A data set, containing twenty-eight phenol ether derivatives as non-covalent proteasome inhibitors, was collected from the same literature to build the QSAR models [20]. These compounds were randomly divided into two groups. The training set of 22 compounds was used to generate quantitative models. The other was a test set of 6 molecules, which was designed to verify the reliability of the created model. The half maximal inhibitory concentration (IC₅₀) of these compounds was known, which was converted to the pIC₅₀ value [pIC₅₀ = –log₁₀(IC₅₀)] as the dependent variable for QSAR modeling. The pIC₅₀ values of 28 phenol ether derivatives covered a wide range from 5.229 to 7.310, which supplied extensive and homogenous data for QSAR analysis. The backbones of non-covalent proteasome inhibitors are shown in Figure 1, which are separated into two series. The backbone of the first series is shown in Figure 1A, and the compounds are from 1 to 16. The backbone of compounds 17–28 in Table 1 is illustrated in Figure 1B. The chemical structures of non-covalent proteasome inhibitors and their biological activity values are shown in Table 1.

The three-dimensional (3D) structure of 28 phenol ether derivatives was constructed in the Sketch module of the SYBYL-X 2.0 software package (Tripos, St. Louis, USA) [24]. Each molecule was minimized to obtain the lowest energy conformation. Energy minimization of all compounds was performed using the Tripos force field with Powell method. Gasteiger Hückel method was used to calculate the atomic charge. In order to get stable conformations, the maximum number of iterations was set to 10,000 and the energy convergence gradient was set to 0.005 kcal/(mol Å) [25–27]. The other parameters adopt the system default.

Molecular alignment operation is not only the basic step to establish a reliable 3D-QSAR model, but the quality of molecular alignment directly affects the prediction ability of the constructed model [28, 29]. To get the best QSAR model, we selected the most active compound 20 (IC₅₀ = 49 nmol/L) as the template molecule for molecular alignment [30, 31]. The chemical structure of the template molecule is shown in Figure 2A and the red part represents the common skeleton. The result of alignment based on the common substructure is shown in Figure 2B.

CoMFA and CoMSIA are two general applied tools of 3D-QSAR [26]. These methods are based on descriptors of 3D structures to analyze the different contributions of steric (S), electrostatic (E), hydrophobic (H), H-bond donor (HD), and H-bond acceptor (HA) fields [32, 33]. These descriptors are directly related to the atomic properties of compounds and the spatial geometry of molecules. Potentials are reflected in their position and expansion in space and their intensity. The CoMFA and CoMSIA models are performed using the QSAR option of SYBYL-X 2.0 software with the default parameters.

The non-covalent proteasome inhibitors were placed in the spatial grid to establish the CoMFA model. A 3D cubic box with 2.0 Å grid spacing in the X, Y, and Z directions was generated by the default values. The spatial lattice was linked with 3D structure, bioactive data, and molecular potential energy, thus providing valuable information for molecular modifications. S and E fields were calculated at each grid point with Tripos force field using a carbon atom probe with sp³ hybridization by means of a van der Waals radius of 1.52 Å and a charge of +1.0. In addition, the default S and E energy cutoffs were both set to 30 kcal/mol to avoid the infinity of energy values inside a compound [34, 35].

CoMSIA and CoMFA applied for the same molecular alignment method, however, the results of CoMSIA were more robust than the CoMFA. CoMSIA model was not limited to S and E fields but covered H, HD, and HA fields, which compensated for some shortcomings of CoMFA model. CoMSIA model was calculated using a sp¹ hybridized carbon atom with the radius of 1.0 Å and net +1.0 charge as a probe atom. The column filter and attenuation factor were set to 2.0 kJ/mol and 0.3, respectively, to reduce noise and speed up analysis [36]. The other parameters took the system default.

The partial least-squares (PLS) method was a multi-regression analysis method based on the ideas of linear transformation [37]. The inhibitor activity was used as the dependent variable and descriptors of potential fields as independent variables for the 3D-QSAR model. In the cross-validation process, the optimum number of components (ONCs) and cross-validation correlation coefficient (Q²) were calculated by leave-one-out (LOO) method. Then the non-cross-validation correlation coefficient (r²), standard error of estimate (SEE), and F-test values (F) were further obtained in the non-cross-validation analysis. The reliable 3D-QSAR models should have low values of SEE and high values of Q², r², and F. These important statistical parameters reflected whether the constructed model was reliable and robust or not [38, 39].

