2.1. Network Pharmacology-Based Analysis

Based on the preliminary work, Kyoto Encyclopedia of Genes and Genomes (KEGG) database was utilized to generate the related proteins to construct a protein-protein interaction network. The hub protein should be identified in the pathway network. It was clear that the inhibition of Mitochondrial 70 kDa heat shock protein (mtHSP70, HAPA9, mortalin) and CK2A2 benefitted from HD resistance. Based on the multitarget idea, the related targets would be identified after network pharmacology-based analysis about YAP1 pathway. The YAP level would decrease in HD brain model [22], which indicates the downshift of YAP could be one of pathogenic factors. And the signal pathway could provide information of up-/down-adjust which can guide us to find the vital protein in the pathway network. Another 2 protein targets, Serine/threonine-protein kinase 3 (STK3) and Serine/threonine-protein kinase LATS1, were studied further.

2.2. 3D Structure Modeling for LATS1 Protein

The protein LATS1 has no enough structure information for structure-based drug design. The origin structure of LATS1 in PDB (5BRK:B) was further modeled to obtain a complete structure [24]. The binding site is set as a natural structure, which indicates that it was credible. The structure was verified in Ramachandran plot [25] and Verify profile-3D.

2.3. Docking Study of Target Proteins

The sequence of human CK2alpha’ protein (residues 1-350) was acquired from the Uniprot Knowledgebase (P19784). The 3D crystal structure of human CK2alpha’ protein, 5M56 (residues 1-350) was obtained from Protein Database Bank (PDB) website [26]. 5M56 was the latest CK2alpha’ prototype structural which has the complete 3D structure used for ligand-protein docking. All compounds from TCM Database@Taiwan, the world largest TCM database [27], were applied to dock with 5M56 by using Accelrys Discovery Studio (DS) software. LigandFit program in DS was performed to conduct molecular docking procedure. The ATP binding area was set as a binding sites referring to the ATP-competitive inhibitors, and one of the known inhibitor was set as a control in our study [28]. HARvard Molecular Mechanics (CHARMm) force field was performed to minimize energy. Affinity would be considered in Dock score, where the ligands with high score in docking would be the candidates to interact with target. Multiple poses would be calculated, 11 scoring patterns would be considered, and the consensus score would be set as a comprehensive consideration.

The same way was descripted above; another two targets screened small molecules in the TCM database as well. The information of HSPA9 (mtHSP70) protein was gained from UniprotKB (P38646). 3D structure of HSPA9 was acquired in PDB (4KBO, 52-431). The binding pocket referred to the nucleotide-binding domain (NBD), which is an active binding region [29]. MST2 (STK3) protein in the PDB (4LGD, 4-491) was selected for studying similarly. A peptide could inhibit the homodimerization of MST2, which could further prevent the activity of MST2 by transautophosphorylation [30]. And the binding site could be referred from it. The LATS1 was docking employing our modeling structure mentioned above.

The screen result of four targets would be analyzed in a network to discover which compounds have effects on multiple targets.

2.4. Quantitative Structure–Activity Relationship (QSAR) Models and Predictive Work

The bioactivity was predicted through QSAR model [31]. The Genetic Function Approximation (GFA) algorithm [32], a method of searching for the optimal solution by simulating the natural evolutionary process, was utilized to look for the proper molecular descriptors and then used Calculate Molecular Property protocol to acquire relative properties in DS software. 2D-QSAR models were conducted through multiple linear regression (MLR) with Matlab and support vector machine (SVM) with libSVM. The way to evaluate these models was the value of square correlation coefficient (R²) which is calculated in the regression. Several molecules with known activity were used for external verification as well. The predictive models were chosen from the models which have high value of square correlation coefficient. And then the predictive models were performed to forecast each candidate drugs after docking.

The docking scores of candidates with CK2A2 were provided, as well as the docking score with other related proteins, mtHSP70, STK3, and LATS1. And the predicted activities or scores through several known compounds or traditional Chinese medicines about HD could also take a reference to the final drugs’ selection. Some proteins (CK2A2 and STK3) also provided a control for comparing.

Thirty-one known ligands [33] for CK2 protein with IC50 information constructed the predicted model by SVM and MLR algorithm. Thirty-five ligands [34] for meHSP70 protein with IC50 information conducted the predicted model as well. The STK3 and LATS1 protein constructed the predicted models through the reported TCM formula, and the active ingredients would be gathered in our TCM database. The aim was to build prediction module through SVM and MLR algorithm to predict better medicines from the clinical medicines models.

