Identification of potential TNF-α inhibitors: from in silico to in vitro studies

Tumor Necrosis Factor Alpha (TNF-α) is a pleiotropic pro-inflammatory cytokine. It act as central biological regulator in critical immune functions, but its dysregulation has been linked with a number of diseases. Inhibition of TNF-α has considerable therapeutic potential for diseases such as cancer, diabetes, and especially autoimmune diseases. Despite the fact that many small molecule inhibitors have been identified against TNF-α, no orally active drug has been reported yet which demand an urgent need of a small molecule drug against TNF-α. This study focuses on the development of ligand-based selective pharmacophore model to perform virtual screening of plant origin natural product database for the identification of potential inhibitors against TNF-α. The resultant hits, identified as actives were evaluated by molecular docking studies to get insight into their potential binding interaction with the target protein. Based on pharmacophore matching, interacting residues, docking score, more affinity towards TNF-α with diverse scaffolds five compounds were selected for in vitro activity study. Experimental validation led to the identification of three chemically diverse potential compounds with the IC50 32.5 ± 4.5 µM, 6.5 ± 0.8 µM and 27.4 ± 1.7 µM, respectively.


Scientific Reports
| (2020) 10:20974 | https://doi.org/10.1038/s41598-020-77750-3 www.nature.com/scientificreports/ yet against TNF-α. Therefore, the identification of small molecules that can inhibit TNF-α regulated pathway presents a promising and current focus area. Thus, in the present study, an integrated virtual screening approach was employed to explore new TNF-α inhibitors. The hits selected through virtual screening were then corroborated by using an in vitro assay. On the basis of TNF-α inhibition activity, compound 1, 2 and 3 showed the most promising inhibitory activity with the IC 50 32.5 ± 4.5 µM, 6.5 ± 0.8 µM and 27.4 ± 1.7 µM, respectively.

Experimental
Ligands preparation. Twenty six reported biologically active inhibitors (IC 50 in the range of 0.001-100 µm) against TNF-α were retrieved from the literature (Supplementary Table S1) [15][16][17][18][19][20][21][22][23][24][25] . All the ligands were sketched in MOE 26 using builder module. Ligands were subjected to protonation and energy minimize using MMFF94 force field 27 . The decoys were generated by submitting the set of active compounds to DUD-E website (http:// dude.docki ng.org). For virtual screening, our In-house database of natural and synthetic molecules was utilized containing ~ 10,000 compounds. The database converted into the 3D format followed by protonation, structure correction and minimization using MOE 26 . Pharmacophore-based virtual screening. In silico pharmacophore-based virtual screening is a sophisticated tool in the modern drug discovery process for the identification of new leads from large database with the desired activity profile 28 . The availability of large compounds library of TNF-α inhibitors and their corresponding biological activity on different cell lines have enabled us to focus on ligand-based pharmacophore modelling approach.
In this approach, conformations of selective active compounds are aligned, and common pharmacophore features are generated where physicochemical functionalities are overlapped. These conformations were calculated from OMEGA, implemented in LigandScout 4.3 29 . The "create shared pharmacophore" function of LigandScout was used to create 12 hypothesis of pharmacophore models comprising diverse chemical features (Hydrogen bond donor, Aromatic, Hydrophobic, Hydrogen bond acceptor and many exclusion volume). Table 1 represented the generated hypothesis models using different combinations of shared features of active compounds. The generated pharmacophore model was validated in order to examine its potential to differentiate between active and decoys compounds in the dataset. For this purpose, an external test set was prepared using 26 confirmed actives and 1750 decoys or inactive for the validation of pharmacophore hypothesis.
Various statistical parameters such as percent ratio of actives, percentage yield of actives, false negatives, false positives, enrichment factor (EF), and area under the ROC curve (AUC) were further calculated. The validated pharmacophore model was used as a 3D query to retrieve the potential compounds from the inhouse natural product and synthetic compounds library (~ 10,000 compounds). Default settings were used for virtual screening (25 conformers generated per entry under omega-fast settings). The resultant compounds were subjected to drug-like assessment employing the ADMET Descriptors and Filter by Lipinski rule. The compounds that have obeyed these criteria were upgraded to the molecular docking studies.

