Discovery of novel anti-inflammatory drug-like compounds by aligning in silico and in vivo screening: The nitroindazolinone chemotype

European Journal of Medicinal Chemistry 46 (2011) 5736e5753

Contents lists available at SciVerse ScienceDirect

European Journal of Medicinal Chemistry journal homepage: http://www.elsevier.com/locate/ejmech

Original article

Discovery of novel anti-inflammatory drug-like compounds by aligning in silico and in vivo screening: The nitroindazolinone chemotype Yovani Marrero-Ponce a, b, c, *, Dany Siverio-Mota b, f, María Gálvez-Llompart c, d, María C. Recio d, Rosa M. Giner d, Ramón García-Domènech c, Francisco Torrens a, Vicente J. Arán e, María Lorena Cordero-Maldonado f, Camila V. Esguera f, Peter A.M. de Witte f, Alexander D. Crawford f a

Institut Universitari de Ciència Molecular, Universitat de València, Edifici d’Instituts de Paterna, P. O. Box 22085, 46071 València, Spain Unit of Computer-Aided Molecular “Biosilico” Discovery and Bioinformatic Research (CAMD-BIR Unit), Faculty of Chemistry-Pharmacy, Universidad Central “Martha Abreu” de Las Villas, Santa Clara, 54830 Villa Clara, Cuba c Unidad de Investigación de Diseño de Fármacos y Conectividad Molecular, Departamento de Química Física, Facultad de Farmacia, Universitat de València, València, Spain d Department of Pharmacology, Faculty of Pharmacy, Universitat de València, València, Spain e Instituto de Química Médica, CSIC, c/Juan de la Cierva 3, 28006 Madrid, Spain f Department of Pharmaceutical Sciences, University of Leuven, Herestraat 49, 3000 Leuven, Belgium b

a r t i c l e i n f o

a b s t r a c t

Article history: Received 27 May 2011 Received in revised form 28 July 2011 Accepted 29 July 2011 Available online 17 August 2011

In this report, we propose the combination of computational methods and in vivo primary screening in zebrafish larvae and confirmatory in mice models as a novel strategy to accelerate anti-inflammatory drug discovery. Initially, a database of 1213 organic chemicals with great structural variability e 587 of them anti-inflammatory agents plus 626 compounds with other clinical uses e was divided into training and test groups. Atom-based quadratic indices e a TOMOCOMD-CARDD molecular descriptors family e and linear discriminant analysis (LDA) were used to develop a total of 13 models to describe the anti-inflammatory activity. The best model (Eq. (13)) shows an accuracy of 87.70% in the training set, and values of Matthews correlation coefficient (C) of 0.75. The robustness of the models was demonstrated using an external test set as validation method, i.e., Eq. (13) revealing classification of 88.44% (C ¼ 0.77) in this series. All models were employed to develop ensemble a QSAR classification system, in which the individual QSAR outputs are the inputs of the aforementioned fusion approach. The fusion model was used for the identification of novel anti-inflammatory compounds using virtual screening of 145 molecules available in our in-house library of indazole, indole, cinnoline and quinoxaline derivatives. Out of these, 34 chemicals were selected, synthesized and tested in a lipopolysaccharide (LPS)-induced leukocyte migration assay in zebrafish larvae. This activity was evaluated based on leukocyte migration to the injury zone of tail-transected larvae. Compounds 18 (3 mM), 24 (10 mM), 25 (10 mM), 6 (10 mM), 15 (30 mM), 11 (30 mM) and 12 (30 mM) gave the best results displaying relative leukocyte migration (RLM) values of 0.24, 0.27, 0.35, 0.41, 0.17, 0. 26 and 0.27 respectively, date that suggest an anti-inflammatory activity of 76, 73, 65, 59, 83, 84 and 73%, respectively. Compound 18 was the most potent but showed high toxicity together with compound 6. Next, we used the tetradecanoylphorbol acetate (TPA)-induced mouse ear oedema model to evaluate the most potent compounds in the zebrafish larvae tail transection assay. All assayed compounds, with the exception of chemical 15, showed anti-inflammatory activity in mice. Compound 12 (VA5-13l, 2-benzyl-1-methyl-5-nitro-1,2-dihydro-3H-indazol-3-one) was the most active and completely abolished the oedema. Compounds 6, 11 and 24 showed inhibition percentages in the range of the reference drug (indomethacin), whereas compounds 18 and 25 reduced the oedema in a lesser extent (inhibition of 73 and 80%, respectively). In addition, all compounds except chemical 15, significantly reduced neutrophil infiltration, measured as myeloperoxidase activity on TPA application test. Compounds 6, 11, 12 and 18 showed values comparable to indomethacin (inhibition percentage of 61), but compounds 6 and 18 were toxic in zebrafish and showed unspecific cytotoxicity in murine

Keywords: TOMOCOMD-CARDD software Atom-based quadratic indices Learning machine-based QSAR Anti-inflammatory database Computational screening Lead generation In vivo anti-inflammatory assay Zebrafish Cytotoxicity Anti-inflammatory compound

* Corresponding author. Institut Universitari de Ciència Molecular, Universitat de València, Edifici d’Instituts de Paterna, P. O. Box 22085, 46071 València, Spain. Tel./fax: þ34 963543156. E-mail addresses: [email protected], [email protected] (Y. Marrero-Ponce), [email protected] (M. C. Recio), [email protected] (V.J. Arán), alexander.crawford@ pharm.kuleuven.be (A.D. Crawford). URL: http://www.uv.es/yoma/ 0223-5234/$ e see front matter
2011 Elsevier Masson SAS. All rights reserved. doi:10.1016/j.ejmech.2011.07.053

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

5737

macrophages at 100 mg/mL, while the remaining compounds 11, 12 and 25 were inactive at most levels. Evidently, this study suggests a new support structure (12, 11 and 24; a nitroindazolinone chemotype) that constitutes a novel promising lead and may represent an important therapeutic alternative for the treatment of inflammatory conditions.
2011 Elsevier Masson SAS. All rights reserved.

1. Introduction “.chemoinformatics is the combination of chemical synthesis, biological screening and data analysis to guide drug discovery and development.” Blake, F. J. Curr. Opin. Chem. Biol. 2004, 4, 407 The development of a new drug is a lengthy and complex process. The identification of an appropriate lead molecule (and its structural optimization) is the most critical component in this phase [1,2]. Over the past few decades, a primary source for novel leads has been the high-throughput screening (HTS) of compound libraries [3]. The advent of virtual screening (either ligand- or structure-based) methods to identify a reduced number of molecules with increased potential for bioactivity to be experimentally evaluated has emerged both as a complementary and alternative method to HTS [4e9]. The ligand-based (LB) methods are supported in the principle of similarity [10] and serve to model the complex phenomena of molecular recognition. Therefore, LB virtual screening (LBVS) has been used to identify novel active compounds in many biological applications. This indicates that ‘similarity’ methods should have substantial ‘selectivity’ in recognizing diverse active compounds [7e9,11,12]. Current efforts to integrate chemoinformatics into “real-life” applications, to improve drug discovery, are currently topics of interest. Following this aim, and because drug discovery is a complex process that requires the evaluation of large amounts of chemical data, it could be said that in silico predictions are suitable to detect the biological activity under study. Therefore, some of our research teams have already reported several cheminformatic studies to drive the selection of novel chemicals as promising new chemical entities (NCEs). In these studies, the TOMOCOMD-CARDD (acronym of TOpological MOlecular COMputational Design Computer-Aided-Rational-Drug Design) method [13] and linear discriminant analysis (LDA) [14] have been used mainly to parameterize all molecules in a database and to develop classification functions, respectively. LDA is one of the most important and simple (supervise, linear and parametric) pattern recognition techniques that could be used to determine which variables discriminate between two or more naturally occurring groups (it is used as either a hypothesis testing or exploratory method-data mining) [14,15]. At present, LDA has become a significant statistical tool and is used in chemometric analysis and drug design studies [12,16e19]. The TOMOCOMDCARDD approach is a novel scheme to the rational ein silicomolecular design and to QSAR/QSPR [20e25]. It calculates several new families of 2D, 3D-Chiral (2.5) and 3D (geometric and topographic) non-stochastic and (simple and double) stochastic (as well as their canonical forms) atom- and bond-based molecular descriptors (MDs) based on algebraic theory and discrete mathematics. They are denominate quadratic, linear and bilinear indices and have been defined in analogy to the quadratic, linear and bilinear mathematical maps [20e25]. These approaches describe changes in the electron distribution with time throughout the molecular backbone and they have been successfully employed in the prediction of several physical, physicochemical, chemical biological and pharmacokinetical properties of organic compounds [21,23,26e40]. However, our research group has not reported any

classification-based QSAR model for anti-inflammatory activity to date. Inflammation is a normal and essential response to any deleterious stimuli that threatens the host and may vary from a localized response to a more generalized one. In the absence of inflammation, wounds and infections would not heal, leading to progressive tissue destruction and thereby compromising the survival of the organism. Nevertheless, inflammatory responses can also be excessive in terms of magnitude and/or duration, and may therefore result in pain, tissue damage, or chronic inflammation when not properly resolved within an appropriate timeframe [41e46]. Therefore, anti-inflammatory drugs have an important clinical role in the control of a response and resolution of an inflammatory process in the host. Although anti-inflammatory agents [glucocorticoids (GCs) and non-steroidal anti-inflammatory drugs (NSAIDs)] [47] are rather common and familiar to most scientists, most current drugs used to treat these inflammatory conditions are decades old and have many limitations, including severe side-effects (toxicity), low-to-medium efficacy or selectivity, price and other important inconveniences. For instance, among other side-effects GCs are responsible of Cushing syndrome, osteoporosis, suppression of hypothalamusepituitaryeadrenal axis and reduced rate of bone growth in children [48e51]. Moreover, NSAIDs inhibit the overproduction of inflammatory mediators [by cyclooxigenase (COX) enzyme inhibition], thus preventing a long-term administration. In addition, these drugs are characterized by their propensity to produce adverse gastrointestinal effects including dyspepsia, gastric erosions and ulceration, as well as bleeding [52]. These drawbacks of the current anti-Inflammatory therapy urge the search for new and safer drugs that would target chronic inflammatory conditions such as osteoarthritis and rheumatoid arthritis [53], disorders affecting million people worldwide [47,54]. In the present report, we will explore the potential of TOMOCOMD-CARDD MDs to seek a QSAR-based ensemble classifier, for anti-inflammatory drug-like compounds from a heterogeneous series of compounds. In the initial step, we selected for the first time a wide-spectrum database of anti-inflammatory drugs. Next, the aforementioned MDs (specifically, the total and local nonstochastic and stochastic quadratic indices) were calculated for this large series of active/non-active compounds, and LDA was subsequently used to fit every individual classification function. Later, we developed a multi-agent QSAR classification system (ensemble classifier), in which the individual QSAR outputs are the inputs of the aforementioned fusion approach. Finally, the fusion model was used for the identification of a novel anti-inflammatory lead-like by using LBVS of small-molecules ‘available’ (with synthetic feasibility) in our ‘in-house’ library. Additionally, we used two in vivo-based assays carried out in zebrafish (Danio rerio) larvae and in rodent models, which are suitable to describe a complete profile of anti-inflammatory activity for new chemicals. Here, we show three different experiments developed for this study. First, we comment the results obtained in the construction of classification models and their assembling by using a fusion-like approach (multiagent-system). Each individual model was evaluated based on the guidelines set up in the principles of the Organization for Economic Cooperation and Development (OECD) [55].

5738

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

Later, we explored the ability of our classification models to find new active compounds carrying out an experiment of a lead generation. Moreover, we search an -in house- dataset of organic chemicals through LBVS, in order to discover novel candidates for anti-inflammatory drug-like compounds. These results encouraged us to develop the novel anti-inflammatory active compounds. Afterwards, the candidates elected by our models (34 in total) were biologically evaluated, by using an in vivo model of acute inflammation in zebrafish. Because of its genetically tractability, zebrafish has emerged as a versatile experimental model to examine mechanisms of human disease as well as for the screening of smallmolecules [56e58]. From a logistical viewpoint, this popularity is linked to the ease of maintaining large colonies of fish, relatively short generation times, and that a single spawning can yield hundreds of experimentally useful offspring. A further appeal is that zebrafish eggs are fertilized externally, allowing early embryonic developmental stages to be accessible for study and observation; optical transparency of zebrafish larvae at these early time points is an additional advantage. On top, the small size of larvae makes them amenable to assay in multiwell plates with the compound simply dissolved in the medium. Also, considerable genomic and genetic resources already exist for zebrafish. Since there is a high level of conservation of genetic pathways and cellular function among the vertebrates, zebrafish can be used to screen for compounds affecting pathways of relevance to human disease [56e58]. In summary, the zebrafish is a well-characterized model organism used in the screening of potential drug candidates to provide invaluable in vivo safety and efficacy data from the earliest stages of drug discovery and throughout the development process. The presence of both innate and adaptive immune systems in zebrafish provides support for the utility of zebrafish as a tool to examine the role of immune cells in normal development and in the pathogenesis of disease states [56,59e61]. For example, Mathias et al. [56] have described a zebrafish chronic inflammation mutant identified in an insertional mutagenesis screen for mutants that exhibit abnormal tissue distribution of neutrophils. More recently, Renshaw et al. [62] established an in vivo model for genetic analysis of the inflammatory response, by generating a transgenic zebrafish line that expresses GFP under the neutrophilspecific myeloperoxidase promoter. This report showed that inflammation is induced after transection of the tail in zebrafish larvae, which subsequently resolves over a time course to similar mammalian systems. Despite progress in characterizing the zebrafish immune system as well as a zebrafish model for chronic inflammatory disease, no anti-inflammatory test has been developed to date. However, in this report we used a newly described anti-inflammatory screening performed in zebrafish larvae thus facilitating the rapid analysis of large numbers of compounds [63]. Finally, we describe the biological characterization in two mouse-based anti-inflammatory tests (including myeloperoxidase assay in mouse ear oedema tissues) which will be presented in order to close the lead discovery cycle (experimental corroboration). In this experiment only the seven most potent chemicals were evaluated in mice in order to confirm the result from zebrafish test. This theoretical-(dry)-to-experimental(wet) integration will be used here in order to identify predictive and experimental in vivo models that permit the ‘rational’ identification of new antiinflammatory drug-like compounds. 2. Results and discussion 2.1. Computational in silico anti-inflammatory activity modelling The development of discriminant functions that allows the classification of organic-chemical drugs as either active or inactive

