All of the constituent data of Huangqi and Huanglian were retrieved from TCM Systems Pharmacology Database and Analysis Platform (TcmSP™, http://ibts.hkbu.edu.hk/LSP/tcmsp.php) (Ru et al., 2014), and then manually supplemented through a wide-scale text-mining method. Meanwhile, four important pharmacology-related properties were also obtained from TcmSP™, including MW, CLogP, nHDon, and nHAcc. The principal component analysis (PCA) of the chemical distribution of Huangqi and Huanglian was built with the above four properties using the SIMCAP+ software package (version 11.0, Umetrics). The variances of PC1, PC2, and PC3 in Figure 1 account for 0.71, 0.23, and 0.04, respectively. The PCA of 34 known drug/drug-like compounds retrieved from DrugBank (http://www.drugbank.ca/) was performed in the same process as above (Supplementary Table S1).

The active ingredients from Huangqi and Huanglian were mainly filtered by integrating oral bioavailability (OB) and drug-likeness (DL). A robust in silico model OBioavail 1.1 that integrated the metabolism (P450 3A4) and transport (P-glycoprotein) information was employed to calculate the OB values of all herbal ingredients (Xu et al., 2012). Those ingredients with OB ≥ 30% were selected. Database-dependent DL evaluation approach based on Tanimoto coefficient (Ma et al., 2011) was applied and shown as T(A, B) = (A×B)/(|A|²+|B|²−A×B). In this equation, A represents the molecular descriptors of herbal compounds, and B displays the average molecular properties of all compounds in DrugBank. Those ingredients with DL ≥ 0.18 were preserved. In this study, the ingredients were adopted as the candidate compounds for further analysis when they met both of these criteria. Besides, owing to the profound pharmacological effects and high contents, those compounds with low OB or DL values were also selected for further research.

To identify the corresponding targets of the active ingredients of Huangqi and Huanglian, several approaches combined with chemometric method, information integration, and data-mining were implemented. First of all, the active ingredients were submitted to various servers viz. DRAR-CPI (Luo et al., 2011), Similarity Ensemble Approach (SEA, http://sea.bkslab.org/) (Keiser et al., 2007), STITCH (http://stitch.embl.de/) (Kuhn et al., 2012), and PharmMapper server (Wang et al., 2017). All active compounds were also sent to Herbal Ingredients' Targets database (HIT) (Ye et al., 2011), Therapeutic Targets Database (TTD, http://bidd.nus.edu.sg/group/ttd/) (Zhu et al., 2012), BindingDB database (http://www.bindingdb.org/bind/index.jsp) (Gilson et al., 2016), DrugBank and Google Scholar to mine C-T, interactions supported by literature. Then, to better dissect the role of Huangqi and Huanglian in DM treatment, all targets obtained from the previous two steps were sent to TTD, Comparative Toxicogenomics Database (CTD, http://ctdbase.org/), Online Mendelian Inheritance in Man (OMIM) database (http://www.omim.org/) and PharmGKB (http://www.pharmgkb.org) (Whirl-Carrillo et al., 2012) to mine whether these targets are related to DM. Noteworthy, only the targets of Homo sapiens were kept for further analysis.

The Molecule Annotation System 3.0 (MAS 3.0, http://bioinfo.capitalbio.com/mas3/) and DAVID (The Database for Annotation, Visualization and Integrated Discovery, http://david.abcc.ncifcrf.gov) webservers were employed to perform GO enrichment analysis for the 50 genes targeted by Huangqi and Huanglian (Huang da et al., 2009). The target organ distribution was determined based on the microarray analyses data of different tissue types lodged in the BioGPS bank (http://biogps.org) (Wu et al., 2009).