Molecular docking.
A campaign of high throughput docking was performed by using MGLTools of Auto-Dock 4.2. The crystal structure of TNF-α with PDB ID 2AZ5 21 was retrieved from Protein Data Bank. Missing residues in the crystal structure were added and refined by using MODELLER 9.18 30 . Polar hydrogen atoms were added while nonpolar were merged and Gasteiger charges were applied. Lamarckian genetic algorithm was used for all the calculations. The coordinates of the crystal structure was saved into pdbqt format for docking calculation. A grid box of centers 28 Å × 26 Å × 30 Å along with dimensions − 19.515 Å × 74.84 Å × 33.894 Å, centered on cognate ligand served as spacing grid which encompassed the entire binding site. The pose with the highest binding affinity and reasonable able conformation were chosen for further analysis. The finger prints of www.nature.com/scientificreports/ protein-ligand interactions were analyzed by using PLIF module in MOE. Furthermore, the compounds were visually analyzed by using Chimera 31 .
In vitro studies. THP Briefly the working concentration of 4 μg/mL in PBS from 720 μg/mL of mouse anti-human TNF-α capture antibodies were used for coating of 96 well ELISA plate. The 100 μl of TNF capture antibody per well was added and plate was incubated overnight at RT. The plate was then blocked by adding 300 μl of reagent diluent in each well and incubated for 1 h at RT. 100 μL/well of collected supernatants was then added each in triplicate and plate was then incubated for 2 h at RT. The 100 μL of detection antibody diluted to 250 ng/mL (working concentration) in reagent diluent was then added to each well and plate was incubated for 2 h at RT. Next 100 μL of 1:200 dilution of streptavidin-HRP in reagent diluent was added in each well in dark and incubated for 20 min at RT. Substrate solution was prepared by mixing color reagent A and color reagent B (provided in the kit) in 1:1 ratio and 100 μL from this mixture was added to each well in dark, plate was incubated for 20 min at RT. The reaction was stopped by adding 50 μL of stop solution (2 N H 2 SO 4 ) and plate was then read at wavelength of 450 nm in ELISA reader (ELX800 NB, DIA LAB, Wr. Neudrof, Austria).
MTT cytotoxicity assay. Cytotoxicity of compounds on NIH-3T3 fibroblast cells was evaluated by MTT colorimetric assay. The cell line was provided by ICCBS Biobank facility which was purchased from (ATCC, Manassas, USA). Briefly 100 μL of 6 × 10 4 cells/mL in DMEM supplemented with 10% FBS were plated into 96-wells flat bottom plate and incubated overnight at 37 ºC in 5% CO2. Different concentrations of test compounds (350-1 µM) were added to the plate in triplicates and incubated for 48 h. 50 µL of 0.5 mg/mL MTT was added to each well and plate was then further incubated for 4 h. MTT was aspirated and 100 µL of DMSO was then added to each well. The extent of MTT reduction to formazan within cells was calculated by measuring the absorbance at 540 nm, using spectrophotometer (Spectra Max plus, Molecular Devices, CA, USA). The cytotoxic activity was recorded as concentration causing 50% growth inhibition (IC 50 ) for 3T3 cells.

