Zishen Jiangtang Pill was provided by Pharmaceutical Department of Shenzhen Traditional Chinese Medicine Hospital (Lot no. 20170509). ZJP consists of night herbs and the mixed proportion of respective herb is illustrated in Table 1. All the botanical names can be checked and validated using http://mpns.kew.org/mpns-portal/?_ga=1.111763972.1427522246.145907734. Herbs used in this study were obtained from Shenzhen Huahui Pharmaceutical Co., Ltd. The plant materials were authenticated by Prof. Shangbin Zhang based on their morphological characteristics. The voucher specimens were deposited in the herbarium of Pharmaceutical Department, Shenzhen Traditional Chinese Medicine Hospital. Assurance of quality control for all the materials was validated according to the Chinese Pharmacopoeia (China Pharmacopoeia Committee, 2015). ZJP was prepared based on the manufacturing process of the local FDA in China. In brief, Astragali Radix and Notoginseng Radix et Rhizoma were prepared together as a superfine powder. The rest seven herbs of ZJP were weighed and extracted in boiling water twice for 1 h. After centrifugation, the supernatant was concentrated under reduced pressure to concentrated solution, and mixed with aforesaid powder to form into pills. The yield of the extract was ∼31.9%. HPLC-MS was conducted to confirm the quality of the ZJP, as indicated in Supplementary Figure 1. Deltamine was purchased from Shiguibao (Shiguibao, China). HPLC grade acetonitrile and formic acid were purchased from TEDIA (Fairfield, OH, United States). Deionized water was prepared using a MilliQ Ultrapure water system (Millipore, Merck, Germany). All other reagents used here were analytical grade.

Six-week-old male C57BL/6 mice were purchased from Nanjing University-Nanjing Biopharmaceutical Institution. Experimental animals were maintained in a specific pathogen-free (SPF) animal facility under a 12 h light-dark cycle, with food and water ad libitum. This study was carried out in accordance with the recommendations of National Institutes of Health Guideline for the care and use of laboratory animals. All animal procedures were conducted with protocol approval from the Ethics Committee of Shenzhen Traditional Chinese Medicine Hospital, Guangzhou University of Chinese Medicine (Shenzhen, China), and all efforts were made to minimize animal suffering.

Mice were provided either normal diet or high-fat diet for 12 weeks. The control group was fed with normal diet (Code LAD3001M, Nantong Trofee Feed Technology Co., Ltd.). The experimental diet was a modified model based on the Van Heek series 60% high-fat obesity model diet (Code TP23400, Nantong Trofee Feed Technology Co., Ltd.), and the feed was produced according to the AIN93 standard. After about 12 weeks of modeling, high-fat diet mice were randomly divided into 4 groups (n = 10 per group): Group 1: model group (high-fat diet only); Group 2: Deltamine group (168 mg/kg/d); Group 3: high dose of Zishen Jiangtang Pill (H-ZJP) group (1.51 g/kg/d); Group 4: low dose of Zishen Jiangtang Pill (L-ZJP) group (0.75 g/kg/d). The same volume of distilled water was given to control and model groups. After 8 weeks treatment, mice were sacrificed. Blood samples were collected from retro-orbital veins.

After drug administration for 8 weeks, oral glucose tolerance test (OGTT) and insulin tolerance test (ITT) were measured (Chaput et al., 2007; Pacifici et al., 2014). Briefly, OGTT was performed in 12-h fasted mice by measuring tail blood glucose after oral glucose administration (2 g/kg) by gavage at 0, 30, 60, and 90 min. ITT was performed using intraperitoneal injection of 0.75 international units (I)/kg insulin in mice fasted for 4 h, and glucose levels were measured after 0, 15, 45 and 75 min. Blood samples were obtained from the tail vein using an automated OneTouch Glucometer (LifeScan, Milpitas, CA, United States). OGTT was measured 4 days before ITT. 2 to 3 days after ITT, the animals were sacrificed. Fasting blood glucose (FBG), fasting blood insulin (FINS), blood triglycerides concentration (TG) and blood total cholesterol concentration (TC) of each group were tested by glucose assay kit (Nanjing Jiancheng Bioengineering Institute, China), rat/mouse insulin ELISA (Merck, Germany), triglycerides assay kit (Nanjing Jiancheng Bioengineering Institute, China) and total cholesterol assay kit (Nanjing Jiancheng Bioengineering Institute, China). The insulin resistance index (HOMA-IR) and quantitative insulin sensitivity check index (QUICKI) were calculated. Calculation formula: HOMA-IR = (FBG, mmol/L) × (FINs, mmol/L)/22.5; QUICKI = 1 / [log (FINs μU/mL) + log (FBG mg/dL)].

