Talanta 175 (2017) 264–272

Contents lists available at ScienceDirect

Talanta

journal homepage: www.elsevier.com/locate/talanta

fl Assessment of anti-in ammatory properties of extracts from MARK (Lonicera sp. L., ) by ATR-FTIR spectroscopy ⁎ R. Nikzad-Langerodia,d, , S. Ortmannb, E.M. Pferschy-Wenzigb, V. Bochkovb, Y.M. Zhaoc, J.H. Miaoc, J. Saukela, A. Ladurnera, E.H. Heissa, V.M. Dirscha, R. Bauerb, A.G. Atanasova,e a Department of Pharmacognosy, University of Vienna, Althanstrasse 9, 1090 Vienna, Austria b Institute of Pharmaceutical Sciences/Pharmacognosy, University of Graz, Graz, Austria c Guangxi Botanical Garden of Medicinal , 189 Changgang Road, Nanning, China d Department of Knowledge-Based Mathematical Systems, Johannes Kepler University Linz, Altenbergerstrasse 69, 4040 Linz, Austria e Institute of Genetics and Animal Breeding of the Polish Academy of Sciences, 05-552 Jastrzebiec, Poland

ARTICLE INFO ABSTRACT

Keywords: Inflammation is a hallmark of some of today's most life-threatening diseases such as arteriosclerosis, cancer, ATR-FTIR spectroscopy diabetes and Alzheimer's disease. Herbal medicines (HMs) are re-emerging resources in the fight against these Traditional Chinese medicine conditions and for many of them, anti-inflammatory activity has been demonstrated. However, several aspects fl Anti-in ammatory activity of HMs such as their multi-component character, natural variability and pharmacodynamic interactions (e.g. quality control synergism) hamper identification of their bioactive constituents and thus the development of appropriate Chemometrics quality control (QC) workflows. In this study, we investigated the potential use of Attenuated Total Reflectance Cell-based assays Fourier Transform Infrared (ATR-FTIR) spectroscopy as a tool to rapidly and non-destructively assess different anti-inflammatory properties of ethanolic extracts from various of the Genus Lonicera (Caprifoliaceae). Reference measurements for multivariate calibration comprised in vitro bioactivity of crude extracts towards four key players of inflammation: Nitric oxide (NO), interleukin 8 (IL-8), peroxisome proliferator-activated receptor βδ/ (PPAR βδ/ ), and nuclear factor kappa-light-chain-enhancer of activated B-cells (NF-κB). Multivariate analysis of variance (MANOVA) revealed a statistically significant, quantitative pattern-activity relationship between the extracts' ATR-FTIR spectra and their ability to modulate these targets in the corresponding cell models. Ensemble orthogonal partial least squares (OPLS) discriminant models were established for the identification of extracts exhibiting high and low activity with respect to their potential to suppress NO and IL-8 production. Predictions made on an independent test set revealed good generalizability of the models with overall sensitivity and specificity of 80% and 100%, respectively. Partial least squares (PLS) regression models were successfully established to predict the extracts' ability to suppress NO production and NF-κB activity with root mean squared errors of cross-validation (RMSECV) of 8.7% and 0.05-fold activity, respectively.

1. Introduction which can be isolated, characterized and routinely quantified by an appropriate analytical method [6]. More frequently however, bioactiv- Many of today's most life-threatening diseases such as arterio- ity is brought about by several structurally related molecules with the sclerosis, cancer, diabetes and Alzheimer's disease involve persistent same mode of action or by the simultaneous action of various inflammatory processes [1–3]. Anti-inflammatory activity has been molecules on different cellular targets [7] (which could lead e.g., to shown for a plethora of plant-derived products. Herbal medicines synergism). Furthermore, bioactive molecules might become unstable (HMs) thus represent a valuable resource in the fight against these upon isolation due to oxidative degradation or other chemical mod- conditions [4,5]. However, HMs are complex cocktails that pose unique ifications as matrix integrity is lost, hampering their quantitative challenges to drug discovery, analytical chemistry, quality control (QC), recovery and analysis [8]. Vibrational spectroscopy has been an product standardization and optimization. Ideally, bioactivity is di- attractive analytical tool for non-destructive material testing and is rectly associated with the presence and amount of a single substance particularly well suited for the analysis of HMs and dietary products in

