An in silico screen and structural analysis identifies bacterial kinase inhibitors which act with β-lactams to inhibit mycobacterial growth

Abstract

New tools and concepts are needed to combat antimicrobial resistance. Actinomycetes and firmicutes share several eukaryotic-like ser/thr kinases (eSTK) that offer antibiotic development opportunities, including PknB, an essential mycobacterial eSTK. Despite successful development of potent biochemical PknB inhibitors by many groups, clinically useful microbiologic activity has been elusive. Additionally, PknB kinetics are not fully described, nor are structures with specific inhibitors available to inform inhibitor design. We used computational modelling with available structural information to identify human kinase inhibitors predicted to bind PknB, and we selected hits based on drug-like characteristics intended to increase the likelihood of cell entry. The computational model suggested a family of inhibitors, the imidazopyridine aminofurazans (IPAs), bind PknB with high affinity. We performed an in-depth characterization of PknB and found that these inhibitors biochemically inhibit PknB, with potency roughly following the predicted models. A novel x-ray structure confirmed that the inhibitors bind as predicted and made favorable protein contacts with the target. These inhibitors also have antimicrobial activity towards Mycobacteria and Nocardia. We demonstrated the inhibitors are uniquely potentiated by β-lactams but not antibiotics traditionally used to treat mycobacteria, consistent with PknB’s role in sensing cell wall stress. This is the first demonstration in the phylum Actinobacteria that some β-lactam antibiotics could be more effective if paired with a PknB inhibitor. Collectively, our data show that in silico modeling can be used as a tool to discover promising drug leads, and the inhibitors we discovered can act with clinically relevant antibiotics to restore their efficacy against bacteria with limited treatment options.

Introduction

Antibiotic resistant bacteria are one of the greatest modern health threats and, if left unchecked, will outpace cancer as a cause of death by 2050¹. Despite renewed calls for novel approaches to antibiotic discovery and development, new antibiotic discovery continues to lag, and is unprofitable to pharmaceutical companies². Mycobacterium tuberculosis (Mtb) is a particular concern as it currently infects two billion people worldwide, and infections with multidrug resistant tuberculosis (MDR) and extensively drug resistant tuberculosis (XDR) are steadily increasing³. Most recently, the NIH has listed non-tubercular mycobacterial (NTMs) infections as an area of particular interest due to rapid increases in these infections with similar resistance concerns⁴. Recently M. abcessus has emerged as a particularly drug resistant non-tubercular mycobacterial infection associated with hospital outbreaks⁵. Mycobacterial antibiotic resistance is multifactorial, including genetic target mutations and an intrinsic antibiotic tolerance due to an impenetrable cell wall and inactivating enzymes such as genetically encoded β-lactamases and aminoglycoside acetyltransferases which specifically restrict the use of the corresponding antibiotics^6–8. Because of high intrinsic and developed resistance, treatment options for mycobacteria have remained limited and stagnant for over forty-five years⁹. A variety of different β-lactamase inhibitors can allow some β-lactams to have clinically useful activity^10–12, but even with a β-lactamase inhibitor, the β-lactam potency is low enough that they are rarely used. Additionally, several other related Gram+ pathogens are becoming increasingly difficult to treat in certain settings. Nocardia spp. are susceptible to a narrow range of drugs, many of which are not tolerated in all patients, decreasing second-line options¹³. Collectively, it is clear that new approaches and new targets are essential to diversify treatment options and combat resistance¹⁴.

Most antibiotic drug development is aimed at targets with no eukaryotic analogs to minimize off target effects. Nevertheless, selectivity is attainable for structurally conserved targets. For example, despite close structurally similar human targets, mycobacterial proteasome proteins can be selectively targeted¹⁵, demonstrating that antibiotic development need not be limited to targets with no human homologs. Similarly, despite human serine/threonine kinases (STKs) having high structural conservation, relatively selective human kinase inhibitors are among the most successful drugs of the 21st century¹⁶. Interest in developing human kinase inhibitors has resulted in a wealth of corresponding structural and biochemical information as well as successful and failed compounds^{17, 18}. Many of these compounds have well developed structure-activity relationship (SAR) data, known toxicity profiles, and are synthetically feasible¹⁹.

Bacterial signal transduction pathways are not targets of current antibiotics²⁰. Still, bacterial eukaryotic-like serine/threonine protein kinases (eSTKs) are structurally similar to human STKs and regulate transcription, metabolism, cell cycle regulation, virulence, and drug resistance^{21, 22}. Penicillin-binding-protein And Ser/Thr kinase-Associated (PASTA) kinases are integral membrane eSTKs that bind peptidoglycan fragments and regulate growth, cell wall maintenance, metabolism, biofilm formation, and β-lactam resistance in a wide range of important Gram positive pathogens^{21, 23–31}. The mycobacterial PASTA kinase, PknB, is essential for cell survival, virulence, and exit from latency which has spurred interest in targeting it for drug development^32–34. Unlike human STKs, bacterial eSTKs have limited published structural and biochemical information. Mtb PknB is the most structurally studied, including some nonspecific inhibitor bound structures^35–39, but kinetic characterization and structurally informed SAR data are not published. Several groups have discovered PknB inhibitors through traditional screening processes, and some of these compounds were optimized to have nanomolar biochemical activity and micromolar-to-millimolar microbiologic activity^39–42.

Since bacterial eSTKs are structurally similar to human STKs, we hypothesized that we could find PknB inhibitors by screening human kinase inhibitors in silico. Thousands of known kinase inhibitor structures are available in digital databases which can be used for computational modelling,^{43, 44} and re-purposing drugs in this way is potentially cost effective and efficient⁴⁵. Although several groups have modeled inhibitors in PknB, these studies have generally investigated known inhibitors to determine preferred molecular interactions^{41, 46, 47}. We found that our in silico structural docking produced valuable active hits, and we followed up on a family of compounds with unique but consistant binding modes compared to predictions on other known inhibitors. We also completed the first kinetic characterization of PknB and were able to find this family of compounds, the imidazopyridine aminofurazans (IPAs), had biochemical activity and some SAR roughly reflected by the computational model. In addition, we were able to determine the structure of an IPA bound to PknB which not only closely matched the computationally predicted binding mode but explained some of the observed SAR and provided insights for future inhibitor development. Similar to the inhibitors identified by others, the IPAs had nanomolar biochemical activity and microbiological activity in the micromolar range. Others have attempted to improve poor microbiological activity by coupling PknB inhibitors with traditional mycobacterial treatments such as isoniazid and ethambutol but have not seen improvement⁴⁰. We hypothesized that since genetic deletion or pharmacologic inhibition of PASTA kinases in related Gram+ bacteria (phylum firmicutes) increases β-lactam susceptibility^{31, 48–50}, this previously untested strategy may also improve microbiologic function of PknB inhibitors in mycobacteria. We found that IPAs successfully potentiated β-lactams in several non-pathogenic and pathogenic mycobacteria as well as a related Gram+ bacteria, Nocardia asteroides. Collectively, our data show that computational models can be integrated with biochemical and structural information to discover and characterize effective PASTA kinase inhibitors. Furthermore the linkage between the recognized PASTA kinase drug target and cell wall stress provides an opportunity to increase inhibitor effectiveness and possibly form the basis for a novel clinical therapy.

