 DMPC molecules (4*4*1) in (x, y, z) directions, respectively, was used as a model membrane monolayer. The size of the monolayer simulation system was selected based on the work done by van der Ploeg and Berendsen (Ploeg and Berendsen, 1982Go). A representation of the model monolayer prior to molecular dynamics simulation modeling is shown in Figure 2Go.



View larger version (20K):
[in this window]
[in a new window]
 
FIG. 1. The chemical structure of a DMPC phospholipid with an arbitrary atom numbering assignment.

 


View larger version (93K):
[in this window]
[in a new window]
 
FIG. 2. An ideal model DMPC monolayer containing 16 molecules.

 
Construction of solute molecules.
The solute hydroxy organic compounds of the training set were constructed using HyperChem software (HyperChem, 1998Go). The AM1 partial atomic charges were then added to these molecules. Structural optimization of each of these training set compounds was carried out using the Chemlab-II software (Pearlstein, 1988Go).

Molecular dynamics simulations.
The molecular dynamics simulations (MDS) were carried out using the MDS package Molsim (Doherty, 1994Go), using a molecular mechanics force with an extended MM2 parameterization (Doherty, 1994Go). The selection of the simulation temperature was based upon the gel to liquid crystalline phase transition for DMPC that occurs at a temperature of 297°K (Bloom et al., 1991Go). The simulation temperature of 311°K was chosen for 2 reasons: (1) It is above the phase transition for DMPC and (2) it is above body temperature. Simulation temperature was held constant by coupling the system to an external constant temperature bath (Berendsen et al., 1984Go). A dielectric constant of 3.5 was used since it reproduces known membrane structural/dynamic features (Jin and Hopfinger, 1996Go).

The model monolayer, without a solute molecule, was slowly heated starting at 20°K, then at 50°K, and from that point in increments of 50°K until a final step was used to achieve the simulation temperature of 311°K. At each temperature increment, 4 picos (ps) of MDS was carried out to allow for structural relaxation and the distribution of the kinetic energy throughout the simulation model. When 311°K was reached, a 50 ps MDS production run was performed to equilibrate the monolayer.

Docking of solute molecules in a DMPC monolayer.
In order to determine the most favorable solute-membrane interaction, as well as to sample the major interaction states between the solute and the model membrane, a modeling schema was proposed in which the solute molecule is docked in 3 different positions inside the membrane model. These 3 positions are different from each other with respect to their chemical environment. Upper region is the polar head group region of the DMPC membrane; lower region is the highly hydrophobic environment near the tails of the hydrocarbon chains; and middle region is the region near the branching of a DMPC molecule to form the 2 aliphatic chains.

The solute molecule is docked in these 3 positions in such a way that the dipole moment of the solute molecule is in the opposite direction to the model membrane's dipole moment. This modeling scheme makes sure that the solute molecule encounters all the possible environments in the biological system, and thus identifies the most energetically stable position for the solute molecule in the membrane. Three corresponding MDS were carried out for each solute molecule with regard to the trial positions of the solute molecule in the monolayer. The energetically favorable geometry of the solute molecule in the monolayer was sought using each of these trial positions.

In order to prevent unfavorable steric repulsive 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 solute molecules was placed at each of the 3 different positions in the monolayer described above with the dipolar portion of the solute "facing" towards the headgroup region. The same heating steps used to equilibrate the isolated monolayer were carried out with solute present. However, a final production run of only 20 ps was performed at 311°K.

Calculation of QSAR descriptors.
The QSAR descriptors are the various properties and features of molecules. A descriptor can be intramolecular—an inherent property of a molecule obtained solely from its chemical structure. The other class of descriptors is the set of intermolecular properties and features that depend on the environment in which the molecules are located. These properties and features are normally computed from the interaction between 2 or more molecules.

The intramolecular descriptors of the molecules in the skin irritation training set were calculated using multiple software packages including TSAR (Oxford Molecular), Cerius 2 (MSI, 1997Go), and Chemlab-II (Pearlstein, 1988Go) and are listed in Table 3Go. The intramolecular descriptor set used in this study includes the large majority of descriptors generally employed in current intramolecular QSAR analyses. The intermolecular descriptor set is given in Table 4Go. Part A of Table 4Go lists the membrane-solute intermolecular descriptors computed from the MDS. Part B of Table 4Go list the solute-solvent descriptors FH2O, FOCT, and Log P. While these 3 descriptors are intermolecular properties, they have been computed using intramolecular computational algorithms.


