A study on quantitative structure–activity relationship and molecular docking of metalloproteinase inhibitors based on L-tyrosine scaffold
© Abbasi et al.; licensee BioMed Central. 2015
Received: 25 October 2014
Accepted: 12 April 2015
Published: 29 April 2015
MMP-2 enzyme is a kind of matrix metalloproteinases that digests the denatured collagens and gelatins. It is highly involved in the process of tumor invasion and has been considered as a promising target for cancer therapy. The structural requirements of an MMP-2 inhibitor are: (1) a functional group that binds the zinc ion, and (2) a functional group which interacts with the enzyme backbone and the side chains which undergo effective interactions with the enzyme subsites.
In the present study, a QSAR model was generated to screen new inhibitors of MMP-2 based on L–hydroxy tyrosine scaffold. Descriptors generation were done by Hyperchem 8, DRAGON and Gaussian98W programs. SPSS and MATLAB programs have been used for multiple linear regression (MLR) and genetic algorithm partial least squares (GA-PLS) analyses and for theoretical validation. Applicability domain of the model was performed to screen new compounds. The binding site potential of all inhibitors was verified by structure-based docking according to their binding energy and then the best inhibitors were selected.
The best QSAR models in MLR and GA-PLS were reported, with the square correlation coefficient for leave-one-out cross-validation (Q2 LOO) larger than 0.921 and 0.900 respectively. The created MLR and GA-PLS models indicated the importance of molecular size, degree of branching, flexibility, shape, three-dimensional coordination of different atoms in a molecule in inhibitory activities against MMP-2.
The docking study indicated that lipophilic and hydrogen bonding interactions among the inhibitors and the receptor are involved in a ligand-receptor interaction. The oxygen of carbonyl and sulfonyl groups is important for hydrogen bonds of ligand with Leu82 and Ala83. R2 and R3 substituents play a main role in hydrogen bonding interactions. R1 is sited in the hydrophobic pocket. Methylene group can help a ligand to be fitted in the lipophilic pocket, so two methylene groups are better than one. The Phenyl group can create a π-π interaction with Phe86.
The QSAR and docking analyses demonstrated to be helpful tools in the prediction of anti-cancer activities and a guide to the synthesis of new metalloproteinase inhibitors based on L-tyrosine scaffold.
KeywordsGA-PLS Metalloproteinase inhibitors MLR Molecular docking QSAR
The matrix metalloproteinases (MMPs) function predominantly as enzymes that degrade structural components of the extracellular matrix (ECM) [1-4]. MMPs play a substantial role in tumor progression and invasion of inflammatory cells. Among MMPs, MMP-2 easily digests the denatured collagens and gelatins [5,6]. It is highly involved in the process of tumor invasion and has been considered as a promising target for cancer therapy [3,7,8]. MMP-2 has a catalytic center) zinc (II) ion (and two hydrophobic domains (S1´ pocket and S1 pocket). S1´ pocket, the key domain of MMP-2, is deeper and narrower than other MMP subtypes and S1 pocket is a solvent exposure domain [3,9,10].
The structural requirements of an MMP-2 inhibitor are: (1) a functional group that binds the zinc ion (zinc-binding group; ZBG), capable of chelating the active site zinc ion; (2) a functional group which interacts with the backbone of enzyme; (3) side chains that undergo effective interactions with the enzyme subsites, such as S1´ pocket and S1 pocket [3,11,12].
Cheng et al. studied the L–hydroxy proline scaffold-based MMP-2 inhibitors in 2008 , and, in order to identify more potent MMP-2 inhibitors, replaced L-hydroxy proline with the L-tyrosine scaffold to form a new integrated structural pattern. They reported that the alteration in substitution pattern at R1, R2 and R3 positions alter MMP-2 inhibitory activity .
In 2012, 30 L–hydroxy tyrosine scaffold-based MMP-2 inhibitors were identified. It seems that finding a relationship between the structure of these compounds and their inhibitory activities in order to design structures with better activities and to predict their activity would be essential.
Quantitative structure activity relationships (QSARs), one of the most extensively used methods in chemometrics, and molecular docking are two of the helpful methods for drug design and prediction of drug activity [14-16]. QSAR models are mathematical equations which generate a communication between chemical structures and their biological activities, while molecular docking is done to specify the structural features that are important for interaction with a receptor.
