Computational classification models for predicting the interaction of compounds with hepatic organic ion importers

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54

DMPK50_proof ■ 1 July 2015 ■ 1/5

Drug Metabolism and Pharmacokinetics xxx (2015) 1e5

Contents lists available at ScienceDirect

Drug Metabolism and Pharmacokinetics journal homepage: http://www.journals.elsevier.com/drug-metabolism-andpharmacokinetics

Regular article

Computational classification models for predicting the interaction of compounds with hepatic organic ion importers Q4

Hwan You a, Kyungro Lee a, Sangwon Lee a, Sung Bo Hwang a, Kwang-Yon Kim b, Kwang-Hwi Cho c, Kyoung Tai No a, b, * a b c

Department of Biotechnology, Yonsei University, Seoul, Republic of Korea Bioinformatics & Molecular Design Research Center, Seoul, Republic of Korea Department of Bioinformatics and CAMDRC, Soong Sil University, Republic of Korea

a r t i c l e i n f o

a b s t r a c t

Article history: Received 19 March 2015 Received in revised form 19 May 2015 Accepted 17 June 2015 Available online xxx

Hepatic transporters, a major determinant of pharmacokinetics, have been used to profile drug properties like efficacy. Among hepatic transporters, importers alter the concentration of the drug by facilitating the transport of a drug into a cell. Despite vast pharmacokinetic studies, the interacting mechanisms of the importers with its substrates or inhibitors are not well understood. Hence, we developed compound binary classification models of whether a compound is binder or nonbinder to a hepatic transporter with experimental data of 284 compounds for four representative hepatic importers, OATP1B1, OATP1B3, OAT2, and OCT1. Support Vector Machine (SVM) along with Genetic Algorithm (GA) was used to construct the classification models of binder versus nonbinder for each target importer. To construct the models, we prepared two data sets, a training data set from Fujitsu database (284 compounds) and an external validation data set from ChEMBL database (1738 compounds). Since an experimental classification criterion between binder and nonbinder has some ambiguity, there is an intrinsic limitation to expect high predictability of the binary classification models developed with the experimental data. The predictability of the classification models calculated with external validation sets were obtained as 77.72%, 84.31%, 84.21%, and 76.38 for OATP1B1, OATP1B3, OAT2, and OCT1, respectively.

Keywords: Importers Drugedrug interaction (DDI) Classification Hepatocyte

Q2

Copyright © 2015, Published by Elsevier Ltd on behalf of The Japanese Society for the Study of Xenobiotics.

1. Introduction Membrane transporters enable the translocation of chemicals into and out of cells with active and passive transporting mechanisms. Some drugs and endogenous compounds with poor membrane permeability utilize the transporters for efficient translocation [1]. The transporters play an important role in drug metabolism and pharmacokinetics in the liver. Moreover, drugedrug or drugefood interactions caused by the perturbation of the transporters function are relevant to pathogenesis factors. Inhibition of transporting activity by certain drugs or food ingredients may raise pharmacokinetic problems with coexisting drug [2e4]. There were many cases where an increase in statin plasma concentration has been observed following the co-administration of cyclosporine in clinical studies. Cyclosporine inhibits the OATPs (Organic Anion Transporting Polypeptide) which are related with

Q1

* Corresponding author. Department of Biotechnology, Yonsei University, Seoul, Republic of Korea. Tel.: þ82 2 393 9551; fax: þ82 2 393 9554. E-mail address: [email protected] (K.T. No).

statin uptake. This is an example of transporter-mediated DDI (DrugeDrug Interaction). DDI in hepatocyte affects the PK (Pharmacokinetics) and PD (Pharmacodynamics) of multiple drugs. They also have potential to affect the efficacy and toxicity of drugs [5]. Theses DDIs are mediated by various transporters in hepatocyte as follows. The transporters are composed of several influx transporters including the Solute Carrier (SLC) superfamily and some efflux transporters including the ATP-binding cassette (ABC) superfamily. The transporters locate themselves in hepatocytes, intestinal epithelia, kidney proximal tubules and brain capillary endothelial cells [1]. In hepatocytes in particular, the importers in the sinusoidal membrane are Naþ-Taurocholate Cotransporting Polypeptide (NTCP), Organic Anion Transporting Polypeptide (OATP) family (OATP1B1, OATP1B3 and OATP2B1), Organic Anion Transporter (OAT) family (OAT2 and OAT7), and Organic Cation Transporter 1(OCT1) [1]. Among them, OATP1B1, OATP1B3, OAT2, and OCT1 are dominant in hepatocytes [6]. Several researchers have studied those four transporters. MeierAbt et al. suggested that the positively charged binding pocket in