is a key step in the present approach for the discovery of new widespectrum anti-inflammatory agents. It is well-known that the general performance and extrapolation power of the learning methods decisively depends on the selection of compounds for the training series, used to build the classifier model [64]. It was therefore necessary to select a training dataset of active and inactive compounds, containing broad structural variability and action modes, as well as therapeutic uses. Therefore, the endpoint (first principle) here is the classification of chemicals into two different experimental classes: anti-inflammatory (1) and non-antiInflammatory (-1) drug-like compounds. Moreover, the antiInflammatory activity is our define QSAR “endpoint”, which can be measured and therefore modelled. To ensure the molecular and pharmacological diversity, we have selected a benchmark dataset composed of a great number of molecular entities, some of them reported as anti-inflammatory [65e67] and others with pharmacological uses [65,66]. Here, we selected a dataset of 1213 organic chemicals with a great degree of structural variability. The 587 anti-inflammatory compounds, considered in this study, are representative of families with different inhibition modes and diverse structural patterns (see Fig. 1). The selection of training and prediction sets was randomly performed. From these 1213 compounds, 919 were randomly chosen to form the training set, 443 out of them being active, and 476 inactive. The remaining sub-series, composed of 144 antiinflammatory compounds and 150 compounds with different biological properties, were prepared as test sets for the external set validation of the models. According to OECD Validation Principle 2, a (Q)SAR should be expressed in the form of an unambiguous algorithm. The intent of this principle is to ensure transparency in the description of the model algorithm. Therefore, while developing a method for predicting anti-inflammatory activity, the first problem we face is how to represent the sample of a molecule. Here we used a defined mathematical algorithm, which is characterized in this case by two atom-based TOMOCOMD-CARDD MD families (non-stochastic ½AP qk ðxÞ and stochastic ½APs qk ðxÞ quadratic indices) [22,27,29,33, 37,68]. These quadratic maps use a complete scheme of atomic properties (AP), which characterizes a specific aspect of the atomic structure (and k mean order, k ¼ 1e15). The weights (atomic-labels) used in this work are those previously proposed for the calculation of the DRAGON descriptors [69], i.e., atomic mass (AP ¼ M), atomic polarizability (AP ¼ P), atomic Mullinken electronegativity (AP ¼ K) plus van der Waals atomic volume (AP ¼ V). All quadratic indices were calculated taking into account all H-atoms in the molecule, H H i.e., ½AP qk ðxÞ and ½APs qk ðxÞ for non-stochastic quadratic indices and their stochastic counterparts, respectively. Two local (L) atomgroup indices for heteroatoms (group ¼ heteroatoms (E): E ¼ S, N, O), not considering ½AP qkL ðxE Þ and considering ½AP qkL ðxE Þ Hatoms in the molecule, were also computed. The representative selection of the training set gives place to the next step related to finding classification functions to discriminate between active and inactive molecules. For this we select the LDA as statistical technique due to it’s broadly use and simplicity. As we describe above, LDA is also a statistical technique with a defined algorithm, therefore on the OECD basis the second principle is proposed as being satisfactorily met. Several classification functions were developed by using the total and group-local non-stochastic and stochastic quadratic indices, as independent variables. These MDs were computed with the TOMOCOMD-CARDD software; according to the weighting schemes proposed above (see also Table SI1). Thus, 13 LDA-based QSAR models were obtained. The first six models were developed by using the non-stochastic total and local atom-based quadratic indices (Eqs. (1)e(6)), and six corresponding equations were

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

CH3 H3C

O

OH CH3

O S

N

O

5739

OH

O

CH3 CH3

O

H N

O

N H

O

Ibuprofen

OH

O

Piroxicam

H Cortisone

O

Aspirin

Cl

O

OH H

HO

O S

N H3C

CH3 O

H N

H3C

Indomethacin

O

H H2N

O S

N

CH3

H

N

H

O

N N Fenamole

O

Ketoprofen

O Diclofenac S S OH OH HO O O O OH Tiferron H N N S N N OH O O Sulfasalazine OH O

HO

CH3

Prednisolone

OH

O

HO

CO

OH

O

HO

OH

CH3

OH

OH OH

COOH

O

S

O N

OH

N

O Cl

O Guaimesal

Amiprilose

O HO

O

N

O

OH

N

Cl

Ditazole

O

N H

O

O

O

N H

CH3

OCH3

N

O

CH3

Cloxacepride

H N

COOH Cl

O

N

O

O

O

OH

O

Talniflumate

HO

S

Tenidap

CF3

H N

O

N

N

H3CO

O

Betamethasone benzoate

NH2

O

H N

Tenoxicam

O

O

O

CH3

H

F

CH3

O

S

N

Aphloiol

Sulindac O

CH2 OH

COO CH3

H

OH

F

Cl

OH

Nimesulide

OH

H N

OH

Tizoprolic acid NO2

OH

O

HO

N

S O O

O

Cl

CH3

O

O

CH3 O

HO

S

N S N H

ICI 115432 Cl

Bay 08276

HP 573 Cl

Fig. 1. Structure of some representative anti-inflammatory agents included in this in silico study.

O

5740

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

achieved by using the stochastic MDs (Eqs. (7)e(12)). The last one was obtained by mixing another set of MDs, for instance nonstochastic and stochastic quadratic indices (Eq. (13)). All equations of these classification models are shown in Table 1. Overall performances of all the obtained models are given in Table 2, together with the Wilks’ statistics (l), the square of the Mahalanobis distances (D2), and the Fisher ratio (F). The selected models show to be statistically significant at p-level <0.001. Table 2 shows most parameters commonly used in medicinal statistics [sensitivity, specificity, Matthews correlation coefficient (C) and false positive rate] for the whole set of developed models. While the sensitivity is the probability of correctly predicting a positive example, the specificity is the probability that a positive prediction is correct. On the other hand, C quantifies the strength of the linear relation between the molecular descriptors and the classifications, usually providing an evaluation of the prediction much more balanced than, for instance, the accuracy [70].

The fitted models 6 and 12, resulting of the combination of weighting schemes for the non-stochastic and stochastic atomlevel quadratic indices, respectively, as well as Eq. (13) (mixing non-stochastic and stochastic indices) exhibit the best results as can be observed in Table 2. The models represented by Eqs. (6) and (12) were classified correctly, 87.60% and 87.38% of accuracy in the training set, respectively. The equations showed fitted C of 0.75 in both cases. However, the best-to-similar result is performed when all set of MDs was used. Eq. (13) showed 87.70% of global good classification and C ¼ 0.75. Receiver operating characteristic curve, ROC, for the training set is shown in the Fig. 2 (Eq. (13)). The area under the curve (AUC) is 0.9395. The posterior probabilities (ΔP%) of the chemicals used in the training set appear in Table SI3 and Table SI4 of the Supporting Information, which summarize the results of the classifications of each compound for all models represented by Eqs. (1)e(13).

Table 1 Discriminant models obtained with total and local non-stochastic and stochastic quadratic indices. LDA-based QSAR models obtained with non-stochastic quadratic indices 4M H 6M H q3L ðxE Þ þ 3:457  104M q3 ðxÞ þ 6:426  107M q7 ðxÞ  2:303  104M qH q6L ðxE Þ Class ¼ 2:971  1:672  104 M qH 3 ðxÞ  9:428  10 0L ðxE Þ þ 2:092  10 2M H 2M H q2 ðxEH Þ þ 7:700  104M qH q1 ðxEH Þ þ 2:041  102M qH 3 ðxEH Þ þ 5:892  10 3L ðxEH Þ þ 4:192  10

(1)

5V H q5 ðxÞ  2:659  103V q0L ðxE Þ þ 7:901  104V q2L ðxE Þ  3:844  104V q2 ðxÞ þ 4:810 Class ¼ 2:936 þ 5:126  105V q3 ðxÞ  2:218  103V qH 1L ðxE Þ  1:202  10 9V H q13 ðxEH Þ  1:879  108V qH  109V qH 10 ðxÞ  1:106  10 9L ðxEH Þ

(2) 3P H Class ¼ 3:013 þ 1:412  102P q3 ðxÞ  1:185  101P qH q5 ðxE Þ  1:190  102P q4L ðxE Þ þ 1:070  101P q3L ðxE Þ þ 5:830  102P q1 ðxÞ  9:971 3L ðxE Þ  1:062  10 10P q15 ðxÞ þ 1:365  102P qH  102P qH 0L ðxE Þ  3:502  10 4L ðxE Þ 2K 2K Class ¼ 3:335 þ 1:511  103K q3 ðxÞ  2:527  103K qH q2L ðxE Þ  2:490  102K qH q0L ðxE Þ  2:231  102K q1L ðxE Þ 3L ðxE Þ þ 2:889  10 2L ðxE Þ þ 3:806  10

(3) (4)

2G 5G H q2L ðxÞ  5:485  102G qH q6 ðxE Þ þ 5:923  101G qH Class ¼ 3:310 þ 4:042  103G q3 ðxÞ  3:564G qH 3L ðxE Þ þ 1:555  10 1L ðxE Þ  3:795  10 1 ðxEH Þ  8:420

 101G qH 0 ðxEH Þ (5) 6V H 6M Class ¼ 2:761 þ 1:417  104V q3 ðxÞ  2:084  103V qH q5 ðxE Þ þ 1:204  102K q2 ðxÞ  2:446  102G qH q7 ðxÞ þ 1:518 1L ðxE Þ  6:430  10 1L ðxE Þ  1:296  10  102P q3L ðxE Þ þ 1:798  108M q10 ðxÞ (6) LDA-based QSAR models obtained with stochastic quadratic indices 2Ms H 2Ms Class ¼ 2:997  6:129  103Ms q9 ðxÞ  5:226  103Ms qH q5 ðxÞ  6:080  103Ms qH q11L ðxE Þ  1:136  102Ms q15 ðxÞ 3L ðxE Þ þ 2:162  10 1L ðxE Þ þ 1:820  10 Ms H 1Ms H q0 ðxEH Þ þ 2:433  101Ms qH q1 ðxEH Þ þ 6:890  103Ms qH  2:007  102Ms qH 7L ðxE Þ  3:404 4 ðxEH Þ þ 1:919  10 1L ðxE Þ 3Vs H 3Vs H Class ¼ 2:932 þ 4:565  103Vs q1 ðxÞ  4:690  103Vs qH q3L ðxE Þ  9:793  103Vs qH q6 ðxÞ  4:035  103Vs qH 0 ðxÞ  1:390  10 1 ðxEH Þ þ 8:805  10 15 ðxÞ

(7)

 3:456  103Vs qH 1 ðxÞ 1Ps H 1Ps H Class ¼ 3:043 þ 7:323  101Ps q1 ðxÞ  2:996  101Ps qH q1 ðxÞ  1:184Ps qH q3 ðxÞ  2:959  101Ps q13 ðxÞ  1:408 1L ðxE Þ  9:994  10 1 ðxEH Þ þ 7:558  10

(8)

1Ps q8L ðxE Þ  101Ps qH 0L ðxB Þ þ 2:063  10 1Ks 1Ks Class ¼ 3:195  3:438  101Ks q3 ðxÞ  6:985  102Ks q1L ðxE Þ  4:275  101Ks qH q2 ðxÞ  1:594  101Ks qH q11L ðxE Þ 1 ðxÞ þ 7:924  10 2L ðxE Þ þ 1:601  10

(9)

1Ks H q4 ðxEH Þ  3:309  101Ks q14 ðxÞ þ 4:705  101Ks qH 2 ðxÞ  1:595  10

(10) 1Gs 2Gs H Class ¼ 2:813  1:964  101Gs q3 ðxÞ  1:069  101Gs q3L ðxE Þ  1:807  101Gs qH q0 ðxÞ  1:157  101Gs qH q1L ðxE Þ 5 ðxEH Þ þ 7:493  10 12 ðxÞ þ 3:245  10 1Gs q6L ðxE Þ þ 1:349  101Gs qH  2:430  101Gs qH 7 ðxÞ 2 ðxE Þ þ 1:646  10 Ps H 2Gs H Class ¼ 2:900 þ 9:039  101Ps q1 ðxÞ  1:510  101Ps q1L ðxE Þ  1:288Ps qH q7 ðxÞ þ 7:220  101Ps qH 1 ðxÞ  1:606 q1 ðxEH Þ þ 4:349  10 3 ðxÞ  7:923 1Ps q13 ðxÞ  2:838  101Gs qH  104Ms qH 9 ðxÞ  2:127  10 1L ðxÞE

(11) (12)

LDA-based QSAR models obtained with whole set of MDs (mixing non-stochastic and stochastic quadratic indices) 6 V H q5 ðxÞ þ 7:768  103 K q2L ðxE Þ þ 2:800  101 Ps q1 ðxÞ  1:781  103 Vs qH Class ¼ 2:706 þ 1:256  104 V q3 ðxÞ  1:650  103 V qH 1L ðxE Þ  6:44210 10 ðxÞ  1:021 1 Ps H q7 ðxÞ  2:102  101 Ps q13 ðxÞ  103 Ms qH 9 ðxÞ þ 2:357  10

(13)

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

5741

Table 2 Prediction performance and statistical parameters for LDA-based QSAR models in the training set. Modelsa

Cb

Accuracy ‘‘QTotal’’ (%)

Specificity (%)

Sensitivity ‘‘hit rate’’ (%)

False positive rate (%)

Wilk’l

F

D2

87.81 86.23 85.55 86.00 85.33 86.23

13.5 11.6 13.0 11.8 13.0 11.1

0.47 0.46 0.47 0.47 0.47 0.45

100.1 119.6 114.5 169.4 146.1 139.4

4.44 4.73 4.53 4.45 4.48 4.90

85.33 86.23 85.33 86.00 85.33 86.23

14.5 13.9 12.6 11.8 12.6 11.6

0.47 0.45 0.46 0.45 0.48 0.45

91.5 157.2 134.5 123.6 111.1 122.7

4.43 4.82 4.72 4.89 4.39 4.86

0.45

122.7

4.86

LDA-based QSAR models obtained with non-stochastic quadratic indices [Eq. [Eq. [Eq. [Eq. [Eq. [Eq.

