Directional preparation of anticoagulant-active sulfated polysaccharides from Enteromorpha prolifera using artificial neural networks

The sulfated polysaccharides from Enteromorpha prolifera (PE) are a potential source of anticoagulant agents. In this study, the PE was degraded by specific degradase and five hydrolysis products with different molecular weights were prepared. The product of 206 kDa is a kind of high rhamnose-containing polysaccharide with sulfate ester (34.29%). It could effectively prolong the activated partial thromboplastin time (APTT), which indicated inhibition of the intrinsic coagulation pathway. The artificial neural network (ANN) was built to realize the directional preparation of anticoagulant-active polysaccharides. Based on monitoring glucose concentration on-line, a visualization system of enzymatic hydrolysis was developed to simplify the operation of ANN. The model could be further applied to predict molecular weights of polysaccharides that possess diverse biological activities.

Anticoagulant activity assay. We prepared five low molecular weight fractions with the method of enzymatic hydrolysis. The anticoagulant activities of all the samples were determined with the classical coagulation APTT, TT, and PT assays in vitro that characterize different stages of the coagulation process. APTT is employed to evaluate the coagulation factors in the intrinsic blood coagulation pathway. PT is used to characterize the extrinsic coagulation factors while an increase in the TT suggests either thrombin inhibition or impaired fibrin 20 . In fact, there was a lack of PT and TT activities in all of the samples.
As seen in Fig. 1, the APTT were prolonged by PE, PE1, PE2, and PE3 in a concentration-dependent manner. The PE3 exhibited the most remarkable APTT activity and the clotting time was 60.56 s at 100 μg/mL. The prolongation of APTT suggests inhibition on the intrinsic coagulation pathway.
The HPGPC chromatogram of PE3 is illustrated in Fig. 2(A). The HPGPC chromatogram appeared as a single and symmetrical peak, which demonstrated PE3 was a kind of homogeneous polysaccharide. The FTIR spectrum of PE3 is shown in Fig. 2(C). The absorption bands at 3433 and 2927 cm −1 were generated by the stretching vibration of O-H and stretching vibration of C-H, respectively. The bands at 1635 and 1420 cm −1 suggested the stretching vibration of C=O and C-O of a carboxyl group, indicating the presence of uronic acid 21 . The bands at 1254 and 853 cm −1 were attributed to the stretching vibration of S=O and C-O-S, and indicated the existence of sulfate esters 22 .
The structural characteristics of the sulfated polysaccharides play an important role in the understanding of anticoagulant activity 23 . According to Kamide et al. 24  significantly increased with the sulfate ester in the C-2 and C-3 sites, but not in the C-6 site. Li et al. 23 reported an anticoagulant-active polysaccharide from Monostroma angicava and sulfate esters substitution occurred at C-3 of (1 → 2)-linked-L-rhamnose residues. Wang et al. 25 prepared three anticoagulant-active polysaccharides from E. linza, with sulfate esters residing at the C-3 of (1 → 4)-linked-L -rhamnose residues. Our research group reported that the backbone of the polysaccharides from E. prolifera consisted of D-  25 reported sulfated rhamnose was possible the anticoagulant compound in green algae. The data above suggest that the anticoagulant activity of sulfated polysaccharide was related to sulfate ester position, monosaccharide composition, and glycosidic linkage. Moreover, the content of sulfate esters also plays an important role in anticoagulant activities. It was reported that anticoagulant activity of the sulfated polysaccharides was partially caused by the strong interaction between the negatively charged sulfate esters and some positively charged peptidic sequences of proteins involved in coagulation process. It is known that Lysine and Arginine residues were present in the heparin-binding site 27 . Therefore, it is possible that the high level of sulfate ester (34.29%) increased the anticoagulant activity of PE3.
