All computational screening methods were performed using the Discovery Studio 2017R2 (DS; BIOVIA-Dassault Systèmes) running on Windows 8.1 operating system in a machine with an Intel^® Xeon^® E5-2699 v3 2.30 GHz octadeca core processor.

The features of a pharmacophore model reflect the drug-target interaction mode. By fitting berberine against a panel of pharmacophore models, the potential targets for berberine were picked out by employing the Ligand Profiler protocol of DS, which is equipped with two available pharmacophore databases, i.e., PharmaDB and HypoDB. Literature retrieval was simultaneously carried out to refine and supplement the profiling results.

The three-dimensional crystal structures of the proven targets for berberine were retrieved from the Protein Data Bank¹, including QacR (PDB ID: 1JUM), BmrR (PDB ID: 3D6Y), RamR (PDB ID: 3VW2), PLA₂ (PDB ID: 2QVD), and double helix DNA d(CGTACG)₂ (PDB ID: 3NP6). The structures were prepared and optimized using the Prepare Protein and Minimization protocols of DS (Chu et al., 2018). After optimization, the binding-site spheres were subsequently defined around the location of berberine. Based on the steric and electronic features in the binding sites of multiple targets with berberine, the MBM was generated using the Receptor-Ligand Pharmacophore Generation module in DS.

The MBM-based screening starts with the analysis of potential targets by the DS package. The optimized berberine was docked into the refined potential targets using LigandFit. The LigandScore can define the binding sites, generate ligand conformations, dock each conformation, save the top docked structures, and apply scoring function to each docked structure for the best binding mode (Chu et al., 2016). The pharmacophore models were subsequently generated using the Receptor-Ligand Pharmacophore Generation module. The binding potential (BP) was calculated using the following equation: BP = n^α × RMSD^β, where BP is also equal to the one-K_D-minus-κ power of ten, n is the number of pharmacophore features matched with MBM, RMSD is the average distance in three-dimensional structure between pharmacophore and MBM. The values of α, β, and κ were determined as 2.599, -5.234, and -5.905, respectively, which was simulated from a series of proven K_D values of berberine. The fit value is a combined measure of the similarity between pharmacophore and MBM, and computed by the following equation.

Here, BP₀ was taken as the minimum BP of proven targets, i.e., BmrR with a value of 49.243. The fit value of BmrR (Fit value = 0.5) was used as a predetermined threshold to select HPTs of berberine from the potential pharmacophores.

The selected poses of berberine with BACE1 (5ENM) and Aβ_1-42 (2NAO) were subjected to 10 ns molecular dynamics (MD) simulations using DS. The stability of the complex was analyzed and confirmed by plotting root mean square deviation (RMSD). The RMSD is a measure of the deviation of the conformational stability of the proteins from backbone structure to the early starting structure and fundamental property investigation in MD. The binding energy of the stable complex was calculated using force field CHARMm which showed the sum of electrostatic and van de Waals interaction terms.

The interaction of berberine (Sigma) with BACE1 (Sigma) and Aβ_1-42 (Sigma) was analyzed by SPR spectroscopy with a Biacore T200 biosensor instrument (GE Healthcare) (Yamasaki et al., 2013). Each target was immobilized onto flow cells in a CM5 chip using an amine-coupling method. Binding analyses were carried out at 25°C and a flow rate of 30 μl min^-1. Berberine in running buffer (1 × PBS, 3 mM DTT and 5% dimethyl sulfoxide, pH = 7.4) was run over each target at gradient concentrations as indicated. An empty flow cell, without any immobilized protein, was used as a reference. The binding curves were analyzed using the steady-state affinity analysis supplied with the BIA evaluation software (GE Healthcare).

The inhibitory property of berberine on BACE1 was evaluated by a fluorescence resonance energy transfer (FRET) assay (Sigma) (Kennedy et al., 2016). Briefly, BACE1 enzyme and substrate were diluted in a Fluorescent Assay Buffer (Sigma) to produce 10× working solutions. Berberine was diluted into different concentrations in deionized water. This assay was performed in 96-well microplates using 100 μl, which comprised of 20 μl substrate solution, 2 μl BACE1 enzyme solution, and 5 μl berberine solution. The reaction was allowed to proceed for 2 h in the dark at 37°C. The product fluorescence before and after the reaction was measured by a POLARstar Omega multimode microplate spectrophotometer (BMG LABTECH) using 320 nm excitation and 405 nm emission wavelengths. The percentage of inhibition was calculated using the following equation: [1 - (S -S₀)/(C -C₀)] × 100, where S and C represent the fluorescence intensities in the presence and absence of berberine, respectively. Inhibition curves were created by graphing the percentage of inhibition versus the inhibitor concentration using linear regression.