(1) (2) (3) (4) (5) (6)

(10)] (9)] (9)] (6)] (7)] (8)]

0.74 0.75 0.73 0.74 0.72 0.75

87.16 87.38 86.29 87.16 86.18 87.60

85.87 87.41 85.94 87.19 85.91 87.82

LDA-based QSAR models obtained with stochastic quadratic indices [Eq. (7) (11)] 0.71 85.42 84.56 [Eq. (8) (7)] 0.72 86.18 85.27 [Eq. (9) (8)] 0.73 86.40 86.30 [Eq. (10) (9)] 0.74 87.16 87.19 [Eq. (11) (9)] 0.73 86.40 86.30 [Eq. (12) (9)] 0.75 87.38 87.41

LDA-based QSAR models obtained with whole set of MDs (mixing non-stochastic and stochastic quadratic indices) [Eq. (13) (9)] 0.75 87.70 87.84 86.46 11.1 Bold values represents the best models using non-stochastic indices, stochastic indices and mixing both kinds of descriptors. a Values in parentheses indicate the quantity of variables of the models. b Matthew’s correlation coefficient.

Other crucial problem, in chemometric and QSAR studies, is the definition of the applicability domain (AD) of a classification or regression model. “Not even a robust, significant, and validated QSAR model can be expected to reliably predict the modelled property for the entire universe of chemicals. In fact, only the predictions for chemicals falling within this domain can be considered reliable and not model extrapolations” [71]. Therefore, the next step of this report was to develop a study to access to chemical’s scope of our models (principle 3: Defined Domain of Applicability). The AD is a theoretical region in chemical space, defined by the model descriptors and modelled response, and thus by the nature of the chemicals in the training set, as represented in every model by specific MDs. Therefore, AD of the QSAR model is “the range within which it tolerates a new molecule” [72]. For MLR and ADL, a multiple predictor problems with normally distributed data, the distance-based measures, like leverage (h) is one of most used (see Experimental Section) [73,74]. The warning leverage, h*, is a critical value or cut-off used to consider the prediction made for the model for a specific compound in the dataset. In order to visualize the AD of a QSAR model, a double ordinate cartesian plot of cross-validated residuals (first ordinate), standard residuals (second ordinate), and leverages (hat diagonal: abscissa) values (h) defined the AD of the model as a squared area

within 3SD band for residuals and a leverage threshold of h* ¼ 0.0261 for anti-inflammatory activity (i.e., Eq. (13)). This plot, so-called Williams scheme, can be used for an immediate and simple graphical detection of both the response outliers (i.e., compounds with standardized residuals greater than three standard deviation units, >3s) and structurally influential chemicals in a model (h > h*). For instance, Fig. 3 shows the Williams plot of Eq. (13) as a simple example and as can be seen, almost all chemicals used lie within this area. Only some few chemicals (29 for the training set) have leverages rather higher than the threshold, but show residuals within the limits. These active and inactive compounds are “outside” of AD for this model, and these chemicals can influence model parameters. Considering this fact, we decided to evaluate the effect generated by the removal of these compounds on the model performance. Interestingly, no significant variation in the model performance was detected when the new parameters e after removal of the chemicals e were studied. Therefore, the influence of these compounds is not critical neither for model parameters and performance. Consequently, their removal is not justified. In conclusion, the model can be used with high accuracy 4

Training Test

3

Std. Residuals

2 1 0 -1 -2 -3 -4 0.00

0.02

0.04

0.06

0.08

0.10

0.12

Leverage

Fig. 2. Receiver operating characteristic curve (ROC) for training set (white points) and random classifier (black points). TPF ¼ sensitivity and FPF ¼ 1 e specificity for different thresholds of discrimination function values [AUC ¼ 0.9395 (Eq. (13))].

Fig. 3. William plot of Eq. (13): outlier will be chemicals corresponding to points with standardized residuals greater than three standard deviation units; influential chemicals are points with high leverage values higher than the threshold or cut-off value h* ¼ 0.0261. The training and test sets are represented by blues circles and green squares, respectively. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

5742

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

in this AD [71,74]. In the next section, we re-take this analysis in order to determine the reliability of prediction for molecules selected as good candidates in virtual screening protocols. The model validation (Principle 4: Statistical Validation) is another key feature in good QSAR practice regarding the diagnostic of developed models. In this sense, a QSAR model should be associated with the appropriate measures of goodness-of-fit, robustness and predictivity [74e79]. The first two are considered as internal validation, while the last is considered as external validation. The evaluation of the performance of models by using external validation (one or more external test sets) can be considered as a superior alternative because the good behaviour of models in internal experiments are a necessary but not sufficient condition the high predictive power of the model. In this regard, predictivity can be claimed only if the model is successfully applied to the prediction of the external test series, which were not used in the model development. Furthermore, in this report we describe the external performance evaluation by using only a prediction set of active and inactive compounds. The chemicals included in this prediction set were never used to fit the QSAR model and then, are cases unknown for external validation only. The robustness of all models was demonstrated with the adequate values of the rather-good classifications above 86%. Eqs. (6) and (12) showed 86.39% and 86.03% of accuracy in the prediction series. These results validate the models for their use in the ligand-based virtual screening, taking into consideration that 75.0% is considered as an acceptable threshold limit for this kind of analysis. Table 3 also depicts the principal parameters used in medicinal statistics for the prediction set. In addition, the best LDAbased QSAR is Eq. (13), with an accurancy of 88.44% vs 87.07% depicted by model (12). Receiver operating characteristic curve, ROC, for the test set is shown in the Fig. 4 (Eq. (13)). The area under the curve (AUC) is 0.9572. The structure of all compounds used in the test set can be seen in Table SI2 of the Supporting Information. Table SI5 and Table SI6 of the Supporting Information summarize the results of the classifications of every compound, for all models represented by Eqs. (1)e(13). The corresponding pharmacological distribution (PDD) diagram [80] is illustrated in Fig. 5. As shown in the figure there is a high

Table 3 Prediction performance and statistical parameters for LDA-based QSAR models in the test set. Modelsa

Cb

Accuracy ‘‘QTotal’’ (%)

Specificity (%)

Sensitivity ‘‘hit rate’’ (%)

Fig. 4. Receiver operating characteristic curve (ROC) for test set (white points) and random classifier (black points). TPF ¼ sensitivity and FPF ¼ 1 - specificity for different thresholds of discrimination function values [AUC ¼ 0.9395 (Eq. (13))].

structural heterogeneity within the compounds placed on the same range of discriminate function values (DP%). Moreover, a new molecule will be selected as active if its calculated values (from Eq. (13), for example) are within the assigned intervals. These thresholds were selected leaving out the zones of limiting property in which the assignment of activity could be doubtful (DP% > 50), to minimize the number of false actives selected. That is, once the PDD is completed the way to follow is easy: the screening for new compounds showing DP% positive values (DP% > 50), for which it is to be expected to find anti-inflammatory activity. Hence OECD Validation Principle 4 is fully met. The last principle 5 (Mechanistic Relevance, if is possible) is rather difficult to address in this report, due to the nature of database used to develop QSARs. 2.2. Dry selection of new leads chemotype Virtual screening of large databases using such models has emerged as an interesting alternative to HTS and an important drug design tool [4,16,81,82]. Aiming to test the ability of our models to discover new leads, we carried out a virtual screening of different compounds which were not screened before as anti-Inflammatorys and whose structure is not similar to the structure of known antiinflammatory compounds.

False positive rate (%)

LDA-based QSAR models obtained with non-stochastic quadratic indices [Eq. (1) (10)] 0.73 86.73 86.11 86.71 12.7 [Eq. (2) (9)] 0.70 85.03 83.33 85.71 13.3 [Eq. (3) (9)] 0.76 87.76 86.11 88.57 10.7 [Eq. (4) (6)] 0.74 87.07 86.81 86.81 12.7 [Eq. (5) (7)] 0.76 88.10 86.81 88.65 10.7 [Eq. (6) (8)] 0.73 86.39 86.11 86.11 13.3 LDA-based QSAR models obtained with stochastic quadratic indices [Eq. (7) (11)] 0.75 87.41 90.28 84.97 [Eq. (8) (7)] 0.77 88.44 89.58 87.16 [Eq. (9) (8)] 0.75 87.41 86.11 87.94 [Eq. (10) (9)] 0.76 88.10 85.42 89.78 [Eq. (11) (9)] 0.75 87.41 86.11 87.94 [Eq. (12) (9)] 0.74 87.07 86.81 86.81

15.3 12.7 11.3 9.3 11.3 12.7

LDA-based QSAR models obtained with whole set of MDs (mixing nonstochastic and stochastic quadratic indices) [Eq. (13)] 0.77 88.44 87.67 88.89 12.0 a b

Values in parentheses indicate the quantity of variables of the models. Matthew’s correlation coefficient.

Fig. 5. Pharmacological distribution diagram (PDD) obtained for the discriminant function (Eq. (13)) from a set of 1213 compounds (587 active and 626 inactive ones). Black and white belts are the active and inactive training sets and dark grey and grey ribbons are the active and inactive test sets, respectively. Expectancy, E, is a measure of the probability of activity or inactivity.

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

The algorithm described above, and the good results obtained prompted us to make in silico evaluations of all the chemicals contained in our ‘in-house’ collections of indazole [83e88], indole [89], cinnoline [90] and quinoxaline [91,92] derivatives (as well as other new related chemicals and their derivatives), several of which have shown interesting properties as trichomonacidal [25,30,88,93], antichagasic [37,88,94], antimalarial [38] and antineoplastic [86e88] drugs. On the basis of computer-aided predictions, we selected 34 potential anti-inflammatory leads (virtual hits). In this experiment, 145 out of such chemicals from our “in house” library were evaluated in silico to determine if they will exhibit anti-Inflammatory activity, by using every individual discriminant functions obtained as an ensemble classifier, CE. That is, here every individual classifier (CI) is fused into the CE through a voting system, where the individual output of CI are used like input of CE, which will have a voting score for the query molecules. Initially, we decided to select the compounds that show CE  10 (w75% of CIs) by “wet” evaluation. In addition, the following criteria were used for the hits’ selection: 1) compounds were selected as hits if the value of posterior probability of possessing anti-inflammatory activity exceeded 50% (ΔP > 50%) by all LDA-based QSAR models (fusion

5743

approach or multi-classification system), and 2) if, among the compounds designed (or that will be obtained) by our chemical team, many similar compounds satisfied criterion 1, then only several representative structures were selected. To provide an intuitive picture, a flowchart to show how these CI are fused into the CE is given in Fig. 6. One series of 34 structural-diverse compounds (see Fig. 7) were selected as anti-inflammatory lead-like compounds, showing a good agreement between the in silico predictions and in vivo assays in several tests (see below). The values of DP% for this subset are depicted in Table 4. However, it is generally acknowledged that QSARs are valid only within the same domain for which they were developed. In fact, even if the models are developed on the same chemicals, the AD for new chemicals can differ from model to model, depending on the specific MDs. One of the main aims of the present work was to develop a model for predicting anti-inflammatory activity at early stages of drug discovery and development. Consequently, one may not pretend to extrapolate the use of these models to other kind of anti-inflammatory compounds, making uncertain predictions in conditions different to those fixed to derive the model [73,74]. Therefore, the chemicals designed in this study were only

Molecule list Imput (Training and test sets)

Molecular Structure (Total & Local MDs)

Anti-inflamatory Activity Class: active (1) or inactive (-1)

Chemometric tool (Linear Discriminant Analysis)

Ensemble classifier: multi-agent predictor/fusion approach

Fuse outputs by voting Final Output

New virtual Leads (Next Step: in vivo anti-inflamatory assays)

Fig. 6. Flowchart illustrating how the individual classifiers (CI) are fused into the ensemble classifier (CE) through a simple voting system. Here we show the fuse discriminant functions by using TOMOCOMD-CARDD MDs into a prediction engine.