Experimental Section

Docking

Coordinates for PknB (1O6Y) were downloaded from the protein data bank and prepared for docking using Sybyl 2.1.1(Tripos). Ligands and waters were removed and Gasteiger-Huckel charges were added. 3D coordinates for ligands in the PKIS libraries were obtained from the Small Molecule Screening Facility at UW-Madison¹⁷. The structure data file for the Selleck library was obtained from Selleck and converted into 3D coordinates using OMEGA (OpenEye Scientific). OMEGA was also used to add AM1BCC charges to the ligands in all libraries. A total of 1054 unique kinase inhibitors from three libraries (Selleck, PKIS 1, and PKIS 2) were prepared. Autodock Tools 1.5.6 was used to prepare the receptor coordinates (PknB) for docking ⁵¹. Docker 1.0.4 script was modified to retain supplied charges in the ligand files and was then used to automate ligand docking to the receptor using Autodock 4⁵¹. Docking was done at the rate of 30 iterations per ligand using a Lamarckian genetic algorithm. Docking results were organized and graphed using Excel (Microsoft) and PRISM (GraphPad). Poses were viewed using Pymol (PyMOL Molecular Graphics System, Version 1.7.6.0 Schrödinger, LLC.) with the Autodock/Vina plugin⁵². Chemical structures for figures were rendered using ChemDraw (Perkin Elmer). Lipinski’s rules were calculated using Sybyl and used to further narrow down selection of our starting compounds⁵³.

Chemicals and Reagents

All chemicals were purchased through Fisher Scientific unless otherwise noted. Miniprep kits and IPTG were from IBI scientific. Restriction enzymes were from New England Bio Labs. DTT and Ampicillin were from Dot Scientific. LB was from Growcells. GS4FF resin was from GE Healthcare. Meropenem was from APP Pharmaceuticals. Magnesium chloride, ATP·Mg, PMSF, chloramphenicol, glycerol, and bis-tris propane were from Sigma. PEG 3350 was from Microscopy Sciences. ATPγ³²P was obtained from Perkin Elmer. Thymine was from Acros Organics. β-cyclodextran was from Alfa Aeser. GSK690693 was from Selleck Chemicals. Other IPA compounds were acquired through collaboration with co-authors at UNC and their synthesis is published^{18, 54}.

Cloning

PknB 1–331 and full length GarA were amplified from Mycobacterium tuberculosis genomic DNA using Phusion polymerase (ThermoFisher) and standard cycling conditions. Amplified DNA was digested and ligated into PGEX-6P vector (GE) and transformed into Top10 competent cells for plasmid storage and BL21-Rosetta cells for protein production. Sequencing was done at the UW BioTechnology Center to confirm the constructs. PknB 1–331 A63I and V72M mutants were generated by following the Quikchange Mutagenesis protocol (Agilent) with minor changes. Primers were designed using Agilent Quikchange Primer Design (IDT) and used in PCRs using PrimeStar polymerase (NEB). PCR product was exposed to EtOH precipitation overnight in 100% EtOH at −20°C. The following day the DNA was collected at 16,060g and washed with 75% EtOH before being resuspended in water. The DNA was then digested with DPNI overnight at 16°C. The digest was directly transformed into Top10 cells and plated on LB agar containing 1mM ampicillin. Colonies were picked and inoculated overnight at 37°C, and DNA was extracted and sequenced. The plasmid with the correct sequence was transformed into BL21-Rosetta cells for protein expression. Codon optimized PknB 1–280 cDNA was ordered as a gBlock (IDT) and expression constructs were prepared as described above.

Protein Purification

Initial affinity purification of all PknB constructs was performed using the following procedure. LB broth containing 1mM ampicillin and 1mM chloramphenicol was inoculated with overnight cultures at 37°C. Cultures were induced with 0.5mM IPTG after reaching an OD₆₀₀ of 0.6–0.8. Protein was overexpressed overnight at 23°C. Cells were harvested at 6,079g and resuspended in lysis buffer [25mM Tris pH 8.0, 150mM NaCl, 1mM DTT, and 10mM MgCl₂] with 1mM PMSF and DNAse. The lysate was sonicated and the soluble fraction was collected at 30,600g and passed over columns with GS4FF resin (GE) at approximately 4 mL of resin per liter of culture. Columns were rinsed with lysis buffer, and protein was removed from the resin with elution buffer [50mM Tris pH 8.0, 5mM NaCl, 10mM MgCl₂, 3mM DTT, and 20mM reduced glutathione (GSH)] and digested with 1/20 w/w PreScission protease. GarA was produced using the same procedure except MgCl₂ was omitted from the buffers.

Digested PknB 1–331 WT and mutants were separated from GST and GSH using anion exchange (Source 15Q, GE) using buffer A [20 mM Tris pH 8.0, 1 mM DTT, 1mM MgCl2] and buffer B [buffer A + 1M NaCl] in a stepped gradient as follows: 0–20 % B for 0.5 column volumes (CV), 20–22.5% B for 5 CV, and 22.5–50% B for 5 CV. PknB was concentrated using 10kDa cutoff columns (Millipore) at 3150g and injected onto a Superdex 75 column (GE) for size exclusion purification using a buffer containing 20 mM Tris pH 8.0, 1 mM DTT, 1mM MgCl2, and 150mM NaCl. Protein was collected at 95% or greater purity and was flash frozen for storage at −80°C.

Digested GarA was partially separated from GST using anion exchange with a gradient of 0–19% B for 0.5 CV and 19–50% B for 5 CV. GarA was passed over GS4FF resin (~1mL resin per 6mg protein) three times to remove GST before concentration and size exclusion. Protein was collected at 95% or greater purity and was flash frozen for storage at −80°C. All buffers were the same as for PknB except MgCl₂ was omitted.

Digested PknB 1–280 for crystallization was further purified using the above procedure, except after anion exchange, the protein was passed over GS4 FF resin (~1mL resin per 6mg protein) to remove traces of GST before concentration and size exclusion. After size exclusion, PknB was concentrated to ~10mg/mL, a small sample was saved to measure activity, and GSK690693 was added to the protein at 3x the molar concentration (330μM) using the following procedure. Two mL of buffer from size exclusion was warmed to 37°C and 612μL of 1mM GSK690693 in DMSO was added slowly added to the warm buffer with frequent mixing. This mixture was added to 2.5mL of PknB and allowed to incubate at room temperature for 15min. After incubation, the sample was spun for 5min at 3150g to remove precipitate then concentrated to 9.8mg/mL. Kinase inhibition was checked using the KinaseGlo® assay as described in the “Kinase assays” section.