The external validation was also a crucial step because it could guarantee the predictive accuracy of 3D-QSAR. A favorable model relied on the high predicted external validation correlation coefficient r²_pred value (r²_pred > 0.6) [40]. According to Roy et al. [41] and Mitra et al. [42], the other four statistical parameters also need to meet certain rules to ensure the robustness of the model. These rules were the following criteria: R2>0.6,0.85≤K≤1.15,[(R2−R 02)/R2]<0.1,R m2>0.5

R² was the square correlation coefficient between the experimental and predicted bioactivity. The formula was as follows: R2=| ∑ (Ytest−Ytest¯)(Ypred(test)−Ypred(test)¯)∑ (Ytest−Ytest¯)2∑ (Ypred(test)−Ypred(test)¯)2 |2

K was the regression of experimental and Pred. and the regression line slope through the origin. The formula was as follows: K=∑ (Ytest*Ypred(test))∑ (Ypred(test))2

R 02 was the square correlation coefficient by calculating the experimental versus Pred. The formula was as follows: R 02=1−∑ (Ypred(test)−kYpred(test))2∑ (Ypred(test)−Ypred(test)¯)2

R m2 was another criterion to analyze the external predictability of the model. The formula was as follows: R m2=R2*(1−R2−R 02) Where Y_test, Ytest¯, Y_pred(test), and Ypred(test)¯ represented experimental, average, predictive and predictive average bioactivity values in the test set, respectively.

Molecular docking is a general method of simulating ligand-protein interactions, through which some vital information about ligand at particular binding site can be provided. For example, we can know the overall geometric conformation of ligands and the microscopic interactions within the pockets [43]. In this study, the Autodock software was used for the docking process [44, 45]. Since the ligand in the crystal structure of 3MG6 [protein data bank (PDB) ID] are selective for the β5 (ChT-L) site of the 20S core particle of the proteasome, our designed molecules have the same mechanism of selective inhibition of the β5 site of the proteasome [46, 47]. Therefore, to elucidate the binding modes of a new series of non-covalent proteasome inhibitors with exquisite potency and selectivity for the 20S β5-subunit, molecular docking calculations were performed. The X-ray crystal structure of the proteasome (PDB ID: 3MG6) was downloaded from the PDB (https://www.rcsb.org/). There were two core operations in the docking process. The first one involved pretreatment of proteins, including the removal of water molecules and original ligand, modification of missing loop regions, the addition of missing atoms and polarized hydrogen, and charges calculations. Then the compounds were docked into the active pockets of protein, which was defined by a 45 × 45 × 45 box centroid of the crystal ligand with a default grid space size of 0.375 Å [48]. All other parameters remained the default values. Finally, ten different molecular conformations were generated for each compound. The docking results were analyzed using PyMol software.

PKs and toxicology are important components of drug preclinical research. PK reflects the patterns of absorption, distribution, metabolism, and excretion of drugs, which are necessary processes for drug metabolism in the human body. According to statistics, most candidate drugs failed in drug development due to poor PK properties or excessive toxicity [49, 50]. Therefore, the ADMET comprehensive evaluation of candidate compounds is conducive to improving the success rate of new drug research and development [51]. In this study, ADMET and drug-like properties of newly designed phenol ether derivatives were obtained from pkCSM online server (http://biosig.unimelb.edu.au/pkcsm/prediction) [52] and SwissADME web tool (http://www.swissadme.ch ) [53]. The predicted ADMET data will provide some theoretical support for further experimental verification.

In order to acquire the best statistical model, we combined different fields to build a wide variety of models, calculated their statistical data, and finally elected the best QSAR model. The results of the CoMFA model and some CoMSIA models utilizing different combination types of fields are shown in Table S1. The Q² and r² values of a good 3D-QSAR model should be greater than 0.5 and 0.9, respectively. For the PLS analysis results of CoMFA-SE (using both S and E fields to build models) model, Q², ONC, r², SEE, and F was 0.574, 10, 0.999, 0.026 and 1,084.404, respectively, suggesting that it had good internal validation capabilities. The predicted external validation correlation coefficient r²_pred value was 0.755. The contribution for S and E fields were 47.1% and 52.9%, respectively. All this statistical data indicated that this model was reliable and robust. Therefore, CoMFA-SE was selected as the final CoMFA model.

