Evaluation of Five Integration Effects Based on Factor Analysis
TCM syndromes cannot be completely cured by a single effect frequently, so we explore two or even more effects in an animal model to improve the research efficiency and fitness with actual symptoms. Mice’s performance of the first constipation model, selection of E1 and E2 indexes and their pharmacodynamic changes were basically consistent with our previous discussion [26]. First black stool time, the number of black stools, fecal weights within 12 h, colonic content weights, organ coefficients of colon and stomach, MTL, SS, VIP and AchE could reflect E1 (Figure S2). Expression levels of TG, Na+-K+-ATPase, TNF-α, IL-1β and grayscale ratios of the aforementioned 9 colonic inflammatory proteins could represent E2 (Figure S3).
The treatment of the second blood stasis syndrome induced by noxious heat also included two aspects of both cooling and activating blood. Model mice appeared obvious symptoms of blood stasis with enlarged spleens and decrescent thymuses (P<0.001), while drug groups’ spleens recovered to some extent. On the one hand, we took anal temperatures, HSP-70, SOD, NO, TNF-α, IL-1β, IL-6 and the same 9 levels of colonic inflammatory proteins as E3 indicators (Figure S4), which resembled E2 that played an anti-inflammatory role from the perspective of Western medicine. Mice maintained opposite hypothermia (P<0.001) in febrile response to LPS, and some drug groups were able to significantly antagonize the toxic reaction. Compared with the model, most groups down-regulated the contents of inflammatory factors (P<0.05), especially phosphorylated proteins and TLR4 (P<0.001) in high-proportioned ethanol-extracted rhubarb groups. LPS could act through the activation of NF-κB signaling pathway mediated by TLR family in connection with pro-inflammatory cytokines [30]. On the other hand, indicators of E4 (Figure S5) covered TT, PT, APTT, FIB, TXB2, 6-keto-PGF1α, ratios of TXB2 to 6-keto-PGF1α, PGE2, ET-1, Mg2+ and Ca2+. The positive group had a better improvement on coagulation four indices, but most of rhubarb groups only showed significant callback to PT and FIB (P<0.05). Ethanol-extracted rhubarb corrected the imbalance of elevatory TXB2/6-keto-PGF1α (P<0.001), which facilitated blood circulation [29]. Prostaglandins could dilate blood vessels whereas endothelins constricted them, which all displayed callback tendency of varying degrees in drug groups.
With regard to the third cholestasis model, as shown in Figure S6, model mice performed visibly declining weights (P<0.01), darkening serum color, intumescent gallbladders and inky bile accompanied by obvious liver damage of which there were dense bleeding points on the surface. The overall state of the positive group presented relatively well recovery, and these lesions also improved to a certain degree in part of rhubarb-treated groups. Whereupon we investigated measurement indexes associated with jaundice including organ coefficients of liver and gallbladder, T-SOD, MDA, GSH, tissue Fe3+, ALT, AST, ALP, serum Fe3+, GST, GGT, TBIL, DBIL and TBA to evaluate E5 (Figure S7). The level regulation of oxidative stress in liver tissues (P<0.05) by rhubarb decocted for a long time might be one of the main mechanisms of removing jaundice. Serum liver function levels were generally elevated in model mice and decreased prominently (P<0.05) in most groups after administration except GGT and DBIL.
Factor analysis was carried out for the above pharmacodynamic indexes of five effects to obtain their integration effect values respectively. Referring to the approach we had described [26], comprehensive weight scores of each group were calculated from the scoring coefficient matrix of indicator variables (Tables S6-10), and then Kolmogorov-Smirnov normal distribution test was utilized to examine the results (Tables S11-15). At last, as illustrated in Table 2, there existed significant differences between the model and control groups. The therapeutic efficacy of positive drugs emerged very obviously in all five effects while the effectiveness of rhubarb with different preparation was not exactly the same. E1 was conspicuous in groups of 10% EW-S ~ 35% EW-L and 80% EW-L ~ ethanol-L. High-concentration ethanol groups possessed better E2 on the whole. E3 of rhubarb extracted by over 20% EW was all evident, particularly for a long time, which conformed to the TCM theory that steaming with wine could be good at clearing blood-aspect heat toxin. In contrast, E4 of rhubarb was relatively little and only worked when it was prepared with water or more than 80% ethanol. 10% EW-L ~ ethanol-L groups showed prominent E5 among which three groups of 35% EW-L, 50% EW-L and 65% EW-L were the best, suggesting that “wine and water co-decoction” might be more conducive to exert the curative effect.