http://dx.doi.org/10.1016/j.dmpk.2015.06.004 1347-4367/Copyright © 2015, Published by Elsevier Ltd on behalf of The Japanese Society for the Study of Xenobiotics.

Please cite this article in press as: You H, et al., Computational classification models for predicting the interaction of compounds with hepatic organic ion importers, Drug Metabolism and Pharmacokinetics (2015), http://dx.doi.org/10.1016/j.dmpk.2015.06.004

55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

DMPK50_proof ■ 1 July 2015 ■ 2/5

2

H. You et al. / Drug Metabolism and Pharmacokinetics xxx (2015) 1e5

OATP1B3 provides selectivity of substrates [7]. Mutational study of three amino acids (R57, K361, and R580) in the OATP1B1 binding pocket supports the hypothesis about the role of positively charged amino acids [8]. Chang et al. tried a meta-pharmacophore approach using limited data composed of diverse or unequal sources and conditions [9]. The meta-model for OATP1B1 produced a hydrophobic feature located at the center and hydrogen bond at the edge. CoMFA (Comparative Molecular Field Analysis) in rat model revealed that the substrate selectivity of OATs was influenced by hydrophobicity and electrostatic interaction [10]. Ahlin et al. suggested that features of transport activity in OCT1 may be related with hydrophobicity, low molecular mass, and positive charge and that the properties of inhibitors may include net positive charge and high lipophilicity [11]. These studies explained the property of each transporters' binding site well yet a systematic approach is essential to precisely predict DDIs in hepatocyte. In this study, we developed the physicochemical property based classification models for four hepatic ion importers, i.e., OATP1B1, OATP1B3, OAT2, and OCT1. Through the combination of classification results from four hepatic ion importer models, we could make an informed decision for DDI in hepatocyte. The classification models were constructed using a support vector machine with physicochemical descriptors selected by genetic algorithm.

2. Methods 2.1. Preparing data set for the classification models development Both substrates and inhibitors data set was built for each of OATP1B1, OATP1B3, OATP2B1, OAT2 and OCT1. The 284 compounds were taken from ADME database (http://jp.fujitsu.com/group/ kyushu/en/services/admedatabase) [12] and 3D structures generated with MMFF forcefield. Since ADME database were constructed from literature, the data were generated with various cell lines and culture conditions. If the compounds for the data sets were only taken from the experiments with the same cell type, such as HEK293 cell, we cannot secure enough data set to build models. For this reason, regardless of cell line type, data was established by compounds with target importers originated from human genome. Despite the alleviation of the data selection conditions, for OATP2B1, only 12 compounds are selected for the training set, OATP2B1 model was not developed. To test the predictability of the models, external validation sets were prepared with the compounds taken from ChEMBL DB (https://www.ebi.ac.uk/chembl) [13], 1804 compounds for OATP1B1, 1709 for OATP1B3, 26 for OAT2, and 245 for OCT1. In OATP1B1 and OATP1B3 data set, most of the compounds came from Tom De Bruyn et al. [14]. Tom De Bruyn et al. used CHO cell line and sodium fluorescein as substrate in inhibition experiment for OATP1B1 and OATP1B3. The inhibitory potential of all 2000 compounds from The Spectrum Collection library was tested at an equimolar substrate-inhibitor concentration of 10 mM. When the inhibition activity was higher than 50%, compounds were considered inhibitors and used as binder class for external validation. On the other hand, the compounds represent lower than 30% inhibition activity were deemed nonbinder class. The compounds with activity from 30 to 50% were excluded in the external validation set due to the ambiguity. 50(OATP1B1), 59(OATP1B3), 7(OAT2), and 46(OCT1) compounds were duplicated with compounds in ADME DB. Some compounds classified into nonbinder in ADME DB were changed to binder class and trained to build the model because they are reported for substrate or inhibitor in ChEMBL DB (3 compounds in OATP1B1, 4 in OATP1B3, 1 in OAT2, and 2 in OCT1, respectively). After preprocessing, 1360 compounds for OATP1B1,

