Title of the Paper (18Pt Times New Roman, Bold) s5

Predicting Permeability of Antimycotics from Calculated Chemodescriptors: A Hierarchical QSAR Approach

SUBHASH C. BASAK AND DENISE MILLS Natural Resources Research Institute University of Minnesota Duluth 5013 Miller Trunk Hwy, Duluth, MN 55811 USA http://www.nrri.umn.edu/Basak/

Abstract: - Quantitative structure-activity relationship (QSAR) models were developed for the prediction of bovine hoof membrane permeability in an effort to gain insight into the rate of penetration of antimycotics through the nail plate. Numerical descriptors based on chemical structure were calculated for a set of 14 drugs, mainly antimycotics. The descriptors were then placed into one of three classes based on level of complexity and demand for computational resources. Models using the various classes of structural descriptors were developed using ridge regression, principal component regression, and partial least squares. Results indicate that permeability of antimycotics can be modeled based on structural descriptors alone, without the need for experimental data. As such, predictions can be made about the permeability of hypothetical compounds of similar structure not yet synthesized. However, additional data are required to allow for reliable modeling of human nail plate permeability.

Key-Words: - Permeability, Antimycotics, Hierarchical Quantitative Structure-Activity Relationship, Chemodescriptors, Hoof membrane, Nail plate

1 Introduction mainly antimycotic agents, using our hierarchical Onchomycosis is treated mainly with oral antifungal QSAR (HiQSAR) approach from calculated medication. However, due to the potential for severe chemodescriptors of these molecules. systemic side effects, there is an interest in the alternative topical application of these drugs. The first step in evaluating the efficacy of this approach 2 Materials and Methods is determining the permeability of therapeutic chemicals through the nail. 2.1 Compounds used in the study Our research team and collaborators have been Unfortunately, data with respect to the human nail involved in the development of quantitative plate are limited; therefore, permeability through structure-activity relationships (QSARs) of the bovine hoof membrane was studied. Permeability efficacies of different classes of antimycobacterial coefficients for nine antimycotics and five other drugs. Since experimental test data and drugs through the bovine hoof membrane were taken physicochemical properties of the majority of from the literature [1]. The structures of these candidate molecules are not usually available, our compounds are provided in Fig. 1. approach has been the development of predictive models based on algorithmically derived 2.2 Chemodescriptors chemodescriptors which can be calculated directly Computer software programs including POLLY from molecular structure without the input of any v.2.3 [2], Triplet [3], and Molconn-Z v.3.5 [4] were other experimental data. The efficacy used to calculate 369 chemodescriptors, including a (pharmacodynamics) of drugs is an important factor large set of connectivity indices [5,6], information in selecting candidate chemicals for preclinical theoretic and neighborhood complexity indices testing. Another set of factors related to the [7,8], electrotopological state descriptors [9], effectiveness of drugs is the ADMET (absorption, descriptors of hydrogen bonding and polarity, as distribution, metabolism, excretion, and toxicity) well as a small set of kappa shape indices [10]. A properties. Therefore, in this paper we attempted to complete list of descriptors may be obtained from an correlate the absorption patterns of a set of 14 drugs, earlier publication [11]. The descriptors were partitioned into hierarchical classes based on level of complexity and demand for N computational resources. The topostructural (TS) CH3 H3C CH3 N CH3 indices are at the low end of the hierarchy, encoding O CH3 N information strictly on atom connectivity within a H3C molecule, without taking any specific chemical 1. Amorolfine information into account. The topochemical (TC) 2. Bifonazole indices, in addition to encoding information on atom connectivity, also encode chemical information such OH Cl O N N as atom and bond type. Collectively, the TS and TC N descriptors are known as topological indices. The geometrical (3D) indices are more complex than the CH3 topological indices, encoding three-dimensional 3. Ciclopirox 4. Clotrimazole aspects of molecular structure. In hierarchical

N QSAR, models are developed based on increasingly OCH Cl 3 OOCH3 complex descriptor sets. In this way, we are able to N Cl O determine whether the addition of the computer- O H3CO intensive molecular descriptors results in significant H C O Cl Cl 3 model improvement or whether the simple, more 5. Econazole 6. Griseofulvin quickly obtainable topological descriptors, alone, are

N sufficient for the development of high-quality N predictive models. O O N N O O H3C H Cl Cl 2.3 Statistical analysis 7. Ketoconazole Prior to statistical analysis, descriptors with a constant value for all or nearly all of the chemicals CH CH3 3 were removed from the data set, as were any N H C N O 3 descriptors with undefined Triplet values, and one of S each perfectly correlated descriptor pair (r = 1.0) as determined by the CORR procedure of the SAS 8. Naftifine 9. Tolnaftate statistical package [12]. A total of 248 descriptors H HO N CH O were retained.3 The descriptors were transformed by the natural logarithm as their values spanned several O N CH3 H H3C O orders of magnitude.