View this table:
[in this window]
[in a new window]
 
TABLE 3 The Set of Intramolecular QSAR Descriptors Used in This Study
 

View this table:
[in this window]
[in a new window]
 
TABLE 4 Intermolecular MI-QSAR Descriptors Used in This Study
 
Construction and testing of QSAR models.
The QSAR models were constructed using the genetic function approximation (GFA; Rogers and Hopfinger, 1994Go), which is a multidimensional optimization method based upon the genetic algorithm paradigm. Multidimensional linear regression (MLR) was used as the GFA fitting function. An optimum QSAR model was assumed when descriptor usage in the GFA model evolution became constant. The quality of a QSAR model has been evaluated using the correlation coefficient of fit, r2, and the leave-one-out cross validation coefficient, xv-r2. All intramolecular and intermolecular descriptors were used in the MI-QSAR descriptor pool to generate MI-QSAR models. Only the intramolecular descriptors were used in the descriptor pool to build the intramolecular QSAR models.


    RESULTS
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 
Intramolecular QSAR Models
The intramolecular descriptors given in Table 3Go were determined for the 22 hydroxy organic compounds of the training set and used to construct intramolecular QSAR skin irritation potency (PII), models that are given in Equations 1–4GoGoGoGo. The definitions of the descriptors are given in Tables 3 and 4GoGo. From the way in which FOCT, FH2O, and LogP are computed, they can be considered as both intra- and intermolecular descriptors in QSAR model building.

Measures of statistical fit were obtained for intramolecular QSAR models with different number of descriptors in order to evaluate the role of the number of descriptors on the significance of the model. The chosen statistical measures of importance in QSAR model construction are the correlation coefficient, r2, and the leave-one-out cross validation coefficient, xv-r2. It was found that the r2 value increases when the model size was increased from 1 descriptor term to 4 descriptor terms. However, r2 values were largely unaffected by a greater (than 4) number of descriptors in an intramolecular QSAR model. However, xv-r2 is maximized for the two-term model, and then decreases for increasing size of the QSAR model. The relatively low values of r2 across the QSAR models, and the low and decreasing values of xv-r2 with increasing model size, suggests these QSAR models are of limited stability and significance.

One descriptor model

((1))

Two descriptor model

((2))

Three descriptor model

((3))

Four descriptor model

((4))

4-tert-Butylphenol and 2-propyl 1-heptanol are outliers in the MI-QSAR models presented below. These 2 compounds were not included in the construction of the intramolecular QSAR models so that the MI-QSAR and intramolecular QSAR models could be meaningfully compared to one another. Thus, the number of compounds used in intramolecular QSAR training set is 20.

MI-QSAR Models
Equations 5Gothrough 8Go represent the best MI-QSAR models as a function of increasing model size.

One descriptor model

((5))

Two descriptor model


((6))

Three descriptor model

((7))

Four descriptor model

((8))

The one-term MI-QSAR model, Equation 5Go, is identical to the one-term intramolecular QSAR model, Equation 1Go. This observation suggests there is no intermolecular descriptor that, by itself, correlates more significantly with PII than LogP. The high r2 value of the four-term MI-QSAR model, Equation 8Go, suggests that this may be an overfit model. The average experimental error in the eye irritation scores, MES (eye irritation potency scores), is at least 20% (Kulkarni et al., 2001Go). It is a reasonable to assume that the average experimental error of the PII measures is as large as that of the MES endpoint. An average error of 20% in the dependent variable corresponds to an upper value in r2 of 0.75 to 0.80 (Kulkarni et al., 2001Go). Thus, the three-term MI-QSAR model is probably the most meaningful QSAR model of all given by Equations 1–8GoGoGoGoGoGoGoGo.

