Activity Prediction and Molecular Mechanism of Bovine Blood Derived Angiotensin I-Converting Enzyme Inhibitory Peptides

Development of angiotensin I-converting enzyme (ACE, EC 3.4.15.1) inhibitory peptides from food protein is under extensive research as alternative for the prevention of hypertension. However, it is difficult to identify peptides released from food sources. To accelerate the progress of peptide identification, a three layer back propagation neural network model was established to predict the ACE-inhibitory activity of pentapeptides derived from bovine hemoglobin by simulated enzyme digestion. The pentapeptide WTQRF has the best predicted value with experimental IC50 23.93 μM. The potential molecular mechanism of the WTQRF / ACE interaction was investigated by flexible docking.


Introduction
Hypertension is a risk factor for cardiovascular diseases including coronary heart disease, peripheral artery disease and stroke. [1,2] Recently, several food-derived bioactive peptides have been found playing a significant role in decreasing blood pressure. Therefore, more and more attention has been paid to peptides from food sources with antihypertensive activity. [3] Most of the antihypertension peptides regulate blood pressure by inhibiting the activity of angiotensin I-converting enzyme (ACE, EC.3.4.15.1). ACE is a zinc-and chloride-dependent metallopeptidase, which belongs to the M2 family of zinc metallopeptidases. [4,5] It converts angiotensin I to angiotensin II (a potent vasoconstrictor) as well as inactivates the vasodilator bradykinin. [6] ACE plays a crucial role in the renin-angiotensin system (RAS), which is well known for its regulation of blood pressure and fluid homeostasis. [7,8] Nowadays, inhibitors of ACE have been considered as first-line therapy for hypertension. [9,10] It has been reported that a number of bioactive peptides, which derived from food sources, have ACE-inhibitory activity. Chibuike C. Udenigwe [11] summarized the major approaches in bioactive peptides research as the classical approach, the bioinformatics approach and the integrated approach. This classification is also suitable for ACE-inhibitory peptides.
In order to circumvent some challenges of the classical approach, the bioinformatics approach has been recently applied towards the discovery of ACE-inhibitory peptides encrypted in food proteins. This approach was recently used to study the distribution of ACE-inhibitory peptides within the primary structure of typical food proteins. [22] Following the identification of bioactive peptides from protein sets by bioinformatics in databases populated following the classical approach, the remainder of the purportedly "inactive" peptides can be analyzed in silico to identify structural patterns that have previously been associated with known bioactivities. [11] Moreover, the strengths of each approach can be combined as deemed fit to enhance the discovery and use of ACE-inhibitory peptides. Bioinformatics software can be used to simulate proteolytic specificities of enzymes in order to establish the peptide database in silico. [11] Quantitative structure-activity relationship (QSAR) studies are widely undertaken for modeling the bioactivities such as the bioactivity of ACE-inhibitory peptides [22][23][24][25][26] and the sweetness of compounds [27]. Neural network, as a kind of artificial intelligence, has been applied to modeling non-linear systems, simulating the chaos bioprocess and predicting the results. It turns out to have higher modeling accuracy and generalization capacity [28] and becomes a potentially effective tool in modeling the QSAR.
Bovine blood, as a by-product generated in great volume in industrial abattoirs, gives rise to several possibilities for their recovery and use. The use of bovine blood as a food component has been widely reported due in part to their high nutritional value and there are many studies have recently demonstrated that bovine blood proteins can be used to obtain bioactive peptides. This is important because it gives an added value to bovine blood. A number of bioactive peptides released from bovine hemoglobin hydrolysates have been reported. [29][30][31][32] It is a highly desirable but difficult task to identify bovine blood derived peptides. In this aspect, prediction model would be a useful technique to highlight potential ACE-inhibitory peptides identification.
It was found that ACE has two homologous domains (the N-domain and the C-domain), each containing an active center. [33] The C-domain of ACE has been proved to be the dominant angiotensin-I converting site, which has a conserved HEXXH zinc-binding motif, for controlling blood pressure and serving cardiovascular functions. [34,35] If inhibitory peptides occupied the active site of the C-domain of ACE and bound to specific amino acid residues, ACE will lose its activity. Therefore, it is possible to reveal the ACE inactivation mechanisms by analyzing the structural consequences of ACE-inhibitor interactions. [34] In this study, we established a database of potential pentapeptides derived from bovine hemoglobin by simulated enzyme digestion, and a prediction model of ACE-inhibitory pentapeptides by back propagation neural network (BPNN). The peptide with best predictive value was synthesized and its IC 50 of ACE was measured. We also sought to elucidate the potential molecular mechanism of how the peptide with best predictive value exerts its ACE-inhibitory effects by automated molecular docking.

