A Virtual Screen Discovers Novel, Fragment-Sized Inhibitors of Mycobacterium tuberculosis InhA

Abstract

Isoniazid (INH) is usually administered to treat latent Mycobacterium tuberculosis (Mtb) infections, and is used in combination therapy to treat active tuberculosis disease (TB). Unfortunately, resistance to this drug is hampering its clinical effectiveness. INH is a prodrug that must be activated by Mtb catalase peroxidase (KatG) before it can inhibit InhA (Mtb enoyl-acyl-carrier-protein reductase). Isoniazid-resistant cases of TB found in clinical settings usually involve mutations in or deletion of katG, which abrogate INH activation. Compounds that inhibit InhA without requiring prior activation by KatG would not be affected by this resistance mechanism and hence would display continued potency against these drug-resistant isolates of Mtb. Virtual screening experiments versus InhA in the GO Fight Against Malaria project (GO FAM) were designed to discover new scaffolds that display base stacking interactions with the NAD cofactor. GO FAM experiments included targets from other pathogens, including Mtb, when they had structural similarity to a malaria target. Eight of the sixteen soluble compounds identified by docking against InhA plus visual inspection were modest inhibitors and did not require prior activation by KatG. The best two inhibitors discovered are both fragment-sized compounds and displayed K_i values of 54 and 59 μM, respectively. Importantly, the novel inhibitors discovered have low structural similarity to known InhA inhibitors and, thus, help expand the number of chemotypes on which future medicinal chemistry efforts can be focused. These new fragment hits could eventually help advance the fight against INH-resistant Mtb strains, which pose a significant global health threat.

INTRODUCTION

Tuberculosis (TB), caused by Mycobacterium tuberculosis (Mtb), kills 1.3 million people each year.¹ According to the World Health Organization, Mtb infects approximately two billion people.¹ Since a third of the global population has a latent Mtb infection, this creates an immense reservoir of disease due to the potential for reactivation.¹ There are 8.3 to 9 million new cases of tuberculosis (TB) annually, and half a million children get TB each year. ¹ Mtb kills more people in the world than any other bacteria. Of all infectious diseases, only HIV kills more people than Mtb, and TB is the leading cause of death for HIV/AIDS patients.¹

Although effective TB drugs have existed for over 60 years, and drug-resistant TB was not a major issue twenty years ago,² cases of multi drug-resistant TB (MDR-TB) and extensively drug-resistant TB (XDR-TB) continue to increase throughout the world in both frequency and distribution.^1–7 MDR-TB cases have nearly doubled in just a few years.⁴ The global treatment success rate for TB is now less than 50%.¹ Each year, one-half million new MDR-TB cases occur (i.e., Mtb infections that are resistant to INH and rifampicin). XDR strains additionally evade fluoroquinolones and at least one of the 2^nd-line injectable drugs (amikacin, capreomycin, or kanamycin).¹ With the emergence of totally drug-resistant TB (TDR-TB) in several countries, no effective treatment options exist for these patients.^{3, 5–8}

Novel InhA inhibitors effective against isoniazid-resistant mutants would be critical for treating MDR and XDR-TB

InhA, an Mtb enoyl acyl-carrier protein reductase, is the primary target of the front-line drug isoniazid (INH).^{9, 10} While it is one of the two most important antitubercular drugs and the only drug used for TB prophylaxis, INH suffers from resistance that continues to increase.^{1, 9, 11, 12} WHO data indicate up to 28% of all TB cases are INH-resistant, and in previously treated TB patients, up to 60% exhibit resistance, making it extremely difficult, time-consuming, and expensive to treat them (if they can be treated at all).^{1, 2, 13}

INH must be activated by Mtb catalase-peroxidase (KatG).^14–16 Most clinically relevant INH-resistant Mtb strains involve mutations in or deletions of katG, which abrogate activation of the INH prodrug.^{17, 18} In some areas, 70% of MDR-TB strains have mutations in katG, as do 100% of sequenced XDR-TB strains.^{19, 20} Although katG mutations are generally responsible for high-level resistance to INH in clinical isolates, those mutations can be enhanced by additional mutations in the promoter region of inhA, which cause low-level INH-resistance by increasing the amount of InhA produced,^{21, 22} and are found in up to 28% of INH-resistant clinical isolates (depending on the location of the study).^21–31 KatG activates INH to enable formation of a covalent adduct with NAD⁺ or NADH.¹⁴ As has been previously demonstrated, novel InhA inhibitors that do not require prior activation by KatG are not vulnerable to this key mechanism of INH resistance.^{11, 17}

Our virtual screen in this study was motivated by the clear need for a novel InhA-targeting drug that is effective against MDR and XDR-TB strains. A next generation InhA inhibitor, specifically lacking significant cross-resistance with INH by not requiring KatG activation, would be a valuable addition to the antitubercular armamentarium and help stem the tide of TB drug resistance. Guiding our inhibitor discovery and design efforts are numerous X-ray crystal structures of InhA bound to co-factor, substrate mimic, or various inhibitors.^{10, 32–45} The pursuit of InhA inhibitors lacking the requirement for activation by KatG has transpired through both structure-based and screening approaches.^{12, 46–49} Notably, the Tonge group ^{17, 46, 47, 49} and Freundlich et al. ¹¹ have both independently evolved potent triclosan-based inhibitors by leveraging X-ray crystal structures of InhA with bound diaryl ethers. However, a significant limitation of the triclosan scaffold has been the requirement for a phenol moiety, which suffers from rapid Phase II metabolism.^50–52 Our virtual screen focused on the NCI library to search for novel scaffolds that inhibit InhA without requiring prior activation by KatG and yet do not contain the problematic phenol group. Although many different factors affect in vivo pharmacokinetics and pharmacodynamics, especially when targeting a pathogen like Mtb that has an unusually thick and waxy cell wall, numerous efflux pumps and detoxification mechanisms, we sought to avoid the known liabilities that some current InhA inhibitors display.