2.5. 3D-QSAR Analysis

3D-QSAR was constructed by known ligands [33] for CK2 protein with Sybyl-X 1.1. These active ligands were superimposed. The Comparative Molecular Field Analysis (CoMFA) [35] and Comparative molecular similarity index analysis (CoMSIA) [36] models were constructed according to their activity and structural characteristics. The models were evaluated through cross-validation (CV). Residual between the measured and predicted values was calculated and the square sum of explained (SSE); F-test, coefficient of determination (R²), and q² for cross-validation were referred as evaluation indexes of 3D-QSAR models. CoMFA analyzes the effects of stereoscopic fields and electric fields, while CoMSIA evaluates stereoscopic fields, electric fields, hydrophobic fields, and hydrogen bonds for acceptors and donors. Various field combinations had been studied separately and the best model would be further studied.

2.6. Pharmacophore Analysis and Cross Validation

Pharmacophore models with 38 compounds [33, 37, 38] were created by hypoGen protocol, which was to analyze spatial influence factors such as hydrogen bonds acceptors, hydrophobic interactions, and π-conjugated effect. The models were verified through leave-one-out cross validation (LOOCV) [39]. Fisher’s randomization test was used to ensure that reliable hypotheses were built. To obtain a 95% confidence level, a total of 19 randomizations were required.

2.7. Molecular Dynamics Simulation (MDs)

In order to verify the results of docking screening, molecular dynamics simulations were performed on all candidate compounds screened above using gromacs5.0.4 program software with world’s top computing resources, Tianhe No. 2 supercomputer for long enough (5000 ns) simulation. The 5000 ns is basically the longest simulation in the current stage (2018). The energy minimization was produced by the steepest descent algorithm with the maximum number of 5000 steps minimization. The NVT equilibration was set in 20 ns with each step of 2fs and constrained with Lincs algorithm. The NPT equilibration was running after NVT in total of 20 ns. All bonds (even heavy atom-H bonds) were constrained with Lincs algorithm at 300K. The long time molecular dynasty was set in 5000 ns (for LATS1 protein nearly 3500 ns because of the computing resources). With the analysis of root-mean-square deviation (RMSD), root mean square fluctuation (RMSF), mean square displacement (MSD), secondary structure of protein, solvent accessible surface area (SASA), radius of Gyration (gyrate), the residue contact map (mdmat), Hydrogen bond distance (H-bond), energy analysis, and torsion angle, even the cluster research, the candidates will be validated to acquire a reasonable conclusion. After all, please be aware of short time MD simulation. Several candidates were “flying away” from protein during 5000 ns, even though they interacted well with targets at the first period.

3.1. Protein-Protein Interactions Network Analysis

From the interaction information about the known targets, some high-related targets could be further confirmed in the research. The decreased expression level of YAP in human HD brains neurons was known [22]. Up-adjust YAP level can be the treatment method.

The hub node in the pathway was important and probably got more interactions with other related proteins. The proteins in the pathway were gathered; it showed the related proteins which could adjust (Figure 2). It could be certain clearly which agonists or inhibitors could be designed from it. The key proteins could be found in the pathway. Here four proteins, CK2A2, mtHSP70, mst2 (STK3), and LATS1 were identified as the targets for the further research.

3.2. Modeling Structure Verification

The modeling structure of LATS1 was further verified by 3D-profile program and Ramachandran plot validation (Figure 3). The 3D-profile verify scores of the most of amino acids were higher than 0, especially the binding site region, which indicates the rationality of the modeling structure. The Ramachandran plot analysis showed whether each amino acid was in a reasonable angle, especially the amino acids of binding site displayed alone (Figure 3(d)).

3.3. QSAR Predicted Model

Appropriate descriptors were selected using the Genetic Function Approximation (GFA) algorithm (Table 1). And the descriptors would further be applied to create 2D QSAR model with support vector machines (SVM) and multiple linear regression (MLR) algorithm. The MLR model is built as follows.

The model was regressed by the compounds from treatment drugs. Blue spheres represented the training set, and the red inverted triangle indicated the test set (Figure 4). External verification and R² were reasonable.