In this report, we have performed a QSAR study and a molecular docking examination on 30 compounds of L-tyrosine derivatives which had been synthesized and evaluated as metalloproteinase (MMP-2) inhibitors .
Materials and methods
All calculations were implemented using an Intel Core i5 2.4 GHz processor, with the windows 7 operating system. Geometry optimization was done by Hyperchem 8.0 software. Descriptor generation was performed by Hyperchem 8.0, DRAGON package and Gaussian 98 W programs. SPSS software (version 11.5) and MATLAB software (version 7.12.0) were used for model creation and validation methods.
Activity data and descriptors generation
Chemical structures of L-tyrosine derivatives and their experimental and predicted activity by MLR and GA-PLS models
Experimental pIC 50
Predicted pIC 50
The output files of Hyperchem 8.0 software (.hin files) were transferred to DRAGON package to calculate four classes of descriptors (0D, 1D, 2D and 3D) including 28 constitutional descriptors, 10 functional groups, 18 atom-centered fragments, 216 topological descriptors, 15 molecular walk counts, 64 BCUT descriptors, 24 Galvestopol.charge indices, 96 2D autocorrelations, 14 charge descriptors, 41 Randic molecular profiles, 27 geometrical descriptors, 150 radial distribution function descriptors (RDF), 160 3D-MoRSE descriptors, 99 WHIM descriptors and 196 GETAWAY descriptors.
The z-matrix files of compounds were also provided by Hyperchem 8.0 program (.zmt files) and then were transferred as input files to the Gaussian98W program . Semi-empirical molecular orbital calculation by PM3 method was performed using Gaussian98W program. Different quantum chemical descriptors were obtained by this method including highest occupied molecular orbital energy (EHOMO), lowest unoccupied molecular orbital energy (ELUMO), and molecular dipole moment. Local charge was obtained by PM3 method in Gaussian98W. Hardness (η = 0.5 (EHOMO + ELUMO)), softness (S = 1/η), electronegativity (χ = 0.5 (EHOMO - ELUMO)) and electrophilicity (ω = χ2/2η) were calculated according to the method proposed by Thanikaivelan et al. .
Data processing and models building
All calculated descriptors were collected in a data matrix, D, the number on rows was representative of molecule numbers and the numbers on columns accounted for descriptors (30 × 1168). At the beginning, the columns which had constant and near constant values were removed from the original data matrix. Since collinear variables disrupt the models in MLR analysis, collinear descriptors needed to be detected and removed. The correlation of descriptors with one another and with activity data was then investigated. In the pairs with collinearity higher than 0.9, one which had the highest correlation versus the activity was retained and the rest were omitted. The number of total descriptors for each molecule reached 291.
The data set (30 compounds) was split into a calibration set and validation set. Validation subset was made of 20% of the total data (here, 6 biological activity data). An MLR analysis was performed by the stepwise regression SPSS (version 12.0) software and the model was built.
Since the data splitting has a considerable influence on the final selected model, a combined data splitting-feature selection (CDFS) strategy was employed . In the CDFS methodology, several subdivisions of calibration and validation set were made (10 times). In each case, the best model was chosen with a correlation coefficient higher than 0.95. The created models were validated by leave-one out (LOO) cross-validation method and Y-randomization test to investigate their predictability.
Molecular docking has become an increasingly main tool for drug discovery. Docking is a method which predicts the preferred orientation of one ligand when bound in an active site to form a stable complex. In this study, molecular docking of L-tyrosine derivatives with MMP-2 was studied by AutoDock 4.2 program  to find their binding site and the best direction based on the binding energy. Pdb file of MMP-2 was obtained from www.pdb.org (PDB: 3AYU) . All water molecules were deleted, polar hydrogens were added and the Kollman charge  was computed. A 120 x 120 x 120 Å size was chosen for the grid box, which covers the whole protein. In all of the ligands, the same as protein, polar hydrogens were added and the Gasteiger charge was computed . After preparation of ligand and protein files, the map files were created. Docking process with 50 runs and maximum number of evaluations 2500000 were performed. The final .dlg files were analyzed and the interaction between ligands and the active site of protein were studied. The ligand-protein interactions were analyzed and visualized by Discovery studio Ver. 3.