Kinase assays

The kinase assays were performed using the KinaseGlo® reagent from Promega. All reactions were done in 50μL volume. The buffer used for all kinase assays was 10mM Tris-HCl pH 7.4, 150mM NaCl, 1mM DTT, and 1mM MgCl₂. For kinetic analysis, plates were prepared with a serially diluted concentration gradient of GarA from 0–104μM across a serially diluted gradient of ATP from 0 to 400μM such that each concentration of GarA was present at each concentration of ATP. 0.25μM of PknB was added to each well and allowed to react at 37°C for 20min, then equilibrated to room temperature for 10min. After incubation, the reaction was stopped by the addition of 50μL of KinaseGlo® reagent, and the signal was allowed to stabilize for 10min at room temperature per the product manual. The plate was read using luminescence detection on a Synergy HT detector (BioTek) and the data were collected using the Gen5 2.0 software (BioTek). The data were processed in Excel (Microsoft) and non-linear regression models were fit in PRISM (GraphPad). The model for the velocity of the reaction plotted as a function of GarA concentration was fit using an allosteric sigmoidal model in preference over a Michaelis-Menten model (p <0.0001) with a shared Hill coefficient. The model for velocities as a function of ATP concentration was fit using the Michaelis-Menten model over the allosteric sigmoidal model (p <0.0001). Curves were fit to plotted apparent V_max and true V_max values were calculated based on the upper limit of each curve. K_cat was calculated based on Vmax using the equation V_max = K_cat * [E_t], where [E_t] is the total enzyme concentration (0.25μM). True K_m values for ATP and GarA were calculated similarly to true V_max values.

Inhibition kinetics were determined using similar parameters as for enzyme kinetics with the following changes. Drugs in 5mM DMSO were diluted in kinase buffer to 3/5 the final concentration using a serial dilution from 20 or 10μM to 0μM. The final DMSO concentration in the reactions was no more than 0.4%. PknB 1–331 was added to the drugs to a final concentration of 0.25μM and allowed to incubate for 10min at 37°C. ATP and GarA were added for a final concentration of 100μM and 42μM respectively, initializing the reaction. The reaction proceeded and was quantified as described for enzyme kinetics. The data were transformed to log scale and non-linear regression was performed in PRISM using the variable slope 4-parameter model for enzyme inhibition to determine IC₅₀. K_i values were determined using the following equation¹⁸:

$$K_i=\frac{{IC}_{50}-\frac{\left[E_t\right]}{2}}{1+\frac{\left[S\right]}{K_m}}$$ Equation 1

Where [E_t] is the total enzyme concentration (0.25μM), [S] is the total substrate (ATP) concentration (100μM), and K_m is the K_m for ATP (42μM).

Radiolabeled kinase assays were done as follows. The reaction buffer was the same as above and reactions were performed in 20μL volumes. PknB was used at a final concentration of 0.5μM, cold ATP was 20μM, ATPγ³²P was 0.1nM (10μCi), GarA was 10μM, and GSK690693 was serially diluted from 0–20μM. GSK690693 and PknB were incubated together for 10min at 37°C then substrates were added and allowed to react for 1hr at 37C. The reaction was stopped by the addition of 10μL SDS loading buffer and 20μL of the resulting mixture was loaded on a 15% gel for SDS-PAGE. The gel was fixed with a solution of 10% acetic acid, 40% methanol, and 5% glycerol for 1hr. The gel was dried for 1hr and autoradiography film (MidSci) was exposed to the gel for 2 minutes.

Crystallization and structure determination

PknB 1–280 for crystallization was produced as described above. Protein was sent to the Hauptman-Woodward institute for screening in 1536 different crystallization conditions in the microbatch-under-oil format at room temperature⁵⁵. Hits were visually assessed using the Macroscope J program (Hauptman-Woodward). The initial hit condition was 0.5 M bis-tris propane pH 6.7 and 15% w/v PEG 3350. This was optimized with a matrix of PEG and buffer concentrations and additives to grow larger, single crystals. Final diffracting crystals were produced as follows: 0.5μL of buffer A [25% PEG 3350, 0.5M bis-tris propane pH 7.5, 1% glycerol v/v] was combined with 0.5μL buffer B [0.2% thymidine w/v and 0.2% β-cyclodextran w/v] were combined with 1.0μL PknB (9.8 mg/mL in size exclusion buffer listed above). The drops were equilibrated at room temperature in the sitting drop format over 50μL of buffer A. Crystal growth was monitored over the course of two weeks and crystals which showed no visible growth in consecutive viewings were harvested and cryopreserved. Crystals were cryopreserved with the following procedure: crystals were resuspended in well buffer containing 0.3mM GSK690693 and slowly transferred to a cryobuffer containing 30% glycerol, well buffer, and 0.3mM GSK690693 in a 1:1 ratio. The solution was removed and the procedure was repeated five times until the final concentration of the solution surrounding the crystal was essentially the same as the cryobuffer. Crystals were mounted in loops (Molecular Dimensions) and quickly flash frozen in liquid nitrogen.

Diffraction data was collected at Sector 21, beamline D at the Advanced Photon Source (APS) at Argonne National Laboratory (ANL) and processed with HKL2000⁵⁶. Model building and refinement was done using the PHENIX software package⁵⁷. Data were assessed using Xtriage, and molecular replacement was done using Phaser-MR with 1O6Y as the search model. Ligand coordinates were generated with eLBOW, and the structure was refined with phenix.refine. The model was manually built and adjusted in Coot and further refined with phenix.refine^{57, 58}. Coordinates were deposited in the Protein Data Bank under accession code 5U94. Figures were created using Pymol, Coot, and Ligplot+^58–60.

Cell Viability Assays-IC50 screen

Mycobacterium smegmatis MC²155 WT (Msmeg) was streaked out onto lysogeny broth (LB) agarose plates from a frozen stock. After incubation at 37°C for 3 days, a colony was picked and inoculated into a 3mL culture of Middlebrook 7H9 media supplemented with 0.08% tyloxapol (v/v) at 37°C and grown overnight. Culture density was measured and then diluted to an OD₆₀₀ reading of 0.3. For IPA IC₅₀ calculations, IPA compounds were serially diluted as appropriate in a 96-well plate. If the addition of meropenem was being tested, 40μL of meropenem (at concentration of 3.125μg/mL) was added to the serially diluted 50μL of IPA compound. If no meropenem was being tested, 40μL of Middlebrook 7H9 media supplemented with 0.08% tyloxapol was added instead. 10μL of Msmeg at OD₆₀₀ of 0.3 was added to the previous mixture to give a final total volume of 100μL. The final concentration of meropenem (if added) was 1.25μg/mL and the final concentration of IPA ranged from 0–100μM. 100μL of Middlebrook 7H9 media with 0.08% tyloxapol (v/v) was used as a negative control. 10μL of Msmeg added to 90μL of Middlebrook 7H9 with 0.08% tyloxapol (v/v) was used as a positive control. The plate was incubated for 16 hours at 37°C before 10μL of resazurin (1mg/mL) was added to each well. After four hours of incubation at 37°C, resazurin fluorescence was read on a microplate reader using an emission wavelength of 530nm and an excitation wavelength of 590nm. The data obtained was then converted into % growth by comparing the control wells (bacteria alone) with the experimental wells. The data were graphed in PRISM (GraphPad) and nonlinear regression was used to calculate IC₅₀ values using the same methods as for enzyme inhibition. ANOVA was performed using PRISM (GraphPad) using the Tukey correction for multiple comparisons. Ineffective inhibitors were analyzed at concentration of 100μM, representing the minimum difference possible based on our assay conditions. Assaying the MIC shift of meropenem was performed by comparing a serial dilution of meropenem with two-fold dilutions of IPAs and percent growth was determined as described above. M. chelonae (clinical isolate from WI State Lab of Hygeine) was grown and tested following the same conditions as M. smegmatis.