Three visualized networks were constructed: (1) Compound-Target network (C-T network). Active ingredients of Huangqi and Huanglian and their corresponding targets were employed to generate the C-T network; (2) Target-Pathway network (T-P network). The pathway information of targets were extracted from KEGG (Kyoto Encyclopedia of Genes and Genomes, http://www.kegg.jp), and then a bipartite T-P network composed of targets and their corresponding putative pathways was built; (3) T-O network. The potential targets and their tissue types were applied to T-O network. All visualized networks were constructed by Cytoscape 3.5.1 (http://www.cytoscape.org/), an open software package project for visualizing, integrating, modeling and analyzing the interaction networks (Smoot et al., 2011).

In order to estimate the contribution of each active ingredient to the antidiabetic effects of Huangqi and Huanglian, a contribution index (CI) based on network based efficacy (NE) weighted by literature was calculated by Equation (1) and equation (2) (Yue et al., 2017):

Where n is the number of targets associated with ingredient j in the C-T network; d_i is the degree of target i associated with ingredient j in the T-P network; c_i is the number of DM-related literature of ingredient i; m is the number of ingredients. For DM-related literature-mining approach, the following keywords were used for DM terms: diabetes, hyperglycemia, and insulin resistance and the common names of active ingredients were also used as search keywords. The numbers of papers having keywords in the title or abstract published in 1990–2017 were obtained from the PubMed database. If the sum of CIs for the top N ingredients was more than 85%, these relevant N ingredients were considered to contribute the most to the antidiabetic effects.

The ingredients in Huangqi and Huanglian were retrieved from TcmSP™ and published literatures. Since those glycosides in Huangqi and Huanglian might be deglycosylated in the intestinal tract associated with gut microbiota, 13 aglycones were also incorporated into the compound library labeled by _qt. Thus, a total of 171 ingredients were retrieved for Huangqi (112) and Huanglian (64), including five common ingredients. The detailed information about these molecules was provided in Supplementary Table S2.

The PCA was further conducted to give visual illustration in chemical distribution. The constituents of Huangqi and Huanglian were diverse and both of them possessed a broad diversity in chemical space (Figure 1), but the majority of them satisfied the Lipinski's rule of five. Moreover, the large overlap between the ingredients in Huangqi and Huanglian and 34 known drug/drug-like compounds for DM demonstrated that many compounds contained in these two herbs had drug potential on DM. Noticeably, since PC2 was highly associated with the value of nHDon, those ingredients of Huangqi (red circle) including astragalosides with PC2 in the range 3~5 possessed larger nHDon values than the antidiabetic drugs from DrugBank (yellow circle), whereas the ingredients of Huanglian (black circle) overlap better with the antidiabetic drugs in Figure 1. Thus, it is meaningful to figure out these active ingredients in Huangqi and Huanglian for DM treatment.

Although a single herb or TCM formula usually contains large numbers of compounds, virtual screening approaches are always of great help to distinguish those active ingredients. In the present work, two ADME (absorption, distribution, metabolism, and excretion)-related models, including OB and DL were employed to screen most of the active ingredients from Huangqi and Huanglian. A few of active compounds that do not meet either of these two criteria were also selected in the cases of high bioactivities and huge amounts. Consequently, a total of 43 active compounds were selected from the 171 compounds of these two herbs (Table 1).