5744

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

OH

O

OH

N

N

N

1 (VA5-5b)

N

N

N

N

OH O2N

O2N

O2N

O2N

N

3 (VA5-8pre) 4 (VA5-5c)

2 (VA5-14g) Cl

OH F O2N

O O CH3

O

N

O2N

O2N

N

HO OH

O N

N CH3

O2N

O2N N

N

N

H

16 (VA2-27)

15 (VA5-5e) H F

14 (VA5-10)

13 (VA5-6) O

N

N

N

N

12 (VA5-13l) CH3 O O2N

N

N N

N

11 (VA5-13h) OH

O OH

N

N

H 10 (VA5-9a)

9 (VA5-5a)

O O2N

N

N

O

N

O2N

OH O

O2N

N

N

N

N

18 (VA3-8a) H

17 (VA4-18) CH3 O N 2 N

O O2N

O

O

O

O2N

O2N

CH3 N

20 (VA2-17)

19 (VAX-22)

N

N

N

N

N

N

N

F 22 (VA7-38)

21 (VAX-23)

23 (VATR-1) O

O O2N N

O2N

24 (VATR-4)

O O2N

CH3 H

N

N

NH N O

Cl

27 (VA1-15)

N

NO2 NO2 30 (VA6-17b)

NH

O

O2N

26 (VA8-38) N N 28 (VA6-5d) H OMe

HN

N

29 (VA6-6b)

F

N

O

O2N

25 (VATR-5)

N

Cl

N

N N

O

O F

O2N

N

CH3

8 (VA5-12b)

7 (VA5-14j)

O2N

O2N

N N

N

N 6 (VA5-15c) H

5 (VA5-8)

O2N

N

N

Cl

OH

OMe

O2N

N

OH

NO2 N

NO2 N N

O

OH

O2N N

N

N N

N 32 (VA-M6)

31 (VA-M1) OMe

NO2 N N

33 (VA-M10)

OH

OMe

O2N

NO2 N

N N

OH

O2N N

N N 34 (VA-M12)

Fig. 7. Structure of 34 compounds (from 145) classified as anti-inflammatory agents by LDA-based QSAR models and which were selected for wet-in vivo-assays.

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

5745

Table 4 Results of ligand-based in silico screening by using CI. Chemicaln

DP%a

DP%b

DP%c

DP%d

DP%e

DP%f

DP%g

DP%h

DP%i

DP%j

DP%k

DP%l

DP%m

1 Va5-5b 2 Va5-14g 3 VA5-8pre 4 VA5-5c 5 VA5-8 6 Va5-15c 7 Va5-14j 8 VA5-12b 9 VA5-5a 10 Va5-9a 11 Va5-13h 12 Va5-13l 13 Va5-6 14 Va5-10 15 VA5-5e 16 Va2-27 17 Va4-18 18 Va3-8a 19 Vax-22 20 Va2-17 21 Vax-23 22 VA7-38 23 VATR-1 24 VATR-4 25 VATR-5 26 VA8-38 27 Va1-15 28 Va6-5d 29 Va6-6b 30 Va6-17b 31 VA-M1 32 VA-M6 33 VA-M10 34 VA-M12

85.49 97.61 97.73 86.36 81.78 83.49 80.18 74.21 79.51 77.84 77.77 78.27 69.84 67.48 79.89 87.33 74.52 91.60 81.27 64.40 40.99 42.62 29.04 28.56 85.91 81.20 78.13 92.70 50.59 99.68 98.15 99.13 99.79 99.89

84.58 96.47 96.98 76.18 74.64 68.48 83.28 72.20 69.83 57.06 64.69 64.45 67.64 54.62 80.65 81.95 71.81 93.70 76.11 36.72 47.45 38.78 54.52 52.47 94.00 73.41 50.81 96.28 89.31 96.37 97.70 98.45 99.81 99.84

87.82 97.42 97.75 85.68 82.89 86.85 87.72 78.64 77.00 74.28 70.87 71.06 65.43 61.56 79.45 85.62 68.19 91.99 77.53 74.82 66.49 51.44 79.16 81.53 97.05 87.46 72.88 97.01 82.62 98.66 99.01 99.19 99.92 99.92

79.19 94.51 95.18 76.21 67.66 69.57 68.92 51.96 63.76 57.72 52.12 52.12 33.15 24.37 73.02 72.54 36.67 91.70 76.34 69.66 14.86 37.88 73.91 77.78 96.44 96.38 62.49 85.97 10.31 99.10 94.80 97.81 99.62 99.85

56.59 85.54 89.41 53.42 46.49 28.69 48.70 35.00 31.85 7.96 28.46 28.60 3.00 21.63 69.40 54.65 51.77 90.15 46.96 43.67 7.58 4.23 29.04 28.56 85.91 56.16 20.40 60.60 5.50 94.71 87.08 91.91 98.78 99.21

73.48 94.74 95.10 72.29 62.45 70.47 69.29 56.39 58.39 53.77 51.45 52.21 32.99 26.66 74.57 73.40 56.87 92.51 74.21 52.90 0.45 34.46 42.85 40.53 92.86 63.21 54.92 89.11 45.45 97.36 93.21 95.23 99.26 99.48

84.23 97.00 97.75 86.43 83.40 80.77 68.42 70.36 77.45 65.41 62.73 67.01 70.33 55.67 87.23 84.91 81.28 95.22 79.84 89.85 53.06 39.30 74.37 84.63 97.73 89.88 66.11 84.78 45.81 99.74 96.82 99.07 99.61 99.92

99.03 99.88 99.92 99.16 98.89 98.37 97.55 97.31 98.47 97.21 97.35 97.56 97.92 96.20 97.40 99.01 96.49 98.96 98.78 99.35 91.21 92.88 86.51 84.37 99.65 88.78 98.59 99.27 97.34 100.00 99.96 99.99 100.00 100.00

87.71 98.67 98.70 87.00 87.53 81.02 73.71 69.63 85.28 78.36 70.66 72.11 74.79 64.35 86.04 89.79 80.24 95.74 87.24 91.10 37.31 41.21 80.96 82.01 98.77 97.11 77.83 91.00 85.66 99.74 97.40 99.52 99.80 99.96

86.20 96.24 96.45 85.71 78.75 70.33 69.72 54.86 77.16 58.47 61.46 62.73 70.53 48.25 75.34 80.81 55.61 89.95 81.22 84.58 40.13 65.64 86.13 90.05 97.65 71.22 59.58 94.32 69.50 98.85 97.56 99.16 99.54 99.92

85.92 94.78 94.69 72.12 60.17 70.89 79.05 45.48 56.16 48.63 35.43 43.80 37.08 28.20 85.16 69.11 50.84 89.06 66.69 68.96 36.30 38.22 61.18 70.78 96.06 87.69 60.39 95.56 42.93 97.50 90.98 94.56 98.57 99.39

86.41 98.10 98.21 84.71 84.65 75.14 69.47 63.18 82.15 72.25 67.46 68.78 70.21 55.58 85.28 88.08 79.09 94.99 84.45 91.29 31.19 38.48 77.42 76.52 98.41 89.45 72.72 90.85 86.52 99.53 95.26 99.22 99.58 99.93

84.38 97.06 97.72 82.87 80.64 81.32 72.53 65.75 74.71 73.68 62.88 64.06 50.02 48.50 83.91 84.32 66.26 96.30 83.59 77.23 20.71 35.83 56.49 57.62 93.32 87.79 74.34 92.46 38.00 98.96 94.78 98.25 99.51 99.86

aem means that DP% values are obtained by Models 1e13, respectively. Here, in order to consider every query molecule as active chemical we used DP%>50% (by 75% of models), because with this cut-off we avoid the unclassified example as well as the risk of false active can be less. n The molecular structures of the compounds represented with codes (numbers) are shown in Fig. 7.

synthetized and posterior in vivo evaluated after they were plotted into the AD of obtained models. For instance, another William plot (Fig. 8) of Eq. (13) (with the training set and chemical series discovered as novel anti-inflammatory leads) was carried out as

4 Training virtual screening

3

Std. Residuals

2

a simple example. As can be noted in Fig. 5, all used compounds lie within this area, which ensures great reliability for the prediction of this kind of leads used in the virtual screening. That is to say, all new leads fall within the AD of the model and so the predictions are reliable. That said above proves the good assessment for the classification of these compounds as novel anti-inflammatory leads. Therefore, this model can be used with high accuracy for new compound predictions in this AD [73,74]. 2.3. Wet evaluation: in vivo screening and confirmation

1

0

-1

-2

-3

-4 0.00

0.02

0.04

0.06

0.08

0.10

0.12

Leverage

Fig. 8. LDA models applicability domain (Eq. (13)) for learning and new lead series (all chemicals included in virtual screening, 145 in total). The training is represented by blue circles and the new compounds are represented by green squares. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

In this section, we describe the main results obtained in the experimental assays (wet evaluation), in three different in vivo tests applied on 34 novel chemicals selected as lead series in our in silico experiment. Here, we developed a wet screening taking into account a hierarchical battery of in vivo tests by using: 1) a novel test with zebrafish larvae as experimental model [63] and 2) two different assays in mouse. The assay in zebrafish was used as a primary rapid screening and the remaining test as confirmatory assays for only best hits. The first stage was to determine the maximum tolerated concentration (MTC) for each compound on zebrafish larvae, with 100 mM as the highest concentration. These results are showed in Table 5 (last column). Briefly, toxicity was not observed when zebrafish were treated with a 10 mM dose of each substance (with the only exception of compound 18, VA3-8a). Moreover, compounds 5, 10, 13 and 14 did not show any toxicity at 100 mM and

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

100

(mM)

0.30 0.73 0.84 0.22 0.86 0.41 0.92 0.80 0.46 0.81 0.46 0.63 0.83 0.94 0.25 0.29 0.80 T 0.77 0.33 0.89 0.48 0.86 0.27 0.35

e e 0.26 0.61 e 0.87 e e 0.65 0.26 0.27 0.74 0.62 0.17 T 0.52 T e e 0.97 e 0.16 e e

e e e e 0.27 e e e e 0.39 T T 0.59 0.66 T T T T e e T e T e e

10b 10b 10b 30b >100 10b 30b 10b 10b >100 30 30 >100 >100 30 10 30 3 10b 10b 30 10b 30 10b 10b

ND ND ND 0.82

0.93 0.89 0.86 0.81

0.93 0.87 0.89 e

0.70 0.50 0.81 e

>100 >100 >100 10b

0.80 0.87 0.92 0.50

0.64 0.69 0.85 0.48

e e e 0.38

e e e 0.28

10b 10b 10b >100

1

3

10

ND ND ND ND ND ND ND ND ND ND ND ND ND ND ND ND ND 0.64 ND ND ND ND ND ND ND N.T. ND ND ND ND N.T. ND ND ND 0.57

0.91 0.94 0.94 ND ND 0.62 ND 0.87 0.88 ND ND ND ND ND ND 0.56 ND 0.40 0.75 0.83 ND 0.59 ND 0.68 0.58

0.84 0.75 0.76 0.78 ND 0.65 0.86 0.87 0.67 ND 0.60 0.67 ND ND 0.46 0.54 0.89 0.24 0.71 0.78 0.74 0.62 0.89 0.67 0.65

ND ND ND 0.85 0.87 0.84 0.72 0.54

ND ¼ Not determinated for low potency. N.T. ¼ Not tested for solubility problems. T ¼ Toxic concentration for zebrafish. Bold values represents indomethacin result. a The molecular structures of the compounds represented with codes (numbers) are shown in Fig. 7. b Maximum concentration tested due to problems of solubility in Danieau’s.

were further used at this concentration. The other chemicals were evaluated at their respective MTC: 30 mM for compounds 1, 4, 7, 11, 12, 15, 17, 21 and 23 as well as 10 mM for chemicals 2, 3, 6, 8, 9, 19, 20, 22, 24, 25, 30 and 32e34. In all experiments, indomethacin was used as positive control (reference drug). Secondly, the anti-inflammatory activity of the compounds was evaluated following the guidelines of the LPS-induced leukocyte migration assay [63]. Briefly, leukocyte migration to the injury zone of tail-transected larvae (tail-tip) was observed, semi-quantified and evaluated as the relative leukocyte migration (RLM) to determine the anti-inflammatory activity of the tested compound. Table 5 and Fig. 9 report the RLM observed in zebrafish for all 34 compounds and for only the most potent chemical, respectively. Most of the tested chemicals exhibited an RLM inferior to 50% at 100 mM or 30 mM. The compounds 18 (3 mM), 24(10 mM), 25 (10 mM), 6 (10 mM), 15 (30 mM), 11 (30 mM) and 12 (30 mM) gave the best results displaying RLM values of 0.24, 0.27, 0.35, 0.41, 0.17, 0.26 and 0.27 respectively, data that suggest a probable anti-inflammatory activity of 76, 73, 65, 59, 83, 84 and 73%, respectively. The RLM of these compounds was comparable with the RLM displayed by the positive control e indomethacin e tested at concentrations of 10, 30 and 100 mM (RLM 0.49, 0.38 and 0.28 respectively). In addition, unspecific cytotoxicity to murine macrophages was tested for the best hits. Results for unspecific cytotoxicity are shown

0.5

0.0 tr ol om .3 0µ M 6 (1 0µ M ) 11 (3 0µ M ) 12 (3 0µ M ) 15 (3 0µ M ) 18 (3 µM ) 24 (1 0µ M ) 25 (1 0µ M )

MTC 30

0.3

on

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 Indomethacin

RLM/Concentration (mM)

C

Chemicala

1.0

In d

Table 5 Activity and toxicity of compounds tested in the LPS-induced leukocyte migration assay in zebrafish.