The procedure for assaying IPA IC₅₀s was done with Nocardia asteroides (clinical isolate from UW-Madison Hsopital) (Nast) as with M. smegmatis with minor changes. Nast was grown from frozen stock into LB where it grew overnight. The following day, Nast was diluted to an OD600 of 0.3. Additionally, ceftriaxone was added in place of meropenem, with the final concentration of ceftriaxone being 0.25μg/mL. The 96-well plate was only incubated for 16 hours at 37°C before 10μL of resazurin (1mg/mL) was added to each well. The resazurin was incubated with Nast for 4 hours at 37°C before being read on the microplate reader.

Mycobacterium bovis, Bacillus Calmette-Guerin 1011 (BCG) inhibition assays were done as follows. BCG was streaked out onto Middlebrook 7H9 agarose plates from a frozen stock. After incubation at 37°C for one week, a BCG colony was picked and inoculated into a 1mL culture of Middlebrook 7H9 media supplemented with 0.08% tyloxapol (v/v) at 37°C. One week later, an additional 9mL of Middlebrook 7H9 media with 0.08% tyloxapol (v/v) was added to the growing culture at 37°C. One week following the adding of additional media, the BCG culture density was measured and then diluted to an OD600 reading of 0.3. IPA compounds were serially diluted as appropriate in a 96-well plate. If the addition of meropenem was being tested, 40μL of meropenem (at concentration of 2.5μg/mL) was added to the serially diluted 50μL of IPA compound. If no meropenem was being tested, 40μL of Middlebrook 7H9 media supplemented with 0.08% tyloxapol was added instead. 10μL of BCG at OD600 of 0.3 was added to the previous mixture to give a final total volume of 100μL. The final concentration of meropenem (if added) was 1μg/mL and the final concentration of IPAs ranged from 0–100μM. 100μL of Middlebrook 7H9 media with 0.08% tyloxapol (v/v) was used as a negative control. 10μL of BCG added to 90μL of Middlebrook 7H9 with 0.08% tyloxapol (v/v) was used as a positive control. The plate was incubated for 5 days at 37°C before 10μL of resazurin (1mg/mL) was added to each well. After an additional day of incubation at 37°C, resazurin fluorescence was read on a microplate reader using an emission wavelength of 530nm and an excitation wavelength of 590nm. The data obtained was then converted into % growth by comparing the control wells (bacteria alone) with the experimental wells.

Cell Viability Assays-MIC shift

Mycobacterium smegmatis (Msmeg) MIC determinations were done according to the microdilution method. In a 96-well format, 50 μL of two-fold serial diluted ß-lactam was added to 40 μL of GSK690693 at specified concentrations. The last column of the 96-well plate received medium without antibiotics. GSK690693 was tested in duplicate at 25 μM, 50 μM, and 100 μM and compared to a 2% (v/v) DMSO control. Kanamycin was used as a control, tested in duplicate, in combination with 100 μM GSK690693 and a 2% DMSO control. SB747651A at 100 μM, in duplicate for each antibiotic, was used as a control and compared to the DMSO controls of the corresponding antibiotic. The strain was streaked out onto Middlebrook 7H9 agarose plates supplemented with 10% ADC from a frozen stock. After incubation at 37°C for two to four days, a Msmeg colony was picked and inoculated into 2.5 mL of Middlebrook 7H9 broth supplemented with 10% ADC Middlebrook and 0.08% tyloxapol (v/v) and grown with shaking at 37°C. The next day, the culture was adjusted to an optical density of 0.3 (OD600), and 10 μL of the suspension was used to inoculate each well of microdilution plates made that day. The plates were incubated for 48 hours, stationary, at 37°C, and then 10 μL of 1mg/mL of resazurin dye was added. The plates were then incubated for an additional 12–15 hours at 37°C. Using a Synergy HT (Biotek) cell growth was quantified using fluorescence of the resazurin (excitation 530nm / read 590nm). Measurements were scaled to the high wells of DMSO control, no antibiotic and formatted as a percent relative to the high wells (100%).

Mycobacterium chelonae was tested using the same experimental setup as Msmeg but with the addition of 8 μg/mL Clavulanate to drug stocks and cultures were incubated at 35°C.

Mycobacterium abscessus PM3386 Δbla⁶¹ (Mab) MIC determinations were done according to the microdilution method. Antibiotics were tested in two-fold serial dilutions with GSK690693 at 50 μM or with a DMSO control in a 96-well format. Other control wells with cells received no antibiotics. The strain was streaked out onto Middlebrook 7H9/ADS agarose plates from a frozen stock. After incubation at 37°C for four days, an Mab colony was picked and inoculated into 10 mL of Middlebrook 7H9 ADS media supplemented with 0.5% tyloxapol (v/v) and grown with shaking at 37°C. The next day, the culture was adjusted to an optical density of 0.5 (OD₆₀₀), diluted 100-fold, and 100 μl of the suspension used to inoculate each well. The plates were incubated for 48 hours, stationary, at 37°C, and then 50 μL of 1:1 mixture of AlamarBlue:10%(v/v) Tween-80 was added to each well and the plate incubated for an additional 24 hours at 37°C. The MIC (μg/mL) was recorded as a range between wells that remain blue and those which were pink.