High-throughput docking virtual screening (VS) studies have been used extensively in both academia and the pharmaceutical industry to discover inhibitors of select drug targets (median hit rate of 13% ⁵³) and are complementary to experimental target-based HTS.⁵⁴ “Docking” flexible models of small molecules computationally probes the energetic landscape governing macromolecular recognition with a target protein, to help guide the discovery and design of novel inhibitors.^55–62 Docking flexible models of potential ligands against atomic-scale models of different protein drug targets may reproduce or predict (a) how tightly these compounds bind; (b) where they prefer to bind; and (c) what specific interactions they form at the binding site.

Many VS studies, including some against InhA, have involved computational studies in the absence of experimental validation of their predictions.^63–69 In contrast, some pioneering VS against InhA have yielded predictions that were experimentally validated with enzyme inhibition assays⁷⁰ and/or whole-cell growth assays against Mtb,^{71, 72} M. vanbaalenii,⁷³ or M. smegmatis.⁷⁴ These previous, experimentally validated VS against InhA helped establish the feasibility of computer-aided drug discovery against this system and laid the foundation for the research we present.

MATERIALS AND METHODS

The Global Online Fight Against Malaria project and its relevance to TB research

IBM’s World Community Grid is a distributed network of over 2 million internet-connected personal computers in over 80 countries, making it effectively one of the largest supercomputers available. It is entirely devoted to humanitarian research at academic and non-profit institutions. The Global Online Fight Against Malaria (or “GO Fight Against Malaria,” GO FAM) was a project on World Community Grid that A.L.P. designed and executed while at TSRI.^75–77 In under two years, the GO FAM project used over 27,385 CPU years to perform virtual screens of 5.6 million commercially available compounds against over 200 structures of targets from 22 classes of validated and potential drug targets for the treatment of malaria and other diseases. Using AutoDock Vina⁶², GO FAM generated a total of 1.16 billion docking results. These docking studies would have taken over a hundred years to complete on typical academic computing resources. The compound libraries screened were: NCI, Enamine, Asinex, ChemBridge, and Vitas-M Labs, with 3-D models obtained from the ZINC server.⁷⁸ To the best of our knowledge, GO FAM is the first project to perform over a billion docking jobs, and it has produced the largest VS data sets against malaria and Mtb.

If a malaria target had structural similarity with valid or potential drug targets from other pathogens whose crystal structures were available, the compounds were docked as well against those cognate proteins, including Mtb, E. coli, Staphylococcus aureus, Yersinia pestis, and Brugia malayi. The Mtb subset of GO FAM involved InhA, DHFR (dihydrofolate reductase), OAR (oxo-acyl ACP reductase, or FabG), and cyclophilin A.

On GO FAM we docked a much larger number of compounds against InhA than all previous VS against it combined.^65–74 The results presented here encompass only 5.6% of the compounds screened on GO FAM against InhA—we began with the NCI library, because NCI compounds are available to researchers for free, through the NCI’s Developmental Therapeutics Program (DTP).

Screening the NCI library of compounds against InhA on GO Fight Against Malaria

The 316,000 pdbqt files generated for the NCI library (and for the other libraries that represent the 5.6 million compounds docked in the GO FAM experiments) are available at: http://zinc.docking.org/pdbqt. AutoDock Vina⁶² 1.1.2 (or “AD Vina”), which was grid-enabled for World Community Grid by IBM staff, was used to dock each compound in the library against the crystallographic conformation of InhA from 2×23.pdb.³⁹ In positive control re-docking experiments, the co-crystallized inhibitor PT70 docked to the target model of 2×23 with an RMSD = 0.49 Å. Additional (successful) positive control re-docking and cross-docking experiments that utilized AD Vina against other crystal structures of InhA bound to different ligands have been published recently elsewhere.⁷⁹ This 2×23 structure of InhA was selected for this study, because it is a complex with PT70, a slow, tight-binding inhibitor of InhA with a 7.8 nM K_i and a residence time of 24 minutes. Displaying a long residence time with a pathogenic target imparts favorable properties in vivo.^80–82 Using the PT70-induced conformation may enable us to identify and develop novel inhibitors that display long residence times with InhA. The ambitiousness of this goal necessitated the resources of the GO FAM project to screen 5.6 million compounds against this and other crystal structures of InhA. Hydrogen atoms were added to the target model using the MolProbity server.⁸³ Crystallographic waters and counterions were deleted from the pdb file. AutoDockTools 4.2⁸⁴ was then used to generate the pdbqt file. A grid box of 30 × 30 × 30 ų centered between the N9 atom of adenine and C1 of the adjacent ribose in the NAD cofactor defined the region that the docked compounds explored. This location and large size were selected to enable the same grid box to be used for all of the different InhA and PfENR targets that were part of these GO FAM experiments. The “exhaustiveness” setting in AD Vina was increased to 20 (from the default of 8), because of the large grid box used.