⁎ Corresponding author at: Department of Pharmacognosy, University of Vienna, Althanstrasse 9, 1090 Vienna, Austria. E-mail address: [email protected] (R. Nikzad-Langerodi). http://dx.doi.org/10.1016/j.talanta.2017.07.045 Received 12 April 2017; Received in revised form 11 July 2017; Accepted 14 July 2017 Available online 18 July 2017 0039-9140/ © 2017 Elsevier B.V. All rights reserved. R. Nikzad-Langerodi et al. Talanta 175 (2017) 264–272 the QC context owing to the possibility to comprehensively map their [24], and MrBayes version 3.0b4 [25]. The individual bootstrap (BS) complex chemical makeup to a unique fingerprint [9]. Infrared (IR) consensus trees were visually inspected on a node to node basis to test and Raman spectroscopy have been successfully used in the past to the congruence among the individual DNA datasets by identifying authenticate HMs [10,11], detect adulterants [12,13] and to quantify contradictory nodes with >70% BS support. The collected Lonicera single substances [14] and substance classes [15,16] within plant samples could be assigned to the following 8 species: Lonicera matrices. Due to highly overlapping information present in spectro- japonica Thunberg (3 accessions, 1 comprising flower buds instead scopic data, method development usually involves a pattern recogni- of leaves), Lonicera hypoglauca Miquel (5 accessions), Lonicera tion step that acts as a filter to identify those parts of the spectrum that confusa Candolle (7 accessions), Lonicera macrantha (D.Don) are related to a given property (e.g. quantity of a substance). Sprengel (10 accessions), Lonicera similis Hemsley (4 accession), Establishing a calibration model then amounts to regressing the Lonicera acuminata Wallich (4 accessions), Lonicera reticulata property of interest against these patterns using a set of samples for Champion (2 accessions) and Lonicera bournei Hemsley (1 accession). which the property has been measured by a reference method. Typical One sample could not be identified. Voucher specimens of all plant reference methods that are widely used in QC are liquid- and gas samples are deposited in the Herbarium of the Guangxi Botanical chromatography. Alternatively, methods that directly link chemical Garden of Medicinal Plants. A list of the used plant samples can be composition to (bio-)chemical activity, like for instance anti-oxidant found in the supporting information (Table S-1). For the preparation of capacity, have been developed and proven useful when bioactivity is extracts, dried leaves were used for all samples except sample Nr. 35, mediated by multiple components with similar physicochemical prop- where flower buds of L. japonica were used for extraction. Extracts erties [17] (e.g. flavonoids). Several authors have used radical scaven- were prepared using 96% ethanol as extraction solvent by accelerated ging assays in combination with IR spectroscopy to build calibration solvent extraction (ASE), using an ASE 200 instrument (Dionex, models for direct determination of total anti-oxidant capacity, which is Sunnyvale, CA, USA). 1 g of dried, powdered plant material was mixed an important quality parameter of several dietary products. For a with pelletized diatomaceous earth (ASE Prep DE, Dionex) and filled concise overview on the topic, we refer to Lu et Rasco [17]. Although into the extraction cells. The following extraction parameters were cheap and reliable, such assays can neither capture synergistic effects used: pressure 68 bar; temperature 80°C; cell preheating 1 min; cell nor account for physiological aspects, e.g., related to uptake and heating 5 min; static extraction 5 min; flush 150%; purge 60 s; cycles 3. metabolism in vivo raising questions concerning their biological The extraction solvent was removed in a CentriVap Benchtop relevance [18]. Cell-based assays, on the other hand, continue being Centrifugal Vacuum Concentrator (Labconco, Kansas City, MO, USA) indispensable tools for drug discovery and are more physiologically under vacuum at a temperature of 40 °C. All extractions were relevant models for pharmacological activity studies, but their routine performed in duplicate except for sample Nr. 34 which could be use for monitoring the quality of HMs and dietary products in terms of extracted only once due to scarcity of plant material, resulting in a bioactivity/toxicity has been limited due to high costs [19]. In the total number of 71 extracts. The dried extracts were finally dissolved in present study, we investigated the application of different cell-based DMSO (at 20mgml−1), aliquoted, and kept at −20 °C until further use. assays for anti-inflammatory activity of crude medicinal plant extracts as a reference method for spectroscopic calibration. Four different cell 2.2. Cell-based in vitro assays models were used to quantitatively measure the anti-inflammatory activity of 71 extracts from different, closely related species of the 2.2.1. NF-kB transactivation assay Genus Lonicera L. (Caprifoliaceae) with respect to suppression of nitric The transactivation of a NF-κB-driven luciferase reporter in oxide (NO) production (in LPS-/IFN-γ-stimulated RAW 264.7 mouse HEK293/NF-κB-luc cells (Panomics, RC0014) was measured as pre- α macrophages), interleukin 8 (IL-8) expression (in TNF- /LPS stimu- viously described [26]. In brief, cells were cultured at 5% CO2 and lated endothelial HUVECtert cells) and nuclear factor kappa-light- 37 °C in cell culture incubators using Dulbecco's modified Eagle's chain-enhancer of activated B-cells (NF-κB) activity (in HEK/NF-κB- medium (DMEM; Lonza, Basel, Switzerland), supplemented with Luc cells) as well as activation of peroxisome proliferator-activated 2 mM glutamine, 100 Uml−1 benzylpenicillin, 100 μgml−1 streptomy- receptor βδ/ (PPAR βδ/ ; in HEK293 cells). The bioactivity data were cin, 100 μg ml−1 hygromycin B, and 10% fetal calf serum (FCS). On the then used to establish chemometric models for quantitative prediction day before the experiment, the cells were stained in serum-free of anti-inflammatory properties. medium supplemented with 2 μM Cell Tracker Green CMFDA (C2925; Invitrogen) for 1 h. Cells were then reseeded in 96-well plates 2. Material and methods in phenol red- and FCS-free DMEM overnight, and pre-treated with the respective extracts (50 μg ml−1) for 30 min prior to stimulation with 2.1. Plant material and extract preparation 2 ng ml−1 TNF-α for 4 h. An equal concentration of DMSO was used as control. After cell lysis the luminescence of the firefly luciferase and the Plant material was collected in Guangxi Province, China in fluorescence of the Cell Tracker Green CMFDA were quantified on a November 2011. In addition, we included one sample collected in the GeniosPro plate reader (Tecan, Grdig, Austria). The luciferase-derived botanical garden Graz (Austria) and one sample purchased in a Chinese signal from the NF-κB reporter was normalized by the Cell Tracker TCM herb market. All samples were authenticated by morphological Green CMFDA-derived fluorescence to account for differences in the analysis and by means of DNA barcoding. For the latter, fresh plant cell number. Extracts were tested in quadruplicates. material was dried with silica gel and total DNA was extracted using the modified CTAB protocol described in [20]. Standard polymerase chain 2.2.2. PPAR βδ/ luciferase reporter gene transactivation assay reaction (PCR) procedures were used to amplify the target DNA The transactivation of a PPAR βδ/ -driven luciferase reporter in regions, i.e. the two cpDNA loci matK and psbA-trnH. PCR products HEK-293 cells (ATCC, USA) was performed as previously described were purified using a gel extraction kit (UNIQ-10; Sangon, Shanghai, [27]. The cells were cultured in DMEM with phenol red, supplemented China). Sequencing reactions were carried out using the DYEnamic ET with 100 μg ml−1 streptomycin, 2 mM L-glutamine, 100 U ml−1 benzylpe- Dye Terminator Cycle Sequencing Kit (Amersham Pharmacia Biotech, nicillin, and 10% FBS. Cells were maintained in cell culturing incuba-

Piscataway, NJ, USA) following the manufacturer's protocols. tors at 37 °C and 5% CO2. For the experiments, the cells were seeded in Sequences were aligned using CLUSTALX version 1.83 [21], then 10 cm dishes, incubated for 18 h, and transfected by the calcium adjusted manually with BioEdit [22]. All DNA datasets were analysed phosphate precipitation method with 4 μg of the reporter plasmid using maximum parsimony (MP), maximum likelihood (ML), Bayesian (tk-PPREx3-luc; kindly supplied by Prof. Ronald M. Evans, Salk inference methods in PAUP version 4.0b10 [23], PhyML version 2.4.3 Institute for Biological Studies, San Diego, CA, USA) [28],4μg from