Mycobacterium tuberculosis H37Rv ΔRD1 ΔpanCD strain mc²6030⁶² (Mtb) was grown for three days by inoculating Middlebrook 7H9/ADS media supplemented with 0.5% glycerol, 0.05% Tween-80, 0.2% casaminoacids, and 24 μg/mL pantothenate with thawed culture to a final optical density (OD₆₀₀) of 0.024. On the third day the culture was adjusted to an OD₆₀₀ of 0.26 and 100 μl of the suspension used to inoculate each well. For IPA MIC calculations, IPA compounds were serially diluted as appropriate in a 96-well plate. For IPA MIC calculations in the presence of meropenem and clavulonate, 25 μL of meropenem (to a final concentration of 1μg/mL) and 25 μL of clavulonate (to a final concentration of 128μg/mL in the case of GSK1007102B or 64 μg/mL in the case of GSK690693) was added to the serially diluted 50μL of IPA compound. 100μL of Mtb at OD₆₀₀ of 0.26 was added to the previous mixture to give a final total volume of 200μL. The plates were incubated for 56 hours, stationary, at 37°C, and then 50 μL of 1:1 mixture of AlamarBlue and 10%(v/v) Tween-80 was added to each well and the plate incubated for an additional 16 hours at 37°C. The MIC (μg/mL) was recorded as a range between wells that remain blue and those which were pink. Due to a lack of availability of GSK1007102B, experiments were only repeated twice. Values from each culture are summarized in Figure S2.

Results

The imidazopyridine aminofurazans (IPAs) are predicted to bind PknB with high affinity

We wanted to test whether in silico modeling could be used to select lead compounds which would be active against PknB. We chose the highest resolution active conformation (“DFG in”) model available (1O6Y) and used Autodock 4 as our search program. We docked over 1000 type I (“DFG in”) kinase inhibitors and found that several were predicted to bind PknB with high affinity (Fig. 1A). The mean binding energy for all docked ligands was −8.87 kcal/mol with a standard deviation (σ) of 1.25 kcal/mol and a range of −13.01 to −4.29 kcal/mol. We focused on 22 lead compounds which docked in the 95^th percentile (2σ). Nine of the 22 compounds fell into three structurally related groups and the rest were dissimilar. Several of the dissimilar compounds are non-specific inhibitors (rapamycin) or are large hydrophobic compounds unlikely to enter cells (zotarolimus). Although Lipinski rules do not predict bacterial cell entry, we wanted to select compounds with a high probability of mammalian cell entry and acceptable pharmacokinetics. Two of these groups contained compounds predicted to have low solubility and were larger than 500 Da (Lipinski rule violators). This pre-filtering reduced our list to two groups: GSK2181306A, and one group containing three compounds: GSK690693, GSK1007102B, and GSK554170A. This second group is structurally similar and was selected for several reasons. The predicted binding mode suggested they made several protein-ligand interactions (hydrogen bonds, stacking interactions, salt bridges) with major interactions along the protein backbone or with conserved residues, making resistance development by single point mutations less likely. They have several hydrogen bond donors and acceptors and are not excessively hydrophobic, making them more likely to enter the cell and easier to formulate^{53, 54}. They also were predicted to bind in PknB’s back pocket, a binding mode not seen in other modeling studies^{41, 46, 47}. These compounds are members of the imidazopyridine aminofurazan (IPA) family which have a well-established SAR with certain human kinases. Previous work by our lab demonstrated limited SAR and microbiological activity from this family with an additional bacterial kinase, PrkA, from L. monocytogenes ³¹. We obtained limited quantities of certain compounds through our collaboration^{17, 18, 54}. Additionally, a member of this family, GSK690693, was used in phase I clinical trials with only moderate side-effects, suggesting that it may be a viable starting point for further development⁶³.

Figure 1. Computationally predicted binding energy, biological IC₅₀, and biochemical inhibition for the IPA family of compounds.

A) A scatter plot showing binding energy values for the in silico screen of kinase inhibitors in PknB. Virtual kinase inhibitor libraries were docked in PknB (1O6Y) using Autodock4. Each point represents one compound with the corresponding estimated binding energy plotted along the Y-axis. Points are separated along the X-axis in alphanumeric order. Lower binding energies indicate more favorable binding to the target. The solid black line indicates the mean energy for all three libraries, with the dashed line representing two standard deviations above the mean. The blue line represents the average binding energy of the IPA family (open symbols). Values presented are the lowest energy of 30 trials. Compounds with colored circles were those used in biochemical and microbiological studies, and colors correspond to those in subsequent figures. B) The main scaffold structure of the IPAs with relevant carbons and R groups numbered. C) Library IPA compounds and related IA compounds are listed along with R group structures and predicted binding energy (first section) biochemical IC₅₀ values and calculated K_i values (second section), and microbiologic data (third section). Biochemical IC₅₀ values were determined as described in the methods, and raw data are shown in Figure S2. K_i values were calculated using Equation 1. Msmeg, Nast, and BCG were grown in various concentrations of IPA compounds with and without a sub-MIC50 dose of meropenem (1.25 μg/mL; Msmeg and BCG) or ceftriaxone (0.25 μg/mL; Nast) and MIC values were determined as described in the methods. Fractional Inhibitors Concentrations (FICs) were calculated based on the ratio of the MIC of the combination treatment to the MIC of IPA alone. Where MIC values were not calculable, “NC” is indicated. One way ANOVA was performed comparing IPAs alone and in combination with a β-lactam, and statistical significance is given by p < 0.05 (*), p < 0.01 (**), p < 0.001 (***), and p < 0.0001 (****). Statistics were calculated using PRISM (GraphPad), and the Tukey correction was used for multiple comparisons. Raw inhibition curves used to calculate MICs and IC₅₀ values are provided in Figures S4–6.

The IPAs consist of an imidazopyridine core with an aminofurazan ring attached to the imido group (Fig. 1B). The core is modified at three different R groups on carbons 4 (R₁), 6 (R₂) or 7 (R₃) with the R₂ and R₃ positions of known IPAs being mutually exclusive: if one has a functional group attached, the other will be a hydrogen only^{18, 54}. The R₁ position was optimized for Akt as a dimethyl propargyl alcohol, whereas human cell entry was modulated with diverse R₂/R₃ groups¹⁸. The IPAs docked in PknB with a range of −11.7 to −9.5 kcal/mol, with more cyclic R₂/R₃ groups demonstrating higher affinity than those with linear groups (Fig. 1C), suggesting a potential SAR.

IPAs are potent biochemical inhibitors of PknB and require a functional group at R₁ for activity

To test whether our modeling predicted biochemical activity, we first characterized PknB kinetics using GarA, which is phosphorylated on T22 by PknB²⁵. Kinetic parameters for PknB were previously estimated at sub-saturating conditions using a construct lacking the juxtamembrane domain⁶⁴, however autophosphorylation of residues in this domain are required for full activation of the kinase³⁵. These residues are autophosphorylated during expression in E. coli, therefore observed in vitro kinase activity is on T22 of GarA only⁶⁵. Various groups have performed kinase assays on PknB using either magnesium⁶⁶, manganese⁴⁰, or both^{39, 67} as the chosen metal cation. The PASTA kinase from Staphylococcus aureus (Stk1) is known to prefer Mn as its cation⁶⁸, therefore we initially characterized our assay using Mg²⁺ and Mn²⁺ seperately to assess if PknB has similar preferences. We found no catalytic differences between Mg²⁺ and Mn²⁺ (Fig. S1), therefore we performed all subsequent assays with Mg²⁺. We characterized the enzyme kinetics under a matrix of conditions where both substrates were saturating in order to determine accurate kinetic parameters (Fig. 2A–F)⁶⁹. Although kinases vary widely in their enzymatic properties, our values (Fig. 2G) were comparable to several other eukaryotic S/T kinases^69–71. Interestingly, when the data were plotted with respect to GarA, the allosteric sigmoidal model was preferred (Fig. 2A). PknB dimerizes both in vitro and in vivo, making it plausible that allosteric cooperativity is occurring,^{35, 72} and the calculated Hill coefficient in our model was 1.9, suggesting that at least two subunits are interacting ⁶⁹.