The molecular weight of sulfated polysaccharide is another important factor influencing its anticoagulant activity. It was reported that the sulfated polysaccharides from Monostroma latissimum required longer chains to achieve the inhibition of thrombin 28 . On the other hand, the spatial arrangement of sulfate ester was reported being important for anticoagulant activity 29 . Therefore, the spatial arrangement of PE3 might be the advantageous conformation and it could promote a better interaction with blood coagulation factors. As can be seen in Fig. 1, PE4 and PE5 with low molecular weights showed no prolongation of APTT. We assumed that some structural change, such as a shift in conformation, caused the extension of enzymatic hydrolysis time.
In conclusion, the potent anticoagulant activity of PE3 depended on the integration of many factors. Realization of accurate preparation of PE3 would enhance its application in pharmaceutical industries.
Artificial neural network for predicting molecular weight distributions. The ANN is a parallel processing network which determined the complex nonlinear relationships between independent and dependent variables 30 . The back-propagation (BP) network is one of the most widely used ANN for multilayered feed-forward networks 31 . The learning rule of BP network is to use the steepest descent method to continuously adjust the weights and thresholds 32 . In this work, a feed-forward neural network trained with an error BP algorithm was employed in constructing and training the ANN. It is noteworthy that the Levenberg-Marquardt algorithm was applied in the training process. This is a combination of the Grade method and the Gauss-Newton method. It was reported that when such an algorithm is used, time-consuming search can be significantly reduced 33 .
According to Kolmogorov theorem, a three-layer network is sufficient to complete any n-dimensional to m-dimensional nonlinear mapping, thereby setting only one hidden layer structure 34 . In the network, enzyme concentration, substrate concentration, enzymatic temperature, enzymatic time, and glucose concentration were assigned as five input layer nodes while molecular weight was output layer node (Fig. 3). The input layer provides the hidden layer with the sum of the weighted input parameters. These weights are adjusted automatically by the BP algorithm during training.
Different combinations of the transfer functions (logsig, tansig, and purelin) were studied to determine the best combination of two transfer functions that will yield accurate results ( Table 2). The performance of these combinations was evaluated on the basis of mean absolute percentage error (MAPE). The MAPE of 0-10%, 10-20%, 20-50%, and >50% represents high prediction accuracy, good prediction, normal accuracy, and bad accuracy, respectively 35 . As can be seen in Table 2, the MAPE of all the combinations is in the range of 0-10%. The tansig and purelin were used as the transfer functions for input and output layer, respectively, because of the minimum MAPE (2.14%). The mean square error (MSE) versus the number of repetitions for training is illustrated in Fig. 4(A). In this case, the MSE of training data is equal to 0.00016, after 173 epochs. As shown in Fig. 4(B), the R value is 0.99923. The model fits well with the actual data when R approaches to 1 36 . Histogram of deviation margin for the optimal ANN is presented in Fig. 4(C). It can be seen that deviation distribution is concentrated around zero, which indicates that the ANN has a remarkable accuracy. A comparison between experimental data and outputs of ANN model is shown in Fig. 4(D). Overall, there is a good fitting between the estimated values and the actual values.
The 10% of the testing sets (Table 3) were used for testing the precision of the ANN model. The mean absolute percentage error (MAPE) of molecular weights for testing sets is 1.86% (Table 3). According to Khoshnevisan et al. 37 , a MAPE value of less than 10% indicates that the best prediction has been achieved. Therefore, this network could be used with high reliability to estimate the molecular weights. The testing performance and regression of ANN    5). As can be seen in Fig. 5(A), testing was stopped after 150 epochs with MSE value of approximately 0.0034. As shown in Fig. 5(B), the R value is 0.99558, which implies good fits between predicted values and the actual values. Therefore, this network could be used with high reliability to estimate the molecular weights. When enzymatic hydrolysis condition is changed, the users need to run MATLAB software and modify parameters in the procedure. However, this will make the manipulation complicated. In order to solve the problem, the visual prediction system was developed to create a convenient user interface (Fig. 6). According to Li et al. 38 , glucose was one of the hydrolysates. Glucose concentration could reflect the enzymatic hydrolysis process to a certain extent. Consequently, the ANN model could predict molecular weight of the sulfated polysaccharide efficiently by monitoring glucose concentration on-line.  Table 3. Comparison of actual and predicted molecular weight on testing sets. Extraction and quantification of PE. The PE was extracted with the improved method as described by Li et al. 38 . The milled alga (60 g) was dipped into 40 volumes of tap water, homogenized and extracted at 100 °C for 1 h. The water extraction solution was centrifuged (4000 × g, 10 min), and the supernatant was dipped in 2 times volume of 95% ethanol to remove pigment. After centrifugation, the sediment was collected, and dissolved in distilled water.