1530 for OATP1B3, 19 for OAT2 and 199 for OCT1 were remained. They were also converted to 3D structure with MMFF forcefield. For model stability, we chose the more balanced data set, ADME DB as training set. Even though the size of the data in ChEMBL DB was larger than ADME DB, ChEMBL DB was unbalanced in ratio between OATPs and the others. The substrate and inhibitor composition of each transporter is summarized in Table 1(Training set) and Table 2(External validation set). The number of binders is not equal to the number of substrates plus inhibitors since some molecules can be both substrate and inhibitor by experiment design. 2.2. Classification model development 2.2.1. Definition of binder and nonbinder The classification criteria between the substrate and inhibitor are hard to simplify into a binary model in that even a single importer shows various binding patterns with multiple binding sites. In addition, an inhibition event of hepatic importers, unlike that of other enzymes, is not well-defined as an experimental protocol. Therefore, a realistic approach classifies whether a compound binds to a target transporter or not. We defined both substrates and inhibitors of a specific transporter as “binders of the transporter”. As a complementary class of “binders of the transporter”, we introduced “nonbinders of the transporter” where the member compounds have no positive experimental data with a specific importer. For example, Compound A, which is inhibitor, substrate, weak inhibitor, and untested compound for OATP1B1, OATP1B3, OAT2, and OCT1, respectively, is defined as binder, binder, nonbinder, and nonbinder for each importer. It shows that about 90% of nonbinders were correctly assumed by curating overlapping compounds between ADME DB and ChEMBL DB. 2.2.2. Descriptor selection In the present work, 102 molecular descriptors were generated with preADMET software [15] and 16 molecular descriptors were calculated using the Solvation Free Energy Density (SFED) model [16]. Descriptors consist of constitutional (3 þ 0), geometrical (26 þ 1), electrostatic (65 þ 12), and physicochemical (8 þ 3) descriptors (the first value in the parenthesis is originated from preADMET and the second value from the SFED model in parentheses). Descriptors were selected to build the model by two ways, manual selection and automatic selection. The first way of manually selecting 83 descriptors was based on the previous studies

Table 1 The composition details of binder (substrate and inhibitor) and nonbinder data set (ADME DB). Transporter

OATP1B1

OATP1B3

OAT2

Binder

131

81

41

Substrate Inhibitor

Nonbinder Sum

66 94

153 284

59 45

203 284

OCT1 24 30

243 284

101

33 94

183 284

The number of binder and nonbinder reflect the change by ChEMBL DB.

Table 2 The composition details of binder (substrate and inhibitor) and nonbinder data set (ChEMBL DB). Transporter

OATP1B1

OATP1B3

OAT2

Binder

247

163

16

Substrate Inhibitor Nonbinder Sum

1113 1360

29 238

1367 1530

21 147

3 19

OCT1 8 11

121

25 99

78 199

Please cite this article in press as: You H, et al., Computational classification models for predicting the interaction of compounds with hepatic organic ion importers, Drug Metabolism and Pharmacokinetics (2015), http://dx.doi.org/10.1016/j.dmpk.2015.06.004