Both r2 and xv-r2 of the MI-QSAR models increase with increasing number of descriptor terms, and are generally larger than the corresponding values of the intramolecular QSAR models. Overall, a comparison of the statistics of intramolecular QSAR and MI-QSAR models for the same number of descriptors demonstrates that the MI-QSAR models are more robust and predictive. Figure 3Go is a plot that compares r2 for the intramolecular QSAR and MI-QSAR models having the same number of descriptors. The correlation coefficients of the MI-QSAR models are larger than the corresponding intramolecular QSAR models when the number of descriptor terms is 3 and 4. The r2 values are essentially constant for models with 4 or more descriptor terms (see Fig. 3Go for both intramolecular and MI-QSAR models).



View larger version (13K):
[in this window]
[in a new window]
 
FIG. 3. A comparison of the r2 values of the intramolecular and MI-QSAR models.

 
Table 5Go lists the observed and predicted PII values for the training set compounds. The predicted PII values in Table 5Go are for the four-term intramolecular QSAR model (Equation 4Go) and for the three-term MI-QSAR model (Equation 7Go). The corresponding plot of the observed and predicted, using Equation 7Go, PII values is shown in Figure 4Go. The 2 outliers for the MI-QSAR model, 4-tert-butylphenol and 2-propyl 1-heptanol, are listed in italics in Table 5Go. Their predicted PII values are based upon using Equations 4 and 7GoGo. 4-tert-Butylphenol is an outlier for the intramolecular QSAR model, but 2-propyl 1-heptanol is fit by Equation 4Go. Table 6Go contains the values of the descriptors of the best QSAR models, Equations 1–8GoGoGoGoGoGoGoGo, for each of the hydroxy organic compounds of the training set.


View this table:
[in this window]
[in a new window]
 
TABLE 5 Comparison of Observed and Predicted PII Values for the Hydroxy Organic Compounds in the Training Set
 


View larger version (21K):
[in this window]
[in a new window]
 
FIG. 4. Plot of the observed and predicted (using the MI-QSAR model given by Equation 7Go) skin irritation potency scores, PII.

 

View this table:
[in this window]
[in a new window]
 
TABLE 6 Observed PII Values and Values of the Most Significant Descriptors of the Best Intramolecular and the Best MI-QSAR Models for the Hydroxy Organic Compounds in the Training Set
 
Statistical Analysis of the QSAR Models
A statistical analysis of the best intramolecular QSAR model, represented by Equation 4Go, was performed by constructing a linear cross-correlation matrix of its descriptors that is given in Table 7Go. Any pair of descriptors having a cross-correlation greater than 0.50, or less than –0.50, is considered a highly correlated pair, and their r2 values are given in italics in Table 7Go. The linear cross correlation matrix of the intramolecular QSAR model shows that 3 out of the 4 descriptors used in Equation 4Go are significantly correlated to one another. A QSAR equation possessing highly cross-correlated descriptors is not considered a stable model for making predictions and/or performing virtual screening. Table 7Go also indicates that the descriptor having highest correlation with PII, the skin irritation potency score, is HOMO followed by the molar refractivity (MR) descriptor. The high cross-correlation between the MR and molecular volume (Vm) descriptors arises from molecular size. It was also observed that the molar refractivity descriptor is significantly correlated to the HOMO and FOCT descriptors.


View this table:
[in this window]
[in a new window]
 
TABLE 7 Linear Cross-Correlation Matrix for PII and Descriptors Used in the Best Intramolecular QSAR Model Represented by Equation 4Go
 
A statistical evaluation of the most significant MI-QSAR model, Equation 7Go was also performed by constructing the linear cross-correlation matrix of its descriptors, which is represented by Table 8Go. None of the pairs of descriptors of Equation 7Go are found to be significantly cross-correlated to one another. Thus, Equation 7Go is judged to be a stable QSAR model. The descriptor in Equation 7Go having highest correlation with skin irritation potency score is HOMO followed by the E(chg + vdw) descriptor.