3.4. 3D-QSAR Analysis

The CoMFA and CoMSIA models of 3D-QSAR were constructed (Figure 5). The value of CoMFA and CoMSIA was calculated further to obtain the predicted activity. And the residuals between observed IC50 and predicted IC50 would be one of the evaluation values of predictive ability (Table 2). Partial least squares (PLS) analysis and validation of CoMFA and CoMSIA model provided the value of correlation coefficient (R²), standard error of estimate (SEE), and F test (F ratio) for evaluating, as well as the cross validation ($$) for assessing (Table 3). Large enough R² and relatively small SEE could explain the rationality of the model. And $$ greater than 0.5 were worth being considered. The information of several 3D-QSAR model CoMFA analysis reminded us where bulky substituent groups were needed (green area) and where no substituent groups were demand (yellow area) (Figure 5(a)). The CoMSIA model suggested what the hydrophobic groups (yellow area), hydrophilic groups (gray area), hydrogen bonds donors (cyan area), and hydrogen bonds acceptors (violet area) needed. Areas that were inappropriate for introducing hydrogen bond acceptors would be warned correspondingly (red area) (Figure 5(b)). External validations of the models were provided to identify its accuracy (Figures 5(c) and 5(d))

3.5. Pharmacophore Analysis and Cross-Validation

The activity predicts model would be verified with cross-validation. The leave one out cross-validation (LOOCV) was used to illustrate the reliability of the model. The information of 10 hypoGen models was shared (Table 4). Cat-Scramble suggested the accuracy of pharmacophore model. The total cost values of 19 random hypotheses were all higher than the initial spreadsheet (Figure 6). The difference between null cost and total cost was vital to assess the confidence of the model. Configuration cost value was 15.59 lower than 17 which at a reasonable interval. Pharmacophore Analysis further evaluated the candidates; compounds 1a and 1b nearly matched all spatial factors (Figure 7).

3.6. Dock Screening Result and Candidate Determination

Top 50 compounds in each target were created in a network to suggest the multitarget effect (Figure 8). The screening rules the vote score based on the docking scores, SVM and MLR predicted activity, and consensus scores (calculated from 11 scores). For example, top 20 percent of each project was set as efficient and scored 1 point, otherwise 0 point. The total score was set as final evaluation (Table 5). The screening result for CK2A2 and mtHSP70 was displayed in Table 6, while the STK3 and LATS1 were shown in Table 7.

2D structure of candidate compounds was displayed in Figure 9. The hydrogen bonds, hydrophobic bonds, etc. were displayed to suggest potential poses of ligands and receptors (Table 8); especially several residues binding with all ligands would be focused on. The hydrogen bonding was a significant reference for these binding analyses of combining ability. 2D diagram for different targets was shown in the interaction bonds between key residues and ligands. Van der Waals force, hydrogen bonds, salt bridge, attractive charges, pi-interaction, etc. displayed the binding potentials (Figure 10). The bond length (Å) was considered as the index of stability. Unfavorable bumps prompted space conflict that is not recommended for introduction (red callout). Probably it was a result that the ligand 3a-STK3 was not stable during MD period.

3.7. MD Analysis

MD results were analyzed in the Gromacs5.0.4 program.

Trajectory files obtained after MD were made as videos (supported video set (available here)) for observing and displaying. They were convenient for our research by improving intuitive information. Total energy and RMSD changes (include proteins and ligands) were provided to analyze whether the receptor-ligand interaction in a proper state (Figure 11). During the MD period, the CK2A2-Flazine (1a), CK2A2-Typhic Acid (1b), mtHSP70-Febrifugine (2a), mtHSP70-Holantosine C (2b), mtHSP70-Cassaine (2d), STK3-3a, STK3-ANP (Control), and LATS1- (+)-Taraxafolin B (4d) were tested whether interactions well for a long time. Most of the complex possessed nonfluctuating total energy, even if the protein interacted with different ligands. Total energy of CK2A2 protein-ligand complex stand at nearly -860000~-850000 kJ/mol, while mtHSP70 was -1175000~-1170000kJ/mol, STK3 was -1550000~-1545000 kJ/mol, and LATS1 was -2600000~-2590000kJ/mol. It was noteworthy that the energy of STK3-control was lower compared to other ligands, which indicates the control complex stayed at a more stable state. Changes in RMSD suggested that the interactions changed in configuration. It was more concerned whether the RMSD is stable at the end period of MD. Ligand RMSD of STK3-control and LATS1-4d fluctuated obviously during MD; nevertheless, both of them did not infect the total energy or protein RMSD; the former probably changed posture at 3500 ns (Figure 11).