Relative leukocyte migration

5746

Fig. 9. The x-axis represents the substances tested (here only the most potent compounds are depicted; for more detail and extension see Table 5). The y-axis represents the percentage of relative leukocyte migration in zebrafish (concentration used for this activity profile is also indicate).

in Table 6. In general, all chemicals showed low unspecific cytotoxicity. Compounds 6 and 18 exhibited unspecific cytotoxicity near to 96% and 73%, respectively, at the higher concentration assayed (100 mg/mL). Only compound 11, 12 and 25 were inactive at all levels. In a second step, we evaluated the most potent compounds in two in vivo tests. Here, we initially tested the efficacy of these chemicals using TPA-induced mouse ear oedema. The overall result archived in this experiment is depicted in Table and Fig. 7. Concerning the TPA-induced ear oedema, all compounds assayed, with the exception of 15 showed anti-inflammatory activity. Compound 12 was the most active and completely abolished the oedema. Compounds 6, 11 and 24 showed inhibition percentages in the range of the reference drug (indomethacin), Table 6 Unspecific cytotoxicity for best hits assayed in macrophages. Chemicalsb

Concentration (mg/mL)a

% Cytotoxicitya  % SD

6

100 10 1 100 10 1 100 10 1 100 10 1 100 10 1 256 mM 128 mM 64 mM 256 mM 128 mM 64 mM

96.4  0.3 9.4  1.4 4.4  0.7 58.4  0.3 43.9  1.8 32.8  4.1 57.5  0.7 21  2.7 11.5  2.4

11

12

15

18

24

25

NT 73.1  2.5 9.4  0.9 3.4  3.53 80.99 71.75 54.14 57.87 64.00 9.77

NT: Non-tested. a Unspecific cytotoxicity at three different concentrations (100, 10 and 1 mg/mL, with the exception of chemicals 24 and 25 which were evaluated at 256, 128 and 64 mM) in murine macrophages cells and standard deviation (SD). b The molecular structures of the compounds represented with numbers are shown in Fig. 7.

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753 Table 7 Anti-inflammatory effect of products on acute TPA-induced ear mouse oedema. Indomethacin and compounds were topically applied, at dosage of 0.5 mg/ear in 20 ml acetone. The weight of the ear punch (6 mm) was measured after 4 h of treatment with the irritant. Group/compound

ΔW  SEM (mg)a

6 11 12 15 18 24 25

2 1 1 11 4 1 3

Vehicle Control Indomethacin

e 17  1 1  0**

      

0** 1** 1** 2 1** 0** 1**

% Ib

MPOc

91 97 104 32 73 92 80

192 228 170 528 225 275 328

20** 27** 19** 75 22** 18** 10*

63 56 67 3 56 46 36

e e 97

209  28** 514  94 203  29**

e e 61

      

% Ib

Data expressed as mean  S.E.M., n ¼ 3. **p < 0.01 and *p < 0.05 with respect to the control group (Dunnett’s t-test). a Increase in ear weight in mg (mean  S.E.M.), n ¼ 6. b Inhibition percentage with respect to the control group treated with only TPA. c Myeloperoxidase assay: A (mOD) ¼ optical density units  103.

whereas compounds 18 and 25 reduced the oedema in a lesser extent (inhibition percentages of 73 and 80, respectively). In addition, all compounds except compound 15, significantly reduced neutrophil infiltration, measured as myeloperoxidase activity on TPA application test. As shown in Table 7 (Fig. 10), compounds 6, 11, 12 and 18 showed similar values to indomethacin (inhibition percentage of 61%), but compounds 6 and 18 were toxic in zebrafish and show unspecific cytotoxicity at 100 mg/mL. Evidently, this study suggests a new support structure (12, 11 and 24; the nitroindazolinone chemotype), which opens the door to posterior chemical modification of these lead compounds to increase their effectiveness. These compounds constitute new promising lead structures. It importantly, is essential to further investigate these compounds to increase the knowledge in the field of structure activity relationships of anti-inflammatory compounds. 3. Concluding remarks The integration (aligning) of dry and wet screening for diverse compound libraries is an essential part of the anti-inflammatory lead discovery effort. The results of our in silico prediction and posterior in vivo screening, by using a battery of assays, are encouraging and show that progress may be made through this kind of approach. Furthermore, here we have shown how the combination of validated QSAR-modelling and LBVS could be successfully used, as innovative technologies to ensure high expected hit rates in the discovery of new bioactive compounds.

5747

Within this set of in house library, we have identified 34 novel chemicals not yet reported (virtual hits) as anti-inflammatory chemotype. The biological evaluation shows that most tested compounds exhibited adequate anti-inflammatory activity. In general, all the best anti-inflammatory chemicals showed low unspecific cytotoxicity, except for compounds 6 and 18. However, the most active compounds, 12, 11 and 24, do not display significant cytotoxicity in macrophage cells. These chemicals show preliminary evidence of good and selective in vivo anti-inflammatory activity, with potential for scaffold optimization. The mode of action of the novel chemicals described in this study is a question that has not yet been addressed. While this is beyond the scope of this report, it is extremely relevant and we are currently following up on the top leads. However, it is well-known that the two arachidonic acid metabolizing enzymes 5-lipoxigenase (5-LPO) and COX are susceptible to inhibition by compounds with low redox potentials such as phenidone and BW755C [95,96]. Redox inhibitors of 5-LPO, as this type of compounds has been loosely termed, generally show poor selectivity for 5-LPO relative to COX. More recently, Bruneau and Delvare [97] demonstrated that structural modification of redox-based compounds can modulate 5-LPO inhibitory potency and selectivity, independently of changes in redox potential. These authors found a key compound (IC1 207968), which combined high selectivity with oral activity as a 5LPO inhibitor. Nevertheless, BW 755C and phenidone were orally active, but were more potent against COX than 5-LPO. Our compounds, 12, 11 and 24 (nitroindazolinone chemotype) show similar-to-superior activity than indomethacin and they do not present significant cytotoxicity in macrophage cells or toxicity in zebrafish larvae, as well as they will be dual inhibitors of COX and LPO. In future outlooks, these models, which relate the chemical structure with a specific endpoint, might be programmed into expert systems helping in the exhaustive search of bioactive molecules within huge chemical libraries. Moreover, the preliminary identification of novel anti-Inflammatory leads in this work is promising and strongly supports the LBVS of additional compound libraries; databases made manly of chemicals with diverse scaffolds are an important strategy to continue exploring. In fact, the assemble classifier presented here will be use to identify new antiinflammatories from a database of well-known drugs already approved for human use for potential ‘off-label’ anti-inflammatory efficacy. The logic of this approach is that hits from such screens are low-hanging fruits that will require less development before they are able to enter clinical trials as anti-inflammatory. Some work in this direction is now in progress and will be published in a forthcoming paper.

Fig. 10. Antinflammatory effect of compounds and indomethacin (0.5 mg/ear) on the TPA-induced mouse ear oedema. Data expressed as mean  S.E.M., n ¼ 6. Dunnett’s t-test: **p < 0.01 respect to the control group treated with only TPA (0.5 mg/ear).

5748

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

4. Experimental section 4.1. Computational methods 4.1.1. Database building Usually, a benchmark dataset consists of a learning (or training) and an independent testing records set. The learning dataset is one of the most important components of a statistical predictor, because it is used for training the predictor’s ‘engine’, whereas the testing information is used for examining the predictor’s accuracy via an external test [98]. A total of 1213 biologically active organic compounds, which showed a great structural variation, were used to integrate the general dataset, from which 587 compounds have reported antiinflammatory activity and the remaining are reported for other pharmacological actions. The database of active compounds includes GCs, NSAIDs, COX-2 selective inhibitors and so on. Fig. 1 shows some representative compounds from this dataset. It is remarkable that the wide structural variability of active compounds in the training and prediction sets assures adequate extrapolation power and increases the possibilities for the discovery of new leads with anti-inflammatory activity. In this way, this dataset will also constitute a useful tool for scientific research in synthesis, natural products chemistry, theoretical chemistry, and other areas related to the field of anti-inflammatory compounds. The other 626 compounds having different clinical uses such as antivirals, sedative/hypnotics, diuretics, anticonvulsants, hemostatics, oral hypoglycemics, anti-hypertensives, cathartics, antihelminthics, and anticancer compounds were chosen like “inactive” compounds through random selection, ensuring great structural variability as well. The classification of these organic compounds as “inactive” (non-anti-inflammatory compounds) does not guarantee that some of them have unknown antiinflammatory activity. All compounds were taken from the Negwer Handbook [65] and Merck Index [66]. The molecular structures of all active compounds used in the database are listed in Table SI2 of the Supporting Information. After the selection of the compounds, a virtual screening of an in house dataset, for which any anti-inflammatory properties were studied, will be carried out in order to discover new lead series. 4.1.2. Representation of the molecular samples Several kinds of representations are generally used in this regard, all well-know like molecular descriptor (MDs) or molecular indices. These parameters are numbers that characterize a specific aspect of the molecular structure [99]. The so-called topological (and topo-chemical) indices are among the most useful MDs known nowadays [100,101]. These theoretical indices are numbers that describe the structural information of molecules through graph-theoretical invariants and can be considered as structureexplicit descriptors [102]. In the present report, a novel 2D TOMOCOMD-CARDD MD family, namely atom, atom-type, and total quadratic indices were used in order to codify the molecular structure of every molecule in the dataset. These MDs are based on the calculation of quadratic maps (quadratic form) in
thought as an interacting-electronic chemical-network in step k. The present approach is based on a simple model for the intramolecular (stochastic) movement of all outer-shell electrons. The theoretical scaffold of these atom-based MDs and its use to represent small-to-medium size organic chemicals, as well as QSAR and drug design studies, have been explained in detail elsewhere [22,27,29,33,37,68]. 4.1.3. Computational program: TOMOCOMD-CARDD approach TOMOCOMD is an interactive program for molecular design and bioinformatics research, developed upon the base of a user-friendly philosophy [13]. In this report, we only used the CARDD (Computed-Aided ‘Rational’ Drug Design) subprogram. All MDs [total and local (both atom and atom-type) non-stochastic and stochastic quadratic indices were calculated in this software. 4.1.4. Chemometric studies The statistical software package STATISTICA [14] was used to develop the LDA models. The LDA is one of the most currently used programs nowadays; moreover, its use in drug discovery and drug design has been widely described [22,28,33,80,103e107]. The aim of LDA, a heuristics algorithm capable of distinguishing between two or more categories of objects, is to find a linear function to allow discrimination between active and inactive compounds [108]. The LDA was selected between many statistical methods to get classification functions due to its simplicity and it was carried out with the STATISTICA software [14]. A forward stepwise search procedure was fixed as the strategy for variable selection. The construction process of the model occurs through many steps in the following way: the variables are entered and evaluated by STATISTICA in the model, the variable with greatest contribution to discriminate between groups is included in the model, and then STATISTICA continues with the next step. The principle of maximal parsimony (Occam’s razor) was taken into account as a strategy for model selection. In connection, we selected the model with a high statistical significance but having as few parameters (ak) as possible. The quality of the models was determined by examining Wilks’ l parameter (U statistic), the square Mahalanobis distance (D2), the Fisher ratio (F), and the corresponding p level [p(F)], as well as the percentage of good classification (accuracy) in the training and test sets. The classification of cases was performed by means of the posterior classification probabilities. The biological activity was coded by a dummy variable “Class”. This variable indicates the presence of either an active compound (Class ¼ 1) or an inactive compound (Class ¼ 1). By using the models, one compound can then be classified as active if %ΔP > 0, where %ΔP ¼ [P (Active)  P (Inactive)]  100. The P(Active) and P(Inactive) are the probabilities with which the equations classify a compound as active or inactive, respectively. Performing the assessment of the obtained models, the sensibility, the specificity (also known as “hit rate”), the false positive rate (also known as “false alarm rate”), and Matthews’ correlation coefficient (C), were calculated; and checked in the training and test sets [109]. Finally, leverage approach [73] was used to evaluate the AD of QSAR models. Through this method, it is possible to verify whether a new chemical will lie within the structural model domain. The leverage h of a compound measures its influence on the model. Therefore, leverage is used as a quantitative measure of the model AD and is suitable for evaluating the degree of extrapolation, which represents a sort of compound “distance” from the model experimental space. Leverage values can be calculated for both training compounds and new compounds. In the first case, they are useful for finding training compounds that influence model parameters to a marked extent, resulting in an unstable model. In the second case, they are useful for checking the applicability domain of the model

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

[74]. The warning leverage, h*, is a critical value or cut-off to consider the prediction made for the model for a specific compound in dataset. The leverage h* can be defined as 3xp0 /n, where n is the number of training chemicals and p0 is the number of model parameters plus one [71,72]. Prediction should be considered unreliable for compounds of high leverage value (h > h*). A leverage greater than the warning leverage h* means that the response of the predicted compound can be extrapolated from the model, and therefore, the predicted value must be used with great care. Only predicted data for chemicals belonging to the chemical domain of the training set should be proposed. However, this fact can be seen from two points of view taken into consideration the set of compounds evaluated. For example, when the leverage value of a compound is lower than the critical value, the probability of accordance between predicted and actual values is as high as that for the chemicals in the training set (good leverage). Conversely, a high-leverage chemical in the test set is structurally distant from the training chemicals (bad leverage); thus it can be considered outside the AD of the model. 4.1.5. Prediction algorithms and ensemble classifier (multi-agent predictor or fusion approach) Here, we used non-stochastic and stochastic quadratic indices to develop classification-based QSAR models in order to classify molecules as anti-Inflammatory or inactive compounds. These MDs have a few parameters that can be “modified” in the calculation process. The number of these uncertain parameters depends on which atom-labels (AP scheme) were used for the prediction engine. It would be much more tedious and time-consuming to determine the optimal values for AP [AP [69]: Atomic mass (AP ¼ M), atomic polarizability (AP ¼ P), atomic Mullinken electronegativity (AP ¼ K), van der Waals atomic volume (AP ¼ V)] and atomic Pauling electronegativity (AP ¼ G) uncertain parameters. In addition, the number of uncertain parameters also depends on which MDs sets are used to represent the chemical samples. For instance, here every model can be fitted by two types of MD sets: 1) non-stochastic MDs (NS) and, 2) stochastic MDs (SS). In order to solve the problem, let us use a [2APþ1NSþ1SSþ1(NS þ SS)]dimensional fusion approach (13 models in total), similar to that which was early intruded in protein research [98]. First, the basic individual classifiers to be generally expressed like CI(NS-AP, SS-AP, NS, SS, NS þ SS) and the predicted classification results for a query molecule M by every classifier can be formulated by,