265 R. Nikzad-Langerodi et al. Talanta 175 (2017) 264–272 the human PPAR βδ/ expression plasmid (pSG5-hPPAR-beta, a kindly a VaCo5-II (Zibrus Technology, Bad Grund, Germany). Dried extracts supplied by Prof. Beatrice Desvergne and Prof. Walter Wahli, Center were re-dissolved in 500 μl water and the procedure repeated in order for Integrative Genomics, University of Lausanne, Switzerland) [29], to remove residual DMSO. Sample preparation was carried out in 96/ and 2 μg green fluorescent protein plasmid (pEGFP-N1, Clontech, 2000 μl deep-well plates (Eppendorf, Hamburg, Germany). ATR-FTIR Mountain View, CA) as internal control. After 6 h, the transfected cells spectra were acquired on a Tensor 27 (Bruker Optics, Ettlingen, were re-seeded in 96-well plates using DMEM without phenol red, Germany) equipped with a MIRacle™ ATR cell (Pike Technologies, supplemented with 100 μg ml−1 streptomycin, 100 U ml−1 benzylpenicil- Fitchburg, USA) and a DTGS detector. Dried extracts were mounted on lin, 2 mM L-glutamine, and 5% charcoal-stripped FBS. The cells were the ZnSe crystal of the ATR unit and spectra recorded with 32 scans further treated with the respective extracts (applied at 10 μg ml−1) or the between 600 cm−1 and 4000 cm−1 at 2 cm−1 resolution. A blank was solvent vehicle (DMSO) and incubated for 18 h. The medium was then recorded before every measurement and subtracted from the spectrum. discarded and the cells were lysed with a reporter lysis buffer (E3971, All spectra were baseline corrected (rubber band method) and vector Promega, Madison, USA). Luciferase activity of the cell lysates was normalized using built-in functions from OPUS software version 5.0 evaluated using a GeniosPro plate reader (Tecan, Groedig, Austria). (Bruker Optics, Ettlingen, Germany) prior to statistical analysis. The luminescence signals obtained from the luciferase activity mea- surements were normalized to the EGFP-derived fluorescence, to account for differences in the transfection efficiency or cell number. Extracts were tested in quadruplicates. 2.4. Statistical analysis 2.2.3. Inhibition of NO production in stimulated RAW 264.7 macrophages Agglomerative hierarchical cluster analysis was carried out in the − The measurement was performed as previously described [30]. 600–1800 cm 1 spectral range using Ward's algorithm along with the RAW 264.7 macrophages were stimulated with lipopolysaccharides Euclidean distance metric from Matlab 2014b (Mathworks, Natick, (LPS) from E. Coli 055:B5 (Sigma) and interferon-γ (IFNγ, mouse USA). Principle component analysis (PCA) and one-way multivariate recombinant E. Coli, Roche Diagnostics) for induction of iNOS gene analysis of variance (MANOVA) were carried out on auto scaled − expression. The effects on NO production were determined by photo- features (600–4000 cm 1) using the functions prcomp and manova metric quantification of nitrite accumulation in cell culture super- from R version 3.2.2 (R Foundation for Statistical Computing, Vienna, natants using the Griess assay: Nitrite levels were compared with a Austria Foundation for R Foundation for Statistical Computing, sodium nitrite standard curve after 16 h of incubation with the Vienna, Austriatatistical Computing, Vienna, Austria). The optimal respective sample as described by Baer et al. with slight modifications number of principle components (PCs) to retain was determined by 7- [31,32]. Activity is referred to nitrite accumulation of cells treated with fold cross-validation as described elsewhere [34] and the scores used as LPS/IFN-γ/DMSO (final concentration of 0.1% DMSO served as dependent variables in one-way MANOVA. Bioactivity was empirically solvent control). Extracts were evaluated at a final concentration of categorized into low, moderate and high activity, avoiding data-driven 50 μg dry extract ml−1. l-NMMA (N-monomethyl-L-arginine), a known cut-point optimization [35], and used as an independent variable to nonselective inhibitor of all NOS isoforms, was used as a positive explore if significant differences exist among the means of these control. Extracts were tested in at least 3 independent experiments, categories with respect to the PCs. The Pillai-Bartlett test was employed each done in duplicate. to assess the significance of between-group differences. Significance levels for pairwise comparisons were adjusted using Bonferroni 2.2.4. Interleukin 8 (IL-8) protein expression quantification correction. Orthogonal partial least squares discriminant analysis Interleukin 8 protein expression was quantified as previously (OPLS-DA) and partial least squares (PLS) regression were carried − described [4]. Telomerase reverse transcriptase technology (hTERT) out on autoscaled variables (600–4000 cm 1) using SIMCA-P 13.0 immortalized human umbilical vein endothelial cells (HUVECtert) [33] (Umetrics, Malmoe, Sweden). All models were optimized by making were cultured in M199 medium supplemented with 20% FCS (both extensive use of built-in model diagnostic facilities. High leverage from Sigma-Aldrich, St Louis, USA), cell growth supplement samples were consecutively excluded from a model if the T2 statistic (PromoCell, Germany), and antibiotics. The experiments were per- and the distance to the model plane (DmodX) were outside the formed in sixtuplicate in 96-well plates in M199 medium containing statistical limits at α =0.05. The process of model fitting and exclusion 5% serum. Plant extracts were tested at a concentration of 50 μg ml−1. of outliers was repeated until any further improvements in the Subconfluent HUVECtert cells were pretreated for 30 min with plant cumulative fraction of predicted variance (Q2) in the dependent extract or inhibitor as indicated, followed by stimulation with 10 ng ml−1 variable, determined by 7-fold cross-validation, could be achieved. A of TNF-α (PeproTech, Rocky Hill, USA) or 100 ng mL−1 of LPS (Sigma- model was considered significant if Q2 > 0.5 after model optimization. Aldrich, St. Louis, USA) for 6 h. Secreted IL-8 was determined in cell The predictive power of discriminant models was evaluated in terms of culture supernatants. IL-8 ELISA was performed using the Quantikine sensitivity (true positive rate) and specificity (true negative rate) of Human CXCL8/IL-8 Immunoassay Kit (R & D Systems, Minneapolis, predictions made on an independent set of test samples which were USA). Supernatants were transferred into 96-well plates (NALGE- randomly chosen prior any statistical treatment of the data and NUNC Int., Rochester, USA) coated with capturing antibody for IL-8 excluded before establishing predictive models. Predictions for which and developed with the respective detection antibody. Peroxidase either the augmented distance to the model (DModXPS+) was beyond activity was assessed with TMB 2-Component Microwell Peroxidase the critical distance (Dcrit) at α =0.05 and/or y <0.5 were considered Substrate Kit (KPL, Gaithersburg, USA), while the optical density (OD) outliers and were thus not included in the assessment of predictive was measured with a SynergyHT Multi-Detection Microplate Reader power. The predictive power of regression models was evaluated based (BioTek Instruments, Winooski, USA) at 450 nm using the OD at on the root mean squared error of cross-validation (RMSECV) and the 620 nm as reference. Extracts were tested in sextuplicates. predictions made on the independent test samples. In addition, the predictive ability of regression models was evaluated by the residual 2.3. ATR-FTIR spectroscopy prediction deviation (RPD) calculated as the ratio of standard deviation of the response in the calibration set and the root mean squared error All samples were obtained at a concentration of 20mgml−1 DMSO. of prediction 50 μl of extract was dissolved in 450 μl distilled water and subse- quently freeze-dried during 12 h under high vacuum (0.05 mbar) using

266 R. Nikzad-Langerodi et al. Talanta 175 (2017) 264–272

ffi 1 ncal 2 component spectra is in general di cult when further knowledge ∑i=1 (−)yyi ncal −1 about chemical composition is missing and thus beyond the scope of RPD = , this paper. Table 1 shows for each species the minimal, maximal and 1 ntest 2 ∑i=1 (−)lyyii average bioactivity across different accessions. In general, bioactivities ntest (1) varied strongly within individual species pointing to a high amount of where y , yl and yi denote mean, predicted and measured response natural variability in terms of chemical composition. This was also value, respectively and ncal and ntest denote the number of samples in reflected by large within-species variability in the extracts ATR-FTIR the calibration and test sets. The upper and lower bounds on the limit spectra which was assessed by means of exploratory data analysis (Fig. of detection (LOD) was calculated as previously described in Allegrini S-2/S-3). Highest activity towards suppression of NO production in et al. [36]: RAW 264.7 cells was observed for species from cluster I (L. japonica/L.