The islet tissues of mice were taken, fixed with 4% neutral paraformaldehyde, made into 4 μm paraffin sections after dehydration, embedding with wax, and slicing. The sections were incubated with specific primary antibodies directed against insulin (Cat No. ab8304, Abcam) and glucagon (Cat No. ab36598, Abcam), respectively. Subsequently, sections were incubated with HRP-conjugated secondary antibodies (Cat NO. KIT-9901, MXB Biotechnologies) and visualized with a fluorescence microscope (Olympus, Japan).

Serum samples were thawed at room temperature, and 100 μl of which was added with 300 μl ice cold acetonitrile. The solution was thoroughly mixed by vortex for 30 s, kept at 4°C for 10 min, then centrifuged (12,000 rpm for 10 min at 4°C). The supernatant (250 μl) was transferred and added with 20 μl IS (2-Chloro-L-phenylalanine, 1 mg/ml), then evaporated to dryness under nitrogen gas at 37°C. The residue was re-dissolved with 100 μl 10% acetonitrile and centrifuged (13,000 rpm for 10 min at 4°C) after being vortexed for 60 s. The supernatant was transferred to auto-sampler vials for further UHPLC-Orbitrap/MS analysis.

To guarantee the quality of the non-targeted bioanalytical data, quality control (QC) samples were used for method validation. Herein, 20 μl of serum samples from each group was pooled to obtain a QC sample, and QC samples were extracted using sample extraction method mentioned above. The QC specimens were analyzed every six samples throughout the whole analysis procedure. Repeatability and stability were performed by preparing and analyzing six QC samples.

Chromatographic separation was performed on a Dionex UltiMate 3000 UHPLC system (Thermo Scientific, San Jose, CA, United States). Waters Acquity BEH C18 column (2.1 × 100 mm, 1.7 μm, Waters, Milford, United States) was applied for all analyses at 30°C. The mobile phase was composed of 0.1% formic acid water (A) and acetonitrile (B) at a flow rate of 0.2 mL/min. The injection volume was set as 2 μl. For positive mode analysis, the gradient conditions were as follows: 0–4 min 10–13% B; 4–6 min 13–50% B; 6–10 min 50% B; 10–14 min 50–85% B; 14–18 min 85% B; 18–20 min 85–100%; 20–25 min 100%; re-equilibrate: 10 min. For negative mode analysis, the gradient conditions were as follows: 0–2 min 15% B; 2–4 min 15–55% B; 4–8 min 55% B; 8–14 min 55–70% B; 14–22 min 70–80% B; 22–24 min 80–100%; 24–28 min 100%; re-equilibrate: 10 min.

Mass spectrometry experiments were accomplished on an orbitrap mass spectrometer (LTQ ORBITRAP VELOS PRO, Thermo Fisher Scientific, San Jose, CA, United States) equipped with an electrospray ionization (ESI) source in both positive and negative modes. The scan range was from 100 to 1500. The MS parameters: electrospray ionization temperature (°C): 350 (ESI⁺)/350 (ESI⁻); Sheath Gas Flow Rate(arb): 35 (ESI⁺)/35 (ESI⁻); Aux Gas Flow Rate: (arb): 10 (ESI⁺)/10 (ESI⁻); Sweep Gas Flow Rate(arb): 0 (ESI⁺)/0 (ESI⁻); I Spray Voltage(kV): 3.2 (ESI⁺)/3.2 (ESI⁻); Capillary Temp(°C): 300 (ESI⁺)/300 (ESI⁻).

Raw data files were pretreated, procedures included peak finding, alignment, filtering and normalization to total area, a three-dimensional data set consisted of sample information, peak intensities, peak retention time (RT) and the mass-to-charge ratio (m/z) was obtained. Retention time and the m/z data were used to be identifier for each ion. Moreover, peaks with missing value (ion intensity = 0) in more than 80% samples were removed in order to obtain consistent variables. Then the resultant data matrices were imported into the SIMCA14.1 (Umetrics, Umeå, Sweden) software for multivariate statistical analysis. The analysis methods containing principal components analysis (PCA), partial least-squares discriminant analysis (PLS-DA) and orthogonal partial least-squares discriminant analysis (OPLS-DA) were used for metabolite profile analysis (Chang et al., 2017).

Orthogonal partial least-squares discriminant analysis model was performed to visualize the metabolic difference between model group and control group. Those variables with VIP > 1 in the OPLS-DA model, as well as the variables with p|corr| value > 0.58 in S-plot, and the variables with confidence interval crossed zero in jack-knifed loading plot were considered as potential biomarkers. The potential biomarkers were identified by using Masslynx (Waters Technologies, United States) combined with the METLIN database¹ and HMDB database². Then, receiver operating characteristic (ROC) curves (SPSS 19.0) were applied to analyze data for evaluating the predictive power of the identified biomarkers. Pathway analysis was depended on KEGG database³, MetaboAnalyst 3.0⁴.

Individual data was expressed as Mean ±_standard deviation (SD). Statistical analyses were performed using one-way ANOVA (version 13.0, SPSS). Statistically significant changes were classified as significant (^∗) where p < 0.05, more significant (^∗∗) where p < 0.01 and highly significant (^∗∗∗) where p < 0.001.

