Two hundred and fifty seven compounds were purified from Selaginella plants, including Selaginella tamariscina, Selaginella pulvinata Maxim, Selaginella braunii Baker, Selaginella delicatula (Desv.) Alston, Selaginella moellendorfii hieron, Selaginella uncinate, Selaginella involven Spring, Selaginella doederleinii Hieron. Total extracts (TE) were extracted using 75% ethanol and then freeze-dried into extractum. The suspensions of saline and freeze-dried extractum after ultrasonic vibration was orally administrated for AD mice.

The learning and memory ability of AD mice were evaluated by Morris water maze test. Male specific-pathogen-free (SPF) grade male ICR mice (body weighing 18–22 g) were purchased from Hunan Provincial Experimental Animal Centers [Changsha, Hunan, China, Certificate No. SYXK (Xiang) 2012-0004] (Sun et al., 2009).

Mice were randomly divided into five groups (10 mice for each group), namely normal control group (NCG), model control group (MCG), low dose group (LDG, 50 mg/kg), middle dose group (MDG, 100 mg/kg) and high dose group (HDG, 200 mg/kg). To build the AD mice model, mice in MCG, LDG, MDG, and HDG were treated with D-gal (120 mg/kg, intraperitoneally) for 56 days (8 weeks), and the mice in normal group were treated with saline of the same volume for 56 days (8 weeks; dorsonuchal subcutaneous injection). After that, the TE suspensions of saline and freeze-dried extractum after ultrasonic vibration was orally administrated to the mice in LDG, MDG, and HDG for 42 days (6 weeks), and the mice in NCG and MCG were orally treated with saline of the same volume for 42 days (6 weeks). Finally, the spatial learning and memory ability of all the mice were tested by Morris water maze.

The equipment of Morris Water Maze were purchased from Anhui Zheng-hua biological equipment corporation and the test process followed to the relevant laboratory manual. Two indexes, the place navigation and spatial probe, were chosen as the main monitor elements to evaluate the spatial learning and memory ability of all the mice. The experimental method is divided into two parts: acquisition phase and probe trial. In the acquisition phase, we randomly put the head of the mouse into the wall of the pool and fix the starting position. After that, the time of finding the underwater platform was recorded. On the day after acquisition phase, the platform was removed and the probe trial began. The time of finding the position where the platform is located, the swimming distance and the number of crossing through the area where the platform is located were recorded as the spatial memory test indexes.

This study was carried out in accordance with the recommendations of “Laboratory Animals-Guideline of welfare and ethics, Ethics Committee of Hunan Provincial Experimental Animal Centers.” The protocol was approved by the “Ethics Committee of Hunan Provincial Experimental Animal Centers.”

According to previous studies published in recent years, we finally found 19 significant proteins related to AD (Cavalli et al., 2008; Fang et al., 2015). For these important AD-related proteins, we collected their ligands that are small, drug-like molecules from Binding database¹. For each protein, activity data were filtered to keep only activity end-point points that have half-maximum inhibitory concentration (IC₅₀), half-maximum effective concentration (EC₅₀) or Ki values. A compound would be considered as a positive sample when its activity value was below 10 μM. Otherwise, this compound would be considered as a negative sample. Following this step, maybe some AD-related proteins have very little number of negative samples. To balance the number between positive samples and negative samples, we randomly selected certain number of compounds from other AD-related proteins to generate the negative samples for these AD-related proteins. The number of these selected negative samples together with inactive samples should be basically equal to the number of the active samples for these AD-related proteins. These prepared positive sets and negative sets were used for the subsequent model building. The detailed information of AD-related proteins and these datasets used for model building can be seen in Supporting Information (Supplementary Material).