Figure 2. Steady state kinetics for PknB.

Steady state kinetics for PknB were performed in an increasing matrix of both substrate concentrations until saturation was achieved. A) Velocities as a function of GarA concentration at various ATP concentrations up to saturation, fit with a sigmoidal model. B) Velocities as a function of ATP concentration at various GarA concentrations up to saturation, fit with the Michaelis Menten model. C) Effect of V_max^ATP with increasing GarA concentration fit with a sigmoidal model. D) Effect of V_max^GarA with increasing ATP concentration fit with the Michaelis Menten model. E) Effect of K_m^ATP with increasing GarA concentration fit with a sigmoidal model. F) Effect of K_m^GarA with increasing ATP concentration fit with the Michaelis Menten model. Experiments were done 2–3 times with 2–3 replicates each. Curves are representative of each experiment. Model fitting and statistics were done using PRISM (GraphPad). G) V_max, k_cat, and K_m were determined as described in the experimental methods. Values along with the SEM for each are given. Catalytic efficiency was calculated using mean values. V_max (and subsequently k_cat) were determined based on the limit of saturating data. K_m values are also reported as values at saturation.

Using drug concentrations from 0–10μM, K_i values ranged from 64nM to 530nM (Fig. 1C). K_i values for GSK1007102B, GSK554170A, GSK690693, and GSK614526A were not statistically different, but were significantly better than GSK943949A, GSK949675A, and GSK902056A. SB-747651A had no inhibitory activity at concentrations tested, suggesting the importance occupying the back pocket, which also agreed with the data in our previous work with PrkA ³¹. Inhibitors with cyclic R₂/R₃ groups varied statistically from linear R₂/R₃ groups (GSK902056A and GSK949675A) by a p <0.01 (Fig. 3A). Inhibition of substrate phosphorylation in a dose dependent manner was visually confirmed using a radiometric assay (Fig. 3B). Collectively, these results demonstrate that our in silico model accurately predicted inhibitors with nanomolar biochemical activity.

Figure 3. IPAs have significantly different PknB inhibition.

A) IC₅₀ values are shown for each compound along with the SEM. One way ANOVA was performed and statistically different groupings are denoted by ticks with arrows indicating the compound from which the group differs. Statistical significance is given by p < 0.05 (*), p < 0.01 (**), p < 0.001 (***), and p < 0.0001 (****). Statistics were calculated using PRISM (GraphPad), and the Tukey correction was used for multiple comparisons. IC₅₀ and K_i values are summarized in Figure 1C, and compound colors correspond to those in Figure 1. B) A qualitative radiometric assay to observe PknB phosphorylation by GSK690693 was performed using ³²P as described in the methods. PknB autophosphorylation was not seen (absent upper bands), but GarA phosphorylation was appreciable and scaled down with increasing inhibitor concentration (lower bands). The purified GarA contains a small amount of GST contamination (< 5%) which is also partially phosphorylated by PknB. C) Protein was normalized using the Bradford assay and visualized using SDS-PAGE for all biochemical reactions. 1μg of protein was loaded on the gel and visualized using Coomassie staining. Raw data inhibition curves used to generate panel A are provided in Figure S2.

GSK690693 co-crystallizes with PknB, binding in an ATP-competitive manner with several critical interactions

Several PknB structures are known, but prior to our work there were none which co-crystallized PknB with a specific kinase inhibitor bound to the active site ^35–39. Additionally, several PknB inhibitors were previously modeled in silico, but structural confirmation of these models was not established^{40–42, 46}. To confirm our model and facilitate future drug development insight, we co-crystallized the catalytic domain of PknB with GSK690693, a compound biochemically equivalent to our best drugs and commercially available in large quantities. PknB-GSK690693 crystals grew in 14 different screening conditions (0.9% hit rate). Upon refinement, well-formed crystals grew fully in 2–5 days. Crystals diffracted to 2.0Å and the model was refined at 2.2Å with an R/R_free of 19.89/23.26% and favorable geometric parameters (Table 1). The overall structure is a bi-lobed globular protein 59.2Å long and 39.5Å across (Fig. 4A). The backbone of the model was almost identical to other published structures at similar resolution (1O6Y (RMSD: 0.271 Å) and 2FUM (RMSD: 0.589 Å)) and similar to the Stk1 structure (4EQM (RMSD: 1.28 Å)) from Staphylococcus aureus ^{36, 38, 73}. The activation loop (residues 164–177) was not visible in the structure. The initial F_o-F_c map after molecular replacement indicated a well outlined area of unmodelled density in the ATP binding pocket of the active site, extending into the pocket behind the catalytic site (Fig. 4B). GSK690693 satisfied the unmodelled density and bound in the same orientation as it binds in Akt2 (3D0E), except for the R₂ piperidine group which extends into solution rather than interacting with the ATP binding floor ¹⁸. Overlaying our structure on the ATP mimic bound structure (1O6Y) revealed that GSK690693 overlapped the ATP binding position extensively, with the aminofurazan ring aligning with the adenine ring of ATP (Fig. 4C). The aminofurazan ring makes two hydrogen bonds to the backbone of the hinge region of PknB (E93 and V95), and the dimethylpropargyl alcohol makes two hydrogen bonds to the back pocket via the backbone of F157 and sidechain of E59 (Fig. 4D & E). The piperidine ring makes a weak hydrogen bond to the backbone of the P-loop (F19) through an ordered water (Fig. 4E). The catalytic lysine (K40) forms a hydrogen bond (2.9Å) with the nitrogen of the pyridine ring (Fig. 4D & E). M145 and M155 make stacking interactions with the imidazopyridine core, and M92 makes a stacking interaction with the aminofurazan ring (Fig. 4E). The ligand is also coordinated through 8 hydrophobic interactions (Fig. 4D).

Table 1. Crystallographic data collection, molecular replacement, refinement, and model statistics for the PknB–GSK690693 co-crystal.

Data were collected on one crystal. Values in parentheses are for the highest resolution shell. The coordinates and data files were deposited into the Protein Data Bank (PDB) under accession code 5U94.