The CoMSIA-HAD model had the highest Q² value (Q² = 0.647) in all CoMSIA series, but it was not chosen as the final CoMSIA model. Because we found that the contribution rate of E field always had a high weight in CoMSIA series. For example, the contribution of E was up to 32.3% in the CoMSIA-SEHAD model with all fields. Therefore, we could not ignore the importance of E in the CoMSIA model. However, the contribution of the HD showed a downward trend with the increase in the number of fields. For instance, it contributed the least (8.1%) in the CoMSIA-SEHAD model. So, the role of HD was not considered at all. Meanwhile, the CoMSIA-SEHA model had the highest Q² value in these five models (CoMSIA-SEHD, CoMSIA-SEHA, CoMSIA-SEDA, CoMSIA-SHDA, CoMSIA-EHAD), hence CoMSIA-SEHA was selected as the CoMSIA model. The CoMSIA-SEHA model had a large Q² value of 0.584 and r² value of 0.989, lower SEE value of 0.077, and F value of 172.183, showing that this model was reliable. The r²_pred value was 0.921, suggesting that this model had a strong external predictive capability. These statistical results demonstrated that the CoMSIA-SEHA model was robust. The contributions of S, E, H, and HA were 12.6%, 35.2%, 27.6%, and 24.5%, respectively. It showed that E played a vital role in this model. To sum up, CoMSIA-SEHA was selected as the best CoMSIA model. The statistical parameters of 3D-QSAR model after PLS analysis are shown in Table 2.

To test the predictive ability of the QSAR model, the test set of six compounds was selected for external verification to evaluate it. The r²_pred of the CoMFA-SE (hereafter referred to CoMFA) and CoMSIA-SEHA (hereafter referred to as CoMSIA) models was greater than 0.6, indicating that the established QSAR model had a decent external predictive ability. The relevant statistical results are shown in Table 2, and the other four statistical parameters calculated from the formulas of Roy et al. [41] and Mitra et al. [42] are also shown in Table 2. The CoMFA and CoMSIA models had R² at 0.684 and 0.897, K at 0.996 and 1.005, (R² − R 02)/R² at −0.459 and −0.110, R m2 at 0.559 and 0.614, respectively. All four statistical parameters met the Roy et al. [41] and Mitra et al. [42] rules, which indicated the created model had stronger external verification. The above results suggested that the CoMFA and CoMSIA models had good external predictive capacity.

The predicted bioactivity values of QSAR models for the training set and test set are shown in Table 1. The scatter plots of experimental and predicted pIC₅₀ values are shown in Figure 3. It could be seen from the picture that red triangle blocks and blue dots were close to the straight line (Y = X), which meant the actual values of whole molecules were almost consistent with the Pred. All these data clearly indicated the excellent stability and the highly predictive characteristics of 3D-QSAR models.

By analyzing the contour maps of CoMFA and CoMSIA models, we can vividly study the physical and chemical properties of the compound to find the key information affecting the bioactivity, which will guide us to design novel molecules. In this study, we mainly discussed the 3D-QSAR in four potential fields. The favorable and unfavorable contributions of all fields default to 80% and 20%, respectively. To better explain the contour maps, the inhibitor of the highest activity (compound 20) was labeled as shown in Figure 2A.

In the S contour maps, the green blocks indicated that the introduction of bulky groups in the region would improve the bioactivity, while the yellow regions mean the opposite. S fields for CoMFA and CoMSIA were shown in Figures 4 and 5, respectively. As shown in Figure 4A, a large-sized yellow block was displayed near the R³ area, while a medium-sized green contour existed in the R¹ region. In addition, in Figure 5A, compound 20 was surrounded by a large yellow contour and relatively small green block in the R³ and R¹ regions, respectively, which was basically consistent with the CoMFA model analysis. It was worth mentioning that a large-sized yellow contour appeared at the R³ opposite position in the CoMSIA model, but not in the CoMFA model. Overall, we found three key clues to enhance the activity of non-covalent proteasome inhibitors in the S fields: the bulky substituents at the R¹ of compound 20, and the small group at the R³ and R³ opposite position, respectively.