Results and discussion
Pharmacophore-based virtual screening. Prior to the pharmacophore model generation, key features of the reported 28 active compounds from 10 different classes (pyrazolones, urea, indole, thiophene, purine, oxime, diaryl heptanoids etc.) were identified by superposing them to determine potential overlapped chemical features with the LigandScout. This procedure has generated 12 hypothesis (Table 1) with three to six potential chemical features. For the selection and validation of the best model, these generated hypotheses were refined and pruned on the basis of the following criteria: (1) the presence of chemical features that possibly interact with tyrosine residues (potentially Tyr119), which is crucial for TNF-α inhibition. (2) ability to select active compounds with good fitness score according to their biological activity with minimum deviation and. (3) ability to picked active compounds from the pool of active and decoys dataset. This criteria declare Hypo_7 as best hypothesis as it yields the pharmacophore fit score range that imitate the activity trend and difference in their magnitude as illustrated in Table 1. The Hypo_7 was in good agreement according to the nature of TNF-α active site residues as it contain three hydrophobic features for the key interaction with Leu57, Tyr59, Tyr119 and Tyr151 and two hydrogen bond acceptor for the interaction with Ser60 and Gln61 (Fig. 1a). Moreover, Hypo_7 align well on two of the highly active TNF-α inhibitors (Fig. 1b). Further the quality of selected model was determined by calculating the enrichment factor i.e. the fraction of actives compounds within a database while the values of other parameters illustrated in Table 2; www.nature.com/scientificreports/ where D is the total number of molecules in the decoy dataset, A is the total number of active compounds, Ha is the number of active hits molecules while Ht is the number of decoy hits molecules. The enrichment factor for the Hypo_7 was found to 12.5. Moreover, ROC curve was also calculated to measure the sensitivity (ability to search true positives) and specificity (ability to avoid false positives) between active and decoys on the basis of pharmacophore fitness score and the value of AUC was found to be 0.83 (Supplementary Fig. S1). For virtual screening in-house database of ICCBS containing ~ 10,000 compounds were virtually screened against the final pharmacophore model. In the result of successive screening ~ 1700 compounds were obtained. Drug like filter was applied on the database: molecular weight < 500 Dalton, no. of H bond acceptor < 5, no. of H bond donor < 10, no. of rotatable bond < 11 and octanol water partition coefficient log p < 5. Assessment of ADMET filtration led to the identification of 700 hit compounds for further analysis. The retrieved hits were further reduced to 400 compounds based on highest pharmacophore fit score in the range of 58.5-55.5 and subjected to molecular docking studies. The overall workflow of pharmacophore-based screening shown in Fig. 2. Molecular docking studies. The binding mode interactions of hits, obtained from pharmacophore based screening, were investigated by molecular docking studies. All the hits were docked in the active site of the target protein within the define grid. The top ranked 142 compounds with the cut off score of > − 6.5 kcal/mol were selected for further analysis. Extracted compounds were subjected to protein-ligand interactions fingerprinting by using PLIF module implemented in MOE. Interaction pattern of 142 compounds were analyzed which led us to the identification of five virtual hits from three different chemical classes. Upon investigating the binding mode, it was revealed that the identified hits reside with the same pattern in the active site of TNF-α as the cocrystallized ligand (SPD304). This suggested that the virtual hits might act in similar way as reference compound SPD304 inhibit the protein activity.
Compound 1 showed good inhibitory potential against TNF-α with the binding affinity of − 8.4 kcal/mol. Detailed molecular interaction pattern of compound 1 demonstrated that chlorobenzene ring, establish π-π stacking interaction with the aromatic ring of TyrB59 and TyrB119 while π-alkyl interaction with LeuA57 and TyrA59. Similarly, fluorobenzene ring involve in mediating π-alkyl interaction with TyrA59, GlnA61, TyrA119, and TyrA151. Moreover protein-ligand contacts were stabilized by two hydrogen bond interactions between hydroxyl group of the ligand and the side chain of SerA60 (Fig. 3a).