Target-specific energetic and interaction filtering of VS results

We characterized the predicted binding mode of each compound in the NCI library, by the number and types of energetically favorable interactions with the InhA active site and the estimated free energy of binding, using a software-automated workflow (python and tc-shell scripts) and established protocols.^85–87 The results were filtered to harvest compounds that displayed critical interactions (based on analyses of the features displayed by nanomolar inhibitors of InhA in existing crystal structures)^{11, 35, 39, 41}. These filters selected compounds for which the top-scoring binding mode/compound displayed base-stacking interactions between the candidate compound and the NAD⁺/NADH co-factor and at least two predicted hydrogen bonds to the active site. The selected compounds also had to display an estimated free energy of binding (FEB) ≤ −8.0 kcal/mol, according to the AD Vina scoring function. This set of docking filters harvested 91 compounds, whose binding modes were then visually inspected (see Figure 1).

Figure 1. Workflow Used to Discover Novel InhA Inhibitors in the Virtual Screen with the NCI Library on GO FAM.

(A) The workflow used energetic and interaction-based filters (e.g., requiring compounds to display an estimated free energy of binding ≤ −8.0 kcal/mol, base stack with the NAD cofactor, and form at least two hydrogen bonds with the active site) to filter the VS results and harvest NCI compounds for visual inspection. Candidates that passed visual inspection were then ordered and tested in InhA inhibition assays. Eight of the sixteen soluble compounds inhibited InhA activity by 27 to 71% at 100 μM. (B) The predicted binding modes for all eight novel InhA inhibitors are displayed as sticks-and-balls with cyan carbons, while the InhA target is shown in magenta. The NAD cofactor is rendered as CPK, and the key residues Gly96, Ser123, Phe149, Tyr158, Thr196, and Met199 are shown as thin sticks.

Visual inspection is a subjective process, but experience in macromolecular recognition and knowledge of the strengths and weaknesses of different modeling approaches used to view, measure, and judge/prioritize the docked results can be helpful. The use of human knowledge (“in cerebro” quality control) to prioritize computer-aided (“in silico”) predictions has been a successful strategy in previous blind docking challenges, such as SAMPL2 and SAMPL4.^87–89 Our visual inspection process incorporated multiple criteria, in an attempt to decrease the number of false positives that often result from virtual screens. Unfavorable aspects of a docking result included compounds that (a) have docked modes displaying distorted geometries (i.e., peptide bonds in the ligand models were allowed to freely rotate during docking, since the target was rigid, but docked modes displaying peptide bonds that were 30 to 90° from cis or trans were rejected); (b) have one or more large hydrophobic groups (e.g., phenyl or t-butyl) exposed to solvent; or (c) display more than three unfavorable electrostatic repulsion interactions with polar or charged groups in the target. Favorable aspects of a docked pose included compounds where (d) a majority of their heteroatoms are involved in favorable electrostatic interactions or hydrogen bonds (as discussed in our recent manuscript)⁷⁹, since heteroatoms will likely need to be added during the optimization process, without violating Lipinski’s rules⁹⁰; (e) the hydrophobic groups displayed van der Waals interactions with non-polar regions of the target (as measured by the AD Vina scoring function⁶², characterized using distance-dependent and atom pair-specific criteria implemented in the Fox software⁸⁵); and (f) aromatic rings displayed pi-pi (base-stacking or T-stacking)^91–93 or pi-cation interactions with the target^92–94, as characterized by the Fox software⁸⁵. The AD Vina scoring function uses ambiguous atom types (i.e., hydrogen bond donors are treated as donor and/or acceptor) and a spherical hydrogen bond potential (instead of an angle-dependent potential).⁶² Consequently, to verify the number of hydrogen bonds that the docking filters detected, the donor-acceptor distance and donor-hydrogen-acceptor angle (which should be between 120–180°) for predicted hydrogen bonds were all measured manually in PMV.⁸⁴ Because rotatable polar hydrogen atoms in the ligands are placed in arbitrary torsion angles by AD Vina,⁶² and since the model of the target was rigid, structural intuition was also employed when deciding which hydrogen bonds the docked mode displayed. Special emphasis was placed on candidates that (g) displayed hydrogen bonds with invariant residues and backbone atoms, since previous studies with MDR HIV protease and MDR P. falciparum DHFR have shown that displaying these features renders the evolution of drug resistance less likely.^95–111

InhA inhibition and kinetics experiments

An InhA inhibition assay was performed on the selected candidate compounds. Briefly, the candidate inhibitor was assayed at 100 μM in a reaction buffer (30 mM PIPES, 150 mM NaCl, 1 mM EDTA, at pH 6.8) containing 30 μM trans-2-dodecenoyl coenzyme A (DD CoA), 250 μM NADH, and 100 nM InhA. The enzymatic activity at 100 μM inhibitor concentration was quantified, and the three compounds that displayed the largest inhibition of enzyme activity were selected for further IC₅₀ measurements. In general, IC₅₀ values were determined by varying the concentration of inhibitors in the aforementioned reaction mixture. The data were analyzed using equation 1, where I is the inhibitor concentration and y is percent activity.

$$y=\frac{100\%}{1+\frac{I}{{IC}_{50}}}$$ Equation 1

To provide mechanistic insight, the top two compounds were chosen for subsequent measurements of their inhibition constant (K_i). The third best inhibitor displayed solubility problems at higher concentrations. The K_i value was calculated by determining the k_cat and K_m (DD CoA) values at several fixed inhibitor concentrations using the same assay condition described above. The data were analyzed using the standard equations for competitive and non-competitive inhibition (equations 2 and 3).

$$v_o=\frac{v_{\mathit{max}}[S]}{K_M\left(1+\frac{[I]}{K_i}\right)+[S]}$$ Equation 2

$$v_o=\frac{v_{\mathit{max}}[S]}{K_M\left(1+\frac{[I]}{K_i}\right)+[S](1+\frac{[I]}{K_i^{\prime }})}$$ Equation 3