The mean square displacement (MSD), solvent accessible surface area (SASA), and Radius of Gyration (gyrate) analysis provided the information about both protein and ligand during MD (Figures 12 and 13). Gyrate shows the more bulky the molecule, the smaller the value and the closer the molecules. In general, gyrate presents a downward trend. And for the STK3 protein STK3-control and STK-3a (red line) all displayed high gyrate, which indicate that the ligand induced fit to the protein and the protein maintain a bigger open binding site (Figure 12(c)). For mtHSP70 protein (Figure 12(b)), the opposite is true; when the protein was in a bulky state, the ligands easily “fly away”; when the protein was tighter, the ligand could get better stay inside. MSD analysis told us the protein change between the original state and simulation state. It was clearly which ligand “fly away”. The value of MSD was logarithmic transformation: Y_MSD=2+log₁₀ (MSD). In fact, several proteins remain changing at the end of simulation, even if we simulate very long time. SASA analysis could help us understand the hydrophobic nature and protein surface state. SASA of CK2A2 was little difference when combined with different ligands as well as LATS1 protein (Figures 12(a) and 12(d)). SASA of STK3 protein (Figure 12(c)) displayed higher hydrophilic when complex with ligand (red line and orange line); it was related to its bulky structure predictably.

Mean square displacement (MSD), solvent accessible surface area (SASA), and Radius of Gyration (gyrate) analysis of each ligand showed in Figure 13. The ligands’ staying with protein complexes during MD would be focused. Ligand gyrate was stable in general. The ligand MSD revealed the ligand change during MD; as for STK3 protein, the ligand 3a (red line) increased at 4000 ns; it was consistent with gyrate drop, which indicates structure contraction at that time. SASA of ligands change little at MD period. One interesting condition is for STK3 protein (Figure 13(c)), where the SASA value of the ligands complex with protein (red line and orange line) was higher than other “fly away” ligands (blue line and green line); probably the bulk protein structure provides more chance for inside ligands to contact with water.

Nonbond relationship provided the binding information between the ligands and receptors; it was an important reference for drug design (Table 8). For example, it told us the Residue R48 of CK2A2 kept the H-bond, electrostatic interaction, and hydrophobic interaction. As the same, E222 of mtHSP70, D164 of STK3, and R618 of LATS1 can be significant for maintaining interaction effect.

Hydrogen bond distance analysis was vital to explain the interaction between targets and proteins. The variation of main H-bond was observed to interpret the stability of receptor-ligand complexes (Figure 14). And the H-bond distance and the occupancy were showed to explain whether these H-bonds were in a proper distances (Table 9). The H-bond occupancy of STK3-control and mtHSP70-Febrifugine was kept high which can predict a good interaction for these complexes. As for LATS1 protein, unfortunately low H-bond occupancy indicated the ligand may leave even if we have not yet observed (Figure 14(e)). The H-bond information informed the significant residues. ARG48 and ASN119 of CK2A2 were great for maintaining complex interactions. ASP59, Gly247, Gly248, and Gly387 of mtHSP70 protein mainly built hydrogen bonds. As for STK3 protein, different ligands bound with different residues of protein (Figures 14(c) and 14(d)), the different interaction mode can be validated in further poses analysis. The observations for LATS1-4d were worrying, which had no stable hydrogen bond.

Root mean square fluctuation (RMSF) can be used to test fluctuation of each residue, which can notify the key residues change during MD. The key residues contributed nonbond interaction which would mainly be focused on. CK2A2 complexes had similar residue fluctuations (Figure 15(a)), which correspondingly showed similar average protein structure (Figure 16). Residues fluctuation of mtHSP70 complexes displayed similar trend, which have the same fluctuation peaks and valleys. The difference between these complexes was the different degrees of change (Figure 15(b)). The only one ligand “fly away” (pink line) had different types of fluctuations, which further verified the role of ligands. As for STK3 protein (Figure 15(c)), ligand 3a (pink line) and control (orange line) displayed different changes mainly due to the loop area outside, which showed in different average structure (Figure 16). The change of LATS1 complexes was more complex due to a lot of residues, but it could be clearly that the fluctuation of key residues like P481, Y597, K608, and R618 influenced the binding interaction.