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

DMPK50_proof ■ 1 July 2015 ■ 3/5

H. You et al. / Drug Metabolism and Pharmacokinetics xxx (2015) 1e5

[7e11]. The descriptors are composed of hydrophobicity-, hydrogen bonding-, and electrostatic interaction-related descriptors those are recognized as important physicochemical descriptors for describing molecular interaction in water. The second way, automatic selection, was performed with Genetic Algorithm (GA). All descriptors were generally scaled by statistical normalization. Two rounds of genetic algorithm analysis were employed for descriptors selection. By first selection, 53 descriptors selected from 16 models (4 models per each transporter) were collected for offering for second selection. 24 descriptors were selected by second selection through parameter optimization (C, kernel, gamma of kernel) for 4 transporters. Descriptors used to build models are listed in the Table S1. 2.2.3. Model development with support vector machine The predictive models were built by a cross validation method, one of the internal validation methods. In k-fold cross validation, the data were randomly partitioned into k subsets. Each subset is excluded in turn so that each compound is excluded from the training data exactly once. Advantage of this method is that all observations are used for both training and validation, and each observation is used for validation exactly once. Robustness of a model can be improved by such a property. Another reason for building the models with cross validation is that the ratio of binder to nonbinder is imbalanced for each transporter. Further, each compound had a class for four importers. Hence random selection to divide compounds into training set and validation set was not suitable. For example, even if compounds were randomly selected as validation set based on a specific transporter, it will not assure that the ratio between training and validation set for binder and nonbinder would be imbalanced in the other transporters. For the above mentioned reason, k-fold cross validation was employed in order to build the four hepatic importer models. A support vector machine, one of the supervised learning methods, was used for classification and regression analysis. It is based on the concept of margin proposed to overcome the limitation of original perceptron and has good generalization capability. It is also a meta-heuristic method that optimizes a problem by iteratively trying to improve the parameter set, such as kernel, kernel's parameter, and soft margin parameter C, with regard to model predictability. To combine LibSVM with genetic algorithm, gamma of kernel, and soft margin parameter C were optimized by the grid optimization approach. The parameters vary from 0 to 10 in gamma of kernel and 0 to 1 in margin parameter C. The C, gamma of kernel, and kernel type for the classification model were reported in Table 3. 2.3. Classification model validation The four hepatic importer models were evaluated with external validation set obtained from ChEMBL DB. Accuracy is used to indicate the robustness of the model. In this study, all machine learning calculations were carried out with RapidMiner 5.2 [17].

3. Results and discussion 3.1. Performance of classification models trained with ADME DB The performances of four GA-SVM models are shown in Table 4. Accuracies (Q) for OATP1B1, OATP1B3, OAT2, and OCT1 were 84.51, 82.04, 95.77, and 96.83, respectively. And True positive rates (Sensitivity:SE) were 83.21, 86.42, 100.0, and 96.04, respectively. Both accuracy and true positive rate were considered when selecting the best performance model among good performance models. Each model was constructed by physicochemical properties studied previously but some compounds failed to classify the class correctly due to a lack of comprehension regarding mechanisms of importers. The unknown mechanism for binding and variation of shape of binding sites due to the induced fit could be the sources of model error. Though the models predict binders versus nonbinders instead of substrate versus inhibitor, the prediction of binding possibility for each transporter could be significant starting point to predict the potential drugedrug or drugefood interactions in drug development. SVM model provides confidence scores that show the possibility of membership for a compound to each class according to distance from hyperplane. There are three compounds in the nonbinder class with a confidence score of binder class greater than 0.75. These are: vinblastine in OATP1B1, olmesartan medoxomil in OATP1B3, and fenofibric acid in OAT2. These three false positives had extremely similar structure with known binders including their physicochemical properties. Structures are reported in Figs. 1e3. Vinblastine has oxygen atoms that could mediate hydrogen bonds and hydrophobic center consisted with fused rings, which is important feature of OATP1B1. But it was not suitable to interact with OATP1B1 that excessive rigidity at the top and bottom side. Olmesartan medoxomil is a prodrug of olmesartan known as substrate of OATP1B3. A prodrug indicates the inactivated form, and it is activated by enzymatic reaction such as hydrolysis. Since the difference between olmesartan and olmesartan medoxomil is an existence of medoxomil only, olmesartan medoxomil might inhibit the OATP1B3 transporting activity or interact with OATP1B3 as substrate. Fenofibric acid has three features: (1) Oxygen atoms in carboxylic acid functional group, (2) Small size of backbone, and (3) Freely rotatable structure. It is similar with representative structure, small and acidic, so makes prediction wrong. Applicable descriptor space was summarized in the Table S2. Four models had high accuracy within this descriptor space. If a compound existed outside of the space, the prediction of the model might have inaccurate performance. 3.2. External validation with ChEMBL DB The four hepatic importer models were evaluated by using a data set from ChEMBL DB. The robustness of the models is shown in Table 5. Table 4 Performances of GA-SVM models (Cross validation). Transporter TP