Comparing the new InhA inhibitors to known InhA inhibitors

To assess the chemical property space of the eight InhA inhibitors found, a Principal Component Analysis (PCA) was performed on the combination of the 157 known InhA inhibitors in TB Mobile 2 and the eight inhibitors discovered.^{112, 113} The sdf of the TB Mobile data set was used for the Principal Component Analysis(PCA), which was calculated using Discovery Studio 3.5¹¹⁴ and used eight interpretable descriptors (AlogP, molecular weight, number of rotatable bonds, number of rings, number of aromatic rings, number of hydrogen bond acceptors, number of hydrogen bond donors, and molecular fractional polar surface area).

The pairwise Tanimoto similarity¹¹⁵ was calculated for each of the new InhA inhibitors discovered in the NCI library versus each of the 154 known InhA inhibitors from TB Mobile ¹¹². This again used the TB Mobile dataset sdf for which MDL descriptors were calculated, followed by Tanimoto similarity using the “find similar molecules by fingerprints” protocol in Discovery Studio 3.5¹¹⁴. This then enabled us to calculate the minimum, maximum and average similarity as measures of proximity to known InhA inhibitors.

The pairwise Tanimoto similarity¹¹⁵ was also calculated for each of the 8 new InhA inhibitors versus each of the 7 other compounds in order to determine the number of distinct scaffolds, according to a cut off > 0.7¹¹⁶ (see Supporting Information Table S1).

Minimum inhibitory concentration (MIC₉₀) assay vs. Mtb

For these whole-cell in vitro studies, new batches of NCI 99389 and 111591 were obtained from the NCI. LC-MS data confirmed that each sample had the expected molecular weight and was >97% pure by HPLC @ 250 nm (see the Supporting Information, Figures S1–S2 and S4–S5). ¹H NMR (600 MHz), using d₆-DMSO for 99389 and d₆-acetone for 111591 as NMR solvents, confirmed that each sample had the expected structure (see the Supporting Information, Figures S3 and S6). MIC₉₀ values of the compounds were determined following the microplate-based Alamar Blue assay (MABA) method as previously described.¹¹⁷ 50 mM compound stock in DMSO was dissolved in sterile Middlebrook 7H9-OADC broth, making a 1 mM pre-test solution. 100 μL pre-test solution was added into wells in column 1 of a sterile polystyrene 96-well round-bottom plate (CLS3795, Corning, NY). Wells in column 2 to 12 received 50 μL of sterile 7H9-OADC broth. Serial two-fold dilutions of compounds were performed, and column 12 was set as drug-free (inoculum-only) control. Final concentrations of compounds were as follows: for INH, 0.012 to 25 μg/mL; for NCI 99389 and 111591, 0.50 to 500 μM. M. tuberculosis wild type strain H37Rv and inhA over-expression strain mc²4914¹⁰ were 1:1000 diluted in 7H9+OADC medium at mid-logarithmic stage of growth (OD_595nm=0.4). 50 μL of diluted bacteria suspensions were inoculated into each well. Plates were sealed with Parafilm and incubated at 37 °C for seven days. 20 μL alamarBlue^® reagent (Invitrogen, Frederick, MD) freshly mixed with 12.5 μL 20% Tween 80 was added into each well, followed by 24-hour incubation in 37 °C. Absorbance was read at 570 nm, with reference wavelength 600 nm, using a microplate reader (ELX808, Biotek Instruments). The MIC₉₀ endpoint was defined as the lowest concentration of the test agent that produced at least 90% reduction in absorbance compared with that of the drug-free control.

Cellular toxicity assay

The cytotoxicity of 99389 and 111591 were determined by the 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazoliuim bromide (MTT) cytotoxicity assay with the Vybrant MTT Cell Proliferation Assay Kit (Molecular Probes).¹³ Vero cells (African green monkey kidney epithelial cells; ATCC) were plated in a sterilized 96-well plate (Costar 3595, Corning, NY) at 2 × 10⁵ cells/well with volume 50 μL and incubated for 3 hours in a 37 °C incubator. 99389 and 111591 were dissolved in DMSO at a final concentration of 10 mg/mL. In a separate 96-well plate, two-fold serial dilutions were performed with Dulbecco’s modified eagle medium (Gibco) supplemented with 10% fetal bovine serum. 50 μL of diluted compounds were added to appropriate test plate wells to generate final concentrations ranging from 0.78 to 100 μg/mL. Column 12 was utilized for a drug-free control. After 48 hours treatment, 10 μL of 12 mM MTT stock solution was added into each well and incubated at 37 °C for 4 hours. 100 μL of 0.1 g/mL SDS•HCl solution was added into each well subsequently to halt the reaction, followed by 4-hour incubation at 37 °C. The absorbance at 570 nm was read with a VersaMax ELISA microplate reader (Molecular Devices). The CC₅₀ was extrapolated by plotting absorbance at 570 nm versus concentration of untreated Vero cells control.