View this table:
[in this window]
[in a new window]
 
TABLE 8 Linear Cross-Correlation Matrix for PII and Descriptors Used in the Best MI-QSAR Equation Represented by Equation 7Go
 
Outlier Analysis and Consensus Models
The predicted skin irritation score, using Equation 7Go for 4-tert-butyl phenol is much higher than the observed score, and the predicted skin irritation potency score for 2-tert butyl phenol is lower than the observed skin irritation potency score. To explore why 4-tert butyl phenol is an outlier in the model given by Equation 7Go, the descriptor values of 4-tert butyl phenol were compared with those of its closest congener in the training set, 2-tert butyl phenol, as given in Table 6Go. It can be seen from Table 6Go that the MI-QSAR descriptor values for the 2 analogs are similar with the largest difference in descriptor measures between 2-tert butyl phenol and 4-tert butyl phenol being E(chg + vdw). 4-tert Butylphenol is predicted to bind less strongly than 2-tert butylphenol to the membrane. However, this loss in membrane binding is not sufficient to lower the predicted PII value of 4-tert-butylphenol anywhere near the observed PII value.

2-Propyl 1-heptanol, the other outlier, has a PII of 7.2, which is the highest skin irritation potency for the aliphatic alcohols in the training set. The average skin irritation potency of aliphatic alcohols in the training set is 3.14 (n = 11). The high potency of 2-propyl 1-heptanol, when compared to other alcohols in the training set, may result from a different skin irritation mechanism as compared to other hydroxy organic compounds in the training set. More insight about the behavior of this molecule might be obtained from a larger training set, or a training set having molecules of greater similarity to 2-propyl 1-heptanol. The 22 hydroxy compounds used in this study were selected on the basis of available PII values that covered a reasonable range of measured irritation and were considered to be reliable measures.

The 22 hydroxy organic compounds in the training set can be classified into the following 3 groups: (1) aliphatic alcohols, (2) alicyclic alcohols, or (3) phenols. The number of compounds in the alicyclic alcohol group is only 3. Hence, the alicyclic alcohols have been grouped with the aliphatic alcohols, and individual MI-QSAR models have been constructed for the aliphatic alcohols and for the phenols. The best three-term MI-QSAR model for the aliphatic alcohols is given by Equation 9Go and the best three-term MI-QSAR model for the phenol group of the training set is represented by Equation 10Go.

Three-term models were constructed for the two classes of alcohols of the training set to permit direct comparison to the three-term MI-QSAR model, Equation 7Go, of the entire training set. The best three-term MI-QSAR model for the aliphatic alcohols is

((9))

The best three-term MI-QSAR model for the phenol analogs is

((10))

EHB is the intramolecular hydrogen bonding energy of the solute when in the membrane, and is defined in Table 4Go, Part A, while SA is the surface area of the solute and defined in Table 3Go. The composite set of PII predictions using Equations 9 and 10GoGo, as well as the observed PII values, are reported in Table 9Go. The corresponding PII plots of the data reported in Table 9Go are given in Figure 5Go. The observed PII of 2-propyl 1-heptanol is predicted with better accuracy using Equation 9Go than by Equation 7Go. Likewise, the PII of 4-tert-butylphenol is better predicted by Equation 10Go than by Equation 7Go. Moreover, Equations 9 and 10GoGo, in composite, better fit the training set data than does Equation 7Go. However, the better fitting of the outliers of Equation 7Go using Equations 9 and 10GoGo may also be due to partial overfitting that is required in order to meaningfully compare Equations 9 and 10GoGo to Equation 7Go.


View this table:
[in this window]
[in a new window]
 
TABLE 9 Prediction of the PII Values of the Aliphatic Alcohols and Phenols Obtained from Equations 9 and 10GoGo
 


View larger version (37K):
[in this window]
[in a new window]
 
FIG. 5. Plot of the observed and predicted (using the MI-QSAR models given by Equations 7, 9, and 10GoGoGo) skin irritation scores, PII.