Table 3 Parameters for GA-SVM model. Transporter

C

g of kernal

Kernal type

OATP1B1 OATP1B3 OAT2 OCT1

0.52 0.64 0.32 0.0

0.05 0.30 2.32 1.05

Radial basis function (RBF) RBF RBF RBF

Round off the numbers to three decimal places.

3

OATP1B1 OATP1B3 OAT2 OCT1

FP

TN

FN SE SP, (TPR), % %

109 22 131 22 70 40 163 11 41 12 231 0 97 5 178 4

83.21 86.42 100.00 96.04

85.62 80.30 95.06 97.27

Q, %

F_measure, Kappa %

84.51 82.04 95.77 96.83

83.21 73.30 87.23 95.57

0.688 0.602 0.848 0.931

TP ¼ True positive, FP ¼ False positive, TN ¼ True negative, FN ¼ False negative, SE ¼ Sensitivity, True positive rate, SP ¼ Specificity, Q ¼ accuracy, SE ¼ TP/(TP þ FN). SP ¼ TN/(TN þ FP). Q ¼ (TP þ TN)/(TP þ TN þ FP þ FN), Fmeasure ¼ 2*precision*recall/precision þ recall, kappa ¼ accuracy  E/1  E, E ¼ (TP þ FN) (TP þ FP) þ (TN þ FP) (TN þ FN)/(TP þ FP þ FN þ TN)2.

Please cite this article in press as: You H, et al., Computational classification models for predicting the interaction of compounds with hepatic organic ion importers, Drug Metabolism and Pharmacokinetics (2015), http://dx.doi.org/10.1016/j.dmpk.2015.06.004

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

DMPK50_proof ■ 1 July 2015 ■ 4/5

4

H. You et al. / Drug Metabolism and Pharmacokinetics xxx (2015) 1e5

OATP1B3 and OAT2 classification models were relatively unstable compare with OATP1B1 and OCT1. It can be inferred through the sensitivity of OATP1B3 and specificity of OAT2. In the case of OAT2, the model was more fitted to the binder class because of the small number of binder class when training the model so binder criteria were relatively loosened and nonbinder compounds were predicted as binder class in external validation. In the case of OATP1B3, 40 compounds tagged with nonbinder class were predicted as binder class. In terms of experimental results, nonbinder class of OATP1B3 has inhibition rate but relatively lower than binders. It can weaken the discriminating power of the model. OATP1B1, OATP1B3, OAT2, and OCT1 showed reliable predictability and robustness with ChEMBL DB. It can be shown that descriptors selected through previous studies are suitable for representing the binding tendency of four hepatic importers. Though the experimental conditions and cell lines were not

Fig. 3. The 2D structure of fenofibric acid known as nonbinder in OAT2.

controlled, compounds obtained from ADME DB also have sufficient diversity and functionality to build the binding model. 3.3. Model validation by Y-randomization The reliability of the binary classification models was assessed by Y-randomization to exclude the possibility of chance correlation. Y-randomization was performed for four classification models wherein classes of training set were randomized. The statistical data on true positive rates for 10 runs are summarized in Table 6. The true positive rate of external set proves that the predictive ability of the four classification models is robust enough. 3.4. Molecular descriptors in each classification model Fig. 1. The 2D structure of vinblastine known as nonbinder in OATP1B1.

Each model included hydrophobicity-, hydrogen bonding-, charged interaction-related and physicochemical descriptors. It was possible to infer the properties of each transporter from selected descriptors. Among transporters, the combination of descriptors had a similar pattern, but a few differences existed. Descriptors responsible for the differences were logD in OATP1B1, acidity in OATP1B3 and basicity in OCT1. The logD descriptor describes hydrophobicity in the specific environment; pH 7.4 and indicates influence of microstate of compounds at the binding site of OATP1B1. In pH 7.4 condition, some compounds experienced conformational and distributional change of electrons regardless of whether the hydrogen is attached to the molecule. So modified

Table 5 Performances for external validation set (ChEMBL DB). Transporter TP

Fig. 2. The 2D structure of olmesartan medoxomil known as nonbinder in OATP1B3.