CI ðNS  AP; SS  AP; NS; SS; NS þ SSÞiM ¼ CNSAP; SSAP; NS; SS; NSþSS ðMÞ˛S

(14)

where the symbol i is an action operator meaning using CI(NS-AP, SS-AP, NS, SS, NS þ SS) to classify M, S representing the union of the two subsets defined (active or inactive). Therefore, the final predicted result should be determined by a fusion approach through the following voting mechanism. Now let us introduce an ensemble classifier CE, which is formed by fusing all sets of the basic individual classifiers CI(NS-AP, SS-AP, NS, SS, NS þ SS) and can be formulated the follows:

CE ¼ C1 ðM; NSÞcC2 ðK; NSÞcC3 ðP; NSÞcC4 ðV; NSÞcC5 ðG; NSÞcC6 ðall AP; NSÞcC7 ðM; SSÞcC8 ðK; SSÞ cC9 ðP; SSÞcC10 ðV; SSÞcC11 ðG; SSÞcC12 ðall AP; SSÞ cC13 ðall AP; NS þ SSÞ

ð15Þ

where the symbol c denotes the fusing operator. Then, the voting score for the query molecules M belonging to the cth class is given by,

pc ¼

2 X

5 X

5749

wAP;MDs DðAP; MDs; Sc Þ

AP ¼ 1 MDs ¼ 1

þ

3 X

(16)

wallAP;MDs Dðall  AP; MDs; Sc Þ; ðc ¼ 1; 1Þ

MDs ¼ 1

where Sc ¼ 1 is for anti-inflammatory leads and Sc ¼ 1 for other drugs, wAP,MDs and wall-AP,MDs are the weighting factors and were set at 1 for simplicity. The delta functions in Eq. (16) are given by,

DðAP; MDs; Sc Þ ¼



1

if CAP;MDs ðMÞ˛Sc

0

otherwise

Dðall  AP; MDs; Sc Þ ¼



1

if CP;MDs ðMÞ˛Sc

0

otherwise

(17)

(18)

thus the query molecule M is predicted belonging to the class (c), or subset Sc, for which the score of Eq. (16) is the highest; i.e.,

m ¼ arg max fpc g ðc ¼ 1; 1Þ c

(19)

where, m is the argument of c that maximizes pc. If there is a tie, then the final predicted result will be randomly assigned (or taken as unclassified) to one of their corresponding subsets, although this kind of tie case rarely happens and, actually, was not observed in the current study. 4.2. Chemistry The 34 chemicals selectionated by in silico study were synthetized like indicated in the original publications of indazole [83e88], indole [89], cinnoline [90] and quinoxaline [91,92] derivatives (as well as other related new chemicals and their derivatives; for instance, chemicals 23e25). 4.3. Wet evaluation 4.3.1. In vivo efficacy studies in zebrafish All chemicals were first dissolved in dimethylsulfoxide (DMSO) 99.5% (GC) SigmaeAldrich), and then diluted in embryo medium (Danieau’s 0.3X (‘E3’): 5 mM NaCl, 0.17 mM KCl, 0.33 mM CaCl2, 0.33 mM MgSO4, 105% methylene blue. The final concentration of DMSO did not exceed 1% in order to avoid damage to the animal. During the course of the experiments, transgenic zebrafish of the line fli-1: EGFP were used. Once the fertilized eggs were collected, they were reared in embryo medium E3 in an incubator at 28  C. At 20 h post fertilization (hpf), 1-phenyl-2-thiourea (PTU) (Sigma-10  (0.03%)) was added to larvae’s medium in proportion 1:9 to prevent the formation of melanophores in zebrafish embryos and larvae [110]. 4.3.1.1. Toxicological evaluation. The Maximum Tolerated Concentration (MTC) was defined as the maximum concentration at which no death or signs of toxicity in the zebrafish larvae. The MTC was initially determined by incubating in the 4-dpf zebrafish larvae with different concentrations of the compounds. During 24 h, 5 larvae per well were placed in a tissue culture plate of 24 wells using 10 mM, 30 mM and 100 mM as concentration in the medium. Larvae were then observed under a light microscope during the time of incubation in treatment groups at 5dpf after 24 h exposure to the tested drugs. 4.3.1.2. Anti-inflammatory experimental procedure: tail transection. In this report, the anti-inflammatory activity was determined by a novel in vivo test (using zebrafish), developed by one of our research teams (Department of Pharmaceutical Sciences, Katholieke

5750

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

Universiteit Leuven, Belgium), taken into consideration some results obtained by Renshaw et al. [62], but making some important changes in the procedures of several steps, such as in the form like quantitative data could be generated from this model. The theoretical topic and main characteristics of this biological test will be seen in a forthcoming paper [63]. However, here we will give some fine points. First, ten larvae from 4 dpf were used per treated and control groups. In each case, the MTC was used as the assay concentration, and indomethacin (Merck Sharp & Dohme) 30 mM was used as control group. These larvae were placed per well in end volume of 1 mL. After 1 h, the larvae were anesthetized by immersion in E3 with tricaine [111], the complete transection of the tail was performed with a sterile scalpel. Once tails were cut, zebrafishses were placed again in a tissue culture plate at assay’s concentrations, together with10 mg/mL solution of lypopolysaccharide to stimulate the leukocyte migration. The time of incubation was from 20 to 22 h. After 7 h, the larvae were fixed in 4% paraformaldehyde for 5 min and then washed twice with PBS. One milliliter of staining solution of Leucognost Pot was added and 10 min later, every zebrafish was observed at a microscope and the leukocyte migration was analyzed [63]. Each assay was triplicated. 4.3.2. In vitro cytotoxicity on macrophage cells Murine J774 macrophages were grown in plastic 25 mL flasks in (RPMI)-1640 medium (Sigma) supplemented with 20% heat inactivated (30 min, 56  C) foetal calf serum (FCS) and 100 IU penicillin/ mL þ 100 mg/mL streptomycin, in a humidified 5% CO2/95% air atmosphere at 37  C and subpassaged once a week. The J774 macrophages were seeded (70,000 cells/well) in 96-well flatbottom microplates (Nunc) with 200 mL of medium. The cells were allowed to attach for 24 h at 37  C and then exposed to the compounds (dissolved in DMSO, maximal final concentration of solvent was 0.2%) for another 24 h. Afterwards, the cells were washed with PBS and incubated (37  C) with 3-(4,5dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) 0.4 mg/mL for 60 min. The MTT solution was removed and the cells solubilized in DMSO (100 mL). The extent of reduction of MTT to formazan within cells was quantified by measurement of OD595 [112]. Each concentration was assayed three times and six cell growth controls were used in each test. The assays were performed in duplicate. Cytotoxic percentages (%C) were determined as follows: %C ¼ [1  (ODp  ODpm)/(ODc  ODm)]  100, where ODp represents the mean OD595 value recorded for wells with macrophages containing different doses of product; ODpm denotes the mean OD595 value recorded for different concentrations of product in medium; ODc represents the mean OD595 value recorded for wells with macrophages and no product (growth controls), and ODm means the mean OD595 value recorded for medium/control wells. The 50% cytotoxic dose (CD50) was defined as the concentration of drug that decreases OD595 up to 50% of that in control cultures [94]. 4.4. Determination of topical anti-inflammatory activity in mouse ear oedema 4.4.1. Animals Groups of six Swiss female mice weighting 25e30 g from Harlan Interfauna Ibérica (Barcelona, Spain) were used. Housing conditions and the in vivo experiments were approved by the Institutional Ethics Committee of the University of Valencia (Spain). 4.4.2. Reagents Biochemicals, chemicals, reagents and materials were purchased from Invitrogen, Cayman Chemical Company, Merck, Panreac, Fluka and SigmaeAldrich.

4.4.3. Tetradecanoylphorbol acetate (TPA)-induced mouse ear oedema Topical anti-inflammatory activity of the compounds was studied using the method described by de Young and de Young [113] and modified by Payá et al. [114]. Oedema was induced on the right ear by topical application of 2.5 mg/ear of TPA in 20 ml acetone (10 ml/side). The untreated left ear was used as control. Compounds and indomethacin (0.5 mg/ear) were dissolved in 20 ml acetone and applied to right ear simultaneously with TPA. Four hours after the inflammation induction the animals were sacrificed and a biopsy (6 mm diameter) of both ears (left and right) was performed. The oedema was measured as an increase in ear weight due to the TPA agent application by difference in weight between both ears. The inflammation inhibition percentage was evaluated from the weight difference between treated and non-treated ears of each animal compared to the control group (vehicle). Details of the method have been described earlier by Giner et al. [115]. 4.4.4. Myeloperoxidase assay in mouse ear oedema tissues We used the method described by de Young et al. [116] with some slight modifications. Ear sections of every treatment were placed in an eppendorf tube with 0.75 mL of 80 mM sodium phosphate buffer (PBS, 1X) containing 0.5% HTAB, homogenized (45 s at 0  C) in a homogenizer (POLYTRON) and decanted into a microfuge tube. The tube was washed with a second 0.75 mL aliquot of HTAB in PBS and added to the tube. The mixture was centrifuged (10300g at 4  C for 20 min) and the supernatant (30 mL  triplicate) was added to a 200 mL of a mixture containing 100 mL of 80 mM PBS, 85 mL of 0.22 M PBS (pH 5.4) and 15 mL of H2O2 0.017% in a 96-well microtiter plate. The reaction was started by the addition of 20 mL of 18.4 mM TMB in 8% aqueous dimethylformamide. The mixture was incubated for 3 min at 37  C and then placed on ice. The reaction was stopped by addition of 30 mL 1.46 M buffer NaOAc/HOAc (pH 3.0). Enzyme activity was determined by measuring absorbance at 630 nm and expressed as the inhibition percentage of MPO levels, determined as the absorbance difference between the control group (vehicle) and the treated group compared to the absorbance observed in the control. Acknowledgements One of the authors (M-P. Y) thanks the program ‘Estades Temporals per a Investigadors Convidats’ for a fellowship to work at Valencia University (VU) in 2011. Dany Siverio-Mota (and M-P. Y.) acknowledges the Lab. Farmaceutische Biologie, Katholieke Universiteit Leuven (Belgium) for kind hospitality during the second semester of 2009. The authors acknowledge also the partial financial support from Ministerio de Ciencia e Innovacion de España (Projects SAF2009-10399, SAF2009-13059-C03-01 and SAF2009-13059-C0302). Also, this work was supported in part by VLIR (Vlaamse InterUniversitaire Raad, Flemish Interuniversity Council, Belgium) under the IUC Program VLIR-UCLV. Finally, but not least, the authors want to express their acknowledgements to Prof. Luis Rios for his availability to experimental tests at the VU and to Prof. Jorge Galvez (VU) for his help and useful comments about the new chemicals. Appendix. Supplementary material Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.ejmech.2011.07.053. References [1] J.A. DiMasi, R.W. Hansen, H.G. Grabowski, J. Healt. Econ. 22 (2003) 151e185. [2] P. Warne, C. Page, Is there a best strategy for drug discovery? Drug News Perspect. 16 (2003) 177e182.