Result and discussion
Multiple linear regression analysis (MLR)
This study made use of an MLR analysis, as a simple regression method. The stepwise regression (using SPSS software) was also utilized to choose the most relevant set of descriptors for each type of the split data. The model coefficients were calculated using calibration data and then were used to predict the biological activity of validation samples. Several models were constructed by running a typical stepwise regression which is ranked based on calibration correlation coefficient (R2c). A model with calibration correlation coefficient higher than 0.9 was selected. The created model needed to be validated. The theoretical validation is generally classified into two groups: internal and external validation. Two of the internal validation methods include Leave-one out cross validation (LOO) and Y-randomization . The offered model was evaluated for both over-fitting and avoidance of chance correlations by the leave-one-out cross-validation (LOO) method and Y-randomization test respectively.
The best ten models were selected for future analysis in MLR
R 2 c a
R 2 p c
Q 2 LOO d
RMS CV e
IC1, RDF135e, Mor24m, RDF035u, E3u, RDF120m, Mor15e
IC1, Mor24m, Mor15e, R7e0, G3p
IC1, Mor24m, Mor15e, dipx, GATS8p, RDF065m
IC1, SP20, RDF115m, RDF115e, Mor28e
IC1, Mor24m, Mor15e, Mor09e, G2u, Mor27u
IC1, Mor24m, Mor15e, Mor09e, G2u, Mor27u
IC1, Mor28e, HATS3p, RDF120m, RDF115m
IC1, Mor28e, HATS3p, RDF115m, RDF120m, G3m
IC1, Mor24m, RDF135e, RDF035v, RDF135m, Mor26m
IC1, Mor26m, G3p, GATS8p, HATS7e
Brief description of the descriptors in ten models
Information content index (neighborhood symmetry of 1-order)
Radial Distribution Function −3.5/ weighted by atomic Sanderson electronegativities
3D-MoRSE – signal 24/ weighted by atomic masses
Radial Distribution Function −3.5/ unweighted
3rd component accessibility directional WHIM index/ unweighted
Radial Distribution Function −12.0/ weighted by atomic masses
3D-MoRSE – signal 15/ weighted by atomic Sanderson electronegativities
3st component symmetry directional WHIM index/weighted by atomic polarizabilities
R maximal autocorrelation of lag 7/weighted by atomic Sanderson electronegativities
Molecular dipole moment at X-direction
Geary autocorrelation -lag 8/ weighted by atomic polarizabilities
Radial Distribution Function −6.5/ weighted by atomic masses
Shape profile no. 20
3D-MoRSE – signal 28/ weighted by atomic Sanderson electronegativities
3D-MoRSE – signal 09/ weighted by atomic Sanderson electronegativities
2st component symmetry directional WHIM index/ unweighted
3D-MoRSE – signal 27/unweighted
Radial Distribution Function −11.5/ weighted by atomic masses
Radial Distribution Function −11.5 / weighted by atomic Sanderson electronegativities
Leverage- weighted autocorrelation of lag 3/ weighted by atomic polarizabilities
3st component symmetry directional WHIM index/ weighted by atomic masses
3D-MoRSE – signal 26/ weighted by atomic masses
The repetition of IC1 in all the ten models indicated that this descriptor has a main effect on L-tyrosine scaffold-based MMP-2 inhibitors. 3D-MoRSE descriptors also have significant effects. The quantum descriptors such as (dipx) that were seen only in one model have lower effects on MMP-2 inhibitors. The observation of the models, as listed in Table 2, revealed that there is a high degree of resemblance between the subset of descriptors (without considering the difference between the chosen descriptors in different models).
In this equation, N represents the number of molecules in the calibration set. R2 c and Q2 LOO are respectively the squared correlation coefficients for calibration and cross-validation. In addition, S.E is the standard error of calibration and RMSCV is the root mean square error of cross-validation. This equation has three common descriptors (IC1, Mor24m and Mor15e), with high calibration statistics and prediction ability.
Topological descriptors, such as IC1, explain the size, degree of branching, flexibility and overall shape. 3D-MoRSE descriptors explain three-dimensional coordination of the different atoms in a molecule. IC1, Mor24m and Mor15e have positive signs which indicate that pIC50 is directly related to these descriptors. The radial distribution function (RDF) descriptors are based on the distance distribution in the molecule. RDF115m has a negative sign which indicates that pIC50 is inversely related to this descriptor.