OATP1B1 OATP1B3 OAT2 OCT1

FP

TN

FN SE (TPR), SP, % %

185 241 872 62 74.90 114 191 1176 49 69.94 13 3 3 0 100.00 99 25 53 22 81.82

78.35 86.03 50.00 67.95

Q, %

F_measure, Kappa %

77.72 84.31 84.21 76.38

54.98 48.72 89.66 80.82

0.415 0.404 0.578 0.501

Please cite this article in press as: You H, et al., Computational classification models for predicting the interaction of compounds with hepatic organic ion importers, Drug Metabolism and Pharmacokinetics (2015), http://dx.doi.org/10.1016/j.dmpk.2015.06.004

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58

DMPK50_proof ■ 1 July 2015 ■ 5/5

H. You et al. / Drug Metabolism and Pharmacokinetics xxx (2015) 1e5

59 60 61 This research is supported by Korea Civil Military Technology 62 Cooperation Center as “Dual Use Technology Cooperation 63 Program”. Q3 64 65 Appendix A. Supplementary data 66 67 Supplementary data related to this article can be found at http:// 68 dx.doi.org/10.1016/j.dmpk.2015.06.004. 69 70 71 References 72 73 [1] Giacomini KM, Huang SM, Tweedie DJ, Benet LZ, Brouwer KL, Chu X, et al. Membrane transporters in drug development. Nat Rev Drug Discov 2010;9: 74 215e36. 75 [2] Neuvonen PJ, Niemi M, Backman JT. Drug interactions with lipid-lowering 76 drugs: mechanisms and clinical relevance. Clin Pharmacol Ther 2006;80: 565e81. 77 [3] Fahrmayr C, Fromm MF, Konig J. Hepatic OATP and OCT uptake transporters: 78 their role for drug-drug interactions and pharmacogenetic aspects. Drug 79 Metab Rev 2010;42:380e401. [4] Burckhardt G, Burckhardt B. In vitro and in vivo evidence of the importance 80 of organic anion transporters (OATs) in drug therapy. In: Fromm MF, 81 Kim RB, editors. Drug transporters. Berlin Heidelberg: Springer; 2011. 82 p. 29e104. [5] MüLler F, Fromm MF. Transporter-mediated drug-drug interactions. Phar83 macogenomics 2011;12:1017e37. 84 [6] Hilgendorf C, Ahlin G, Seithel A, Artursson P, Ungell AL, Karlsson J. Expres85 sion of thirty-six drug transporter genes in human intestine, liver, kidney, and organotypic cell lines. Drug Metab Dispos Biol Fate Chem 2007;35: 86 1333e40. 87 [7] Meier-Abt F, Mokrab Y, Mizuguchi K. Organic anion transporting polypeptides 88 of the OATP/SLCO superfamily: identification of new members in non89 mammalian species, comparative modeling and a potential transport mode. J Membr Biol 2005;208:213e27. 90 [8] Weaver YM, Hagenbuch B. Several conserved positively charged amino acids 91 in OATP1B1 are involved in binding or translocation of different substrates. 92 J Membr Biol 2010;236:279e90. [9] Chang C, Pang KS, Swaan PW, Ekins S. Comparative pharmacophore modeling 93 of organic anion transporting polypeptides: a meta-analysis of rat Oatp1a1 94 and human OATP1B1. J Pharmacol Exp Ther 2005;314:533e41. 95 [10] Kaler G, Truong DM, Khandelwal A, Nagle M, Eraly SA, Swaan PW, et al. Structural variation governs substrate specificity for organic anion transporter 96 (OAT) homologs. Potential remote sensing by OAT family members. J Biol 97 Chem 2007;282:23841e53. 98 [11] Ahlin G, Karlsson J, Pedersen JM, Gustavsson L, Larsson R, Matsson P, et al. Structural requirements for drug inhibition of the liver specific human organic 99 cation transport protein 1. J Med Chem 2008;51:5932e42. 100 [12] Fujitsu ADME database. 101 [13] ChEMBL Database (19 release, July, 2014). [14] De Bruyn T, van Westen GJ, Ijzerman AP, Stieger B, de Witte P, Augustijns PF, 102 et al. Structure-based identification of OATP1B1/3 inhibitors. Mol Pharmacol 103 2013;83:1257e67. 104 [15] preADMET v2.0. 105 [16] Lee S, Cho KH, Acree WE, No KT. Development of surface-SFED models for polar solvents. J Chem Inf Model 2012;52:440e8. 106 [17] Mierswa I, Wurst M, Klinkenberg R, Scholz M, Euler T. YALE: rapid prototyping 107 for complex data mining tasks. In: Ungar L, Craven M, Gunopulos D, Eliassi108 Rad T, editors. KDD'06: proceedings of the 12th ACM SIGKDD international conference on knowledge discovery and data mining. New York, NY, USA: 109 ACM; 2006. p. 935e40. 110 [18] Lee S, Cho K-H, Lee CJ, Kim GE, Na CH, In Y, et al. Calculation of the solvation 111 free energy of neutral and ionic molecules in diverse solvents. J Chem Inf Model 2010;51:105e14. 112 [19] Jeremy M, Berg JLT, Stryer Lubert. Biochemistry. 5th ed. New York: W 113 H Freeman; 2002. 114 115 116 Acknowledgment