The average protein structures would be superimposed to observe whether the binding structures are alike, which can predict these structures related to their inhibition. CK2A2 and mtHSP70 were nearly superimposed on one structure, respectively. STK3 can be superimposed except the outside loop. The two LATS1 protein structures cannot merge into one because one ligand of structure “flies away”. It manifested the structure which acting with ligand was completely different from the origin protein (Figure 16).

The protein residues contact map provided 3D information about the residues distance of proteins in 2D matrix. Average amino acid distance was shown in different color, which indicates the looseness of protein. The distance between key residues was focused on. It demonstrated a tight conformation for SK2A2 protein. The key amino acids R48 and N119 of CK2A2 are all close to other residue that can construct a “hotspot” cavity. Similarly, key residues D59, G247, etc. of mtHSP70 constitute an active cavity site. The amino acids S37, K56, E70, and D164 of STK3 are also in a reasonable range. Different ligands for CK2A2 or mtHSP70 protein would get similar 3D structure shown in residues distance matrix, respectively, which further validate our inference above.

Secondary structure of protein provided important information on structural stability. Maintaining A-helix and β-sheet during MD period can keep structural stability. Basically secondary structure was stable, including the key residues. A change was observed like residue 176-181 of mtHSP70-Febrifugine since 3000 ns, where A-helix changed to β-turn. It is possible that complex structure changes to a more stable configuration since this moment influenced by the ligand (Figure 17).

Cluster analysis was taken for each complex at the last period. The dock poses, cluster poses during the last 500 ns period, and the last poses were provided (Figure 18). Cluster poses were got from the big class of cluster map. The poses between dock and the last were changed in some complex like LATS-ligand; in the large cavity of the action site, the ligand is completely steered, showing another way of acting. The CK2A2 and mtHSP70 were relatively fixed. And the H-bond relationship was shown which could clearly know the interaction modes of key residues like Asp59, Glu222, Asp 244 in the mtHSP70, etc. For the same receptor, different ligand employed different residual contributions; however, the common force mode between different complexes could help us find the key residues like Asp48 and Asn119 in the CK2A2 protein.

The change in the torsion angle provides important clues to the stability of hydrogen bonds. If a single key that can be rotated originally is fixed in a relatively small range of variation, it may be fixed by space or electrically fixed. Relatively speaking, finding its related hydrogen bonds also shows that the hydrogen bonds are relatively stable (Figure 19). Torsion 1 showed 180 degrees of twist, but in fact two O elements on the carboxyl were equivalent when carboxyl dissociates; hydrogen bonds which could form after 180-degree rotation were still equivalent to origin hydrogen bond; what is certain was that this was a weak hydrogen bond. Torsion 2 had a great influence on the overall conformation of the ligand molecule connecting two planes, which showed rotation of furan ring plane; its orientation can bring hydrogen bonding or hydrophobic interaction; approximately 30 degrees of change is relatively stable to conform a H-bond or hydrophobic interaction. Torsions 4-8 are all related to H-bond. The torsion angle here remained a 20-45-degree rotation; there may be hydrogen bond formation. Torsion 4 may be limited by spatial orientation as a tetra-alkyl end. Less than 20 degrees of twist in torsions 9 and 11 indicate the N can conform H-bonds; it was due to the limited effect of the ring, although it is not an aromatic ring. Torsion 10 can show the potential effect of hydroxyl. The change of torsion 12 was about 60 degrees. Mainly studying single bonds near O or N, torsions 13 and 15 are very stable with about just 10- or 20-degree rotation. Oxygenated areas around single rings were likely to form several binding forces. Torsions 18-22 rotated in a small range. The tricyclic and double bond limited made several bonds stable that may be advantageous. Most single bonds of molecular 3a rotated freely except the two single bonds 23 and 24. This may suggest that compound 3a spatially interacts with the receptor in a different way during simulation trajectory. Torsions 25-31 showed the changes of compound 3d, which further illustrated the good interaction between STK3-Control complexes. So the strategy for STK3 target was to de novo modify after screening referring to the control compound structure. O element around torsions 33-34 has the possibility of hydrogen bonding, but the hydroxy group on the phenyl ring failed to form a fixed hydrogen bond. The span of torsion 32 was nearly 90 degrees that this single would keep moving obviously.

Pathway analysis could provide various pathways that the ligands could entry using Caver 3.0 [40]. The possibility of ligands entering the binding region was diverse. More entry paths mean greater possibilities for integration of ligands and proteins. Almost all the targets provide lots of access routes for ligands, which is advantageous for binding (Figure 20).