Reagents and materials
Acetonitrile (HPLC grade) and ethanol (HPLC grade) were purchased from Merck (Darmstadt, Germany). 2-NBDG probe was purchased from Selleck Chemicals (Houston, USA). Formic acid (MS grade) was from Sigma-Aldrich (St. Louis, MO, USA). The reference compounds of coptisine (≥ 98.0%), berberine (≥ 98.0%), berberubine (≥ 98.0%), epiberberine (≥ 98.0%), chlorogenic acid (≥ 98.0%), palmatine, jatrorrhizine (≥ 98.0%), demethyleneberberine (≥ 98.0%), 5-O-caffeoylquinic acid (≥ 98.0%), 3, 5-O-dicaffeoylquinic acid (≥ 98.0%) were obtained from Chengmust Biological Technology Co., Ltd. (Sichuan, China). Ultra-pure water was produced by a Milli-Q water purification system (Milford, MA, USA).
Preparation of sample solutions
Different batches of three herbal samples were collected from herbal medicine market in China, including 6 batches of Coptis chinensis, 5 batches of Astragalus membranaceus, and 6 batches of Lonicera japonica. All the samples were authenticated by Prof. Peng Li and deposited at State Key Laboratory of Quality Research in Chinese Medicine, University of Macau (Macau, China).
The samples were homogenized and sieved through a No. 40 mesh. Total of 26 JQJT samples were prepared by combining the powders of three herbal medicines at a ratio of 10.3:15.4:61.8 (Coptis chinensis: Astragalus membranaceus: Lonicera japonica). The JQJT samples were extracted by ultra-sounded for 60 min with 10%, 50% and 90% ethanol, respectively. The extracts were centrifuged, and the supernatant was filtrated through a 0.22 µm filter (Millipore, USA). An aliquot of 10 μL resulted filtrate was subject to UPLC-LTQ-Orbitrap analysis.
Untargeted metabolomic profiling
The chromatographic separation was performed on a C18 column (4.6 mm × 250 mm, 5 µm, Shiseido Co., Ltd., Tokyo, Japan) using Dionex UltiMate 3000 UPLC system (Thermo Scientific, SanJose, CA, USA). The column was maintained at 27 °C and eluted with mobile phase consisting of 0.1% formic acid water (A) and acetonitrile (B) under the following gradient program: 8%-12% B (0-5 min), 12-18% B (5-10 min), 18-22% B (10-15 min), 22-25% B (15-20 min), 25-32% B (20-35 min), 32-65% B (35-50 min), 65-95% B (50-60 min) at a flow rate of 1.0 mL/min. The column was equilibrated for an additional 5 min at 8% B after a gradient run. And 25% of column effluent was introduced into the ESI source via a post-column flow splitter (Analytical Scientific Instruments, CA, USA).
Mass spectrometry was performed on an LTQ-Orbitrap system (Thermo Fisher Scientific, Bremen, Germany) equipped with a heated electrospray ionization (ESI) source operating in the positive ionization mode. The key operating parameters of MS were as follows: spray voltage of 3.2 kV, sheath gas of 20 (arbitrary units); capillary temperature of 350 °C; auxiliary gas of 10 (arbitrary units), sweep gas of 2 (arbitrary units), and capillary voltage of 25 V. Full scan mass spectra were acquired in the mass range of m/z 200 to 1500 with a resolving power of 30,000. Data-dependent MS/MS fragmentation was performed at a mass resolution of 15,000. Dynamic exclusion was used to avoid repeated MS/MS analysis with the exclusion time of 30 s. The collision-induced dissociation (CID) was used with normalized collision energy of 35 %. The data acquiring, and processing was performed using Thermo Xcalibur 2.1 (Thermo Fisher Scientific) workstation.
Metabolomics data processing
Raw data (.raw format) acquired with Xcalibur workstation were converted to the mzXML data using the MSConvert software. XCMS online, an open-source deconvolution software, was employed to pretreat the obtained mzXML [20]. After peak extracting, filtering and alignment, the dataset containing the integrated peak intensity, m/z, and retention time was obtained. MetaboAnalyst 3.0 online algorithm, a web-based tool designing for data normalization, visualization and interpretation, was applied for further multivariate analysis of the peak table using the statistical analysis module. Advanced data processing function including missing value estimation, data filtering and data normalization are available in this module [21].