 

    DISCUSSION
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 
One of the best MI-QSAR models determined in the study of the ECETOC data set for eye irritation (Kulkarni et al., 2001Go) has the form

((11))

MES is the Draize molar eye irritation score and has the same range of values as PII of the hydroxyl organic compound training set. The three-term MI-QSAR model for skin irritation given by Equation 7Go can be reexpressed as

((12))
where FOCT is the 1-octanol solvation free energy of a solute and

((13))

The form and descriptors of Equation 11Go are remarkably similar to those of Equation 12Go for PII. In both MI-QSAR models irritation is predicted to increase as E(chg + vdw) becomes more negative, that is as solute-membrane binding becomes stronger. Skin irritation appears to possess a more complex dependence on the interaction of a solute and membrane than eye irritation. The regression coefficient for E(chg + vdw) in Equation 12Go is larger in magnitude than in Equation 11Go, suggesting solute-membrane binding increases PII to a greater extent than MES. However, Equation 12Go also includes the FOCT term that is taken to represent increasing membrane solubility of a solute as FOCT becomes more negative. The positive regression coefficient for FOCT in Equation 12Go indicates that PII decreases as membrane solubility increases. Thus, FOCT appears to modulate the role of E(chg + vdw) on PII in Equation 12Go, and these 2 terms are, consequently, grouped together as membrane-solute interaction terms.

An increase in aqueous solubility, as measured by FH2O becoming more negative in value, is postulated to increase both types of irritation. The regression coefficients for the FH2O terms in both MI-QSAR models, Equations 11– 12GoGo, are quite similar suggesting both types of irritation are about equally dependent upon the aqueous solubility of the solute.

Equation 11Go has LUMO as a descriptor while Equation 12Go contains HOMO suggesting that the chemical reactivity of a solute plays a role in the expression of irritation potency. It is not possible to discern why 1 of these 2 electronic measures is selectively preferred in a given MI-QSAR model. However, it is noted that HOMO and LUMO are highly correlated [r = 0.79] to one another over the combined skin irritation and ECETOC eye irritation training sets. Thus, the preferred selection of LUMO, or HOMO, in a MI-QSAR model for a given training set of compounds, may simply be a consequence of a slightly better statistical fit of the data. A reviewer of this manuscript has suggested that the phenols might be metabolized to reactive oxidative species (free radical or catechol, hydroquinone) and that alcohols are metabolized to aldehydes or ketones (e.g., cyclohexanol to cyclohexanone) and these metabolites may be chemically reactive. The HOMO and LUMO descriptors may reflect the propensity for the corresponding metabolic activations.

A very significant observation can be made from the skin irritation and eye irritation MI-QSAR models as well as the discussion in the preceding paragraphs. A common molecular mechanism of irritant action seems to exist for these 2 types of toxic endpoints. This mechanism of action involves 3 factors. First, irritation increases as the available concentration of irritant for membrane uptake from an aqueous medium increases. FH2O and Log P are 2 descriptors that model this concentration factor. Next, the strength of interaction of the irritant with the phospholipid-rich regions of the cellular membrane is directly proportional to irritation potency. This irritation factor is modeled in the skin and eye irritation MI-QSAR models by E(chg + vdw). Lastly, both skin and eye irritations are seemingly dependent upon on the chemical reactivity of the irritant. It is not possible to pinpoint the specific type of chemical reaction(s) involved for a given class of irritation and/or chemical class of irritant. However, LUMO and HOMO repeatedly show up in the MI-QSAR models as representations for chemical reactivity and postulated metabolic pathways have been suggested above. Future MI-QSAR studies of toxic endpoints will be extremely interesting to analyze in terms of seeing the extent to which this proposed molecular mechanism of action for irritation endpoints can be further generalized. It is certainly possible that other factors, explicitly involving, for example, the acidity of the phenols and/or the nonacidity of the aliphatic alcohols, may be involved in skin irritation. Acidity is included in the MI-QSAR methodology only in terms of solvation and solute-membrane interactions.

A significant operational finding from this work is that MI-QSAR analysis provides reliable models that can serve as predictive quantitative virtual screens to evaluate skin irritation. The training set used in this study is small, and limited to alcohols. Certainly more, as well as structurally diverse, compounds for the training set would be welcomed to extend the utility of the corresponding MI-QSAR models for skin irritation. However, the similarity of these MI-QSAR models to those models developed for eye irritation, as described above, leads the assumption that these skin irritation models incorporate all the salient features of a general virtual screen for skin irritation.