For the MMP-2 inhibition activity, a higher value of molecular size, degree of branching, flexibility, shape, three-dimensional coordination of the different atoms in a molecule (IC1) and a lower value of radial distribution function, RDF115m, are beneficial to the activity.
R 2 and Q 2 LOO values after ten Y-randomization tests
Q 2 LOO
Q 2 LOO
Genetic algorithm partial least squares (GA-PLS)
Since the variable selection method, that is the stepwise regression, may not be a suitable selection, variable selection methods such as the genetic algorithm, which demonstrate much better outcomes in comparison with the stepwise regression, are more favorable [25,26].
In genetic algorithm, if its corresponding descriptor is contained, a gene receives a value of 1 in the subset; otherwise, it takes a value of zero. The number of genes at each chromosome is equivalent to the number of descriptors. The population size varied between 80 and 125 for different GA runs. The number of genes with the values of 1 was relatively lower than the number of genes with the values of 0. The genes with the values of 1 were maintained. The chromosomes with the smallest numbers of chosen descriptors (total number of descriptors for each molecule reached 105) and the highest fitness were selected as the intended model. The predicted model was tested by leave-n-out cross-validation . A leave-one-out cross-validation was triggered and the value of Q2 LOO was obtained 0.850.
This equation has three MLR descriptors (IC1, Mor24m and Mor15e) with high calibration statistics and prediction ability. IC1 and Mor24m have a display positive effect and Mor15e and Mor32e have a display negative effect on inhibitory activity in GA-PLS. Unlike MLR, The radial distribution function (RDF) descriptors have no effect on inhibitory activity in GA-PLS.
As observed from Equation 2, the variables used in this equation can explain 93.2% of the variances in the inhibitory activity of the MMP-2 inhibitors used in this study.
The Y-randomization test was performed to evaluate the robustness of the created model in GA-PLS and 10 models were generated. All of the models were expected to have lower R2 and Q2 LOO values than the original GA-PLS model, as shown in Table 4.
Applicability domain of the model
Response outliers are compounds that have standard residual points higher than ± 2.0 standard deviation units and a leverage value higher than the warning leverage, which is 0.75 and 0.5 for MLR and GA-PLS respectively. As can be seen in Figure 2, all studied molecules in calibration and validation set lie with high degree of confidence in application domain of the developed models in both methods.
Molecular docking studies
The obtained binding energy by AutoDock
Binding energy (kcal/mol)
Binding energy (kcal/mol)
R1 is sited in hydrophobic pocket. Methylene group can help the ligand to be fitted in lipophilic pocket, so two Methylene groups are better than one. Phenyl group can create a π- π interaction with Phe86. Therefore, phenethyl (C6H5CH2CH2) is better than benzyl (C6H5CH2).
To design L-tyrosine based inhibitors and prediction of their activity against MMP-2 based on AutoDock 4.2 and QSAR studies, addition to a higher value of molecular size, degree of branching, flexibility, shape and three-dimensional coordination of the different atoms in a molecule, hydrophobicity and hydrophilicity of R1, R2 and R3 are highly important. When R1 is more hydrophobic, and R2, is more hydrophilic, there is a stronger inhibition against MMP-2.
In this study, quantitative relationships between molecular structure and inhibitory effect of L-tyrosine scaffold based MMP-2 inhibitors were investigated by MLR and GA-PLS. In MLR, A combined data splitting-feature selection method (CDFS) was offered to develop a quantitative structure–activity relationship model. It was found that topological parameter (IC1) has a main effect on the inhibitory activity of the compounds, among the different QSAR models. By this simple procedure, a multilinear 5-parametric model was created out of 22 descriptors. This method yielded acceptable results for the prediction set which was measured by cross-validation and Y-randomization. The findings indicate that the linear model produced by CDFS methodology could reproduce more than 92% of variances in the inhibitory activity. In addition to MLR, genetic algorithms, which demonstrated much better outcomes in comparison with stepwise regression, were used. By GA-PLS, the model with higher cross-validation statistics was created and the results indicated that IC1, Mor24m, Mor15e and Mor32e are main descriptors. It can be concluded from the two methods that higher values of molecular size, degree of branching, flexibility, shape and three-dimensional coordination of the different atoms in a molecule are particularly important
The docking study revealed that two hydrogen bonds between all of the inhibitors and the active site of MMP-2 (Leu82 and Ala83) are formed, as well as a π-π interaction between phenyl group and Phe86.