−2 −2 1/2 macrantha/L. hyploglauca/L. similis) with up to 90% inhibition over LODmin = 3.3[SEN var(xh ) +0 SEN var( xh ) +0 var( y )] (2) min min cal vehicle control while some of the lowest activities were observed for −2 −2 1/2 accessions from the same species (e.g. L. hypoglauca: 11%). Similar LODmax = 3.3[SEN var(xh ) +0max SEN var( xh ) +0max var( ycal )] (3) observations were made for IL-8, PPAR βδ/ , and NF-κB, although SEN denotes the sensitivity which is given by the reciprocal (euclidean) extracts displaying the highest bioactivity were distributed over repre- length of the PLS regression vector (i.e. 1/∥ b ∥). h0min and h0max denote sentatives from clusters I/II, I/III and I/II/III, respectively indicating minimal and maximal values of the zero-leverage (i.e. the estimated that spectral patterns related to bioactive constituents have a marginal leverage of a sample for which the target y=0) which are estimated contribution to the overall shape of the spectra. This observation from the calibration set according to suggests that anti-inflammatory activity is mediated by minor rather y than bulk constituents (e.g. chlorophyll), which is often the case for h = cal 0min N 2 plant extracts. In order to further support this hypothesis, cluster ∑ y (4) i=1 i analysis was undertaken at the level of individual instead of average and spectra to see if extracts with overall similar spectra also tend to exhibit fi ⎛ ⎡ ⎤⎞ similar bioactivity. However, this could not be con rmed (Fig. S-3, ⎛ y ⎞2 ⎜ ⎢ ⎜ cal ⎟ ⎥⎟ Table S-4). Altogether these results indicate that i) the in vitro anti- h0max =max⎜hhcal + 0min 1−⎟ , ⎢ ⎝ y ⎠ ⎥ fl ⎝ ⎣ cal ⎦⎠ (5) in ammatory activity is mediated by constituents present in the extracts at low concentration and ii) different species from the Genus where ycal is the mean target value and hcal holds the leverages of the Lonicera share structurally similar constituents exhibiting bioactivity calibration samples. Geometrically, h0min is the projection of the center towards the here tested pharmacological targets. This might partially of the scores space onto the zero-leverage plane and h0max is the explain why different Lonicera species are currently listed as equivalent projection from the calibration set yielding the largest zero-leverage. anti-inflammatory drugs in the Chinese Pharmacopoeia [38]. var(x) and var(ycal) in (2) denote the variance in spectroscopic measure- As a next step, we investigated whether patterns in the extracts' IR ments and reference (bioactivity) measurements, respectively. The spectra exist that can explain the quantitative differences in the former was estimated by averaging the variance over all technical, bioactivities they exhibit towards the four pharmacological targets. the latter by averaging over biological replicates. We therefore adopted a strategy described by Derenne et al. [39] employing principle component analysis (PCA) for pattern extraction 3. Results and discussion in combination with multivariate analysis of variance (MANOVA) to test the significance of spectral differences between empirically defined 3.1. Explorative analysis of ATR-FTIR spectra and pharmacological bioactivity classes (Table 2). Overall tests were significant for all four data targets, indicating that bioactivity is quantitatively manifested in the extracts' IR spectra (Table 3). Post-hoc tests revealed significant Our study includes several accessions of the three Lonicera species differences between samples exhibiting low/moderate and high activity that are listed in the Chinese Pharmacopoeia and are used in but not between samples exhibiting low and moderate activity towards Traditional Chinese medicine (TCM) for the treatment of conditions suppression of NO production (Table 3). Conversely, for PPAR βδ/ and involving toxic inflammation [37]: Thunberg, IL-8 pairwise differences were significant only between extracts Lonicera hypoglauca Miquel and Lonicera confusa Candolle, for exhibiting low and moderate activity. Sample size has a strong impact simplicity further referred to as L. japonica, L. hypoglauca, and L. on the outcome of statistical tests and by far fewer samples exhibited confusa, respectively. In addition, we included extracts from Lonicera high compared to low activity towards NO, IL-8 and PPAR βδ/ .We macrantha (D.Don) Sprengel, Lonicera similis Hemsley, Lonicera therefore tested if spectra from samples exhibiting low activity acuminata Wallich, Lonicera reticulata Champion and Lonicera significantly differed from those that exhibited moderate and high bournei Hemsley (L. macrantha, L. similis, L. acuminata, L. reticula- activity by pooling the latter two classes together. Finally, the opposite ta, L. bournei) that exhibited significant bioactivity in at least one of the scenario was also investigated, pooling together samples exhibiting test systems related to inflammation employed in this study (Table 1). moderate and low activity and testing them jointly against the high Fig. 1 shows the average spectra from the extracts of each species. activity class. Our data revealed significant differences for all three Cluster analysis of the extended fingerprint region (600–1800 cm−1) targets in the low vs. moderate/high activity scenario, while in the suggested following grouping of the species: L. macrantha, L. japonica, opposite scenario (high vs. low/moderate activity) all comparisons L. hypoglauca and L. similis constitute cluster I. The second cluster (II) were significant except for PPAR βδ/ . Here, the small number of highly comprise spectra from L. acuminata and L. reticulata both showing active extracts (n=4) was too low for reliable statistical testing characteristic sharp peaks at 1140 cm−1 and 1290 cm−1 and a double (Table 3). Finally, we found significant differences between spectra peak at 950 cm−1, the former being attributed to stretching vibrations from extracts that exhibited high and moderate activity towards NF-κB. of ester moieties, the latter originating from olefinic compounds. The third cluster (III), involving L. bournei and L. confusa, display the 3.2. Predictive modelling of bioactivity smoothest spectra with the carbonyl-/ester region around 1730 cm−1 and several weak absorptions between 1300 cm−1 and 1600 cm−1 being We next established predictive models for the discrimination of similar to custer II while patterns between 1100 cm−1 and 1300 cm−1 samples according to the significant scenarios in PCA-MANOVA using resembling spectra from group I. Visual interpretation of multi- orthogonal partial least squares discriminant analysis (OPLS-DA) and

267 R. Nikzad-Langerodi et al. Talanta 175 (2017) 264–272

Table 1 Summary statistics of in vitro bioactivity. Species listed in the Chinese Pharmacopoeia [35] are highlighted in bold. The number of accessions tested for bioactivity is indicated in brackets. The minimal, maximal and average bioactivity is shown for each species. The 3 highest bioactivities for each pharmacological target are highlighted in bold. Roman letters I-III indicate cluster membership according to Fig. 1.