Aβ_1-42 was initially dissolved in HFIP (Sigma), evaporated, and re-dissolved in DMSO (Sigma) to a final concentration of 5 mM. To examine the effect of berberine on Aβ_1-42 aggregation, peptide solutions containing Aβ_1-42 monomers (10 μM) with and without berberine were incubated at 37°C for 48 h.

The morphological changes of Aβ_1-42 aggregation in the presence and absence of berberine were characterized by TEM using a JEM-1011 transmission microscope (JOEL, Tokyo, Japan). Each sample was spotted onto glow-discharged, formvar-coated 300 mesh copper grids (Ted Pella Inc.), incubated for 1 min, dried, and then negatively stained with 2% uranyl acetate (Chu et al., 2016).

The aggregation kinetics of Aβ_1-42 were monitored by incorporation of ThT. ThT solution was diluted in phosphate buffer (10 mM phosphate, 150 mM NaCl, pH = 7.0) to a final concentration of 20 μM. The fluorescence intensity for 20 μM Aβ_1-42 incubated with and without berberine was measured at 37°C using a POLARstar Omega multimode microplate spectrophotometer (BMG LABTECH) under kinetic fluorometric mode. Measurements were carried out at an excitation wavelength of 450 nm and an emission of 485 nm. To account for background fluorescence, ThT intensity from solution without Aβ_1-42 was subtracted from solution containing Aβ_1-42. Raw data were fitted using BMG LABTECH’s MARS Data Analysis Software.

Thirty 120-day male APP/PS1 transgenic mice were purchased from the Jackson Laboratory (Bar Harbor, ME, United States). All mice were housed at 25°C under a 12-h light-dark cycle and given free access to standard laboratory food and water. The mice were randomly assigned into three groups: control group (n = 10), 50 mg/kg berberine (n = 10) and 100 mg/kg berberine (n = 10). Berberine was administrated to these mice in their drinking water for 4 months, starting from 4 months of age just before they developed cognitive impairment and key pathologic features. The dose of berberine was chosen according to previous studies with no gender differences (Kong et al., 2004). All experiments were conducted in compliance with the Animal Care and Institutional Ethical Guidelines and approved by the Biomedical Ethical Committee of Peking University.

All mice at 8 months of age were subjected to the Morris water maze task for 5 days for evaluation on their learning and memory abilities (Vorhees and Williams, 2006). The apparatus consisted of a circular white metal pool (160 cm in diameter and 50 cm in height) filled with 26 cm deep water at constant temperature (22 ± 1°C) throughout the experiment. The water pool was divided into four quadrants by the water maze software and had a translucent acrylic platform (12 cm in diameter and 24 cm in height) placed in the center of the northwest quadrant and 1.0–2.0 cm below the water surface. The acquisition trial for the mice was the last training trials to find the hidden platform at a target quadrant. The escape latency was recorded by the finding-platform time. At day 5, the mice were subjected to the probe trial to access spatial memory. The time in the target quadrant and crossing the former platform area were recorded for testing spatial memory.

Five-micrometer-thick sagittal paraffin sections of mouse brain were mounted on glass slides. The tissue sections were deparaffinized and rehydrated through graded ethanol washes. Antigen retrieval was conducted in 10 mM citrate buffer (pH = 6.0) and boiled for 2 min. The slides were incubated with 3% hydrogen peroxide for 10 min, and then blocked by 10% normal goat serum for 10 min at room temperature. After removing excess blocking buffer, indirect immunohistochemical staining was performed with anti-Aβ_1-42 antibody (1:100, Covance Research Products, Berkeley, CA, United States) at 37°C for 1 h, washed thoroughly, and then incubated with biotinylated anti-mouse IgG antibody (1:100, Dako, Denmark) at 37°C for 45 min. The slides were detected using diaminobenzidine.