10. Paracetamol 11. Phenacetin For comparative purposes, three distinct regression methodologies were used to develop the OH predictive models based on theoretical molecular OH O OH descriptors.Cl Ridge regression (RR) [13], principal H3C N H N N components regression (PCR) [14], and partial least squaresCl (PLS) [15,16] are all appropriate methods in N O N O O2N OH situations where the number of descriptors is large CH3 with respect to the number of observations and when 12. Diprophylline 13. Chloramphenicol there is a high degree of intercorrelation among the

HO OH descriptors. Each of these regression methods O I O utilizes the entire set of available descriptors, in HO OH N N contrast to subset regression wherein a small subset H H of descriptors is selected from an available pool for I I HO H modeling and all others are discarded. NH H3C For the sake of brevity, we have reported the 2 O cross-validated R for each of the models rather than 2 Fig. 1. Compounds 1 used4. Lopa m inido l the present study, the models, themselves. The R c.v. is obtained using antimycotics (1-9) and other drugs (10-14) the cross-validated sum of squares approach wherein each compound, in turn, is held out and the remaining n-1 compounds are used to determine the model coefficients which are then used to predict the Acknowledgements: activity value of the held-out compound. This is contribution number 378 from the Center for Comparable to validating a model with a reserved Water and the Environment of the Natural test set, this method assesses the final model using Resources Research Institute. Research was test compounds that were not involved in the model supported in part by Grant F49620-02-1-0138 from 2 fitting. As such, a model with a large R c.v can be the United States Air Force. The U.S. Government is relied upon to provide an accurate prediction of authorized to reproduce and distribute reprints for future cases that are similar to those used to calibrate Governmental purposes notwithstanding any the model [17]. It is important to note the copyright notation thereon. The views and 2 2 2 distinction between R and R c.v. R tends to increase conclusions contained herein are those of the authors upon the addition of any descriptor, often leading to and should not be interpreted as necessarily overly optimistic results which are not reliable. On representing the official policies or endorsements, 2 the other hand, R c.v tends to decrease upon the either expressed or implied, of the Air Force Office addition of irrelevant descriptors and provides a of Scientific Research or the U.S. Government. reliable measure of model predictability.