Cluster Species Bioactivity

NO (% Inhibition) IL-8 (% Inhibition) PPARb/d (fold activity) NF-kB (fold activity)

Min Max Mean Min Max Mean Min Max Mean Min Max Mean

I L. japonica (6) 53 70 58 16 47 30 0.97 1.50 1.15 0.04 0.01 0.03 L. macrantha (10) 17 71 36 51 60 55 0.85 1.62 1.21 0.30 0.01 0.10 L. hypoglauca (10) 11 90 61 2 94 48 1.04 1.21 1.11 0.24 0.01 0.09 L. similis (8) 24 90 55 55 75 68 1.07 1.55 1.23 0.30 0.03 0.11

II L. acuminata (8) 30 67 44 23 59 36 0.84 1.28 1.10 0.15 0.06 0.09 L. reticulata (4) 30 40 35 45 83 64 0.85 1.09 0.98 0.58 0.02 0.29

III L. bournei (2) 31 46 39 64 67 66 0.96 0.99 0.98 0.50 0.42 0.46 L. confusa (14) 12 36 21 24 69 48 0.85 1.98 1.18 0.39 0.01 0.17

Table 4 Bioactivity of extracts used to test the predictive ability of chemometric models. The bioactivity class for each pharmacological target is indicated in brackets. I: Low activity. II: Moderate activity. III: High activity.

Test sample Species Bioactivity nr. NO (%) IL-8 (%) PPARb/d NF-kB (fold) (fold)

1 L. macrantha 53 (II) 70 (III) 2.04 (III) 0.06 (III) 2 L. macrantha 56 (II) 71 (III) 2.03 (III) 0.06 (III) 3 Not Identified 36 (I) 23 (I) 1.98 (III) 0.37 (II) 4 L. japonica 10 (I) 26 (I) 1.67 (III) 1.20 (I) (flower buds) 5 L. japonica 13 (I) 35 (I) 2.06 (III) 1.40 (I) (flower buds) Fig. 1. Mean ATR-FTIR spectra and cluster analysis. Species names corresponding to 6 L. japonica 27 (I) 6 (I) 1.52 (III) 0.14 (III) each spectrum are indicated. Dendrogram based on agglomerative hierarchical cluster 7 L. japonica 24 (I) 6 (I) 1.75 (III) 0.16 (II) analysis (Ward's algorithm) using the Euclidean distance metric is shown on the right. Roman letters I-III indicate putative clusters. 0.56) and low activity (Q2 = 0.73 and 0.62) towards NO and IL-8. Table 2 Inspection of the loadings vector of the discriminating latent Empirical bioactivity classes. The number of samples in each class is indicated in variables revealed distinct patterns underlying high activity towards brackets. NO and IL-8 suggesting involvement of different molecular species to the observed pharmacological effects (Fig. 3). Major differences Pharmacological target Bioactivity class encompass absorptions in the regions around 3250 − 4000 cm−1, High Moderate Low 2000 − 2500 cm−1 and within the fingerprint region (600 − 1500 cm−1) where assignment is difficult due to presence of overtone-, combina- NO (% Inhibition) >70(10) 50 − 70(12) <50(42) tion- and overlapping vibrations. Absorption in the first region can be IL−8 (% Inhibition) >65(14) 50 − 65(26) <50(24) PPARb/d (Fold Activity) >1.5(4) 1.25 − 1.5(8) <1.25(52) largely attributed to stretching vibrations of -OH (hydroxyl) groups −1 NO (Fold Activity) <0.15(46) 0.15 − 0.5(16) >0.5(2) while absorptions around 2000 − 2500 cm typically originate from unsaturated hydrocarbons. Both regions displayed high absolute load- ings on the discriminant vector for activity towards NO but not for IL- tested their performance on an independent set of test samples 8. Peaks within the carbonyl (C˭O) strechting vibration region (Table 4). Fig. 2 shows the scores plots of discriminant OPLS models (1650 − 1750 cm−1) displayed high absolute loadings for both bioactiv- for which good generalization was expected based on cross-validation ities indicating presence of either carbonyl, carboxylic acid and/or ester statistics (i.e. Q2 > 0.5). Good generalizability was found for models moieties in the underlying active principle. tailored to the identification of extracts exhibiting high (Q2 = 0.66 and

Table 3 Multivariate analysis of variance (MANOVA) of bioactivity classes. 8 principle components were considered. p <0.001 (***), p <0.01 (**), p <0.05 (*). n.s.: Not significant. Scenarios in which the number of samples in one class was smaller than 5 were not considered.

Pharmacological target Significance

Global Low/Moderate Low/High High/Moderate Low/ (Moderate + High) High/ (Moderate + Low)

NO *** n.s. *** * *** *** IL−8 * * n.s. n.s. ** * PPARb/d ** ** n.s. n.s. *** – NF-kB * ––* – n.s.

268 R. Nikzad-Langerodi et al. Talanta 175 (2017) 264–272

Fig. 2. OPLS discriminant models for the identification of extracts exhibiting high and low activity towards NO and IL-8. tp[1]: First predictive component. to[1]: First orthogonal component. The number of samples in each model (n), cumulative explained (R2) and predicted variance (Q2) computed from 7-fold cross-validation are indicated in each plot. Ellipses indicate 95% confidence regions.