The results were conducted using Student’s t-test with SPSS 13.0 software (Chu et al., 2014). The data were expressed as means ± standard deviation (SD) of three independent experiments. Values of p < 0.05 were considered to be statistically significant.

Berberine is a natural isoquinoline alkaloid drawing increased attention for its favorable therapeutic effects on various diseases. The clinical efficacy of berberine is determined by its activity across multiple targets, including QacR, PLA₂, BmrR, RamR, and d(CGTACG)₂ DNA. Since the co-crystal structures of these targets in complex with berberine have been reported in the PDB database, these targets are selected as proven targets to generate MBM model for berberine. Two binding sites obtained for berberine consist in RamR, including site I and site II (Figure 1A). The co-crystal structures locate berberine close to the negatively charged binding sites (Figure 1B). The distinguishing feature appears to be the presence of buried acidic residues, such as Glu, Asp, and nucleic acids, which provide electrostatic interactions with the positively charged nitrogen (N⁺) of berberine. Distances between the N⁺ of berberine and the key anions in QacR, PLA₂, BmrR, RamR site I, RamR site II and d(CGTACG)₂ DNA are 4.29, 5.28, 4.88, 3.56, 3.75, and 3.79 Å, respectively (Figure 1B). Berberine has a large hydrophobic surface, which is ideal for interacting with the aromatic and aliphatic residues, constituting the hydrophobic pockets (Figure 1C). The binding pose of berberine and surrounding residues within the multi-target binding pockets are nearly identical (Figure 2A). The side chains from residues provide π–π and π-alkyl interactions to aid in electrostatic neutralization for the positive charge of berberine (Supplementary Figure S1). Molecular interactions reveal that berberine can establish strong attractive charge contacts with the hydrophobic moieties of neighboring residues within 7 Å (Figure 2B). Accordingly, pharmacophores for multiple targets were generated based on the atomic details of the binding features (Figure 2C). Through essential pharmacophore superimposition, a predetermined MBM model of berberine was developed for further target screening (Figure 2D).

Multi-target binding motifs-based screening started with the choice of potential targets for berberine by target fishing with two available pharmacophore databases, i.e., PharmaDB and HypoDB. Literature retrieval was simultaneously carried out to refine and supplement the profiling results. A complete list of 86 potential targets involved with 38 diseases was summarized in the Supplementary Table S1. According to the MBM-based comparison function, a fit value equal to or greater than 0.50 was used as a heuristic threshold to select HPTs of berberine from the hit pharmacophores (Table 1). Among the 13 candidates, there are two HPTs associated with AD, including BACE1 and Aβ_1-42, which indicated that berberine tended to be beneficial for the patients with AD. To validate the stability of berberine with BACE1 (5ENM) and Aβ_1-42 (2NAO), we performed standardized MD stimulations through Pipeline Pilot using the CHARMm component in DS. As shown in the Supplementary Figure S2, the most plausible models of berberine with BACE1 and Aβ_1-42 were stable.

The Divine Farmer’s Classic of Materia Medica, finished in the Han Dynasty (206 BC-220 AD), documented that “Chinese goldthread can prevent the loss of memory with extended treatment,” which is described as an “upper herb.” The major active principle, i.e., berberine, has recently been demonstrated as an effective therapeutic remedy to prevent or delay the pathogenesis of AD. However, the underlying mechanisms of berberine against AD are poorly understood.

According to MBM-based screening, BACE1 and Aβ_1-42 were predicted as HPTs for berberine, which lead to the aberrant production and aggregation of amyloid plaques in Alzheimer’s brains. Through application of molecular modeling, berberine was docked into the electronegative binding pocket of BACE1 (Supplementary Figure S3). The N⁺ of berberine engaged in electrostatic interactions with the key anion (Asp⁸⁰) of BACE1, and the phenyl groups created π–π stacking with the Tyr¹¹⁹ active site residue (Figure 3A). A pharmacophore model was subsequently generated based on the interactions between berberine and BACE1 (Figure 3B). By fitting the pharmacophore against MBM, BACE1 was picked out as a HPT for berberine with a value of 0.788156 (Table 1). The complex of BACE1 with compound 10 (5ENM) was calculated using force field CHARMm which showed the binding energy to be -110.2 kcal/mol (Wu et al., 2016). The docked complex of BACE1 with berberine was more structurally stable and energetically favorable than the compound 10, with the binding energy reaching to -123.6 kcal/mol. Furthermore, the binding affinity of berberine to BACE1 was tested by Surface Plasmon Resonance (SPR). The equilibrium dissociation constant (K_D) was calculated from SPR data to be 1.261 μM (Figure 3C). Moreover, we demonstrated that berberine effectively inhibited the activity of BACE1 with an IC₅₀ of 62.96 μM, thereby decreasing the generation of Aβ_1-42 (Figure 3D). Thus, BACE1 was identified as a new target of berberine.