Compound 2, 3 and 4 were belong to chemical class of thiourea with binding affinities − 7.1, − 7.0 and − 6.6 kcal/mol respectively. Compound 2 and 3 were more active as compared to compound 4 based on TNF-α inhibition activity. Upon evaluating the molecular interactions, it was suggested the compound 2 and 3 showed similar type of interactions (Fig. 3b,c). The hydrophobic cleft formed by the Leu59 and Tyr59, Tyr119 and Tyr151 around both compounds provide additional stability to better fit the ligands in the active site of TNF-α. In case of compound 2, π-alkyl interactions was observed between nitrobenzene ring and LeuA57, LeuB57 and TyrB59 while other nitrobenzene ring exhibiting fluorine group showed π-alkyl interaction with TyrA119. Similarly, in case of compound 3 π-alkyl interaction observed between fluorobenzene ring and TyrA59 and TyrA119 while nitrobenzene ring in the ligand mediate π-alkyl interaction with TyrB59 and TyrB119. Moreover, both www.nature.com/scientificreports/ compounds observed to establish hydrogen bond contacts with GlyA121. In case of compound 4, any interaction with the hydrophobic cleft is not observed, which may account for the low potency of this compound. Detailed binding mode analysis revealed that compound 5 mediated similar type of interaction as observed for SPD304 (Fig. 3d). Binding affinity of compound 5 with the TNF-alpha was found to be − 7.4 kcal/mol. Benzene ring of the ligand interact with the side chain of LeuA57, LeuB57 and TyrA59 by mediating π-alkyl interactions while methyl group of the ligand establish alkyl-alkyl interaction with TyrB119. Similarly, the nitrogen of hydrazine group in the ligand establish three hydrogen bond contacts with SerB60 and LeuB120. Four compounds among all the docked compounds showed strong hydrophobic interactions along with the hydrogen bond contacts with the crucial residues of the TNF-α which suggested that these compounds might be serve as direct inhibitors of the target protein.
Bioassay validation. Compounds obtained from in silico studies were evaluated for their TNF-α inhibition activity through in vitro studies using Pentoxifyllin as standard drug. Table 3 list the TNF-α inhibitory activity for the standard and shortlisted compounds obtained from virtual screening. Shortlisted compounds were the derivatives of benzophenone, flurbiprofen and thiourea. Compound 1, a benzophenone derivative synthesized and reported by Arshia et al. 32 from our institute, showed significant inhibition against TNF-α with the IC 50 value of 32.5 ± 4.5 µM. Benzophenone molecule possesses good anti-inflammatory activity e.g. ketoprofen contain benzophenone group is one of the marketed anti-inflammatory drugs 33,34 . Compound 2-4 were the derivatives of thiourea, previously reported by Bilquees et al. 35 , from our institute all these compounds showed potent inhibition (6.5 ± 0.8, 27.4 ± 1.7 and 280.6 ± 9.6 µM respectively), except compound 4 as compare to standard pentoxifyllin (IC 50 = 340.6 ± 7.54 µM), against TNF-α. Similarly compound 5 was the derivative of Flurbiprofen; the most important NSAIDs (non-steroidal anti-inflammatory drugs), which is widely used to treat arthritis 36 , previously synthesized by Momin et al. 37 , from our institute. Compound 5 showed inhibition against TNF-α with the IC 50 value of 117.7 ± 1.1 µM. Moreover, all compounds were found to be non-toxic on NIH-3T3 cell line when compared to the standard drug cyclohexamide (Table 3). In vitro studies were in good agreement with in silico studies and revealed that all the tested compounds inhibit TNF-α produced from lipopolysaccharide (LPS) activated THP-1 cells, which explain the potencies of these compounds.

Conclusion
Inhibition of TNF-α has emerged as a potential therapeutic to treat tumor and especially autoimmune diseases. Currently, no orally active FDA approved drug against TNF-α is reported. Small-molecule drugs that can regulate TNF-α levels or activity may provide an economic alternative to antibody therapeutics. In the current www.nature.com/scientificreports/ investigation we aim to identify novel small molecule from inhouse database that obey the pharmacophoric features of TNF-α inhibitors. The statistical evaluation of the developed pharmacophore model highlights its ability to discriminate between active and decoys. The screened compounds were subjected to molecular docking to investigate their binding mode. Subsequently, 5 compounds were identified with high docking score and demonstrated key interaction as was noticed for reference compound. Experimental validation indicated that three of these compounds exhibit strong TNF-α inhibitory potential with the IC 50 32.5 ± 4.5 µM, 6.5 ± 0.8 µM and 27.4 ± 1.7 µM, respectively. The identified inhibitors have potential to act as anti-inflammatory agent and may serve as a starting point in developing novel drugs. www.nature.com/scientificreports/ www.nature.com/scientificreports/