Table 6 Y-randomization data for four hepatic classification models. Model

OATP1B1

OATP1B3

OAT2

OAT1

Original Y-randomizationa

74.90 29.70 ± 8.04

69.94 53.69 ± 5.38

100.00 0.43 ± 1.36

81.82 28.59 ± 2.84

a The Y-randomization values represent the mean ± deviation of sensitivities from 10 independent runs.

hydrophobicity of the compounds was used to determine the binding affinity for OATP1B1. OATP1B3 models had an acidity descriptor that represented the summation of acidity on acidic functional groups. The difference was only minimum acidity of the functional groups included in the summation. Unlike viewpoint of transporter researchers, acidity descriptors were selected in the OATP1B3 model. Selected acidity descriptors did not mean the more acidic compounds bound to OATP1B3 because they used as deterministic descriptors to separate binders from nonbinders. Contrary to acidity, basicity of OCT1 expressed the characteristics of cation transporter. It was a distinct feature obtained from relative complement of OCT1 binders. Moreover the OATP1B1 model had a SFED descriptor [18]. SFED represents the effect of a solution on a chemical or biological process [16]. Because the descriptor was calculated based on the water solvent it indicated hydration free energy of compounds describing the effect of water in hepatic transporters.

4. Conclusion In this study, we found descriptors for explaining binding affinity of each transporter (OATP1B1, OATP1B3, OAT2, and OCT1) and developed binary classification model for four hepatic ion importers with high predictive accuracy using physicochemical properties based GA-SVM. The results of cross validation and external validation showed high accuracy and stability of our models, and they indicated the different features of hepatic ion importers. By applying the classification model, potential of drugedrug interaction or drugefood interaction was predicted when two or more compounds were treated at the same time. Furthermore, the model classifying the substrate versus inhibitor failed to build because of the unknown action mechanism of transporters for each compound. To be specific, some substrates are also inhibitors in parallel and there are multiple binding sites in each transporter. In addition, there are 4 inhibitory mechanisms; competitive inhibition, uncompetitive inhibition, mixed inhibition, and noncompetitive inhibition [19]. It is difficult for a ligand based approach to handle the above problems in a diverse data set. In the future, if the crystal structure of the transporter is revealed and binding site of the transporter and mechanism of the inhibitor are studied further, a more improved classification model for broad data set would be developed.

5

Please cite this article in press as: You H, et al., Computational classification models for predicting the interaction of compounds with hepatic organic ion importers, Drug Metabolism and Pharmacokinetics (2015), http://dx.doi.org/10.1016/j.dmpk.2015.06.004