RESULTS AND DISCUSSION

This pilot study of a small subset (i.e., 5.6%) of the GO FAM results vs. InhA has led to the discovery of several novel inhibitors (see Figures 2–4). 19 compounds were ordered, but three of them were insoluble at 100 μM in DMSO and could not be assayed (see Table 1). Eight of the 16 soluble candidates were modest inhibitors of InhA (i.e., showing 27% to 71% inhibition at 100 μM). Using the same assay conditions and protocol (i.e., without pre-incubating the compound and InhA), the very potent positive control compound PT70 displayed 75% inhibition at 100 nM (i.e., the well-optimized lead compound PT70 was >1000 times more potent than our best new inhibitor). If scaffolds are defined as small molecule chemotypes that possess less than 70% similarity according to Tanimoto values¹¹⁶, these eight inhibitors represented five novel scaffolds against InhA (see Figure 4 and Supporting Information Table S1). As required by the docking filters, all eight compounds are predicted to base stack with the NAD cofactor (see Figures 2–3), instead of forming a covalent adduct with it. As demonstrated by the InhA activity assays, they do not require prior activation by Mtb KatG. “Hits” were classified in a way that involved the potency for a particular size of compound. “Fragment hits” (novel inhibitors with a MW < 300 Daltons)¹¹⁸ tend to display a potency in the 100’s of μM to low mM range (according to K_i or K_d).^{88, 118–124} Consequently, we defined fragment hits as novel inhibitors with a MW < 300 g/mol that displayed a K_i value < 100 μM. Two of our inhibitors are “fragment hits”; they are structurally distinct from known InhA inhibitors (see Figures 2, 5 and Table 2); and they displayed K_i values of 54.1 ± 5.4 and 59.2 ± 8.7 μM, respectively (see Figure 6). Thus, using a stringent metric, 2 out of 16 compounds were novel fragment hits against InhA (i.e., a 12.5% hit rate), representing two promising new scaffolds.

Figure 2. Predicted binding modes of the two most potent new InhA inhibitors discovered in GO FAM experiment 5.

The docked modes produced by AutoDock Vina are displayed as ball-and-sticks with cyan carbon atoms, and the InhA target (2X23.pdb) is displayed as magenta ribbons. The NAD cofactor is displayed in CPK, and the key residues Gly96, Ser123, Phe149, Tyr158, Thr196, and Met199 are shown as thin sticks. A close-up view of the predicted binding mode of the top fragment hit, NCI 99389 (K_i^app = 54.1 ± 5.4 μM), is shown in (A), while the full view is displayed in (B). In (C) the docked mode of NCI 99389 is compared to the experimentally-determined binding mode of PT70, the inhibitor that crystallized with InhA in 2×23.pdb, which is displayed as ball-and-sticks with magenta carbons. In (D) the predicted binding mode of the 2^nd most potent fragment hit, NCI 111591 (K_i^app = 59.2 ± 8.7 μM), is displayed.

Figure 4. Summary of the 2D structures, docking scores, and InhA inhibitory activities of the eight new inhibitors discovered.

The most potent new InhA inhibitor discovered is shown in the top-left corner, while the least potent new inhibitor is displayed in the bottom-right corner. These eight new inhibitors correspond to five novel scaffolds versus InhA (i.e., NCI 111588 – 111591 represent analogs of one scaffold, according to a Tanimoto cut-off of 0.7; see Supporting Information Table S1). FEB signifies the estimated free energy of binding value from AutoDock Vina’s scoring function, in kcal/mol. The Lig. Eff. is the calculated ligand efficiency from AutoDock Vina, in kcal/mol/heavy atom. The % inhibition of InhA activity was produced when each compound was present at 100 μM. The region of each compound that was predicted to base stack with the NAD cofactor is highlighted with a red circle.

Table 1.

Compound ID’s, 2D structures, and InhA inhibition data.

NCI ID	Structure	Kiapp (μM) & mechanism	InhA inhibition at 100 μM inhibitor
660846			2.7%
683622			9.2%
609097			10.3%
314884			10.3%
371850			Not soluble
11142			4.5%
75300			Inhibitor precipitated in buffer
75301			Inhibitor precipitated in buffer
99389 (ZINC01654204)		54.1 ± 5.4 competitive	71.0% (estimated ~ 40 μM) IC50
111588 (ZINC0135077)			29.1%
111589 (ZINC04994329)			33.7%
111590 (ZINC00129134)			30.9%
111591 (ZINC01703321)		59.2 ± 8.7 noncompetitive	42.8%
112144 (ZINC04878446)		205.6 ± 46 (precipitation observed at high concentrations)	34.6%
130836			9.1%
135809 (ZINC01722139)			29.1%
196166 (ZINC01734860)			27.3%
213837			10.9%
293934			9.1%

Note: Each compound was initially tested at a single concentration. Only the top inhibitors (highlighted with bold font) were subjected to additional experimental characterization. For the top eight new inhibitors, both the NCI ID and the ZINC ID are displayed.

Figure 3. Predicted binding modes of the least potent new InhA inhibitors discovered in GO FAM experiment 5.

The docked modes produced by AutoDock Vina are displayed as ball-and-sticks with cyan carbon atoms. The InhA target (2×23.pdb) is displayed as magenta ribbons, with the NAD cofactor as CPK. The key residues Gly96, Ser123, Phe149, Tyr158, Thr196, and Met199 are shown as thin sticks. The predicted binding modes of the following new InhA inhibitors are depicted: (A) NSC 112144 (K_i^app = 205.6 ± 46 μM), (B) NSC 111589, (C) NSC 111590, (D) NSC 111588, (E) NSC 135809, and (F) NSC 196166.

Figure 5. Comparing the chemical space of the new InhA inhibitors to known InhA inhibitors.