With respect to Aβ_1-42, berberine was docked into a large hydrophobic surface of the oligomer (Supplementary Figure S4). Molecular modeling revealed that berberine interacted with the residues by establishing hydrophobic bonds with Phe¹⁹ and Val²⁴ (Figure 4A). The binding motif of berberine in Aβ_1-42 was outlined from MBM-based screening with a fit value of 0.906099 (Figure 4B and Table 1). The docked complex of Aβ_1-42 with berberine was calculated using force field CHARMm which showed the binding energy to be -73.6 kcal/mol. Moreover, the SPR responses indicated that berberine bound to Aβ_1-42 oligomer with a K_D value of 1.491 μM (Figure 4C). The aggregation kinetics of Aβ_1-42 in the absence or presence of berberine were monitored by incorporation of Thioflavin T (ThT). ThT fluorescence intensity showed that berberine was able to inhibit the seeding capacity of Aβ_1-42 in a dose-dependent manner (Figure 4D). Furthermore, transmission electron microscope (TEM) was employed to observe Aβ_1-42 aggregates. TEM image of Aβ_1-42 alone appeared characteristic fibrils after 48 h incubation at 37°C (Figure 4E). The fibrillization was disturbed when co-incubated with increasing concentration of berberine, indicating that Aβ_1-42 tended to be a new target of berberine (Figure 4E).

Furthermore, we studied the role of berberine in the prevention and treatment of AD. Spatial learning ability was observed on the basis of the time required to find the hidden platform. The berberine-treated APP/PS1 transgenic mice effectively improved their spatial learning ability in terms of daily escape latency through the 5-day consecutive training, compared with the control group (Figure 5A). However, no significant difference was assessed in swimming speed between the berberine-treated and the control groups (Figure 5B). Following the 5-day training, probe trials were conducted to assess both short-term (24 h) and long-term (72 h) memory tests on the 6th and 8th days, respectively. Significant differences were observed between the berberine-treated and the control mice (Figure 5C). In addition, the berberine-treated mice showed a significant improvement in the time spent searching for the pre-placed platform in both the short-term (p < 0.05) and long-term (p < 0.05) memory tests (Figure 5D). More importantly, the formation of amyloid plaque in the brain tissue was significantly inhibited for all the 50 mg/kg berberine (n = 10) and 100 mg/kg berberine (n = 10) treated mice, compared to the control group (n = 10) (Figure 5E). No significant difference was observed between 50 mg/kg berberine (n = 10) and 100 mg/kg berberine (n = 10) treated mice.

Based on relationship between drugs and targets, we built a space model for drug-target network. In the drug-target space, MBM represents a two-dimensional plane across all targets for one drug, which reveals that (i) single drug against multiple targets; (ii) multiple drugs against single target (Figure 6A). The resulting interactomes will connect all drug targets into a highly interlinked space, with strong local clustering of targets for each drug based on MBM modeling. To study the space, we developed a novel approach on the basis of enormous information on drug-target interactions. The rational design strategy includes the choice of potential targets, the generation of a dedicated MBM model, the MBM-based screening of HPTs, and the identification of new targets (Figure 6B). The candidates lead continue in an iterative process of reentering MBM determination and reevaluation for further optimization (Figure 6B). Key principles in this emerging field were tested through integrating the network pharmacology for berberine based on MBM screening (Figure 6C). Overall, 13 HPTs were predicted, with 2 targets verified experimentally, including BACE1 and Aβ_1-42 involved with AD (Table 1). Our results indicated that berberine was beneficial to decrease the generation of Aβ_1-42 via inhibiting the enzymatic activity of BACE1, and disrupt the aggregation of Aβ_1-42 into amyloid plaques (Figure 6D). Further studies on these HPTs will elucidate the underlying mechanism of berberine against polypharmacological profiles.