Materials and Chemicals
Angiotensin converting enzyme from rabbit lung, hippuryl-L-histidyl-L-leucine, hippuric acid and HPLC grade acetonitrile, trifluoroacetic acid were purchased from Sigma Chemical Co. (St. Louis, Mo., U.S.A.). All the other reagents were analytical grade.

Molecular Descriptors Calculation
Ten molecular descriptors were calculated for analogs of peptides including structural, spatial, thermodynamic and electronic. Structural descriptors included the number of rotatable bond (Rotbond), the number of hydrogen bond acceptor (H bond acceptor), the number of hydrogen bond donor (H bond donor), the number of ring (Ring) and the number of arcomatic ring (AR). [49][50][51] The molecular fractional polar surface area (PSA) [52] was used as spatial descriptor. The thermodynamic descriptors were taken to describe the hydrophobic character (ALog P: logarithm of the partition coefficient in octanol / water and Log D: logarithm of the partition coefficient in octanol / water in pH 7.4) and refractivity (MR: molar refractivity). [49] The atomic polarizabilities (Apol) were calculated as electronic descriptors. [50] Peptide structures were generated by Accelrys Discovery Studio 3.5 software (Accelrys Inc., San Diego, USA) and the energy was minimized with CHARMm program using steepest descent and conjugate gradient techniques. All descriptors were calculated by calculate molecular properties protocol.

Modeling by Back-Propagation Neural Network (BPNN)
To construct a BPNN, the Levenberg-Marquardt algorithm [53] was used to train the network. With all data normalized, the descriptors and the pIC 50 of dataset were introduced as input and output values, respectively. In order to avoid overfitting, the dataset was randomly classified for training, validation, and test sets. To verify the suitable number of nodes in hidden layer and transfer function, the mean square error (MSE) and pearson correlation coefficient (R) of each BPNN model were calculated.

Peptide Synthesis
The peptide was synthesized (ChinaPeptides Co. Ltd., Shanghai, China) by the solid phase procedure peptide using FMOC protected amino acids synthesis. The purity (95%) and the molecular masses of the peptides were determined by HPLC and mass spectrometry.

Assay for ACE-Inhibitory Activity
The ACE-inhibitory activity of the peptide with best predict value was measured according to Liu, et al. [54] Docking The 3-D structure of human ACE was imported from Protein Data Bank (1O8A.pdb, a crystal structure of human ACE). The structure of peptide was generated by DS and the energy was minimized with CHARMm program. Automated molecular docking was performed using the flexible docking tool of DS software in the presence of cofactors (zinc and chloride ions). The binding site with a radius of 15 Å and a coordinates x: 45.0463, y: 38.6842, z: 45.8268. Evaluation of the molecular docking was performed according to the scores and total potential energy in order to obtain the best pose of peptide. The DS software was used to identify the hydrogen bonds as well as the hydrophobic, hydrophilic, electrostatic and coordination interactions between residues present within ACE active site.

Structural Features Analysis
Most of the ACE-inhibitory peptides derived from egg white hydrolysates in our lab previously [55][56][57] were pentapeptides. Therefore, we are interested in ACE-inhibitory pentapeptide and try to build a prediction model based on a pentapeptides database (Table 1). It shows that 75% of the pentapeptides taken from literature have 1 or 2 hydrophobic amino acid residues. A half of these pentapeptides contain aromatic or hydrophobic amino acid residues at the C-terminus, which are beneficial to ACE-inhibitory activity. [24] In addition, 66.67% of these pentapeptides contain hydrophilic amino acid residues in the second amino acid residues from the C-terminus. The aforementioned common structural features can be used to instruct ACEinhibitory peptides' design. The molecular descriptor (Tables 2, 3) calculation is a logic and mathematical procedure which converts the chemical information of a molecule into some useful data. Most of the number ranges of descriptors of Pentapeptides derived from hemoglobin simulated enzyme digestion included in the scope of the reported pentapeptides, except the molar refractivity of GHGAK. The range of MR of reported pentapeptides is 111.291 to 186.874, while the GHGAK is 107.89. Nevertheless, the predicted IC 50 of GHGAK is the fourth lowest among the seven hemoglobin pentapeptides, because the BPNN is a nonlinearity model, leading to a nonlinear relationship between the input and the output.