In Huangqi, only 26 ingredients passed through the strict filtering criteria, and most of them exhibited potent pharmacological activities. For examples, calycosin (M120, OB = 47.75%, and DL = 0.24) displayed therapeutic effects on diabetic complication (Xu et al., 2011); formononetin (M125, OB = 69.67%, and DL = 0.21) showed significant antihyperglycemic activity (Qiu et al., 2017). Other characteristic isoflavonoids in Huangqi accounted for large contents and were also preserved. Specifically, calycosin-7-O-β-D-glucoside (M121, OB = 10.05%, and DL = 0.81), formononetin-7-O-β-D-glucoside (M126, OB = 11.52%, and DL = 0.78), (6αR, 11αR) 3-hydroxy-9,10-dimethoxypterocarpan-3-O-β-D-glucoside (M123, OB = 36.74%, and DL = 0.92), together with calycosin and formononetin in Huangqi were measured up to 0.44–1.76 mg/g (Wu et al., 2005). Surprisingly, astragalosides, the main active and characteristic compounds in Huangqi, exhibited low OB values (Gu et al., 2004; Ma et al., 2017). Among them, astragaloside IV (AIV, M98, OB = 2.20%, and DL = 0.15) has been determined as the quality marker of Huangqi in Chinese Pharmacopoeia (The State Pharmacopoeia Commission of China, 2015) and exhibited significant hypoglycemic effect (Lv et al., 2010). Astragaloside II (M94, OB = 0.79%, and DL = 0.13) and isoastragaloside I (M104, OB = 37.80%, and DL = 0.14) could alleviate insulin resistance and glucose intolerance by enhancing the expression of an insulin-sensitizing adiponectin (Xu et al., 2009). Besides, the contents of astragaloside I (M92, OB = 46.79%, and DL = 0.11), astragaloside II, astragaloside III (M96, OB = 31.83%, and DL = 0.10) and AIV in Huangqi were 0.78, 0.35, 0.20, and 0.26 mg/g, respectively (Zu et al., 2009). Importantly, astragalus injection (the major components are astragalosides) has been used in China to treat DM with good clinical effects (Nie et al., 2014). Thus, astragalosides were also selected for targeting. Astragalus polysaccharides (M171) as a polysaccharide mixture from Huangqi (accounted for 90.00–288.75 mg/g) was also preserved because it could ameliorate insulin resistance and restore glucose homeostasis, in part, via gut microbiota (Zou et al., 2009; Zhang et al., 2013; Liu et al., 2014).

Searching for the targets of candidate drugs solely by the experimental approaches is overspending, labor-intensive, and time-consuming. In the present work, an integrated in silico approach was introduced to identify the target proteins for the active ingredients of Huangqi and Huanglian. Predictive models were used including DRAR-CPI, SEA, STITCH and PharmMapper server, and databases were mined including HIT, TTD, BindingDB database, DrugBank and Google Scholar. Finally, 50 DM-related targets were determined, interacting with the selected 43 active ingredients of this herb combination (Table 2). Of note, we have implemented molecular docking and surface plasmon resonance (SPR) assay to explore the reliability of the interactions between the active ingredients and their putative targets. As shown in Supplementary Table S3, Supplementary Figures S3, S4, it could be concluded that the predicted targets were convincible to explore the network interactions of Huangqi and Huanglian.

Analysis of interaction network regulation of 50 targets was performed using MAS 3.0. As shown in Figure 2, biological process (BP, GO:0008150), molecular function (MF, GO:0003674), and cellular component (CC, GO:0005575) accounted for 64.82, 24.10, and 11.08%, respectively. Further, BP, MF, and CC enrichment analysis were performed by DAVID bioinformatics resources. The top 10 significantly enriched terms in BP, MF, and CC categories (P < 0.05, P-values were corrected using the Benjamini-Hochberg procedure) were listed in Figure 2, indicating that Huangqi and Huanglian may regulate glucose homeostasis, insulin secretion, and nitric oxide biosynthetic process via enzyme binding, insulin binding, and kinase binding in the cytosol, caveola, and plasma membrane so as to exert antidiabetic potential. It is interesting to note that a large number of targets were associated with a variety of BP terms such as regulation of insulin secretion, glucose homeostasis, cellular response to insulin stimulus, lipopolysaccharide-mediated signaling pathway, and positive regulation of nitric oxide biosynthetic process, which are closely related to the pathogenesis of DM. KEGG pathway enrichment analysis was also performed by DAVID bioinformatics resources and displayed in Supplementary Figure S5. Furthermore, network and functional association of 50 targets of Huangqi and Huanglian was mapped in Supplementary Figure S6 using GeneMANIA (http://www.genemania.org/).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.