Table 2
Integration effects of rhubarb on the basis of factor analysis (n = 6~8, mean ± SD).
Groups
|
E1
|
E2
|
E3
|
E4
|
E5
|
Control
|
(2.08 ± 0.28)
|
(-2.53 ± 1.96)
|
(-1.24 ± 3.53)
|
(1.89 ± 0.60)
|
(-1.07 ± 2.78)
|
Model
|
(-5.48 ± 1.96)###
|
(5.39 ± 3.77)###
|
(4.40 ± 3.62)#
|
(-0.89 ± 1.86)##
|
(5.40 ± 2.36)###
|
Positive
|
(1.65 ± 3.91)***
|
(0.01 ± 2.12)***
|
(-0.88 ± 2.32)**
|
(1.70 ± 1.78)*
|
(-0.48 ± 3.09)**
|
water-S
|
(-0.58 ± 2.56)***
|
(-0.32 ± 3.04)**
|
(2.65 ± 3.50)
|
(0.62 ± 1.77)
|
(3.60 ± 1.53)
|
water-L
|
(-0.35 ± 1.67)***
|
(0.12 ± 0.42)**
|
(1.22 ± 2.90)
|
(1.92 ± 2.34)*
|
(2.70 ± 3.69)
|
10%-S
|
(0.87 ± 0.80)***
|
(-0.83 ± 1.89)***
|
(2.27 ± 2.72)
|
(1.60 ± 3.66)
|
(3.22 ± 1.77)
|
10%-L
|
(0.02 ± 2.53)***
|
(0.30 ± 0.38)**
|
(0.91 ± 2.04)
|
(0.77 ± 0.63)
|
(1.10 ± 4.00)*
|
20%-S
|
(0.29 ± 2.42)***
|
(0.64 ± 3.21)**
|
(1.65 ± 3.36)
|
(0.08 ± 1.24)
|
(1.27 ± 3.26)*
|
20%-L
|
(0.52 ± 2.09)***
|
(1.93 ± 2.98)*
|
(-0.82 ± 3.59)*
|
(0.01 ± 1.80)
|
(0.45 ± 2.58)**
|
35%-S
|
(0.09 ± 3.00)***
|
(1.22 ± 3.76)*
|
(-0.60 ± 4.00)*
|
(-0.98 ± 2.97)
|
(0.35 ± 3.84)**
|
35%-L
|
(0.80 ± 2.55)***
|
(1.48 ± 3.52)*
|
(-0.50 ± 1.44)*
|
(-0.33 ± 0.74)
|
(-0.97 ± 1.73)***
|
50%-S
|
(-0.55 ± 2.25)***
|
(-0.11 ± 2.62)**
|
(-0.27 ± 3.83)*
|
(-0.94 ± 3.69)
|
(-0.21 ± 3.89)**
|
50%-L
|
(0.13 ± 2.98)***
|
(1.10 ± 2.48)*
|
(-0.82 ± 1.42)*
|
(-0.17 ± 2.32)
|
(-1.01 ± 2.29)***
|
65%-S
|
(-1.84 ± 2.63)**
|
(-1.10 ± 3.09)***
|
(-0.61 ± 2.24)*
|
(-0.69 ± 3.38)
|
(-0.32 ± 2.39)***
|
65%-L
|
(-0.29 ± 3.30)**
|
(-0.53 ± 2.24)***
|
(-1.92 ± 2.01)**
|
(-0.66 ± 2.06)
|
(-1.55 ± 3.29)***
|
80%-S
|
(-0.67 ± 3.18)**
|
(-1.38 ± 2.57)***
|
(-0.12 ± 3.36)*
|
(-0.52 ± 2.21)
|
(0.14 ± 3.88)**
|
80%-L
|
(0.47 ± 3.49)***
|
(-0.76 ± 3.02)**
|
(-1.71 ± 2.02)**
|
(0.89 ± 2.31)
|
(-0.37 ± 3.48)**
|
90%-S
|
(1.00 ± 3.00)***
|
(-1.43 ± 2.71)***
|
(-2.25 ± 2.69)**
|
(1.58 ± 1.49)*
|
(-0.76 ± 3.76)**
|
90%-L
|
(0.47 ± 2.38)***
|
(-2.11 ± 1.25)***
|
(-2.40 ± 3.29)**
|
(0.66 ± 1.79)
|
(-0.42 ± 3.00)***
|
ethanol-S
|
(1.42 ± 3.08)***
|
(-3.39 ± 1.99)***
|
(-1.96 ± 2.34)**
|
(1.22 ± 2.91)
|
(-0.49 ± 3.86)**
|
ethanol-L
|
(2.19 ± 3.40)***
|
(-2.61 ± 0.61)***
|
(-1.25 ± 1.54)**
|
(0.98 ± 1.44)
|
(0.08 ± 3.96)**
|
# P<0.05, ## P<0.01, ### P<0.001 compared with the Control; |
* P<0.05, ** P<0.01, *** P<0.001 compared with the Model. |
The Establishment of Quantity-Effect Correlation Method on the Basis of BP Neural Network