    ACKNOWLEDGMENTS
 
A.J.H. and K.K. are pleased to acknowledge the financial support of the Procter & Gamble Company. Resources of the Laboratory of Molecular Modeling and Design at UIC and The Chem21 Group, Inc. were used in performing this work.


    NOTES
 
1 To whom correspondence should be addressed. Fax: (312) 413-3479. E-mail: hopfingr{at}uic.edu. Back


    REFERENCES
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 REFERENCES
 
Balls, M., Berg, N., Bruner, L. H., Curren, R. D., deSilva, O., Earl, L. K., Esdaile, D. J., Fentem, J. H., Liebsch, M., Ohno, Y., Prinsen, M. K., Spielman, H., and Worth, A. P. (1999). Eye irritation testing: The way forward. ALTA 27, 53–57.

Berendsen, H. J. C., Postma, J. P. M., Gunsteren, W. F. V., DiNola, A., and Haak, J. R. (1984). Molecular dynamics with coupling to an external bath. J. Chem. Phys. 81, 3684–3690.[ISI]

Bloom, M., Evans, E., and Mouritsen, O. (1991). Physical properties of the fluid lipid-bilayer component of cell membranes. Q. Rev. Biophys. 24, 293–397.[ISI][Medline]

Bronaugh, R. L., Stewart, R. F., and Congdon, E. R. (1985). Methods for in vitro percutaneous absorption studies. IV: The flow-through diffusion cell. J. Pharm. Sci. 74, 64–68.[ISI][Medline]

Doherty, D. C. (1994). Molsim Version 3.0 User's Guide. Chem21 Group, 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.

Fentum, J. H., Archer, G. E. B., Balls, M., Botham, P. A., Curren, R. D., Earl, L. K., Esdaile, D. J., Holzhutter, H. G., and Liebsch, M. (1998). The ECVAM international validation study on in vitro tests for skin corrosivity. 2. Results and evaluation by the management team. Toxicol. in Vitro 12, 483–524.[ISI]

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.[ISI][Medline]

HyperChem (1998). HyperChem, Release 4.5 for MS Windows. Hypercube, Waterloo, Ontario.

Jin, B., and Hopfinger, A. J. (1996). Characterization of lipid membrane dynamic by simulation. Quantitative representation of lipid fluidity. Comput. Theor. Polymer Sci. 6, 95–101.[ISI]

Kulkarni, A., and Hopfinger, A. J. (1999). Membrane-interaction QSAR analysis: Application to the estimation of eye irritation by organic compounds. Pharm. Res. 16, 1245–1253.[ISI][Medline]

Kulkarni, A., Hopfinger, A. J., Osborne, R., Bruner, L. H., and Thompson, E. D. (2001). Prediction of eye irritation from organic compounds using membrane-interaction QSAR analysis. Toxicol. Sci. 59, 335–345.[Abstract/Free Full Text]

Liebsch, M., Traue, D., Barrabas, C., Spielmann, H., Uphill, P., Wilkins, S., Wiemann, C., Kaufmann, T., Remmele, M., and Holzhutter, H. G. (2000). The ECVAM prevalidation study on the use of EpiDerm for skin corrosivity testing. ALTA 28, 371–401.

Mopac (1990). Mopac 6.0 Release Notes. Frank J. Seiler Research Laboratory, United States Air Force Academy, Fort Collins, CO.

MSI (1997). Cerius 2 V.3.0 Users Guide. Molecular Simulations, San Diego.

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.[ISI]

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.[ISI]

RTK Web site (2001). Available at http://www.rtk.net/tscatsinputstandard.html. Accessed December 8, 2001.





This Article
Abstract
FREE Full Text (PDF)
Alert me when this article is cited
Alert me if a correction is posted
Services
Email this article to a friend
Similar articles in this journal
Similar articles in ISI Web of Science
Similar articles in PubMed
Alert me to new issues of the journal
Add to My Personal Archive
Download to citation manager
Search for citing articles in:
ISI Web of Science (3)
Disclaimer
Request Permissions
Google Scholar
Articles by Kodithala, K.
Articles by Robinson, M. K.
PubMed
PubMed Citation
Articles by Kodithala, K.
Articles by Robinson, M. K.