Before drug treatment, ZJP was firstly chemically standardized. An HPLC-MS method was developed to reveal the profile of ZJP extract and quantify the main ingredients in the extract. By using the respective individual standard, ten chemical markers were identified from ZJP extract (Supplementary Figure 1) and the minimum amount in mg/g of dried extract was 0.0070 for β-ecdysone, 0.1163 for calycosin 7-O-glucoside, 0.0366 for acteoside, 0.1901 for naringin, 0.4649 for notoginsenoside R1, 0.1567 for ginsenoside Re, 0.7127 for ginsenoside Rg1, 0.4280 for icariin, 0.8445 for ginsenoside Rb1, 0.1868 for astragaloside IV and 0.0633 for schisandrin. The chemical analysis of ZJP extract here served as the quality control for the reproducibility of the below animal study.

To investigate the anti-diabetic efficacy of ZJP on animal model, the diabetic mice induced by high-fat diet were employed for this study. The timeline of current study is presented in Figure 1A. After drug administration for 8 weeks, OGTT and ITT were performed to confirm the efficacy of drug treatment (Supplementary Figure 2). Compared with normal group, mice from the high-fat diet group showed increased levels of serum insulin, glucose, and total cholesterol by ∼2.5-fold, 1.6-fold, and 2.4-fold, respectively (Figure 1B–D). The increased levels of insulin, glucose and total cholesterol could be significantly decreased by high dose of ZJP treatment compared with model group, and ZJP treatment at low dose showed some improvement but not in a significant level (Figure 1B–D). The triglyceride levels did not show obvious change in both model and drug treatment groups (Figure 1E). In addition, the insulin resistance index (HOMA-IR) and quantitative insulin sensitivity check index (QUICKI) were calculated. The HOMA-IR and QUICKI results further supported the improvement of high-fat diet group by drug treatment (Figure 1F,G). Deltamine treatment served as a positive control, which could improve insulin sensitivity as compared to the diabetic group.

Immunohistochemical analysis demonstrated that pancreatic islets in model mice exhibited strikingly different appearance, with poor insulin staining and increased glucagon staining, and that of ZJP treatment mice were showed to be improved compared with model group (Figure 2A,B).

The overlapping total ion current (TIC) chromatograms (Supplementary Figure 3) of QC samples showed acceptable variations occurred during large-scale sample analysis. Meanwhile, six selected extracted ion chromatograms (EICs) in QC samples were employed to evaluate system repeatability and stability. Taking sample in positive mode as an example, retention times, peak areas and mass accuracies of these six selected peaks demonstrated acceptable RSDs. RSDs of these six peaks were 0.036–0.088% for retention times, 3.25–15.60% for peak areas, and 0.8 × 10⁻⁴%–1.0 × 10⁻⁴% for mass accuracies (Supplementary Tables S1, S2).

By using the developed UHPLC-Orbitrap/MS method, a total of 8000/8000 peaks (ESI⁺/ESI⁻) were detected in serum samples. The vectors of data from control and model groups were collected as a matrix for PCA and OPLS-DA analysis to discover potential variations. The score plots of PCA demonstrated a clear separation between control and model groups in the metabolic profiles (Figure 3A,B). Similar findings were found in the score plots of OPLS-DA (Figure 3C,D). For OPLS-DA analysis, the model showed excellent classification and predictive ability between control and model groups with the intercept R²Y above 0.99 and Q² above 0.85 (Figure 3).

A total of 26 potential biomarkers were found and identified in plasma samples (Table 2). The ROC curve analysis demonstrated the detailed area under the curve (AUC), p-value, 95% confidence interval and the error value of the 26 identified potential biomarkers for diabetes prediction (Figure 4). The 26 metabolites showed excellent diagnostic properties with average area under the curve at 0.78–1.00 and p-value at 0–0.034. Here, the data indicated that the identified biomarkers exerted a powerful diagnostic ability.

In order to reveal the beneficial role of ZJP for treating diabetes, PLS-DA analysis was conducted to obtain the changes of metabolic trajectory after drug treatment. As indicated in Figure 5, metabolic trajectory of model group was gradually being moved toward to that of control group after administration of ZJP or deltamine. Moreover, the relative quantities of 26 potential metabolites were obtained using peak intensities from the resultant data matrices, amongst which 24 metabolites were robustly affected by ZJP compared with model group (Figure 6). Here, ZJP treated-mice restored to normal levels altered by the high-fat diet, which included down regulation of biomarkers.

The aforesaid databases in Method Section were used for analyzing the 26 metabolites related metabolic pathways. The topology map analysis described the impact of 26 metabolites between control and model groups on metabolic pathway. As indicated in Supplementary Figure 4, these metabolites were mainly involved in glutathione metabolism, steroid hormone biosynthesis and glycerophospholipid metabolism. These three principal pathways therefore were suggested to be closely involvement in the anti-diabetic action mechanism of ZJP.