A Principal Component Analysis (PCA) was performed on the combination of the 157 known InhA inhibitors in the TB Mobile 2 data set and the 8 novel InhA inhibitors discovered. Three PCs explain 84.8% of the variance observed. The 157 InhA inhibitors in the TB Mobile 2 data set are displayed in magenta. The two most potent new InhA inhibitors discovered are depicted in green, and the other six novel InhA inhibitors identified are in blue. A red circle highlights the location of NCI 111591. The PCA indicates that the 8 new InhA inhibitors have similar chemical properties to known InhA inhibitors, but they are generally not within the main clusters of these previously characterized InhA inhibitors.

Table 2.

Tanimoto Similarity to 154 Known InhA Inhibitors in the TB Mobile data set

Compound IDs	Average Similarity	Maximum Similarity	Minimum Similarity
1) NCI 99389 ZINC01654204	0.3087	0.4130	0.1458
2) NCI 111591 ZINC01703321	0.3681	0.4634	0.1311
3) NCI 112144 ZINC04878446	0.4503	0.5738	0.15
4) NCI 111589 ZINC04994329	0.4307	0.5732	0.1642
5) NCI 111590 ZINC00129134	0.4296	0.5522	0.1935
6) NCI 111588 ZINC0135077	0.4542	0.5897	0.1406
7) NCI 135809 ZINC01722139	0.2731	0.4366	0.1061
8) NCI 196166 ZINC01734860	0.3551	0.4833	0.1724

Figure 6. Kinetic data on the two most potent InhA inhibitors discovered indicate NCI 99389 is a competitive inhibitor, while 111591 is noncompetitive.

IC₅₀ values were measured for the top 2 fragment hits, followed by a detailed mechanistic study to measure the K_i values. NCI 99389 showed a Ki^app of 54.1 ± 5.4 μM and a competitive binding mechanism, indicating that the inhibitor competed with the CoA substrate and bound directly to the enzyme. Conversely, NCI111591 had a Ki^app of 59.2 ± 8.7 μM and a non-competitive binding mechanism, suggesting a more complex scenario where the inhibitor could bind to both the holo enzyme and to the substrate-enzyme complex.

The most promising new scaffold, NCI 99389, is a 4,6-diaminopyrimidine and was predicted to form the following quaternary interactions in the docking studies: base stacking with the nicotinamide ring of the NAD cofactor (similar to the crystallographic binding mode of PT70)³⁹; T-stacking with the side-chain of Phe149; hydrogen-bonding with the hydroxyl of Tyr158 (similar to PT70)³⁹; hydrogen-bonding with the 2′-hydroxyl of the ribose adjacent to the nicotinamide ring of NAD (similar to PT70)³⁹; hydrogen-bonding with the sulfur in the side-chain of Met199; favorable electrostatic interactions with the hydroxyl of Tyr158; and favorable electrostatic interactions with the phosphate group proximal to the ribose of NAD (see Figure 2.A). As a reference, in addition to the aforementioned hydrogen bonding and base stacking interactions that the crystallographic binding mode of PT70 displays, PT70 also contains a six carbon alkyl tail that packs very well into a hydrophobic pocket of InhA. Despite using a large grid box in these docking calculations (i.e., allowing the ligands to sample a large volume of the protein), docking calculations suggest a clear preference for 99389 to bind in a similar fashion to the crystallographic binding mode of PT70 (see Figure 2.C). Experimental structure verification will be necessary to firmly establish the actual binding mode. Since this compound is an un-optimized fragment hit and lacks the long alkyl tail that PT70 contains, it might not produce the same “closed” conformation that PT70 induces^{39, 125} (which might give 99389 a different binding mode than what these docking calculations predicted), and/or it might have to pay a larger enthalpic penalty to induce the ligand-bound conformation that InhA forms (which, when combined with the numerous hydrophobic contacts that it lacks, might also explain part of the >1,000-fold difference in potency versus PT70). NCI 99389 displayed an IC₅₀ of approximately 40 μM against InhA (see Table 1). In subsequent kinetic experiments (see Figure 6 and Table 1), it displayed an apparent K_i of 54.1 ± 5.4 μM, and it demonstrated a competitive mechanism of inhibition with respect to substrate.

The second most promising new scaffold discovered, NCI 111591, is a 5-amino-1H-1,2,3-triazole and had a docked mode that displayed the following quaternary interactions with InhA: base stacking with the nicotinamide ring of the NAD cofactor; hydrogen-bonding with the hydroxyl of Tyr158; hydrogen-bonding with the 2′-hydroxyl of the ribose adjacent to the nicotinamide ring of NAD; hydrogen-bonding with the carbonyl oxygen of the nicotinamide ring; favorable electrostatic interactions with the sulfur in the side-chain of Met199; favorable electrostatic interactions with the phosphate group proximal to the ribose of NAD; and favorable electrostatic interactions with the 2′-hydroxyl of the ribose that is adjacent to the nicotinamide ring of NAD (see Figure 2.D). When present at 100 μM, NCI 11591 inhibited InhA activity 42.8%. It displayed an apparent K_i of 59.2 ± 8.7 μM. As shown in Figure 6, 111591 followed a non-competitive mechanism of inhibition (i.e., it can bind when the substrate is “not on or on,” meaning it could potentially bind to the holo enzyme, to the InhA:substrate complex, and/or to the InhA:NAD⁺:product complex). To further confirm the inhibition mode, data were fit to both Equation 2 and Equation 3 for comparison. The K_i values generated from the non-competitive fit and competitive fit were 59.2 ± 8.7 μM and 54.0 ± 51.0 μM, respectively. The much greater error for the competitive fit unambiguously confirmed the inhibition mechanism as non-competitive.