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753 [3] C.B. Mazza, K. Rege, C.M. Breneman, N. Sukumar, J.S. Dordick, S.M. Cramer, High-throughput screening and quantitative structure-efficacy relationship models of potential displacer molecules for ion-exchange systems, Biotechnol. Bioeng. 80 (2002) 60e72. [4] M.H.J. Seifert, K. Wolf, D. Vitt, Virtual high-throughput in silico screening, Biosilico I (2003) 143e149. [5] J. Bajorath, Integration of virtual and high-throughput screening, Nat. Rev. Drug Discov. 1 (2002) 882e894. [6] J. Galvez, R. Garcia-Domenech, J.V. de Julian-Ortiz, R. Soler, Topological approach to drug design, J. Chem. Inform. Comp. Sci. 35 (1995) 272e284. [7] M.J. Duart, R. Garcia-Domenech, J. Galvez, P.A. Aleman, R.V. Martin-Algarra, G.M. Anton-Fos, Application of a mathematical topological pattern of antihistaminic activity for the selection of new drug candidates and pharmacology assays, J. Med. Chem. 49 (2006) 3667e3673. [8] M.J. Duart, G.M. Anton-Fos, P.A. Aleman, J.B. Gay-Roig, M.E. Gonzalez-Rosende, J. Galvez, R. Garcia-Domenech, New potential antihistaminic compounds. virtual combinatorial chemistry, computational screening, real synthesis, and pharmacological evaluation, J. Med. Chem. 48 (2005) 1260e1264. [9] R. Garcia-Domenech, J. Galvez, J.V. de Julian-Ortiz, L. Pogliani, Some new trends in chemical graph theory, Chem. Rev. 108 (2008) 1127e1169. [10] H. Kubinyi, Chemical similarity and biological activities, J. Braz. Chem. Soc. 13 (2002) 717e726. [11] H. Eckert, J. Bajorath, Molecular similarity analysis in virtual screening: foundations, limitations and novel approaches, Drug Discov. Today 12 (2007) 225e236. [12] M. Gálvez-Llompart, R.M. Giner, M.C. Recio, S. Candeletti, R. García-Domenech, Application of molecular topology to the search of novel NSAIDs. Experimental validation of activity, Lett. Drug Des. Disc. 7 (2010) 438e445. [13] Y. Marrero-Ponce, V. Romero, TOMOCOMD-CARDD software. TOMOCOMD (TOpological MOlecular COMputational Design) for Windows, version 1.0 is a preliminary experimental version; in future a professional version can be obtained upon request to Y. Marrero: [email protected] or [email protected], Central University of Las Villas, Santa Clara, Villa Clara, 2002. [14] STATISTICA (Data Analysis Software System) version 6. StatSoft, Inc, 2001. www.statsoft.com. [15] H. van de Waterbeemd, Discriminant analysis for activity prediction. in: H. van de Waterbeemd (Ed.), Chemometric Methods in Molecular Design. VCH Publishers, Weinheim, 1995, pp. 265e288. [16] E. Estrada, E. Uriarte, A. Montero, M. Teijeira, L. Santana, E. De Clercq, A novel approach for the virtual screening and rational design of anticancer compounds, J. Med. Chem. 43 (2000) 1975e1985. [17] H. Gonzalez-Diaz, E. Uriarte, R. Ramos de Armas, Predicting stability of Arc repressor mutants with protein stochastic moments, Bioorg. Med. Chem. 13 (2005) 323e331. [18] A. Garcia-Garcia, J. Galvez, J.V. de Julian-Ortiz, R. Garcia-Domenech, C. Munoz, R. Guna, R. Borras, New agents active against Mycobacterium avium complex selected by molecular topology: a virtual screening method, J. Antimicrob. Chemother. 53 (2004) 65e73. [19] M.T. Cronin, A.O. Aptula, J.C. Dearden, J.C. Duffy, T.I. Netzeva, H. Patel, P.H. Rowe, T.W. Schultz, A.P. Worth, K. Voutzoulidis, G. Schuurmann, Structure-based classification of antibacterial activity, J. Chem. Inf. Comput. Sci. 42 (2002) 869e878. [20] Y. Marrero-Ponce, Total and local quadratic indices of the molecular pseudograph’s atom adjacency matrix: applications to the prediction of physical properties of organic compounds, Molecules 8 (2003) 687e726. [21] Y. Marrero-Ponce, Total and local (Atom and atom-type) molecular quadratic indices: significance-interpretation, comparison to other molecular descriptors, and QSPR/QSAR applications, Bioorg. Med. Chem. 12 (2004) 6351e6369. [22] Y. Marrero Ponce, Linear indices of the “molecular pseudograph’s atom adjacency matrix”: definition significance-interpretation, and application to QSAR analysis to flavone derivatives as HIV-1 integrase inhibitors, J. Chem. Inf. Comput. Sci. 44 (2004) 2010e2026. [23] Y. Marrero-Ponce, A. Huesca-Guillen, F. Ibarra-Velarde, Quadratic indices of the “molecular pseudograph’s atom adjacency matrix” and their stochastic forms: a novel approach for virtual screening and in silico discovery of new lead paramphistomicide drugs-like compounds, J. Mol. Struct. (Theochem) 717 (2005) 67e79. [24] Y. Marrero-Ponce, F. Torrens, R. García-Domenech, S.E. Ortega-Broche, V. Romero Zaldivar, Novel 2D TOMOCOMD-CARDD descriptors: atom-based stochastic and non-stochastic bilinear indices and their QSPR applications, J. Math. Chem. (2006). doi:10.1007/s10910-10008-19389-10910. [25] Y. Marrero-Ponce, A. Meneses-Marcel, J.A. Catillo-Garit, Y. Machado-Tugores, J.A. Escario, A. Gómez-Barrio, D. Montero Pereira, J.J. Nogal-Ruiz, V.J. Arán, A.R. Martínez-Fernández, F. Torrens, R. Rotondo, Predicting antitrichomonal activity: a computational screening using atom-based bilinear indices and experimental proofs, Bioorg. Med. Chem. 14 (2006) 6502e6524. [26] Y. Marrero Ponce, M.A. Cabrera Perez, V. Romero Zaldivar, H. Gonzalez Diaz, F. Torrens, A new topological descriptors based model for predicting intestinal epithelial transport of drugs in Caco-2 cell culture, J. Pharm. Pharmaceut Sci. 7 (2004) 186e199. [27] Y. Marrero-Ponce, A. Montero-Torres, C.R. Zaldivar, M.I. Veitia, M.M. Perez, R.N. Sanchez, Non-stochastic and stochastic linear indices of the ‘molecular

[28]

[29]

[30]

[31]

[32]

[33]

[34]

[35]

[36]

[37]

[38]

[39]

[40]

[41] [42] [43] [44] [45] [46]

[47] [48]

5751

pseudograph’s atom adjacency matrix’: application to ‘in silico’ studies for the rational discovery of new antimalarial compounds, Bioorg. Med. Chem. 13 (2005) 1293e1304. Y. Marrero-Ponce, R. Medina-Marrero, F. Torrens, Y. Martinez, V. RomeroZaldivar, E.A. Castro, Atom, atom-type, and total non-stochastic and stochastic quadratic fingerprints: a promising approach for modeling of antibacterial activity, Bioorg. Med. Chem. 13 (2005) 2881e2899. Y. Marrero-Ponce, R.M. Marrero, F. Torrens, Y. Martinez, M.G. Bernal, V.R. Zaldivar, E.A. Castro, R.G. Abalo, Non-stochastic and stochastic linear indices of the molecular pseudograph’s atom-adjacency matrix: a novel approach for computational in silico screening and “rational” selection of new lead antibacterial agents, J. Mol. Model. 12 (2006) 255e271. Y. Marrero-Ponce, Y. Machado-Tugores, D.M. Pereira, J.A. Escario, A.G. Barrio, J.J. Nogal-Ruiz, C. Ochoa, V.J. Aran, A.R. Martinez-Fernandez, R.N. Sanchez, A. Montero-Torres, F. Torrens, A. Meneses-Marcel, A computer-based approach to the rational discovery of new trichomonacidal drugs by atomtype linear indices, Curr. Drug Discov. Technol. 2 (2005) 245e265. Y. Marrero-Ponce, M. Iyarreta-Veitia, A. Montero-Torres, C. Romero-Zaldivar, C.A. Brandt, P.E. Avila, K. Kirchgatter, Y. Machado, Ligand-based virtual screening and in silico design of new antimalarial compounds using nonstochastic and stochastic total and atom-type quadratic maps, J. Chem. Inf. Model. 45 (2005) 1082e1100. Y. Marrero-Ponce, J.A. Castillo-Garit, F. Torrens, V. Romero-Zaldivar, E. Castro, Atom, atom-type, and total linear indices of the “molecular pseudograph’s atom adjacency matrix”: application to QSPR/QSAR studies of organic compounds, Molecules 9 (2004) 1100e1123. Y. Marrero-Ponce, J.A. Castillo-Garit, E. Olazabal, H.S. Serrano, A. Morales, N. Castanedo, F. Ibarra-Velarde, A. Huesca-Guillen, A.M. Sanchez, F. Torrens, E.A. Castro, Atom, atom-type and total molecular linear indices as a promising approach for bioorganic and medicinal chemistry: theoretical and experimental assessment of a novel method for virtual screening and rational design of new lead anthelmintic, Bioorg. Med. Chem. 13 (2005) 1005e1020. Y. Marrero-Ponce, J.A. Castillo-Garit, E. Olazabal, H.S. Serrano, A. Morales, N. Castañedo, F. Ibarra-Velarde, A. Huesca-Guillen, E. Jorge, A. del Valle, F. Torrens, E.A. Castro, TOMOCOMD-CARDD, a novel approach for computeraided ‘rational’ drug design: I. Theoretical and experimental assessment of a promising method for computational screening and in silico design of new anthelmintic compounds, J. Comput.-Aided Mol. Design 18 (2004) 615e634. Y. Marrero-Ponce, M.A. Cabrera, V. Romero-Zaldivar, M. Bermejo, D. Siverio, F. Torrens, Prediction of intestinal epithelial transport of drug in (Caco-2) cell culture from molecular structure using in silico approaches during early drug discovery, Internet Electron J. Mol. Des 4 (2005) 124e150. Y. Marrero-Ponce, M.A. Cabrera, V. Romero, E. Ofori, L.A. Montero, Total and local quadratic indices of the “molecular pseudograph’s atom adjacency matrix”. Application to prediction of Caco-2 permeability of drugs, Int. J. Mol. Sci. 4 (2003) 512e536. A. Montero-Torres, M.C. Vega, Y. Marrero-Ponce, M. Rolon, A. Gomez-Barrio, J.A. Escario, V.J. Aran, A.R. Martinez-Fernandez, A. Meneses-Marcel, A novel non-stochastic quadratic fingerprints-based approach for the ‘in silico’ discovery of new antitrypanosomal compounds, Bioorg. Med. Chem. 13 (2005) 6264e6275. A. Montero-Torres, R.N. Garcia-Sanchez, Y. Marrero-Ponce, Y. MachadoTugores, J.J. Nogal-Ruiz, A.R. Martinez-Fernandez, V.J. Aran, C. Ochoa, A. Meneses-Marcel, F. Torrens, Non-stochastic quadratic fingerprints and LDA-based QSAR models in hit and lead generation through virtual screening: theoretical and experimental assessment of a promising method for the discovery of new antimalarial compounds, Eur. J. Med. Chem. 41 (2006) 483e493. A. Meneses-Marcel, Y. Marrero-Ponce, Y. Machado-Tugores, A. MonteroTorres, D.M. Pereira, J.A. Escario, J.J. Nogal-Ruiz, C. Ochoa, V.J. Aran, A.R. Martinez-Fernandez, R.N. Garcia Sanchez, A linear discrimination analysis based virtual screening of trichomonacidal lead-like compounds: outcomes of in silico studies supported by experimental results, Bioorg. Med. Chem. Lett. 15 (2005) 3838e3843. G.M. Casanola-Martin, M.T. Khan, Y. Marrero-Ponce, A. Ather, M.N. Sultankhodzhaev, F. Torrens, New tyrosinase inhibitors selected by atomic linear indices-based classification models, Bioorg. Med. Chem. Lett. 16 (2006) 324e330. C.A. Dinarello, Interleukin-1, Interleukin-1 receptors and Interleukin-1 receptor antagonist, Int. Rev. Immunol. 16 (1998) 457e499. C. Nathan, Natural resistance and nitric oxide, Cell 82 (1995) 873e876. K.J. Tracey, Tumor necrosis factor (cachetin) in the biology of septic shock syndrome, Circ. Shock 35 (1991) 123e128. M.J. Fenton, D.T. Golenbock, I2S-binding proteins and receptors, J. Leukoc. Biol. 64 (1998) 25e32. M.P. Glauser, The inflammatory cytokines. New developments in the pathophysiology and treatment of septic shock, Drugs 52 (1996) 9e17. K.J. Tracey, H. Wei, K.R. Manogue, Y. Fong, D.G. Hesse, H.T. Nguyen, G.C. Kuo, B. Beutler, R.S. Cotran, A. Cerami, Cachectin/tumor necrosis factor induces cachexia, anemia, and inflammation, J. Exp. Med. 167 (1998) 1211e1227. H.P. Rang, M.M. Dale, J.M. Ritter, R.J. Flower, Pharmacology, sixth ed. Churchill Livingstone, China, 2007. W.A. Check, M.A. Kaliner, Pharmacology and pharmacokinetics of topical corticoid derivatives used for asthma therapy, Am. Rev. Respir. Dis. 144 (1990) 44e51.