For each protein, a series of high confidence SAR models were built by Random Forest (RF) and different fingerprint representations (FP2, MACCS, Daylight, ECFP2, ECFP4, and ECFP6). RF was introduced by Breiman and Cutler for regression and classification modeling in 2001 firstly (Breiman, 2001). The method is based upon an ensemble of decision trees, from which the prediction of a classification task is provided as the majority vote of the predictions of all trees. Recent studies have suggested that RF offers several striking features which make it very attractive for QSAR/QSPR studies. These include relatively high accuracy of prediction, built-in descriptor selection, and a method for assessing the importance of each descriptor to the model (Cao et al., 2011a,b; Yun et al., 2016). RFs were trained using the RF library in the statistical computing environment, R. All the fingerprints were calculated by some tools developed by our group: ChemDes, BioTriangle webserver and ChemoPy package (Cao et al., 2013; Dong et al., 2015, 2016). To improve the prediction ability of the SAR model, we assembled all fingerprint models to obtain the consensus models with average output. All the assembled models were validated by 5-fold cross validation and test set validation to demonstrate their prediction performance. In this part, some popular statistic parameters were applied to evaluate the performance of these classification models: true positive (TP); false negative (FN); true negative (TN); false positive (FP); sensitivity (SE); specificity (SP); accuracy (ACC); area under receiver operating characteristic curve (AUC). These classification evaluation parameters are defined as follows:

After a series of modeling and validation processes, we aimed to obtain reliable SAR models for above-mentioned AD-related proteins. And then, 257 compounds purified from Selaginella plants were predicted by these robust and practical models and their inhibitory activities were identified preliminarily for further study.

For the compounds that have been regarded as active ingredients by the multi-target SAR models, the in vitro target enzyme inhibitory activity test was applied to verify their actual activity for AD treatment. The inhibitory activities were determined by fluorimetric method on Infinite M200 Multi scan Spectrum (Tecan, Swiss). Each concentration was analyzed in triplicate and IC₅₀ values were determined by nonlinear regression of inhibition vs. log concentration plots, using GraphPad Prism 7 for Windows, Version 7.00 (GraphPad Software Inc.). BACE1 fluorescence resonance energy transfer assay kits were purchased from the Pan Vera Co and Monoamine Oxidase B (MAO-B) inhibitor screening kits were purchased from Bio Vision Inc.

In the BACE1 inhibition test, the assay was performed in 384-well plates. The assay solution was consisted of 10 μL test compounds (concentrations: 0.017, 0.050, 0.167, 0.500, 1.667, 5.000, 16.667, and 50.000 μM), 10 μL BACE1 substrate and 10 μL BACE1 enzyme. LY2811376 was selected as the reference compound with IC₅₀ = 0.242 μM and the blank buffer was set as the negative control. The mixture was incubated for 60 min at room temperature. At the end, 10 μL BACE1 stop solution was added to stop the reaction and the fluorescence was detected at the Ex/Em = 545/585 (12 nm bandwidth) settings on Multi scan Spectrum.

In the MAO-B inhibition test, the assay was performed in 96-well plates. The assay solution was consisted of 10 μL test compounds(concentrations: 0.2, 0.4, 0.8, 1.6, 3.2, 6.4, 12.8, and 25.6 μM), 37 μL MAO-B assay buffer, 1 μL MAO-B substrate, 1 μL developer and 1 μL OxiRed Probe. Selegiline was used as the reference control with IC₅₀ = 0.028 μM and the blank buffer was set as the negative control. The mixture was incubated for 10 min at 37°C. The fluorescence was measured at Ex/Em = 535/587 nm kinetically at 37°C for 10–40 min. Two points (T1 and T2) in the linear range of the plot were chosen and the corresponding fluorescence values (RFU1 and RFU2) were obtained to calculate the slope for all samples. The Calculation of % relative inhibition was following the manual of MAO-B inhibitor screening kit.

To further verify the results of multi-target SAR prediction and enzyme inhibitory experiments, the molecular docking process was applied to simulate the binding position and binding affinity between the active compounds and target proteins. Generally speaking, docking is a computer simulation modeling technique used to predict the interaction between a ligand and a receptor active site, and is an important tool in structure-based drug design. The technique of docking is to position the ligand in different orientations and conformations within the binding site to calculate optimal binding geometries and energies. In this part, the molecular operating environment (MOE, version 2014.) was applied to carry out the molecular docking process. MOE's dock application searches for favorable binding modes between small- to medium-sized ligands and a not-too-flexible macromolecular target. For each ligand, a number of placements called poses are generated and scored. The score can be calculated as either a free energy of binding including among others solvation and entropy terms, or enthalpy based on polar interaction terms including metal ligation, or as qualitative shaped-based numerical value. According to the score values, ligands with different conformations can be ranked and the optimal structural conformation will be affirmed (Wang J. et al., 2015). To make the interactions with the binding site easy to see, the ligand interaction was carried out. It will automatically be loaded with a 2D diagram of the original ligand and a schematic representation of the binding site residues, with the important interactions between ligand and binding site shown. In this study, we selected two proteins as the docking acceptors: BACE1 (PDB ID: 1TQF) (Cobum et al., 2004); MAO-B (PDB ID: 2V5Z) (Binda et al., 2007). As a control, the original ligand included in the crystal structure should also be docked. A series of parameters were set: Dock: rescoring 1 = ASE; retain = 100; rescoring 2 = ASE; retain = 100. Configure force field: final gradient = 0.0001; maximum iterations = 1,000; force constant = 10; radius offset = 0.4. For rest parameters, the default treatment was applied.