In E field contour maps shown in Figures 4B and 5B for CoMFA and CoMSIA, respectively, the red and blue blocks indicated that negative and positive groups in these regions could strengthen the inhibition of the compound. A medium-sized blue and a large blue contour appeared at the R¹ and R² in Figure 4B, which suggested that the original group replaced by the positive substitutions would increase the bioactivity. In the R³ region, there was a medium-sized red contour indicating that the negative group here was beneficial to the inhibition of the compound. As for the CoMSIA, the situation was similar to the CoMFA model, so the details were not discussed.

For the CoMSIA model, the H field contour map was shown in Figure 5C. The yellow block indicates that the H group introduced in the area can enhance the activity of compounds, while the magenta block indicates that the hydrophilic groups introduced in the area are helpful for increasing bioactivity. It could be seen from Figure 5C that a yellow contour was embedded in the R¹ area, indicating that the H group here was advantageous. There was a huge magenta block in the R³ area, showing that the hydrophilic group here was favorable.

In the HA field of the CoMSIA model shown in Figure 5D, magenta blocks indicate that the introduction of HA groups is valuable to increase the activity of the compound, while red blocks are the opposite. There was a medium-sized red color block in the R² area, and a large red contour embedded in the R³ area, indicating that the HD groups in the R² and R³ regions were advantageous for improving the bioactivity of the compound.

By analyzing the 3D-QSAR of a series of non-covalent proteasome inhibitors, some key clues about structural modifications to improve bioactivity were obtained: 1) The R¹ area was mainly surrounded by the green contour in the S fields of the CoMFA and CoMSIA models, the blue block in the E field, and the yellow block in the H field of the CoMSIA model. These results showed that the addition of bulky, positive, or H groups in the R¹ region was beneficial. 2) The R² region was mainly surrounded by a red contour in the HA field of the CoMSIA model, indicating that the introduction of the HD here was helpful. 3) The R³ area had the appearance of color block in four fields, namely, the yellow block in the S field, the red contour in the E field, the magenta block in the H field, and the red contour in the HA field, which represented the addition of small group, negative charge, hydrophilic, or HD group in this area was favorable. The main SAR information of non-covalent proteasome inhibitors is illustrated in Figure 6.

Based on the above clues, a series of novel non-covalent proteasome inhibitors were designed and evaluated using the highest bioactivity compound 20 as the template (Figure 6). Finally, twenty-four new inhibitors (compound D01–D24) were screened out with higher Pred. than the template. The chemical structures of these novel inhibitors and their pIC₅₀ values predicted by the constructed QSAR models are shown in Table 3.

In order to better explain the relevant docking results, three representative compounds for detailed descriptions were selected. The results of docking are clearly illustrated in Figure 7. How the least biologically active compound 15, the most biologically active compound 20, and the newly designed compound with the highest predictive activity D24 bound with the receptor protein 3MG6 are shown in Figure 7A, B, and C, respectively. At the same time, their molecular docking scores were −6.3 kcal/mol (compound 15), −8.3 kcal/mol (compound 20), and −8.9 kcal/mol (compound D24), respectively. The lower the molecular docking score, the stronger the binding ability of the ligand to the receptor [54]. And it can be seen that the size of the binding force is consistent with the size of the biological activity value. The binding mode of compound 15 which is the least active molecule in the dataset is clearly shown in Figure 7A. This inhibitor afforded some interactions with the proteasome. It could be seen that the oxygen atom of compound 15 near the R² region presented a hydrogen bond with alanine 22 (−O^…HN, 3.0 Å ). Furthermore, the oxygen atom was also bound with the amino acid residue of asparagine 24 by the formation of a hydrogen bond (−O^…HN, 3.1 Å ). In addition, there was a π-π stacking between tryptophane 25 and the benzene ring of compound 15. Several H interactions in the binding site were also found. However, when the docking result of compound 20 (the most active inhibitor) was analyzed, there were stronger interactions in the binding pocket which could explain why the activity of compound 20 was higher than compound 15. As could be seen from Figure 7B, the oxygen atom near the R² region of compound 20 (in the same position as compound 15) formed a hydrogen bond with tryptophane 25 (−O^…HN, 3.4 Å ). The second one was that the NH group of compound 20 near the R³ region interacted with alanine 27 (−O^…HN, 3.8 Å ). The third one was the oxygen atom bound with glycine 128 and there were other hydrogen bond interactions with key residues like serine 112, serine 118, and aspartic acid 114, respectively. Compared with compound 15, compound 20 had more hydrogen bond interactions with receptor which would be vital for the binding stability of inhibitor in the active site. Meanwhile, there was the same π-π stacking interaction as compound 15. In addition, the H interaction between compound 20 and key residues such as tryptophane 25, histidine 98, and aspartic acid 114 further enhanced the bioactivity. These multi-conjugate effects revealed that compound 20 (IC₅₀ = 49 nmol/L) had stronger stability in the active pocket of the receptor and higher activity than compound 15 (IC₅₀ = 5,907 nmol/L), which is consistent with the experimental result.