The predicted binding modes for the six less potent (and thus less promising) new InhA inhibitors discovered are presented in Figure 3. They were all predicted to base-stack with the nicotinamide ring of the NAD cofactor and to form a hydrogen bond with the hydroxyl of Tyr158. Five of the six inhibitors (i.e., all except NCI 111590) docked to form both a hydrogen bond and an additional favorable electrostatic interaction with the 2′-hydroxyl of the ribose adjacent to the nicotinamide ring of NAD, but the predicted pose of 111590 only formed a favorable electrostatic interaction with that hydroxyl. 112144 was also predicted to make a hydrogen bond with the carbonyl of Gly96, but it was predicted to display unfavorable electrostatic repulsion with both phosphate groups of the NAD cofactor. The amino and ether oxygen atoms of 111589 were predicted to form both favorable and unfavorable electrostatic interactions, respectively, with the sulfur in the side-chain of Met199, and it displayed an internal hydrogen bond between that amino and ether oxygen. The conserved regions of the scaffold in 111590 and 111588 had docked modes that superimposed, and they formed favorable electrostatic interactions with the sulfur in the side-chain of Met199 and with the hydroxyl of Ser123. If a slight conformational change occurs in the side-chain of Phe149, they could both T-stack with it. Docking of 135809 suggested favorable electrostatic interactions could be formed with the phosphate group adjacent to the ribose of the NAD cofactor, but it formed unfavorable electrostatic repulsion with both the sulfur in the side-chain of Met199 and the carbonyl of the nicotinamide ring of the NAD cofactor. The weaker InhA activity of 135809 compared to the other compounds with the most similar structures to it (see Figure 4 and Supporting Information Table S1) might be due to these unfavorable electrostatic repulsions and perhaps to the less stable/less likely protonation state of the central thiadiazole ring (i.e., of the different independent models for the different protonation states of this compound that were screened, the model that passed the docking filters had a protonated thiadiazole, but ChemDraw calculations predict that this compound should not be protonated at neutral pH). 196166 was the least potent InhA inhibitor, and its docked mode displayed three unfavorable electrostatic repulsions with the phosphate adjacent to the ribose of the NAD cofactor. The terminal pyridine also displayed unfavorable electrostatic repulsion with the sulfur in the side-chain of Met161, but the pyridine that is part of the two fused rings might T-stack with Phe149.

The 8 inhibitors discovered were compared to a set of 157 previously-characterized InhA inhibitors available in TB Mobile 2.^{112, 113} Principal Component Analysis (PCA) showed that these 8 InhA GO FAM inhibitors were generally not part of the main clusters of known InhA inhibitors, but they do have similar chemical properties to them (see Figure 5). Each inhibitor was also individually compared to all known InhA inhibitors in the TB Mobile data set to obtain sets of pairwise Tanimoto coefficients, which were then averaged. Their average Tanimoto similarities ranged from 0.27 to 0.45 (with 1.0 indicating identical 2D structures). The two most promising new scaffolds discovered, NCI 99389 and 111591, had a maximum Tanimoto similarity to a known InhA inhibitor of 0.413 and 0.463, with a minimum similarity of 0.146 and 0.131, and average similarity values of 0.309 and 0.368, respectively. These distinct cheminformatic analyses support these 8 GO FAM compounds as novel InhA inhibitors based on current literature data.

The top two fragment hits were then studied with (a) whole-cell in vitro Mtb growth experiments, using both the wild type strain H37Rv and the inhA-over-expressing strain mc²4914,¹⁰ and with (b) Vero cell cytotoxicity experiments (see Table 3). Unfortunately, the top InhA inhibitor discovered, 99389, displayed an MIC₉₀ of 500 μM against wild type Mtb. The MIC > 500 μM it exhibited against the inhA-overexpressor strain of Mtb suggests that InhA might be the primary target against whole-cell Mtb. However, this compound lacked sufficient efficacy against Mtb and displayed considerable cytotoxicity with Vero cells, with a CC₅₀ < 3.0 μM. As expected, the positive control INH displayed almost a 10-fold decrease in potency against the inhA overexpressor: MIC₉₀ of 0.4 μM (0.05 μg/mL) against wild type Mtb and 3 μM (0.4 μg/mL) against strain mc²4914. The second most promising InhA inhibitor discovered, NCI 111591, displayed an MIC₉₀ of 125 μM against both wild type Mtb and the inhA-overexpressing strain of Mtb. Although 111591 was 4-fold more potent against whole-cell Mtb, the lack of a shift in its potency against the inhA overexpressor indicates that InhA is likely not its primary target in Mtb. In the PCA that compared the new inhibitors to the 157 known InhA inhibitors in TB Mobile 2^{112, 113}, 111591 was further from the main clusters of known InhA inhibitors, while 99389 was at the edge of the central cluster (see Figure 5). NCI 111591 was less toxic against Vero cells, with a CC₅₀ of 26 μM; however, it still displayed an insufficient selectivity index of 0.21.

Table 3.

Summary of Anti-Mtb Efficacy and Mammalian Cell Cytotoxicity

NCI ID	Structure	Vero cell cytotoxicity: CC50 (μM)	MIC90 vs Mtb(μM)
			H37Rv wild type	mc24914 inhA overexpressor
99389		< 3.0	500	> 500
111591		26	125	125

Although several InhA inhibitors with nM potency have been reported previously,^{11, 15, 41, 126} novel chemotypes that inhibit InhA are needed for the following reasons: (a) the presence of the phenol group in triclosan derivatives ^{11, 41, 126} poses a metabolic liability for in vivo applications, ^50–52 (b) other advanced InhA leads have fared poorly when administered in mouse models of TB,¹⁵ and (c) drug-resistant strains of Mtb (i.e., MDR-TB, XDR-TB, and TDR-TB) continue to evolve and spread throughout the world.^1–7 In addition, since 28% to 60% of TB cases are INH-resistant, ^{1, 2, 13} and InhA is one of the most validated targets for treating TB, new chemotypes that inhibit InhA without displaying cross-resistance with INH could eventually seed the development of urgently needed new drug combinations for the treatment of active and latent TB infections.