Learning and memory ability of AD mouse was evaluated by Morris water maze test in which the navigation and space exploration are used as indexes. There were five groups of mice under study and the behavioral results can be seen in Table 1 and Figure 1. In the Table 1, the residence time and residence distance of each quadrant, the total distance and the number through platform for each group of mice were listed. From the table, we can see that the residence time of MCG was significantly decreased (P < 0.05) in 1st and 4th quadrant compared with NCG. The stay intervals in 1st quadrant for MCG group were significantly lower than NCG (P < 0.05). However, the residence distance in 3rd quadrant increased prominently (P < 0.05). HDG showed longer distance in 1st quadrant (P < 0.05) and opposite trends in 3rd quadrant (P < 0.05) compared with MCG after 42 days' dosage. With respect to the total distance and numbers through platform, they were significantly reduced for MCG (P < 0.05) compared with NCG. They were increased significantly for HDG (P < 0.05) compared with MCG after 42 days' dosage. What's more, the crossing through number for MDG were also significantly increased (P < 0.05).

Figure 1 shows the navigation and space exploration for different groups of mice. From the figure, we can see that the spatial learning and memory ability of NCG were significantly increased, but there was an opposite trend for MCG and mice in MCG mainly ran along the cell wall. What's more, all the low, medium and high dose of TE can significantly increase the numbers of exploration platform. Considering both results in Table 1 and Figure 1, TE of Selaginella has a remarkable improvement on learning and memory function for AD mice. This result inspires us to further explore the effect of the chemical ingredients from Selaginella on AD treatment.

Based on the results of the Morris water maze test, the TE of Selaginella plants show a potential benefit for AD treatment. To quickly screen the active ingredients from a number of compounds preliminarily, multi-target SAR models associated with AD-related proteins were constructed as described before. In this part, we finally obtained a series of ensemble predictive models for AD-related proteins. Their statistic results of 5-fold cross validation and test set validation were listed in Table 2. From this table, we can see that for each predictive model, the accuracy is good enough not only for cross validation (0.808–0.955) but also the test set validation (0.846–0.970). With respect to other statistic parameters, the similar results were obtained and it can be strong evidence for the good predictive ability of this model. Therefore, we have reasons to believe that these ensemble models are robust and practical and can be used to predict the inhibitory activity for a new compound in the early stage of drug discovery.

To evaluate the probability of inhibitory activity for the 19 AD-related targets, 257 compounds were purified from the TE including 143 flavonoids, 9 selaginellins and some other compounds. Before the inhibitory activity prediction by SAR models, the preliminary druggability evaluation was carried out to exclude some compounds that have no beneficial property for further drug development process. In this part, we mainly evaluated the molecular weight and two important ADME (absorption, distribution, metabolism, elimination) properties for druggability by corresponding QSAR models developed by our group: logD_7.4 (the distribution coefficients at pH = 7.4) (Wang J. B. et al., 2015; Wang et al., 2017), and logPapp (the Caco-2 membrane permeability) (Wang et al., 2016). Based on previous studies, a good drug candidate should have a logD_7.4 value smaller than 5, a logPapp value larger than −5.15 and a molecular weight smaller than 500. After excluding compounds that perform very poorly in at least two of three aforementioned properties, there were 238 compounds left for further activity screening.