BPNN model Building
There were three layers in the BPNN model (input layer, hidden layer and output layer), while the number of nodes in hidden layer was not certain. So that we trained BPNN models with different number of nodes in hidden layer and different transfer functions to obtain the suitable model structure (Fig. 1, Fig. 2 and S1 Table, S2 Table). For the log-sigmoid & purelin transfer function, the highest MSE appears when the number of hidden layers is twelve. However, in terms of the tan-sigmoid & tan-sigmoid transfer function, the MSE is relatively low when the number of hidden layers is between four and fifteen, and the lowest MSE (0.0587 ± 0.0351) is obtained when it reaches seven. Besides, seven hidden layers with the log-sigmoid & purelin transfer function bring the highest determination coefficient (square of R, 0.3819 ± 0.2781). Therefore, we decided to select 10-7-1 as the topological structure of BPNN and the tan-sigmoid transfer function as the transfer function between both input layer & hidden layer and hidden & output layer. To improve the accuracy of the final model, we set the aim as R > 0.9 and trained 10-7-1 BPNN with tan-sigmoid transfer function & tan-sigmoid transfer function several times. The MSE and correlation coefficients of this model were in acceptable ranges as shown in Table 4. These correlation coefficients indicate that there is a strong correlation between the predicted and experimental result and the MSE (0.162) is acceptable. The plots of experimental versus predicted values (Fig. 3) confirmed the discussed results. Hence, we chose this one as the final model. To our knowledge, it was the first time applying BPNN to predict the IC 50 of ACE of pentapeptide. In our previous work [58], a BPNN model was built to predict the IC 50 of ACE of tripeptide. The MSE of the tripeptide model (0.2148) is higher than this pentapeptide model (0.0162). Meanwhile, the R of the tripeptide model (0.854) was less than this new model (0.9176). These differences possibly resulted from the difference of descriptors.

Activity Prediction and Experimental Verification
After simulated enzyme digestion of hemoglobin, seven pentapeptides among all the cleavage fragments were obtained. The descriptor calculated value and the predicted IC 50 were shown in Table 3. The pentapeptide WTQRF and AAWGK contain two or three hydrophobic amino acid residues, which is in agreement with the common features of the 24 reported pentapeptides, and show a low predicted IC 50 . The AHRYH, WTQRF and FTPVL contain hydrophpbic or aromatic amino acid residues in the C-terminus which contribute to ACE inhibition. For the second amino acid residues from C-terminus, the PTTKT, WTQRF and AAWGK contain hydrophilic amino acid residues, which also match the common features of the 24 reported pentapeptides. Obviously, WTQRF conforms to all the aforementioned common features. It include three hydrophilic amino acid residues (Thr, Gln, Arg) and two hydrophobic amino acid residues with aromatic (Trp) / heterocyclic (phe) ring in their side chain. It molecular weight is 736.83 g / mol. The hydrophobic amino acid residues in both of the N-and C- terminus as well as the hydrophilic amino acid residues in the second amino acid residues from C-terminus may make main contribution to ACE inhibition. The result of predicted IC 50 indicated that the peptide WTQRF has the best inhibitory activity of ACE. The experimental IC 50 of WTQRF is 23.93 μM, which is similar to a pentapeptide identified from egg white protein in our lab before (RVPSL, IC 50 20 μM). [54] The difference between predicted and experimental IC50 may bring from data normalization.

Molecular Docking
Interactions between ACE and inhibitor are fundamental to ACE-inhibitory processes. In this work, docking is used to predict the preferred orientation of WTQRF to ACE when they bound to each other to form a stable complex. The free energy of binding represents the binding affinity between ACE and WTQRF. The molecular docking study of WTQRF presenting within the ACE catalytic site in the presence of the cofactor showed the best pose with potential binding energy:-749.747 kj / mol. It is indicated that the ligand bind tightly to the receptor and the ACE-WTQRF complex is stable. As we know, ligand-protein affinities are influenced by noncovalent intermolecular interactions between the two molecules. The best pose of WTQRF was stabilized mainly by hydrophobic and hydrophilic interactions (Fig. 4), electrostatic interactions (Electrostatic energy:-154.728 kj / mol; Van der Waals energy:-898.743 kj / mol) and hydrogen bond (  [34,59] The WTQRF were positioned to interact with the HEXXH. . .E motif. Furthermore, the interactions between inhibitors and Zn 2+ at the ACE active site also play a significant role in modulating catalysis. [60] It is believed that the shorter distance between the Zn 2+ and the carbonyl oxygen of the peptide, the greater the degree of ACE inhibition. [34] Conclusion In this present work, a back propagation neural network model was built to predict the IC 50 of ACE of pentapeptides. The topological structure was 10-7-1 and the transfer function was tan-sigmoid transfer function & tan-sigmoid transfer function. We also built a database of potential pentapeptides derived from bovine hemoglobin by simulated enzyme digestion and found that WTQRF has the highest predictive value with experimental IC 50 23.93 μM. The molecular docking result indicated that Tyr 360 and Arg 522 gave a significantly contribution to the stabilization between WTQRF and ACE. The result also demonstrated that the short distance between the Zn 2+ and the carbonyl oxygen of the peptide are desirable to the ACE- inhibitory activity. The enzyme kinetics investigation as well as docking and molecular dynamics simulation analysis will be our further study.
Supporting Information S1