The novel fragment hits we discovered against InhA displayed low μM potency, with K_i values of 54.1 and 59.2 μM for the two most promising new scaffolds. Using stringent criteria to define a hit (a K_i value < 100 μM for a fragment-sized compound), this VS had a 12.5% hit rate. Thus, from a computational chemistry perspective, this VS was a success, since the median hit rate from hundreds of published VS is ~ 13%.⁵³ More importantly, these novel InhA inhibitors are all predicted to base stack with the NAD cofactor (instead of forming a covalent adduct with it). Our InhA inhibition assays demonstrated that these novel inhibitors do not require prior activation by Mtb KatG, which means that they should not be susceptible to the main mechanism of INH resistance found in clinical settings. In addition, the novel inhibitors discovered all lack the presence of the phenol group that poses a metabolic liability for triclosan derivatives. Consequently, the two most promising new chemotypes discovered might eventually enable the development of new InhA inhibitors that are effective against MDR-TB, XDR-TB, and TDR-TB. Since it is slightly weaker than NCI 99389, we consider NCI 11591 to be the second most promising new scaffold against InhA discovered in the present study. However, given its superior MIC against Mtb and its lower toxicity against Vero cells, NCI 111591 seems to be a more promising scaffold versus whole-cell Mtb.

Although these new chemotypes discovered against InhA are somewhat weak inhibitors (i.e., they are “fragment hits” and not drug-sized “leads”), they were identified from a diverse, commercial library using freely available compounds. In addition, they are both fragment-sized compounds (i.e., MW < 300 g/mol)¹¹⁸. Fragment-based hit discovery was not our initial goal, but it was the ultimate result of this pilot study. Unlike traditional high-throughput screens, fragment-based drug discovery is founded on screening a smaller number of smaller-sized compounds, to advance the goal of discovering novel fragment hits with K_i or K_d values in the high μM to low mM range.¹¹⁹ Those initial novel fragments can then be optimized using structure-based and medicinal chemistry strategies to develop potent leads, some of which have advanced to become a clinical candidate¹¹⁸ or an FDA approved drug.¹²⁷ In the pioneering “SAR by NMR” study, the novel fragment hits discovered displayed K_d values of 100 μM to 9.5 mM.¹²⁴ Although a few studies have discovered fragment hits with potencies as great as 24 μM,¹²⁸ 49 μM,¹²⁸ 60 μM,¹²⁸ or 80 μM,¹²⁹ most fragment hits have potencies (i.e., K_i or K_d values) in the 100 to 300 μM range, and many are in the low mM.^{88, 118–120, 122, 123, 128–130} Thus, although our two fragment hits have a very weak potency when compared to optimized lead compounds (such as PT70), our hits have not yet been optimized against the InhA target or against whole-cell Mtb, and they are actually more potent than most novel fragment hits.

Our most promising new fragment hit against InhA has a fairly simple structure and should be amenable to structure-based, medicinal chemistry-guided optimization. Due to their lack of whole-cell efficacy against Mtb and their inadequate selectivity indexes for Vero cell cytotoxicity, both of our top fragment hits will need considerable optimization before advancing to the lead compound stage. Several previous studies performed structure-activity-relationship (SAR) experiments to guide the development of more potent InhA inhibitors.^{41, 43, 47, 48, 126} Those studies suggest that our most promising new fragment hit could perhaps be developed into a more potent InhA inhibitor by appending appropriate functionality off one or both of the phenyl rings to make additional energetically favorable interactions with residues such as Gly96, Phe97, Phe149, Met155, Pro156, Ala157, Pro193, Ala198, Met199, Ile202, Val203, Leu207, Gln214, Ile215, Leu218, or the NAD cofactor.

The similarity in the substructures that are predicted to base stack with the NAD cofactor (see Figure 4) suggests that click chemistry, especially target-guided click chemistry,^131–134 might be a useful approach to aid the discovery and development of novel InhA inhibitors. Target-guided click chemistry is based on the principle that the azide and alkyne-containing fragments will interact with each other and click together (to form the triazole ring) only if those fragments are correctly positioned and have high affinity and long residence times in the target enzyme.^131–134 Since inhibitors that display long residence times with pathogenic targets can have more favorable properties in vivo,^{39, 80–82, 125} and since our computational results indicate that the triazole ring is predicted to form key base stacking and hydrogen bonding interactions with InhA (instead of just serving as a linker), we suggest that in situ click chemistry should be investigated in future studies against InhA.

CONCLUSIONS

Our Virtual Screen of the NCI library with a published crystal structure (PDB ID: 2×23³⁹) led to the discovery of two promising and novel fragment hits that inhibit InhA activity. This pilot study demonstrated the utility of the (public domain/open access) GO FAM docking data against InhA and of our approach to its analysis. Novel inhibitors of the key TB drug target InhA were discovered in an efficient manner, requiring the experimental assessment of fewer than 20 candidate compounds. These open access GO FAM data against InhA and other targets for treating TB and malaria represent a valuable resource for the drug discovery community.