PDB code	5U94
Data Set
Space group	C2221
Wavelength (Å)	1.1272
Resolution (Å)	50–2.05 (2.09–2.05)
Unique observations	20754 (748)
Redundancy	12.7 (7.2)
Rsymm (%)	8.7 (0)
Rpim (%)	3.4 (4.8)
Completeness (%)	96.5 (77.7)
I/sig(I)	23.6 (0.9)
Molecular Replacement
Search model	1O6Y
Refinement
Resolution (Å)	41.69–2.20 (2.60–2.20)
No. reflections (free)	17624 (1685)
Completeness (%)	99 (95)
R-work (%)	20.1
R-free (%)	23.7
Number of atoms (total)	4102
Protein	3807
Water	123
RMSD bond lengths (Å)	0.002
RMSD bond angles (°)	0.565
Average B-factors (Å2)	83
Wilson B (Å2)	51.3
Ramachandran plot
Preferred regions (%)	96.1
Allowed regions (%)	3.90
Outliers (%)	0.00

Figure 4. Structure of the PknB-GSK690693 co-crystal.

A) The overall kinase structure is nearly identical to previous published structures (1O6Y, 1MRU) with overall dimensions as indicated. Secondary structures and ligands are labeled. PknB is shown in N-C spectrum, GSK690693 as magenta sticks, and ATP overlaid from 1O6Y as grey sticks. B) The difference map in electron density after molecular replacement indicates the location of GSK690693 in the active site (green unmodeled density, 3σ). C) Close-up of the active site, outlined in light grey. GSK690693 (Inhibitor; magenta) is shown in the active site extending into the back pocket (BP) and overlapping the position of ATP from 1O6Y (grey). D) A ligand plot of GSK690693 interactions with PknB highlights several interactions. Hydrophobic residues are labeled in black with red hashes corresponding to red hashes around atoms in the ligand responsible for the interaction. Hydrogen bonding residues are shown in purple stick with green labels and distances indicated in angstroms. GSK690693 is shown in cyan stick. E) A 3D stereo image of the interactions between GSK690693 and PknB. Relaxed-eye stereo uses the left two images and cross-eye stereo uses the right two images. The P-loop should appear above the ligand. Hydrogen bonds are indicated with black dashes and the distance in angstroms is given. Residues participating in hydrogen bonding or stacking interactions are labeled. V72 and A63 are also labeled. Figures were rendered using Pymol (A, B, & E), Coot (C), and Ligplot+ (D).

The position of GSK690693 in the PknB co-crystal structure corresponds to the most favored position calculated by Autodock (Fig. 5A–C), with the imidazopyridine-aminofurazan core and the dimethylpropargyl alcohol in nearly identical positions in both the predicted and crystal structures. The model placed the R₂ piperidine group inside the P-loop, whereas in the crystal structure it adopts a more extended conformation; however, this area is likely highly dynamic in solution. Additionally, the alignment was similar to the crystal structure of GSK690693 in Akt2 (3D0E) (Fig. 5D). Overall, these data demonstrate GSK690693 binding in an ATP-competitive fashion, in agreement with our in silico modeling and previous structural biology for GSK690693 in Akt2.

Figure 5. Docking poses agree well with the crystallographic model.

A) The most energetically favorable position of GSK690693 (green sticks) in PknB (blue lines) as determined by Autodock. B) GSK690693 (cyan sticks) as modelled in PknB (blue lines) based on crystallographic data. C) The overlay of the docked pose and the actual position in the co-crystal shows generally good agreement. The active site cavity is outlined in transparent blue surface in all panels. The R2 group shows more variability in the docking models and is less well resolved in the crystal structure. D) GSK690693 crystallized in AKT2 (3D0E) in a similar position as in PknB. GSK690693 is shown as green sticks while AKT2 is shown in green lines with the ATP binding pocket outlined. Additionally, the back pocket of PknB is larger than that of AKT. M229 of Akt is positioned closer to the pocket than M92 of PknB, restricting the entrance to AKT’s back pocket and there is more space in the back of the pocket due to extra steric bulk from Akt L204 as compared to PknB A63.

Adding steric bulk in the PknB back pocket reduces inhibition by GSK690693

As most of the key hydrogen bond interactions in the back pocket are coordinated by backbone atoms or critical residues, and the numerous stacking and hydrophobic interaction residues individually contribute only a small fraction to binding, we focused on inhibiting entry to the back pocket to validate our model and observed SAR. We chose mutations V72M to restrict the back pocket entrance and A63I to sterically occupy the back pocket (Fig. 6A). The mutant enzymes were kinetically equivalent (with respect to apparent K_m^ATP at 100μM ATP) to the WT enzyme (Fig. 6B), allowing us to use biochemical inhibition as a method to compare GSK690693 binding. Although the V72M mutant was statistically equivalent to wild type enzyme with GSK690693 IC₅₀ values of 483nM and 341nM respectively, occupying the back pocket was significantly more disruptive, as the IC₅₀ for the A63I mutant was approximately 10-fold higher (IC₅₀ = 3679nM) (Fig. 6C). These results, combined with the ineffectiveness of SB747651A, support our hypothesis that binding the back pocket is critical for PknB inhibition by IPAs.

Figure 6. The back pocket mutant, A63I, is not efficiently inhibited by GSK690693.

A) Point mutants were chosen to both restrict entrance to (V72M) and collapse (A63I) the back pocket. Modelling these mutations on the structure showed the intended effects. The original structure back pocket (outlined in red mesh) predicted a smaller entrance (magenta surface) with a V72M substitution, and and a collapsed back pocket (blue surface) with the A63I substitution. The original residues are shown in black lines or sticks, and mutated residues are showin in respectively colored sticks. B) The A63I mutant PknB has a higher IC₅₀ than WT or a V72M mutant. The difference in IC₅₀ is statistically significant at p < 0.05 (indicated by *). C) The three mutants did not have statistically different K_m^ATP values, indicating the change in inhibition is not due to changes in ATP competition. D) Protein used for all biochemical assays was normalized using the Bradford assay and checked for accuracy using SDS-PAGE. 1μg of protein was loaded on the gel and visualized using Coomassie staining. Raw data inhibition curves used to generate panel B are provided in Figure S3.

IPAs inhibit bacterial growth and potentiate β-lactams against several bacteria

Effective biochemical inhibitors of PknB exist, but efficacy against whole mycobacteria has stagnated in the low micromolar range with biochemical optimization not correlating with microbiologic activity for unclear reasons^{40, 42}. Since pharmacologic inhibition of non-essential PASTA kinases can potentiate β-lactam activity against L. monocytogenes and Methicillin Resistant Staphylococcus in culture^48–50, we hypothesized that they may be able to act similarly in mycobacteria. We initially tested wild-type M. smegmatis (Msmeg) as a safe and comparable surrogate for mycobacterial drugs in early development stages⁷⁴. For our initial microbiologic experiments, we assayed variable doses of IPAs against a fixed sub-MIC50 dose of β-lactam and calculated IPA MIC values in order to assess SAR. We subsequently used an expanded matrix with the biochemically best IPA (GSK690693) to determine its ability to potentiate β-lactams.