3 Results and Discussion References: Results in Table 1 indicate that the easily calculable [1] D. Mertin and B. C. Lippold, In-vitro topological descriptors explain most of the variance permeability of the human nail and of a keratin in the permeability data. There is significant membrane from bovine hooves: Prediction of improvement in model quality upon the addition of the penetration rate of antimycotics through the 2 the TC to the TS descriptors, with an associated R c.v. nail plate and their efficacy, J. Pharm. value of 0.816. The addition of three-dimensional Pharmacol., Vol. 49, 1997, 866-872. descriptors, however, did not result in improved [2] S. C. Basak, D. K. Harriss, and V. R. Magnuson, model quality. This is in line with our earlier POLLY v.2.3, Copyright of the University of observations on QSARs of physicochemical Minnesota, 1988. properties, e.g., vapor pressure [18,19], boiling point [3] P. A. Filip, T. S. Balaban, and A. T. Balaban, A [20], etc., as well as biological properties, viz., new approach for devising local graph mutagenicity [21-23]; pharmacokinetically invariants: Derived topological indices with low important properties such as blood:air partition coefficient [24-26], tissue:air partition coefficient degeneracy and good correlational ability, J. [11]; Ah receptor binding [27]; and estrogen receptor Math. Chem., Vol. 1, 1987, 61-83. binding affinity of chemicals [28]. It is noteworthy [4] Molconn-Z Version 3.5, Hall Associates that the data set correlated here is small; but larger Consulting, Quincy, MA, 2000. data sets are not available in the published literature. [5] M. Randic, On characterization of molecular We will attempt to develop HiQSARs for the branching, J. Am. Chem. Soc., Vol. 97, 1975, permeability of larger sets of antimycotic agents 6609-6615. when they become available. [6] L. B. Kier and L. H. Hall, Molecular Connectivity in Structure-Activity Analysis, Research Studies Press, Letchworth, Table 1. Regression results (N = 14) Hertfordshire, U.K., 1986. [7] S. C. Basak, A. B. Roy, and J. J. Ghosh, Study of the structure-function relationship of 2 R c.v. pharmacological and toxicological agents using Descriptors RR PCR PLS information theory, (X. J. R. Avula, R. Bellman, TS 0.628 0.601 0.450 Y. L. Luke, and A. K. Rigler, Eds), 1979, 851- TS+TC 0.816 0.735 0.697 856. Rolla, Missouri, University of Missouri- TS+TC+3D 0.814 0.730 0.694 Rolla. [8] A. B. Roy, S. C. Basak, D. K. Harriss, and V. R. TS 0.628 0.601 0.450 TC 0.793 0.752 0.657 Magnuson, Neighborhood complexities and 3D 0.622 0.385 0.466 symmetry of chemical graphs and their biological applications. In, Mathl. Modelling Sci. Tech. (X.J.R. Avula, R.E. Kalman, A.I. Liapis, Assessment of the mutagenicity of aromatic and E.Y. Rodin, Eds.). Pergamon Press, 1983, amines from theoretical structural parameters: A pp. 745-750. hierarchical approach, SAR QSAR Environ. Res., [9] L. B. Kier and L. H. Hall, Molecular Structure Vol. 10, 1999, 117-129. Description: The Electrotopological State. [22] S. C. Basak and D. Mills, Prediction of Academic Press, San Diego, CA, 1999. mutagenicity utilizing a hierarchical QSAR [10] L. B. Kier and L. H. Hall, The kappa indices approach, SAR QSAR Environ. Res., Vol. 12, for modeling molecular shape and flexibility. In, 2001, 481-496. Topological Indices and Related Descriptors in [23] S. C. Basak, B. D. Gute, and G. D. Grunwald, QSAR and QSPR (J. Devillers and A.T. Balaban, Relative effectiveness of topological, Eds.). Gordon and Breach Science Publishers, geometrical, and quantum chemical parameters Amsterdam, 1999, pp. 455-489. in estimating mutagenicity of chemicals. In, [11] S. C. Basak, D. Mills, D. M. Hawkins, and H. Quantitative Structure-activity Relationships in A. El-Masri, Prediction of tissue:air partition Environmental Sciences VII (F. Chen and G. coefficients: A comparison of structure-based Schuurmann, Eds.). SETAC Press, Pensacola, and property-based methods, SAR QSAR FL, 1998, pp. 245-261. Environ. Res., Vol 13, 2002, 649-665. [24] S. C. Basak, D. Mills, D. M. Hawkins, and H. [12] SAS Institute, Inc. In SAS/STAT User Guide, El-Masri, Prediction of human blood:air partition Release 6.03, Cary, NC, 1988. [13] A. E. Hoerl and R. W. Kennard, Ridge coefficient: A comparison of structure-based regression: Biased estimation for nonorthogonal and property-based methods, Risk Analysis, Vol. 23, 2003, 1173-1184. problems. Technometrics, Vol 8, 1970, 27-51. [25] S. C. Basak, D. M. Hawkins, and D. Mills, [14] W. F. Massy, Principal components regression Predicting blood:air partition coefficient of in exploratory statistical research, J. Am. structurally diverse chemicals using theoretical Statistical Assoc., Vol. 60, 1965, 234-246. molecular descriptors. In, Advances in [15] I. E. Frank and J. H. Friedman, A statistical Molecular view of some chemometrics regression tools, Similarity (R. Carbo-Dorca and P.G. Mezey, Technometrics, Vol. 35, 1993, 109-135. Eds.). Kluwer, Amsterdam, in press. [16] H. Wold, Soft modeling by latent variables: [26] S. C. Basak, D. Mills, H. A. El-Masri, M. M. The nonlinear iterative partial least squares Mumtaz, and D. M. Hawkins, Predicting approach. In, Perspectives in Probability and blood:air partition coefficients using theoretical Statistics, Papers in Honor of M. S. Bartlett molecular descriptors, Environ. Toxicol. (J.Gani, Ed.). Academic Press, London, 1975. Pharmacol., Vol. 16, 2004, 45-55. [17] D. M. Hawkins, S. C. Basak, and D. Mills, [27] S. C. Basak, D. Mills, M. M. Mumtaz, and K. Assessing model fit by cross-validation, J. Balasubramanian, Use of topological indices in Chem. Inf. Comput. Sci, Vol. 43, 2003, 579-586. predicting aryl hydrocarbon receptor binding [18] S. C. Basak and D. Mills, Quantitative potency of dibenzofurans: A hierarchical QSAR structure-property relationships (QSPRs) for the approach, Indian J. Chem., Vol. 42A, 2003, estimation of vapor pressure: A hierarchical 1385-1391. approach using mathematical structural [28] S. C. Basak, D. Mills, and B. D. Gute, descriptors, J. Chem. Inf. Comput. Sci., Vol. 41, Predicting Bioactivity and Toxicity of Chemicals 2001, 692-701. [19] S. C. Basak, B. D. Gute, and G. D. Grunwald, from Mathematical Descriptors: A Chemical- Use of topostructural, topochemical, and cum-Biochemical Approach. In, Advances in geometric parameters in the prediction of vapor Quantum Chemistry (D.J. Klein and E. Brandas, pressure: A hierarchical approach, J. Chem. Inf. Eds.). Elsevier, in press. Comput. Sci., Vol. 37, 1997, 651-655. [20] S. C. Basak, B. D. Gute, and G. D. Grunwald, A comparative study of topological and geometrical parameters in estimating normal boiling point and octanol-water partition coefficient, J. Chem. Inf. Comput. Sci., Vol. 36, 1996, 1054-1060. [21] S. C. Basak, B. D. Gute, and G. D. Grunwald,