The information above could be used to design new inhibitors, and to show higher inhibitory activities and chemical synthesis of new inhibitors.
Quantitative structure activity relationship
Multiple linear regression
Combined data splitting-feature selection
- Q2 LOO :
Square correlation coefficient for leave-one-out cross-validation
- R2 c :
Calibration correlation coefficient
- R2 p :
Prediction correlation coefficient
standard error of calibration
- RMSCV :
root mean square error of cross-validation
radial distribution function descriptors
Genetic algorithm partial least squares
- Cheng X-C, Wang R-L, Dong Z-K, Li J, Li Y-Y, Li R-R. Design, synthesis and evaluation of novel metalloproteinase inhibitors based on L-tyrosine scaffold. Bioorg Med Chem. 2012;20:5738–44.View ArticlePubMedGoogle Scholar
- Lerchenberger M, Uhl B, Stark K, Zuchtriegel G, Eckart A, Miller M, et al. Matrix metalloproteinases modulate ameboid-like migration of neutrophils through inflamed interstitial tissue. Blood. 2013;122:770–80.View ArticlePubMed CentralPubMedGoogle Scholar
- Page-McCaw A, Ewald AJ, Werb Z. Matrix metalloproteinases and the regulation of tissue remodelling. Nat Rev Mol Cell Bio. 2007;8:221–33.View ArticleGoogle Scholar
- Sapna G, Gokul S, Bagri‐Manjrekar K. Matrix metalloproteinases and periodontal diseases. Oral Dis. 2013;20:538–50.View ArticlePubMedGoogle Scholar
- Bouloumié A, Sengenès C, Portolan G, Galitzky J, Lafontan M. Adipocyte produces matrix metalloproteinases 2 and 9 involvement in adipose differentiation. Diabetes. 2001;50:2080–6.View ArticlePubMedGoogle Scholar
- Visse R, Nagase H. Matrix metalloproteinases and tissue inhibitors of metalloproteinases structure, function, and biochemistry. Circ Res. 2003;92:827–39.View ArticlePubMedGoogle Scholar
- Hu Q, Yu L, Liao Q. Methylation of matrix metalloproteinases-2 promoter in endometrial cancer invasion. Beijing Da Xue Xue Bao. 2012;44:911–5.PubMedGoogle Scholar
- Wang Y, Hu C, Dong R, Huang X, Qiu H. Platelet-derived growth factor-D promotes ovarian cancer invasion by regulating matrix metalloproteinases 2 and 9. Asian Pac J Cancer Prev. 2011;12:3367–70.PubMedGoogle Scholar
- Maskos K. Crystal structures of MMPs in complex with physiological and pharmacological inhibitors. Biochimie. 2005;87:249–63.View ArticlePubMedGoogle Scholar
- Zucker S, Cao J. Selective matrix metalloproteinase (MMP) inhibitors in cancer therapy. Cancer Biol Ther. 2009;8:2371–3.View ArticlePubMed CentralPubMedGoogle Scholar
- Di Pizio A, Laghezza A, Tortorella P, Agamennone M. Probing the S1′ Site for the Identification of Non‐Zinc‐Binding MMP‐2 Inhibitors. Chem Med Chem. 2013;8:1475–82.View ArticlePubMedGoogle Scholar
- Wojtowicz-Praga SM, Dickson RB, Hawkins MJ. Matrix metalloproteinase inhibitors. Invest New Drugs. 1997;15:61–75.View ArticlePubMedGoogle Scholar
- Cheng X-C, Wang Q, Fang H, Tang W, Xu W-F. Design, synthesis and evaluation of novel sulfonyl pyrrolidine derivatives as matrix metalloproteinase inhibitors. Bioorg Med Chem. 2008;16:5398–404.View ArticlePubMedGoogle Scholar
- Hansch C, Hoekman D, Gao H. Comparative QSAR: toward a deeper understanding of chemicobiological interactions. Chem Rev. 1996;96:1045–76.View ArticlePubMedGoogle Scholar
- Hansch C, Kurup A, Garg R, Gao H. Chem-bioinformatics and QSAR: a review of QSAR lacking positive hydrophobic terms. Chem Rev. 2001;101:619–72.View ArticlePubMedGoogle Scholar