The advanced BP-ANN mimics neurons based on the structure and function of the biological brain with high prediction accuracy and optimization ability [31, 32]. It can be intelligently applied in the fitting correlation analysis to help us determine the contribution degree of each chemical component to the corresponding efficacy. So, we tried to objectively associate as many components detected with efficacies as possible. The reasonable quantity-effect correlation method was established by repeatedly debugging parameters as follows.
(1) The number of hidden layers: Theoretically, the more hidden layers, the stronger the ability to fit functions is. However, in fact, many layers are apt to cause overfitting as well as increase training difficulty of convergence. On the premise of setting 108 nodes in the input layer and outputting 1 node, we explored the fitting consequence at 1, 2 and 3 layers. The number of nodes in different hidden layers was randomly set to 10, and relative errors between the predicted and true values were consequently compared by means of repeat for 20 runs. It could be seen in Figure 3 that, although the relative error values of double hidden layers were not optimal, they were more stable and not easy to appear local extremum. Three hidden layers had maximum negative values combined with observing the predicted values of test samples (Table S16). So, we adopted 2 hidden layers for the next step.
(2) The number of neurons per hidden layer: According to empirical formulas and the setting principle of hidden neurons, in the range of 108 input nodes to 1 output node, the number of neurons was set every 20 values (i.e., 20, 40, 60, 80, 100), and then added 5, 10 and 15 from 0 to 20, 70 and 75 from 60 to 80, 130, 160 and 190 behind 100. The second hidden layer kept pace with the first one. Five sets of data from three test samples were acquired by 5 runs under each number of neurons (Table S17). Synthetically in view of the predicted, true and their relative error values, we set the number of neurons at 20 due to its more steady relative errors on the whole.
(3) Learning rate: In general, it tends to choose small adaptive learning rates to ensure the stability of the system, but too small learning rates will lead to a very long training time. In order to obtain better results, we ran test samples for five times severally when the learning rate was 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7 and 0.8 (Table S18). It was found that there were no significant trend changes. Relatively speaking, the learning rate of 0.02 could make BP algorithm with gradient descent preferable.