found good predictive performance underpinning the usefulness of OPLS-DA models in a screening context. Notably, both test samples exhibiting moderate activity towards NO (samples 1 and 2, both L. macrantha) were beyond the model scope (i.e. outliers) in the high vs. moderate/low activity discriminant model while the orthogonal dis- tance (DModXPS+) of 1 to the opposite model (low vs. moderate/high activity) was within the acceptable region and its bioactivity class was correctly predicted, underpinning the benefit of taking predictions from both models into consideration when analysing new samples. A similar observation was made for predictions of bioactivity towards IL- 8: In the high vs. moderate/low activity model, samples 3, 4, 5 and 6 were correctly identified as not being highly active (i.e. true negatives) while 1, 2 (both highly active) and 7 were outliers. Samples 1 and 2 were outliers also in the opposite model (i.e. low vs. moderate/high activity), whereas 7 was correctly predicted to exhibit low activity. The fact that samples 1 and 2 are outliers to both models indicates that their spectra contain unmodelled information and that the composition of these extracts might thus differ considerably from the composition of the samples in the training set. Finally, considering predictions from Fig. 3. Loadings of the first discriminant component of OPLS-DA models for the both models, test samples 5 and 6 were classified to neither the high discrimination of high vs. moderate/low activity samples with respect to IL-8 and NO. nor the low activity class providing valuable information about their bioactivities (i.e. that they do not exhibit high activity) in spite of the In order to validate the discriminant models, we investigated the fact that both samples were misclassified (false negatives) in the low vs. predictions on a set of independent test samples (Table 5). Overall, we moderate/high activity model (Table 5). Test samples 6 and 7 were

Table 5 Discriminant OPLS test set predictions for bioactivity towards NO and IL-8. TP: True positives, FP: False positives, TN: True negatives. FN: False negatives. Outliers: Test samples for which either DModXPS + > Dcrit (α=0.05) and/or predicted y-values were <0.5 for both classes. Sensitivity=TP/(TP + FN). Specificity=TN/(TN + FP).

Bioactivity Scenario #TP #FP #TN #FN Outliers Sensitivity Specificity

NO High vs. Moderate/Low 0 0 5 0 2 – 100% Low vs. Moderate/High 5 0 1 0 1 100% 100%

IL-8 High vs. Moderate/Low 0 0 4 0 3 – 100% Low vs. Moderate/High 3 0 0 2 2 60% –

Total 8 0 10 2 8 80% 100%

269 R. Nikzad-Langerodi et al. Talanta 175 (2017) 264–272 prepared from L. japonica collected in Austria whereas all samples Table 6 from the training set were prepared from accession originating from Summary statistics for PLS regression models. Guangxi, China (Table S-1). Yet, all models correctly predicted that Bioactivity RMSEP RPD LOD /LOD these samples do not exhibit high activity with respect to suppression min max of NO production and induction of IL-8 expression. Furthermore, we NO 12.7 1.54 18.9/23.0 have validated our models using samples that were prepared from NF-κB 0.58 0.13 0.56/0.54 flower buds instead of leaves (test samples 4 and 5) to investigate the effect of ”incorrect” sample preparation which is a frequent issue in the QC of HMs. Interestingly, the orthogonal distance of both samples was within the 95% tolerance region for all models indicating that the extracts share similar main constituents with leaf extracts. Finally, bioactivity was correctly predicted for test sample 3 which was prepared from an unidentified accession. In contrast to activity towards NO and IL-8, no meaningful discriminant models could be established for any of the significant scenarios for PPAR βδ/ and NF-κB activity indicating that the spectra do not support robust separation of the empirically defined bioactivity classes. However, treating in vitro bioactivity as a dichotomous feature is not optimal since subsumption of within-class variability leads to considerable loss of information [38]. We therefore also investigated the possibility to treat bioactivity as a continuous feature and modelled the data using PLS regression. Good generalization properties were obtained for models that predicted bioactivity towards NO and NF-κB with Q2 statistics of 0.79 and 0.66, respectively (Fig. 4). The limits of detection (LOD) estimated from the calibration samples were in the range of 18.9 − 23.0% suppression for the former and 0.56 − 0.54-fold activity (corresponding to 44 − 46% inhibition) for the latter (Table 6). In contrast, any attempt to establish predictive models for activity towards IL-8 and PPAR βδ/ yielded no Fig. 5. Regression vectors of PLS models for prediction of activity towards NO and NF- significant latent variable (LV) indicating that the spectra do not κB. contain enough structure to enable accurate prediction of bioactivity. Fig. 5 shows the regression coefficients of PLS models for the below the upper bound of the detection limit of the calibration model, prediction of suppression of NO production and inhibition of NF-κB whereas bioactivity of sample 7 was correctly predicted to be above the activity. For activity towards NO, high absolute beta coefficients were LOD. The residual prediction deviation (RPD) was >1.4 which further observed in the hydroxyl- and saturated hydrocarbon regions which is indicates acceptable quality of the model. On the other hand, bioactiv- in good agreement with the fact that the same regions showed strong ity towards NF-κB was accurately predicted for samples 1, 2, 6 and 7 loading on the discriminant vector between high and low/moderate whereas poor predictions were obtained for test samples 3, 4 and 5. activity samples. Beta coefficient plots further indicated contributions For the latter two, the distances to the model plane (DModXPS+) were from the carbonyl region to both NO and NF-κB activity. Besides the inside the acceptable region at α =0.05. However, their actual bioac- small contributions from the hydroxyl- and saturated hydrocarbon tivities (1.2 and 1.4-fold activation) are far beyond the range of region to NF-κB activity, the broad peak around 1200 − 1250 cm−1 seen reference values from the training samples indicating that the model for the NO model is absent in the NF-κB regression vector. This peak does not extrapolate well to opposite bioactivities, which seems might originate either from C-O vibrations within ester moieties or intuitive. Despite a RMSEP around the detection limit and a poor from alcoholic C-H vibrations, however assignments within the finger- RPD, the model has still proven useful to decide whether an extract print region remain per se highly speculative. from the test set will exhibit high activity or not with the only exception Test set predictions of bioactivity towards NO were surprisingly being test sample 3 for which the poor prediction quality could not be accurate for samples 3-7, which were all correctly predicted to exhibit anticipated from its outlyingness with respect to the training set. low activity. 2 was correctly identified as moderate activity sample, Altogether, we found that multivariate classification and regression whereas sample 1 was slightly overestimated by the model. The techniques allowed modelling the extracts ATR-FTIR spectra for the measured as well as the predicted bioactivity of samples 4 and 5 was prediction of bioactivity towards NO, IL-8, and NF-κB. In contrast, no

Fig. 4. PLS regression models for NO and NF-κB activity. Blue circles: Training samples. Red circles: Test samples. The root mean squared error of cross-validation (RMSECV), R2 and Q2 statistics, the number of latent variables (LV's) included in the model and the number of samples (n) in the training set are indicated in each plot. Blue region: Low activity. Green region: Moderate activity. Red region: High activity (according to Table 2).