Preparation for degradase for PE (DPE) and Enzyme assay. Alteromonas sp. A321 was cultivated and its extracellular supernatant was extracted according to Li et al. 38 . Then, the supernatant was brought to 60% (W/V) saturation with (NH 4 ) 2 SO 4 . The precipitate collected was dissolved in 20 mM phosphate buffer saline (pH 7.5), yielding DPE. The degradase activity was assayed by the method described by Fenice, Selbmann, Zucconi, and Onofri 39 . Characterization. The molecular weights of PE, PE1, PE2, PE3, PE4, and PE5 were determined by using an Agilent 1260 HPLC system (Wilmington, USA) equipped with PL Aquagel-OH 50 column (0.75 cm × 30 cm) and a refractive index detector (RID). The column was eluted with 0.2 M NaNO 3 and 0.01 M NaH 2 PO 4 at a flow rate of 0.8 mL/min. The molecular weights of all samples were estimated by referencing the calibration curve made from dextran standards. The total sugar was determined by phenol-sulfuric acid method using glucose as standard 40 . Sulfate ester content was estimated according to the method described by Therho and Hartiala 41 . The monosaccharide Anticoagulant activity assays. Activated partial thromboplastin time (APTT), thrombin time (TT), and prothrombin time (PT) assays were carried out according to the method described by Pawlaczyk et al. 43 Saline and the unfractionated heparin were used as negative and positive controls, respectively. All clotting assays were performed with three individual replicates to estimate the mean clotting time. The anticoagulant activity was expressed as the clotting time.

Preparation of polysaccharides with different molecular weights.
Experimental design. Enzymatic temperature (25,30, and 35 °C), enzymatic time (1, 2, 3, 4, 5, 6, 7, 8, 9, 10, and 11 h), substrate concentration (4, 6, 8, 10, and 12 mg/mL), and enzyme dose (8.10, 9.72, and 12.15 U) were considered according to the situation of enzymolysis. In order to construct and train the ANN model, two hundred and thirty-one experimental groups were designed referring to the uniform orthogonal design method. ANN modeling. The training of ANN was accomplished by adapting the strengths or weights of the connections among the input, intermediate and output neurons, which were capable of storing memory and information. By achieving the learning ability, ANN produced the desired responses according to the given decision variables 44 .
In this study, the ANN model was developed using the neural network toolbox in MATLAB version 8.1. The 231 experimental data points were employed to build the ANN model. They were divided into training and testing sets with a ratio of 90% and 10%, respectively. Flowchart of the optimal neural network is presented in Fig. 7. The equation below was used for data normalization: where X i , X, X min , and X max represented the normalized value, actual value, minimum value, and maximum value, respectively. The BP network and Levenberg-Marquardt optimization algorithm was applied in the training process. The MSE between the output values and targeted values are calculated and returned back to the hidden layer. Weights between inputs and hidden layer are also calculated, which are updated between the hidden layer and the output layer. Statistical analysis. All experiments were repeated in triplicate, and the data were reported as mean ± standard deviation and evaluated by one-way ANOVA. Statistical analysis was performed using MATLAB mathematical software (version 8.1).