5752

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753

[49] D.M. Geddes, Inhaled corticoids: benefit and risks, Thorax 47 (1992) 404e407. [50] T. Yasuhiro, O. Hiroko, T. Hiroshi, Glucocorticoid-induced osteoporosis, Biomed. Pharmacother. 58 (2004) 500e504. [51] H. Schӓcke, W.D. Dӧcke, K. Asadullah, Mechanisms involved in the side effects of glucocorticoids, Pharmacol. Ther. 96 (2002) 23e43. [52] M.W. James, C.J. Hawkey, Assessment of non-steroidal anti-inflammatory drug (NSAID) damage in the human gastrointestinal tract, Br. J. Clin. Pharmacol. 56 (2003) 146e155. [53] J.G. Hardman, L.E. Limbird, A. Goodman-Gilman, Bases Farmacológicas de la terapéutica Goodman and Gilman, nineth ed. The McGran-Hill Companies, Inc, Atlampa, 1996. [54] D. Bagchi, B. Misner, M. Bagchi, S.C. Kothari, B.W. Downs, R.D. FafarD, H.G. Preuss, Effects of orally administered undenatured type II collagen against arthritic inflammatory diseases: a mechanistic exploration, Int. J. Clin. Pharmacol. Res. 22 (2002) 101e110. [55] OECD, Guidance Document on the Validation of Quantitative StructureeActivity Relationship [QSAR] Models, , In: OECD Environment Health and Safety Publication, Series on testing and assessment, vol. 69 (2007).http:// www.oecd.org/dataoecd/55/35/38130292.pdf Paris. [56] J.R. Mathias, M.E. Dodd, K.B. Walters, J. Rhodes, J.P. Kanki, A.T. Look, A. Huttenlocher, Live imaging of chronic inflammation caused by mutation of zebrafish Hai1, J. Cell Sci. 120 (2007) 3372e3383. [57] E.E. Patton, L.I. Zon, The art and design of genetic screens: zebrafish, Nat. Rev. Genet. 2 (2001) 956e966. [58] L.I. Zon, R.T. Peterson, In vivo drug discovery in the zebrafish, Nat. Rev. Drug Discov. 4 (2005) 35e44. [59] M.O. Crowhurst, J.E. Layton, G.J. Lieschke, Developmental biology of zebrafish myeloid cells, Int. J. Dev. Biol. 46 (2002) 483e492. [60] N.S. Trede, D.M. Langenau, D. Traver, A.T. Look, L.I. Zon, The use of zebrafish to understand immunity, Immunity 20 (2004) 367e379. [61] D.M. Langenau, L.I. Zon, The zebrafish: a new model of T-cell and thymic development, Nat. Rev. Immunol. 5 (2005) 307e317. [62] S.A. Renshaw, S.A. Loynes, D.M.I. Trushell, S. Elworthy, P.W. Ingham, M.K.B. Whyte, A transgenic zebrafish model of neutrophilic inflammation, Blood 108 (2006) 3976e3978. [63] M.L. Cordero-Maldonado, D. Siverio-Mota, L. Vicet-Muro, L. Pieters, C.V. Esguerra, P.A.M. de Witte, A.D. Crawford, Zebrafish bioassay-guided isolation of anti-inflammatory natural products from a Latin American medicinal plant, submitted for publication. [64] A. Golbraikh, A. Tropsha, Predictive QSAR modeling based on diversity sampling of experimental datasets for the training and test set selection, Mol. Divers. 5 (2002) 231e243. [65] M. Negwer, Organic-Chemical Drugs and their Synonyms. Akademie-Verlag, Berlin, 1987. [66] The Merck Index. Chapman and Hall, 1996. [67] J.S. Glasby, Encyclopedia of Antibiotics. Woodhouse, Manchester, 1978. [68] Y. Marrero-Ponce, R.M. Marrero, F. Torrens, Y. Martinez, M.G. Bernal, V.R. Zaldivar, E.A. Castro, R.G. Abalo, Non-Stochastic, Stochastic Linear, Indices of the Molecular Pseudograph’s Atom-Adjacency Matrix: a Novel Approach for Computational in silico Screening and "Rational" Selection of New Lead Antibacterial Agents, J. Mol. Model (2005) 1e17. [69] V. Consonni, R. Todeschini, M. Pavan, Structure/response correlations and similarity/diversity analysis by GETAWAY descriptors. 1. Theory of the novel 3D molecular descriptors, J. Chem. Inf. Comput. Sci. 42 (2002) 682e692. [70] P. Baldi, S. Brunak, Y. Chauvin, C.A. Andersen, H. Nielsen, Assessing the accuracy of prediction algorithms for classification: an overview, Bioinformatics 16 (2000) 412e424. [71] P. Gramatica, Principles of QSAR models validation: internal and external QSAR, QSAR Comb. Sci. 26 (2007) 694e701. [72] L. Eriksson, J. Jaworska, A.P. Worth, M.T. Cronin, R.M. McDowell, P. Gramatica, Methods for reliability and uncertainty assessment and for applicability evaluations of classification- and regression-based QSARs, Environ. Health Perspect. 111 (2003) 1361e1375. [73] A.C. Atkinson, Plots, Transformations and Regression. Clarendon Press, Oxford, 1985. [74] E. Papa, F. Villa, P. Gramatica, Statistically validated QSARs, based on theoretical descriptors, for modeling aquatic toxicity of organic chemicals in Pimephales promelas (fathead minnow), J. Chem. Inf. Model. 45 (2005) 1256e1266. [75] A. Golbraikh, M. Shen, Z. Xiao, Y.D. Xiao, K.H. Lee, A. Tropsha, Rational selection of training and test sets for the development of validated QSAR models, J. Comput. Aided Mol. Des. 17 (2003) 241e253. [76] A. Golbraikh, A. Tropsha, Beware of q2!, J. Mol. Graph Model. 20 (2002) 269e276. [77] S. van Damme, P. Bultinck, Software news and update a new computer program for QSAR-analysis: ARTE-QSAR, J. Comput. Chem. 28 (2007) 1924e1928. [78] S. Van Damme, W. Langenaeker, P. Bultinck, Prediction of bloodebrain partitioning: a model based on ab initio calculated quantum chemical descriptors, J. Mol. Graphics Modell. 26 (2008) 1223e1236. [79] M. Shen, Y. Xiao, A. Golbraikh, V.K. Gombar, A. Tropsha, Development and validation of k-nearest-neighbor QSPR models of metabolic stability of drug candidates, J. Med. Chem. 46 (2003) 3013e3020. [80] J. Gálvez, R. García-Domenech, C. de Gregorio Alapont, J. de Julián-Ortiz, L. Popa, Pharmacological distribution diagrams: a tool for de novo drug design, J. Mol. Graph. 14 (1996) 272e276.

[81] E. Estrada, S. Vilar, E. Uriarte, Y. Gutierrez, In silico studies toward the discovery of new anti-HIV nucleoside compounds with the use of TOPSMODE and 2D/3D Connectivity indices. 1. pyrimidyl derivatives, J. Chem. Inf. Comput. Sci. 42 (2002) 1194e1203. [82] M.G. Ford, W.R. Pitt, D.C. Whitley, J. Mol. Graph. Mod. 22 (2004) 467e472. [83] V.J. Arán, J.L. Asensio, J.R. Ruiz, M. Stud, The heterocyclization of n0 , n0 disubstituted 2-halobenzohydrazides to 1,1-disubstituted indazol-3ylioxides, J. Chem. Res. (S) 218 (1993) 1322e1345. [84] V.J. Arán, J.L. Asensio, J.R. Ruiz, M. Stud, J. Chem. Res. (M) (1993) 1322e1345. [85] V.J. Arán, J.L. Asensio, J.R. Ruiz, M. Stud, Reactivity of 1,1-disubstituted indazol-3-yliooxides: synthesis of some substituted indazolols and indazolinones, J. Chem. Soc. Perkin Trans. 1 (1993) 1119e1127. [86] V.J. Arán, M. Flores, M. Muñoz, J.R. Ruiz, P. Sánchez-Verdú, M. Stud, Liebigs Ann. (1995) 817e824. [87] V.J. Arán, M. Flores, M. Muñoz, J.A. Páez, P. Sánchez-Verdú, M. Stud, Liebigs Ann. (1996) 683e691. [88] V.J. Arán, C. Ochoa, L. Boiani, P. Buccino, H. Cerecetto, A. Gerpe, M. Gonzalez, D. Montero, J.J. Nogal, A. Gomez-Barrio, A. Azqueta, A. Lopez de Cerain, O.E. Piro, E.E. Castellano, Synthesis and biological properties of new 5nitroindazole derivatives, Bioorg. Med. Chem. 13 (2005) 3197e3207. [89] J.R. Ruiz, V.J. Arán, J.L. Asensio, M. Flores, M. Stud, Synthesis of quaternary indoxyl derivatives by intramolecular cyclization of some substituted acetophenone, Liebigs Ann. Chem. (1994) 679e684. [90] V.J. Arán, J.L. Asensio, J. Molina, P. Muñoz, J.R. Ruiz, M. Stud, Approaches to 1,1-disubstituted cinnolin-3-ylio oxides: synthesis and reactivity of a new class of heterocyclic betaines, J. Chem. Soc. Perkin Trans. (1997) 2229e2235. [91] S. de Castro, R. Chicharro, V.J. Arán, Synthesis of quinoxaline derivatives from substituted acetanilides through intramolecular quaternization reactions, J. Chem. Soc. Perkin Trans. (2002) 790e802. [92] R. Chicharro, S. de Castro, J.L. Reino, V.J. Arán, Synthesis of tri- and tetracyclic condensed quinoxalin-2-ones fused across the C-3-N-4 bond, Eur. J. Org. Chem. (2003) 2314e2326. [93] O.M. Rivera-Borroto, Y. Marrero-Ponce, A. Meneses-Marcel, J.A. Escario, A. Gómez-Barrio, A.V. A, M.A. Martins-Alho, D. Montero Pereira, J.J. Nogal, F. Torrens, F. Ibarra-Velarde, V. Vera Montenegro, A. Huesca-Guille, N. Rivera, V. Vogel, Discovery of novel trichomonacidals using LDA-driven QSAR models and bond-based bilinear indices as molecular descriptors, QSAR Comb. Sci. (2008). doi:10.1002/qsar.200610165. [94] M. Celeste-Vega, A. Montero-Torres, Y. Marrero-Ponce, M. Rolón, A. GómezBarrio, J.A. Escario, V.J. Arán, A.R. Martínez-Fernández, A. Meneses-Marcel, New ligand-based approach for the discovery of antitrypanosomal compounds, Bioorg. Med. Chem. Lett. 16 (2006) 1898e1904. [95] G.J. Blackwell, R.J. Flower, Prostaglandins 16 (1978) 417. [96] G.A. Higgs, R.J. Flower, J.R. Vane, Biochem. Pharmacol. 28 (1979) 1959. [97] P. Bruneau, C. Delvare, Indazolinones, a new series of redox-active 5lipoxygenase inhibitors with built-in selectivity and oral activity, J. Med. Chem. 34 (1991) 1028e1036. [98] K.-C. Chou, H.-B. Shen, Recent progress in protein subcellular location prediction, Anal. Biochem. 370 (2007) 1e16. [99] H. Van de Waterbeemd, R.E. Carter, G. Grassy, H. Kubinyi, Y.C. Martin, M.S. Tute, P. Willett, Annu. Rep. Med. Chem. 33 (1998) 397. [100] M. Karelson, Molecular Descriptors in QSAR/QSPR. John Wiley & Sons, New York, 2000. [101] A.R. Katritzky, E.V. Gordeeva, Traditional topological indices vs electronic, geometrical, and combined molecular descriptors in QSAR/QSPR research, J. Chem. Inf. Comput. Sci. 33 (1993) 835e857. [102] R. Todeschini, V. Consonni, Handbook of Molecular Descriptors. Wiley-VCH, Germany, 2000. [103] Y. Marrero-Ponce, M.A. Cabrera, V. Romero, D.H. González, F. Torrens, A New topological descriptors based model for predicting intestinal epithelial transport of drugs in Caco-2 cell culture, J. Pharm. Pharm. Sci. 7 (2004) 186e199. [104] Y. Marrero-Ponce, R. Medina, E.A. Castro, R. de Armas, H. González, V. Romero, F. Torrens, Protein quadratic indices of the “macromolecular pseudograph’s R-carbon atom adjacency matrix”. 1. Prediction of Arc repressor alanine-mutant’s stability, Molecules 9 (2004) 1124e1147. [105] Y. Marrero-Ponce, R. Medina-Marrero, J.A. Castillo-Garit, V. Romero-Zaldivar, F. Torrens, E.A. Castro, Protein linear indices of the “macromolecular pseudograph’s R-Carbon atom adjacency matrix” in bioinformatics. 1. Prediction of protein stability effects of a complete set of alanine substitutions in Arc repressor, Bioorg. Med. Chem. 13 (2005) 3003e3015. [106] D.J. Klein, Graph theoretically formulated electronic-structure theory, Internet electron J. Mol. Des. 2 (2003) 814e834. [107] J.V. Julián-Ortiz, J. Gálvez, C. Muñoz-Collado, R. García-Domenech, C. Gimeneo-Cardona, Virtual combinatorial synthesis and computational screening of new potential anti-herpes compounds, J. Med. Chem. 42 (1999) 3308e3314. [108] H. van de Waterbeemd, in: H. van de Waterbeemd (Ed.), Chemometric Methods in Molecular Design, VCH Publishers, Weinheim, 1995, pp. 265e328. [109] A. Baldi, E. Dragonetti, T. Battista, A.M. Groeger, V. Esposito, G. Baldi, D. Santini, Detection of circulating malignant cells by RT-PCR in long-term clinically disease-free i stage melanoma patients, Anticancer Res. 20 (2000) 3923e3928. [110] J. Karlsson, J. von Hofsten, P. Olsson, Generating transparent zebrafish: a refined method to improve detection of gene expression during embryonic development, Mar. Biotechnol. 3 (2001) 522e527.

Y. Marrero-Ponce et al. / European Journal of Medicinal Chemistry 46 (2011) 5736e5753 [111] M. Westerfield, The Zebrafish Book: a Guide for the Laboratory Use of Zebrafish (Brachydanio rerio). University of Oregon Press, Eugene, OR, 1995. [112] Y. Hattori, N. Nakanishi, Effects of cyclosporin A and FK506 on nitric oxide and tetrahydrobiopterin synthesis in bacterial lipopolysaccharide-treated J774 macrophages, Cell. Immunol. 165 (1995) 7e11. [113] J.M. Young, L.M. De Young, Cutaneous models of inflammation for the evaluation of topical and systemic pharmacological agents. in: J. Spector, N. Back (Eds.), Pharmacological Methods in the Control of Inflammation. Alan R. Liss, New York, 1989, pp. 215e231.

5753

[114] M. Payá, M. Ferrándiz, M.J. Sanz, G. Bustos, R. Blasco, J.L. Ríos, M.J. Alcaraz, Study of the antioedema activity of some seaweed and sponge extracts from the Mediterranean coast in mice, Phytother. Res. 7 (1993) 159e162. [115] R.M. Giner, M.L. Villalba, M.C. Recio, S. Máñez, M. Cerdá-Nicolás, J.L. Ríos, Eur. J. Pharmacol. 389 (2000) 243e252. [116] L.M. De Young, J.B. Kheifets, S.J. Ballaron, J.M. Young, Edema and cell infiltration in the phorbol ester-treated mouse ear are temporally separate and can be differentially modulated by pharmacologic agents, Agents Actions 26 (1989) 335e341.