- Sabet R, Mohammadpour M, Sadeghi A, Fassihi A. QSAR study of isatin analogues as in vitro anti-cancer agents. Eur J Med Che. 2010;45:1113–8.View ArticleGoogle Scholar
- Frisch M, Trucks G, Schlegel H, Scuseria G, Robb M, Cheeseman J, et al. Gaussian 98, revision A. 7. Pittsburgh, Pa: Gaussian, Inc; 1998. p. 12.Google Scholar
- Thanikaivelan P, Subramanian V, Raghava Rao J, Unni Nair B. Application of quantum chemical descriptor in quantitative structure activity and structure property relationship. Chem Phys Lett. 2000;323:59–70.View ArticleGoogle Scholar
- Hemmateenejad B, Shamsipur M, Miri R, Elyasi M, Foroghinia F, Sharghi H. Linear and nonlinear quantitative structure–property relationship models for solubility of some anthraquinone, anthrone and xanthone derivatives in supercritical carbon dioxide. Anal Chim Acta. 2008;610:25–34.View ArticlePubMedGoogle Scholar
- Morris GM, Huey R, Lindstrom W, Sanner MF, Belew RK, Goodsell DS, et al. AutoDock4 and AutoDockTools4: Automated docking with selective receptor flexibility. J Comput Chem. 2009;30:2785–91.View ArticlePubMed CentralPubMedGoogle Scholar
- Hashimoto H, Takeuchi T, Komatsu K, Miyazaki K, Sato M, Higashi S. Structural basis for matrix metalloproteinase-2 (MMP-2)-selective inhibitory action of β-amyloid precursor protein-derived inhibitor. J Biol Chem. 2011;286:33236–43.View ArticlePubMed CentralPubMedGoogle Scholar
- Singh UC, Kollman PA. An approach to computing electrostatic charges for molecules. J Comput Chem. 1984;5:129–45.View ArticleGoogle Scholar
- Azizian H, Bahrami H, Pasalar P, Amanlou M. Molecular modeling of Helicobacter pylori arginase and the inhibitor coordination interactions. J Mol Graph Model. 2010;28:626–35.View ArticlePubMedGoogle Scholar
- Shahlaei M, Madadkar-Sobhani A, Saghaie L, Fassihi A. Application of an expert system based on Genetic Algorithm–Adaptive Neuro-Fuzzy Inference System (GA–ANFIS) in QSAR of cathepsin K inhibitors. Expert Syst Appl. 2012;39:6182–91.View ArticleGoogle Scholar
- Barros A, Rutledge D. Genetic algorithm applied to the selection of principal components. Chemometr Intell Lab Syst. 1998;40:65–81.View ArticleGoogle Scholar
- Liu S-S, Liu H-L, Yin C-S, Wang L-S. VSMP: a novel variable selection and modeling method based on the prediction. J Chem Inform Comput Sci. 2003;43:964–9.View ArticleGoogle Scholar
- Fernández M, Caballero J. Modeling of activity of cyclic urea HIV-1 protease inhibitors using regularized-artificial neural networks. Bioor Med Chem. 2006;14:280–94.View ArticleGoogle Scholar
- Tropsha A, Gramatica P, Gombar VK. The importance of being earnest: validation is the absolute essential for successful application and interpretation of QSPR models. QSAR Comb Sci. 2003;22:69–77.View ArticleGoogle Scholar
- Mansourian M, Fassihi A, Saghaie L, Madadkar-Sobhani A, Mahnam K, Abbasi M. QSAR and docking analysis of A2B adenosine receptor antagonists based on non-xanthine scaffold. Med Chem Res. 2015;24:394–407.View ArticleGoogle Scholar
- Feng Y, Likos JJ, Zhu L, Woodward H, Munie G, McDonald JJ, et al. Solution structure and backbone dynamics of the catalytic domain of matrix metalloproteinase-2 complexed with a hydroxamic acid inhibitor. BBA Proteins Proteom. 2002;1598:10–23.View ArticleGoogle Scholar
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.