(4) Running effect: We adjusted the display frequency to 100, training times to 1000 and the minimum error of training target to 0.0000001 for further accurate prediction of this method. After repeated debugging, the final running program was demonstrated in Figure S8 (a). All training samples and corresponding integration effect values were substituted into respective optimized neural networks. The training would stop when the mean squared error was less than its set point or the gradient reached its set value, and it was forced to end as well when errors did not decrease but rise continuously for 6 times by validation checks of the generalization capability. As a result, the fitting correlation effects were all excellent (Figure 4).
(5) Decision weight calculation: The above neural network training results only reflected the relationship among neurons. What’s more, we needed to clarify the relative importance of each input factor to output information. Hence, the following neural network learning algorithms were sequentially used to work out decision weights of input components.
(The input layer: n = 108, the first hidden layer: m = 20, the second hidden layer: p = 20, the output layer: l = 1.)
Impact of the o unit in the second hidden layer on the k output relative to all units:
\({F}_{ok}=\left|{w}_{ok}\right|/??\left|{w}_{ok}\right|\) o = 1, 2…p, k = 1, 2…l
Impact of the j unit in the first hidden layer on the o unit in the second hidden layer relative to all units:
\({F}_{jo}=\left|{w}_{jo}\right|/??\left|{w}_{jo}\right|\) j = 1, 2…m, o = 1, 2…p
Impact of the i unit in the input layer on the j unit in the first hidden layer relative to all units:
\({F}_{ij}=\left|{w}_{ij}\right|/??\left|{w}_{ij}\right|\) i = 1, 2…n, j = 1, 2…m
Thus, impact of the i input on the k output:
$${F}_{ik}={F}_{ij}*{F}_{jo}*{F}_{ok}$$
Decision weights of input components could be expressed as:
$${S}_{i}={F}_{ik}/?{F}_{ik}$$
The Main Bioactive Components for Five Effects on the Basis of Association Analysis of Components and Effects
We analyzed the components whose contribution was greater than 0.01 ranking from the ultimate decision weight Table S19, which were 30 in E1 (Nos. 42 > 21 > 35 > 37 > 34 > 46 > 11 > 98 > 36 > 27 > 59 > 88 > 89 > 18 > 108 > 78 > 105 > 72 > 6 > 55 > 106 > 45 > 61 > 51 > 90 > 32 > 41 > 95 > 38 > 107), 28 in E2 (Nos. 42 > 7 > 107 > 83 > 29 > 28 > 61 > 98 > 108 > 8 > 80 > 35 > 58 > 81 > 40 > 15 > 92 > 33 > 106 > 36 > 50 > 79 > 11 > 100 > 57 > 19 > 101 > 39), 35 in E3 (Nos. 24 > 18 > 12 > 86 > 72 > 37 > 51 > 80 > 26 > 55 > 34 > 21 > 16 > 46 > 97 > 42 > 22 > 33 > 88 > 28 > 92 > 27 > 2 > 67 > 70 > 59 > 36, 44 > 90 > 1 > 82 > 6 > 73 > 4, 30), 32 in E4 (Nos. 40 > 81 > 56 > 75 > 69 > 53 > 9 > 64 > 100 > 58 > 49 > 29 > 74 > 83 > 90 > 52 > 63 > 42 > 66 > 106 > 67 > 48 > 31 > 62 > 71 > 89 > 87 > 92 > 20 > 60 > 46 > 36) and 33 in E5 (Nos. 19 > 55 > 62 > 12 > 33 > 52 > 22 > 35 > 87 > 64 > 69 > 95 > 59 > 42 > 88 > 46 > 43 > 25 > 107 > 72 > 29 > 104 > 71 > 94 > 89 > 99 > 92 > 81 > 36 > 45 > 58 > 108 > 102). Table 3 summarized the total contribution of each category of rhubarb components. Thereinto, the top 10 of total weights in five effects and of each effect should be focused on respectively (Table S20-25).
Table 3. Total contribution of each category of rhubarb components with weights over 0.01.