In order to better demonstrate that the new designed non-covalent proteasome inhibitors had higher biological activity than the template molecule (compound 20), the molecular docking method was applied for studying their binding mode with the receptor. New designed compound D24 with the predicted most biological active values was taken as an example. In Figure 7C, we noticed that compound D24 had the same hydrogen bond interaction as compound 20 where they interacted with the same residue (tryptophane 25), but the distinction between them was the hydrogen bond distance. For compound D24, the oxygen atom was far away from the surrounding residue, which was bad for ligand-protein interactions. It was noteworthy that there was a stronger 2.8 Å-hydrogen bond interaction between the NH group of compound D24 (at R² position) and phenylalanine 113 (−O^…HN, 2.8 Å ), which made the ligand-protein interaction binding more stable. Additionally, the pyridazine ring of compound D24 formed another stable hydrogen bond with serine 118 (−O^…HN, 2.9 Å ). Except that, compound D24 was bound with arginine 125 and phenylalanine 113, respectively, to generate other hydrogen bond interactions. Generally speaking, compound D24 had a closer hydrogen bonding distance than the template molecule, so it could more tightly bind with the receptor to have better inhibitor activity. In addition, we found that all three representative compounds formed π-π stacking with an indole ring of tryptophane 25, which implied it may be a key residue in the active pocket. Of course, the H interactions also existed in the compound D24 to enhance the binding with the protein. In summary, all of these results clearly indicated that compound D24 had high stability at the binding site, followed by a high inhibitory activity.

ADMET is a complex process and each link may have a great impact on the effectiveness and safety of drugs, so it is an important index to evaluate the drug properties of compounds in the process of new drug development and clinical use. Therefore, the online predictive evaluation of the newly designed compound (D01–D24) was performed. The results of statistical parameters for ADMET prediction of the newly designed non-covalent proteasome inhibitors are shown in Table 4. The results of drug-likeness and synthesis accessibility are shown in Table 3.

The intestinal absorptions of the newly designed compounds (D01–D24) are all greater than 30%, except compounds D14 and D21, which indicates that they will be highly absorbed by the human intestine. In particular, the highest intestinal absorption of compound D1 is 92.849%, suggesting that the novel non-covalent proteasome inhibitors have better absorption capacity as an oral drug. Enzymatic metabolism of the newly designed compound is assessed by whether it is a substrate or inhibitor of CYP. Considering that there are different enzymes in the family of CYP enzymes in human cells, CYP2D6 and CYP3A4 are two remarkable isoforms of P450 responsible for drug metabolism. As can be seen from Table 4, all the designed compounds are neither substrates nor inhibitors of CYP2D6. Compounds D1, D2, D5, D6, and D10 are substrates of CYP3A4, suggesting that they can be metabolized by CYP3A4. The other compounds are not CYP3A4 inhibitors, meaning that they may not affect normal metabolism. In terms of drug clearance, the Pred. indicates that all newly designed compounds could be cleared by combined hepato-renal clearance. Candidate drugs are supposed to be non-toxic to humans, and the AMES test is a bacteria short-term test that determines whether a chemical has mutagenic potential. Fortunately, none of the designed compounds are toxic to AMES and induce skin sensitization. The designed compounds are hepatotoxic, which may impair the normal function of the liver. According to the results of ADMET prediction, we may theoretically assume that the newly designed molecules have good PK properties.

Drug-likeness qualitatively assesses the likelihood of a compound becoming an oral drug from the perspective of bioavailability. As shown in Table 3, all novel compounds fulfill the Lipinski rule, which means they meet the criteria for being oral drugs. In addition, it is worth noting that the newly designed compounds D01–D24 have a synthetic accessibility range of 2.61–4.01, suggesting that they are not difficult to synthesize. These results could provide researchers with valuable information on drug synthesis and screening.