The normalized data were conducted to perform PCA using statistical analysis module in Metaboanalyst for discrimination of the metabolomic differences among different herbal medicines and JQJT samples, respectively [22]. The resulting data were imported into the SIMCA-P 13.0 platform (Umetrics AB, Umeå, Sweden) for the partial least square (PLS) analysis to screen chemical features with anti-diabetic activity. R2 value, the determination coefficient, was used to estimate the predictive ability of the established model.
Bioactivity assay
Glucose consumption
Glucose consumption was determined by using a murine hepatocyte AML12 cell line obtained from American Type Culture Collection (Rockville, MD, USA). AML12 cells were cultured in 96-well plates and starved in DMEM/F12 (17.5 mM glucose) for 6 h, then treated with JQJT solutions (2 mg/mL) or metformin (1 mM, the positive control). After 16 h, glucose concentrations in the culture supernatant were determined and glucose consumption were calculated as described previously [23].
α-Glucosidase inhibition
The inhibition effect of JQJT preparation against α-glucosidase enzyme was measured according to previous study [24]. α-Glucosidase and substrate (4-nitrophenyl α-d-glucopyranoside, p-NBDG) were purchased from Sigma-Aldrich. Briefly, α-glucosidase and p-NBDG were dissolved in potassium phosphate buffer (67 mM, pH 6.8). Reaction mixture containing 10 μL of sample solutions (1.3 mg/mL), 40 μL of α-glucosidase (0.25 U/mL) solution was pre-incubated at 37 °C for 15 min. Then, 135 μL of 4 mM p-NBDG was added to the mixture and incubated for 30 min at 37 °C. The enzymatic reaction was terminated by the addition of 75 μL of sodium carbonate solution (0.2 mol/L). The α-glucosidase inhibition activity of tested samples was evaluated by measuring the absorbance (Abs) at 405 nm. Acarbose (2000 µg/mL) was used as the positive control. The inhibitory potency against α-glucosidase was calculated as follows: Inhibition% = [(Abs control – Abs sample) / Abs control].
Glucose Uptake Assay
Cellular glucose uptake was determined using the fluorescent probe 2‐NBDG according to previous method [25]. Palmitate was dissolved in ethanol and mixed with fatty acid-free bovine serum albumin (BSA) stirred at 50 °C for 2 h [26]. L6 myotubes were seeded into 96-well cell culture plates, and pre-treated with test samples (2 mg/mL) or the positive control-metformin (1 mM) for 2 h, followed by incubation with 0.5 mM palmitic acid-BSA conjugate solution for 16 h. Cells were starved with free-serum medium for 2 h, and then incubated with or without 100 µM 2-NBDG for 30 min. The intensity of fluorescence was measured after incubation with 100 nM insulin for 10 min.
Chemometric analysis
Chemical features ranking and pre-selecting using ReliefF
The peak matrix extracted from the raw data was first proceeded using ReliefF-based feature selection algorithm [27]. A few representative features which fitter better with the bioactivity were selected for subsequent chemometric analysis. ReliefF algorithm was performed within the MATLAB (R2016a, Mathworks, Natick, USA).
BP-ANN model
ANNs consisting of an input layer, one or more hidden layers, and one output layer are collections of inter-connected processing elements or neurons. The weights (w) and biases (b) are employed to evaluate the strengths of connections between inputs, hidden layers and output layers [28]. The BP algorithm is usually utilized in the establishment of ANN model to discovery variables that contribute target information [29]. As illustrated in Fig. S1, BP algorithm minimizes the error in predictions and produces satisfactory results by adjusting each weight of the networks, which is suitable for modeling complex nonlinear relationships [28].
Here, the inputs for the BP-ANN were the pre-selected peaks, while the bioactivity was set as output. It is known that the type of BP algorithm, the initial weights and bias, and the mu factor have considerable influence on the performance of networks [15]. The inappropriate parameters will lead to an over-fit model and take more training time. In this work, above mentioned parameters were optimized to achieve the best fitness of the BP-ANN model using response surface methodology (RSM), according to our previous study [30]. ANN tool of MATLAB was used to build the BP-ANN model.
Structural elucidation of selected compounds
The Xcalibur 2.1 was used to process the raw MS data. The compounds responsible for the bioactivity were screened and further identified. Mass Frontier 7.0 (Thermo Fisher Scientific, San Jose, CA, USA) was employed for the structural identification and the proposed of fragmentation patterns. Relevant published literature and the retention times were used to confirm the results.