270 R. Nikzad-Langerodi et al. Talanta 175 (2017) 264–272 useful model could be established for PPAR βδ/ activity despite 115 (5) (2005) 1111–1119. ff fl fi ff [3] F.L. Heppner, R.M. Ransoho , B. Becher, Immune attack: the role of in ammation signi cant di erences between low and moderate activity and low in Alzheimer disease, Nat. Rev. Neurosci. 16 (6) (2015) 358–372. and moderate + high activity extracts found with MANOVA (Table 3). [4] S. Vogl, A.G. Atanasov, M. Binder, M. Bulusu, M. Zehl, N. Fakhrudin, E.H. Heiss, PCA-MANOVA turned out as a useful tool for explorative analysis of P. Picker, C. Wawrosch, J. Saukel, G. Reznicek, E. Urban, V. Bochkov, V.M. Dirsch, B. Kopp, The herbal drug Melampyrum pratense L. (Koch): the role of inflamma- spectral and pharmacological data allowing to identify those empirical tion in Alzheimer disease, Evid.-Based Complement. Altern. Med. 2013 (4) (2013) cut-points where bioactivity classes differed significantly. Non-signifi- 1–10. cant dichotomies yielded no meaningful OPLS-DA models. However, [5] B. Waltenberger, A.G. Atanasov, E.H. Heiss, D. Bernhard, J.M. Rollinger, low p-values did not necessarily imply good class separability or J.M. Breuss, D. Schuster, R. Bauer, B. Kopp, C. Franz, V. Bochkov, M.D. Mihovilovic, V.M. Dirsch, H. Stuppner, Drugs from nature targeting generalizability of discriminant models. This is in good agreement inflammation (DNTI): a successful Austrian interdisciplinary network project, with the fact that p-values are no direct measure of class dissimilarity Mon. für Chem. - Chem. Mon. 147 (3) (2016) 479–491. in MANOVA. It is widely accepted that dichotomization can cause [6] S. Li, Q. Han, C. Qiao, J. Song, C.L. Cheng, H. Xu, Chemical markers for the quality control of herbal medicines: an overview, Chin. Med. 3 (1) (2008) 7. serious problems in data analysis, however as outlined by Shentu and [7] F.R. Stermitz, P. Lorenz, J.N. Tawara, L.A. Zenewicz, K. Lewis, Synergy in a Xie transforming continuous response variables into categorical ones medicinal plant: Antimicrobial action of berberine potentiated by 5-methoxyhyd- can be advantageous in case of measurement uncertainty which is a nocarpin, a multidrug pump inhibitor, Proc. Natl. Acad. Sci. USA 97 (4) (2000) 1433–1437. well known issue with biological assays [38]. We found that dichot- [8] S.K. Oettl, J. Hubert, J.-M. Nuzillard, H. Stuppner, J.-H. Renault, J.M. Rollinger, omization was a prerequisite to obtain predictive models for activity Dereplication of depsides from the lichen Pseudevernia furfuracea by centrifugal towards IL-8 with acceptable generalization properties, whereas partition chromatography combined to 13C nuclear magnetic resonance pattern – κ fi recognition, Anal. Chim. Acta 846 (2014) 60 67. dichotomization of NF- B activity yielded no signi cant discriminant [9] S. Sun, J. Chen, Q. Zhou, G. Lu, K. Chan, Application of mid-infrared spectroscopy models. Therefore we conclude that both, regression and classification in the quality control of traditional chinese medicines, Planta Med. 76 (17) (2010) approaches should be considered when modelling IR spectra with 1987–1996. [10] M. Laasonen, T. Harmia-Pulkkinen, C.L. Simard, E. Michiels, M. Räsänen, bioactivity related data. Finally, we found that combining several H. Vuorela, Fast identification of EchinaceapurpureaDried roots using near- models significantly increased the scope of individual discriminant infrared spectroscopy, Anal. Chem. 74 (11) (2002) 2493–2499. models in terms of the number of test samples for which the orthogonal [11] J. Mao, J. Xu, Discrimination of herbal medicines by molecular spectroscopy and distance (DModXPS+) was within the tolerance region. This result chemical pattern recognition, Spectrochim. Acta Part A, Mol. Biomol. Spectrosc. 65 (2) (2006) 497–500. suggests that building an ensemble of models using different cut-points [12] W.-J. Dong, Y.-N. Ni, S. Kokot, Quantitative analysis of two adulterants in is a promising dichotomization strategy when the approach of model- Cynanchum stauntonii by near-infrared spectroscopy combined with multi-variate – ling a continuous response fails. Therefore, this strategy should be calibrations, Chem. Pap. 66 (12) (2012) 1 9. [13] V.A. Huck-Pezzei, L.K. Bittner, J.D. Pallua, H. Sonderegger, G. Abel, M. Popp, subject to further investigations in the future. G.K. Bonn, C.W. Huck, A chromatographic and spectroscopic analytical platform Taken together the present work has demonstrated the potential for the characterization of St John's wort extract adulterations, Anal. Methods 5 (3) use of ATR-FTIR spectroscopy for the quantitative assessment of in (2013) 616. fi fi fl [14] H. Schulz, M. Baranska, Identi cation and quanti cation of valuable plant vitro anti-in ammatory activity of Lonicera. In particular, our data i) substances by IR and Raman spectroscopy, Vib. Spectrosc. 43 (1) (2007) 13–25. revealed a significant quantitative pattern-activity relationship between [15] C.-O. Chan, C.-C. Chu, D.K.-W. Mok, F.-T. Chau, Analysis of berberine and total the here investigated extracts IR spectra and their in vitro bioactivities, alkaloid content in Cortex Phellodendri by near infrared spectroscopy (NIRS) compared with high-performance liquid chromatography coupled with ultra-visible ii) yielded chemometric models tailored to predict bioactivity towards spectrometric detection, Anal. Chim. Acta 592 (2) (2007) 121–131. NO, IL-8 and NF-κB with reasonable accuracy, and iii) allowed reliable [16] M.N. Escamilla, F.R. Sanz, H. Li, S.A. Schönbichler, B. Yang, G.K. Bonn, assessment of bioactivity of extracts prepared from accessions from C.W. Huck, Rapid determination of baicalin and total baicalein content in ff Scutellariae radix by ATR-IR and NIR spectroscopy, Talanta (2013) 1–7. di erent origins (i.e. China, Austria), species (i.e. L. japonica, L. [17] X. Lu, B.A. Rasco, Determination of antioxidant content and antioxidant activity in macrantha) and even plant organs (i.e. flower buds, leaves). The latter foods using infrared: spectroscopy and chemometrics: a Review, Crit. Rev. Food point underpins that the here presented approach was capable of Sci. Nutr. 52 (10) (2012) 853–875. capturing the pharmacological activity-related spectral patterns in a [18] K.L. Wolfe, R.H. Liu, Cellular Antioxidant Activity (CAA) assay for assessing antioxidants, foods, and dietary supplements, J. Agric. Food Chem. 55 (22) (2007) wide array of biologically diverse samples which goes beyond previous 8896–8907. studies where metabolomic fingerprint data has been exploited for the [19] P. Banerjee, B. Franz, A.K. Bhunia, Mammalian cell-based sensor system, Adv. – prediction of bioactivity of samples from a single species only [40]. Biochem. Eng./Biotechnol. 117 (Chapter 21) (2010) 21 55. [20] J.J. Doyle, J.L. Doyle, A rapid DNA isolation procedure for small quantities of fresh Finally, we want to point out that although correlation between spectral leaf tissue, Phytochem. Bull. 19 (1987) 11–15. patterns and bioactivity does not necessarily imply a direct, causal [21] J.D. Thompson, T.J. Gibson, F. Plewniak, F. Jeanmougin, D. Higgins, The relationship they possibly provide valuable information that might help CLUSTAL_X windows interface: flexible strategies for multiple sequence alignment aided by quality analysis tools, Nucleic Acids Res. 25 (24) (1997) 4876–4882. guiding discovery of the underlying active principle(s). In light of the [22] T.A. Hall, BioEdit: a user-friendly biological sequence alignment editor and analysis here presented results we conclude that ATR-FTIR in combination with program for Windows 95/98/NT, Nucleic Acids Symp. Ser. 41 (1999) 95–98. pharmacological assays holds promise for the assessment of HMs [23] D.L. Swofford, PAUP*: Phylogenetic Analysis Using Parsimony (and other meth- fi ods) 4.0.b5. beyond mere authentication and quanti cation of marker compounds. [24] S. Guindon, O. Gascuel, A simple, fast, and accurate algorithm to estimate large phylogenies by maximum likelihood, Syst. Biol. 52 (5) (2003) 696. Acknowledgements [25] F. Ronquist, J.P. Huelsenbeck, MrBayes 3: Bayesian phylogenetic inference under mixed models, Bioinformatics 19 (12) (2003) 1572. [26] E.-M. Pferschy-Wenzig, A.G. Atanasov, C. Malainer, S.M. Noha, O. Kunert, The authors acknowledge the support by the Austrian Science Fund D. Schuster, E.H. Heiss, N.H. Oberlies, H. Wagner, R. Bauer, V.M. Dirsch, (FWF) grant S107 “Drugs from Nature Targeting Inflammation” and Identification of isosilybin a from milk thistle seeds as an agonist of peroxisome – the FWF project P25971-B23. proliferator-activated receptor gamma, J. Nat. Prod. 77 (4) (2014) 842 847. [27] A.G. Atanasov, M. Blunder, N. Fakhrudin, X. Liu, S.M. Noha, C. Malainer, M.P. Kramer, A. Cocic, O. Kunert, A. Schinkovitz, E.H. Heiss, D. Schuster, Appendix A. Supplementary data V.M. Dirsch, R. Bauer, Polyacetylenes from Notopterygium incisum-new selective partial agonists of peroxisome proliferator-activated receptor-gamma, PLoS One 8 (4) (2013) e61755. Supplementary data associated with this article can be found in the [28] S.A. Kliewer, K. Umesono, D.J. Noonan, R.A. Heyman, R.M. Evans, Convergence of online version at http://dx.doi.org/10.1016/j.talanta.2017.07.045. 9-cis retinoic acid and peroxisome proliferator signalling pathways through heterodimer formation of their receptors, Nature 358 (6389) (1992) 771–774. [29] G. Lazennec, L. Canaple, D. Saugy, W. Wahli, Activation of peroxisome proliferator- References activated receptors (PPARs) by their ligands and protein Kinase A activators, Mol. Endocrinol. 14 (12) (2009) 1962–1975. [1] U. Weiss, Inflammation, Nature 454 (7203) (2008) 427. [30] X. Liu, O. Kunert, M. Blunder, N. Fakhrudin, S.M. Noha, C. Malainer, [2] K.E. Wellen, G.S. Hotamisligil, Inflammation, stress, and diabetes, J. Clin. Investig. A. Schinkovitz, E.H. Heiss, A.G. Atanasov, M. Kollroser, D. Schuster, V.M. Dirsch,