As described before, the inhibitory activity of these 238 compounds were predicted by the multi-target SAR model. The predictive result for a new compound was outputted as a probability value. For each compound that was classified as active ingredient by SAR models, if its probability value >0.5, it is considered to be active, otherwise, it is inactive. From the predictive result, it can be seen that 54 flavonoids and 4 selaginellins present a good inhibitory correlation with MAO-B, 21 flavonoids may show BACE1 inhibitory activity. However, to improve the reliability of prediction, we apply the prediction probability of 0.8 as a cut-off value to select the active compounds for some related targets. As a result, 18 compounds with a probability value larger than 0.8 were extracted. These compounds were prepared for further validation in the inhibitory activity test and their detailed information can be seen in the Supporting Information (Supplementary Material).

Based on the prediction outputs, we focused on the screening of BACE1/MAO-B dual inhibitory activity of flavonoids and selaginellins. The enzyme BACE1 is considered as a prime target to design therapeutics for AD mainly because of that the catalysis process by BACE1 is the rate-limiting step in APP proteolysis and the BACE1 knock-out mice lacking Aβ production survives with normal physiology (Roberds et al., 2001). As the majorβ-secretase enzyme that initiates the generation of Aβ, BACE1 is undoubtedly a prime target for anti-Aβtherapy in AD (Ohno, 2016). The increase of MAO-B activity is associated with gliosis, which can result in higher levels of H₂O₂ and oxidative free radicals (Nebbioso et al., 2012). Thus, the MAO-B inhibitors are potential candidates for anti-AD drugs due to their capacity to regulate neurotransmitters and inhibit oxidative damage in the central nervous system.

Considering the quantity of existing compounds in our laboratory, 13 compounds were chosen to carry out the inhibitory activity validation experiments. Their chemical structures were displayed in Figure 2 and their IC₅₀ values can be seen in Table 3. To evaluate the inhibitory activity of these compounds, a threshold value of IC₅₀ = 10 μM was applied. If a compound has a IC₅₀ values smaller than 10 μM, it would be considered to be active. Otherwise, it is inactive. We can find that nine of them show good inhibition on BACE1 with IC₅₀ values ranged from 0.7454 to 7.578 μM and five of them show good inhibition on MAO-B with IC₅₀ values ranged from 2.913 to 8.813 μM. Among them, S-8, S-5, S-13, and S-12 all have significant dual BACE1/MAO-B inhibitory activities with IC₅₀ values in the micromole magnitude and S-8 has been proved to be the most potent against BACE1 and MAO-B with IC₅₀ values of 0.7454 and 3.619 μM, respectively. Among them, S-5 and S-8 are biflavones, S-12 and S-13 are selaginellins. The inhibitory curves for these four compounds were summarized in Figure 3.

As described before, for each structural conformation of S-5, S-8, S-12, and S-13, a score value was obtained to evaluate the binding affinity between this active compound and each target protein (BACE1 and MAO-B). Generally, a lower score is better. Therefore, the optimal conformation can be decided from a series of generated conformations for each compound according to their score values. Combining the result of ligand interaction, four active compounds and their docking results were listed in Table 4. From the table, we can see that four compounds indeed all have some degree of interaction with BACE1 and MAO-B compared with their original ligands. For BACE1, the most active molecule is S-8 for which the score value is −32.7 and the main binding force is the hydrogen bond force and pi-bond force with ASP (232A) and THR (231A). As to the rest three molecules, the mainly binding force are also the hydrogen bond and pi-bond force with different residues. For MAO-B, the most active molecule is S-5 for which the score value is −44.0 and the main binding force is the hydrogen bond force and pi-bond force with CYS (397A) and GLY (13A). In summary, the molecular docking results were consistent with the results of aforementioned inhibitory experiments that S-8 has the strongest inhibition activity for BACE1 and S-5/S-8 performs better than S-12/S-13 in the inhibition of MAO-B. Therefore, as the conclusion obtained from above in intro inhibitory test, S-5, S-8, S-12, S-13 all have significant dual BACE1/MAO-B inhibitory activities and S-8 promises to be the most potent against BACE1 and MAO-B. The docking results and corresponding 2D ligand interaction diagram of S-8 bound to BACE1 and MAO-B can be seen in Figures 4, 5. The detailed information of all conformations and docking results for other three compounds can be seen in the Supporting Information (Supplementary Material).

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.