The hypothesis of “additive effect” of medicinal herbs was put forward in 2014 [33, 34]. That is, the assemblage or superposition of “effective forms” is the core chemical essence of medicinal herbs. Compounds with identical mother nuclei structure are grouped together and may have the same pharmacological target. Both the “additive effect” of various components on single target and the “synergy effect” on multi-targets, even universality and individuality characters of different efficacies, can be well reflected in the Table 3 above, which is also the essence of fuzzy identification our team proposed for new interpretations to pharmacodynamic components and action mechanisms of a medicinal herb.
In this study, we quantified the “additive effect” of these chemical groups. Through specific values, we can see that combined anthraquinones, flavanol and its polymers might be the universality character to the multi-functional properties of rhubarb, reflecting the correlation among multiple efficacies to some extent. For example, chrysophanol glucosides rank the 18th, 4th, 4th, 8th and 10th severally in E1~5. Their aglycones, namely free anthraquinones, possess strong bacteriostasis and notably affect the activation of lipid inflammatory mediators [35]. There are currently many evidences that rhubarb anthraquinones can treat constipation, ischemic lesions [36], sepsis [37] and other inflammation. Especially now in the suppression of hepatobiliary diseases [38, 39], Kehuanglidan Capsule has completed the treatment of chronic viral hepatitis jaundice in the drug clinical trial registration and information public platform (www.chinadrugtrials.org.cn), and its dominant ingredients are the rhubarb anthraquinones. Some literature speculated that these anthraquinone glycosides might revert to free type ones through metabolic transformation in vivo to exert various pharmacological actions. In addition, flavanols such as Cianidanol and (-)-Epicatechin rank in the top 10 of E1~5. Catechins contribute to improve metabolic changes caused by a high-fat diet [40], and can maintain normal blood circulation by reducing the adhesion of platelets [41] and efficaciously shorten icteric period by lowering serum bilirubin levels [42]. Procyanidins are the polymers of flavanols. They also have great contribution for E1, E2, E3 and E5, showing protective effects on the digestive system for the treatment of gastrointestinal diseases, cholecystitis, constipation or diarrhea, etc. In particular, oligomeric procyanidins are the most in the top 10 of E2, which are the internationally recognized natural antioxidant to scavenge free radicals.
With regard to individuality characters of different efficacies, stilbene glycosides are important to E2~5 that are closely linked to anti-bacteria, antiphlogosis and other aspects [43]. Stilbene compounds are always a momentous part of rhubarb chiefly for anti-hyperlipidemia and antioxidation [38]. Resveratrol glucosides are the representative ranking high in these efficacies, which can alleviate LPS stimulation by inhibiting TNF-α and macrophages from producing NO [44]. Sennosides belong to anthranone dimers not only for defaecation (E1), but also for detoxifying to treat bacillary dysentery, epidemic hemorrhagic fever (E3) and postpartum milk return in clinic (E4). Free anthraquinones, mainly Emodin, Physcion, Rhein or their derivates, act directly in E1 and E5. Chromones appear in the top 10 of E1 and E2, which have been proven to remove cholesterol, inflammation and so on [38]. A kind of gallotannins ranks the 9th in E3. Gallotannins are generally characterized by potent antivirus, astringency and hemostasis [38]. E3 of rhubarb is related to clinical blood loss syndrome, and grey relational analysis has attested to a big impact of tannins and anthraquinones on the hemostatic function [45]. Butyrylbenzene and its glycosides may play a certain protective role on the cardiovascular system (E4).
To sum up, from the perspective of both components whose contribution > 0.01 and the top 10 components, we revealed the universality and individuality characters of five effects. As intuitively shown in Figure 5, combined anthraquinones, flavanol and its polymers might be the universality character to the multi-functional properties of rhubarb while their individuality characters were listed severally. There are some inadequacies of lacking experimental validation on effective constituents, which would be designed in our subsequent work. And if these active components act alone or exist in other drugs, whether they can also contribute to the same efficacy still needs to be further studied.