271 R. Nikzad-Langerodi et al. Talanta 175 (2017) 264–272

R. Bauer, Polyyne hybrid compounds from notopterygium incisum with peroxisome categorizing or dichotomizing continuous variables is inadvisable: an example from proliferator-activated receptor gamma agonistic effects, J. Nat. Prod. 77 (11) the natural history of unruptured aneurysms, AJNR. Am. J. Neuroradiol. 32 (3) (2014) 2513–2521. (2011) 437–440. [31] H.P. Baer, K. Schmidt, B. Mayer, W.R. Kukovetz, Pentamidine does not interfere [36] F. Allegrini, A.C. Olivieri, Iupac-consistent approach to the limit of detection in with nitrite formation in activated RAW 264.7 macrophages but inhibits consti- partial least-squares calibration, Anal. Chem. 86 (15) (2014) 7858–7866. http:// tutive brain nitric oxide synthase, Life Sci. 57 (21) (1995) 1973–1980. dx.doi.org/10.1021/ac501786u (pMID: 25008998). [32] V.B. Konkimalla, M. Blunder, R. Bauer, T. Efferth, Inhibition of inducible nitric [37] D. Bensky, S. Clavey, E. Stöger, Chinese Herbal Medicine, Materia Medica, 2004. oxide synthase by bis(helenalinyl)glutarate in RAW264.7 macrophages, Biochem. [38] C.P. Commission, Chinese Pharmacopoeia, Chemical industry press, 2005. Pharmacol. 79 (11) (2010) 1573–1580. [39] A. Derenne, V. Van Hemelryck, D. Lamoral-Theys, R. Kiss, E. Goormaghtigh, [33] H.B. Schiller, A. Szekeres, B.R. Binder, H. Stockinger, V. Leksa, Mannose 6- Biochimica et Biophysica Acta, BBA - Molecular Basis of Disease1832 (1) (2013) phosphate/insulin-like growth factor 2 receptor limits cell invasion by controlling pp. 46–56. v3 integrin expression and proteolytic processing of urokinase-type plasminogen [40] N.J.C. Bailey, Y. Wang, J. Sampson, W. Davis, I. Whitcombe, P.J. Hylands, activator receptor, Mol. Biol. Cell 20 (3) (2009) 745–756. S.L. Croft, E. Holmes, Prediction of anti-plasmodial activity of Artemisia annua [34] P. Eshghi, Dimensionality choice in principal components analysis via cross- extracts: application of NMR spectroscopy and chemometrics, J. Pharm. Biomed. validatory methods, Chemom. Intell. Lab. Syst. 130 (C) (2014) 6–13. Anal. 35 (1) (2004) 117–126. [35] O. Naggara, J. Raymond, F. Guilbert, D. Roy, A. Weill, D.G. Altman, Analysis by

272