IPAs alone did not inhibit Msmeg growth at 100μM or below, except for GSK1007102B which had an IC₅₀ of 99μM. Three out of eight IPA compounds were able to potentiate a sub-MIC50 dose of meropenem to inhibit Msmeg growth, with GSK1007102B being the most potent followed by GSK690693 and GSK943949 (Fig. 1C). Importantly, IPA potentiation was specific to β-lactams as sub-MIC50 doses of isoniazid did not potentiate their microbiologic activity (Fig. 7), consistent with previous reports that isoniazid and ethambutol did not improve the microbiologic activity of several structurally diverse, biochemically effective PknB inhibitors⁴⁰. Similarly, we tested the ability of GSK690693 to potentiate meropenem in Msmeg by determining the MIC shift of meropenem at various concentrations of inhibitor and found that the bacteria was up to 8 fold more susceptible to meropenem, reducing the MIC from 4 μg/mL to 0.5 μg/mL in an IPA dose dependent manner (Fig. 8A). We tested IPAs separately with both a β-lactam and kanamycin, a non-β-lactam third line treatment option for resistant mycobacterial infections⁷⁵. We found that IPAs did not potentiate kanamycin efficacy (Fig. 8A) and do work with any carbapenem tested (Fig. S8), further suggesting that this mechanism is β-lactam specific.

Figure 7. IPA compounds do not synergize with isoniazid to inhibit growth of M. smegmatis.

Msmeg was grown with four times less than the lethal concentration of isoniazid and various concentrations of A) GSK690693 and B) GSK1007102B. Isoniazid did not potentiate either of these IPAs. FIC values are calculated based on the ratio of the MIC of the combination treatment to the MIC of IPA alone indicated in the graphs. Where there is no MIC determined for both the single and combination treatment, the FIC is not calculatable (NC).

Figure 8. IPA compounds are potentiated by β-lactams in M. smegmatis, M. chelonae, and M. abscessus in culture.

A) Msmeg,B) Mche, and C) Mab were grown in various concentrations of IPA compounds with the indicated β-lactams and MIC values were determined as described in the methods and graphed as the maximum value of the range. Msmeg and Mche were wild type strains and Mab was a Δbla knockout. Clavulanate was used for Mche only at a concentration of 8 μg/mL for all conditions. One way ANOVA was performed and statistically different groupings are denoted by brackets indicating the compound from which the group differs. Symbols above bars indicate statistical difference from respective DMSO controls. Statistical significance is given by p < 0.05 (*), p < 0.01 (**), p < 0.001 (***), and p < 0.0001 (****). Statistics were calculated using PRISM (GraphPad), and the Tukey correction was used for multiple comparisons. Compound colors correspond to those in Figure 1. FIC values are given above the bars and were calculated based on the ratio of the MIC of the combination treatment to the MIC of β-lactam alone. All three species show significant potentiation by β-lactams relative to kinase inhibitor activity only. Raw data inhibition curves used to generate panels A & B are provided in Figure S7 and values called for panel C are given in Table S1.

The catalytic domains of PASTA kinases in the actinobacteria and firmicutes are highly conserved (Fig. S9). To determine whether this strategy is broadly applicable, we treated Nocardia asteroides (Nast) with our IPAs. GSK1007102B, GSK554170A, GSK943949A, and GSK690693 were microbiologically effective without a β-lactam with IC₅₀ values of 24–100 μM, while the others had no measurable activity (Fig. 1C). This might suggest that the nocardial PknB is essential, as the inhibitors would not work in the absence of a β-lactam if the protein was not-essential. Similar to Msmeg, these four inhibitors significantly potentiated a β-lactam 4–10 fold relative to β-lactam alone (Fig. 1C). Taken together, the data suggest that IPAs could be broadly applicable across closely related PASTA kinase containing pathogens and that potentiation of β-lactams offers a significant improvement over IPA administration alone.

Although PknB is well conserved among mycobacteria (Fig. S9), differences in cell wall composition, efflux, and metabolism may complicate inhibitor accumulation in different species. The Mtb vaccine strain derived from Mycobacterium bovis, Bacillus Calmette-Guerin (BCG) is highly similar to pathogenic Mtb; therefore, we tested our panel of IPAs against BCG in culture. We found that IPAs had better activity against BCG than against Msmeg, and even than Nast in some cases, and this effect did not require the presence of a β-lactam (Fig. 1C). Interestingly, the gap between biochemical inhibitors and microbiologic inhibition against BCG was much smaller, but the dimethylpropargyl alcohol was required, consistent with some amount of PknB activity being needed. Whether mechanisms of accumulation or activity of IPAs may be hitting other kinases between BCG and the other mycobacterium tested here is unclear but the potentiation is bidirectional.

NTMs vary in cell wall composition and have different clinical relevancies and treatments, therefore we wanted to confirm whether IPAs would be effective at potentiating a β-lactam against other mycobacterial pathogens. We assayed the effect of GSK690693 against M. chelonae (Mche) and initially did not see any potentiation. Mche appears to have higher intrinsic resistance to meropenem, therefore we hypothesized that adding clavulanate would protect meropenem long enough to see an effect. With 8 μg/mL clavulanate, we found that GSK690693 could potentiate meropenem two-fold and retained some dose dependence but leveled off quicker than in Msmeg. The higher intrinsic resistance combined with the need for clavulanate, may suggest that this is the rate limiting step of inhibition in Mche. We also tested GSK690693 against M. abscessus (Mab) β-lactamase mutant (negating the need for a β-lactamase inhibitor) and found it has no effect potentiating kanamycin; however, it did significantly potentiate the activity of ampicillin 8-fold (Fig. 8C). This agrees with the previous data that suggests this potentiation is specific to β-lactams and the 8-fold increase in effectiveness is similar to that seen in both Msmeg and Nast. We calculated the Fractional Inhibitory Concentration (FIC) Index and found that most combinations of IPA and beta lactams were in the range that defines synergy <0.5 as opposed to additive effects (0.5–1.0) and clearly better than the FIC index of an IPA with isoniazid (Table 1 and Fig. 7).

Finally, we were able to obtain an auxotrophic Mtb H37Rv strain, and we tested our remaining limited amount of compounds. GSK690693 was not active against Mtb either alone or in the presence of sub-MIC50 doses of meropenem (Fig. 9), but GSK1007102b was able to attenuate the growth of Mtb by itself and its activity was shifted upon the addition of sub-MIC50 doses of meropenem in the presence of clavulonic acid. These data suggest that certain IPAs have activity against Mtb and a tripartite therapy may lower the needed dose of any one compound.

Figure 9. GSK1007102B has activity against auxotrophic M. tuberculosis and is potentiated with meropenem.

Auxotrophic Mtb was grown in the presence of a sub-MIC50 dose of meropenem with clavulonic acid and either GSK690693 or GSK1007102B. GSK690693 was not lethal to Mtb with or without meropenem, but GSK1007102B was lethal on its own and was potentiated with meropenem. One way ANOVA was performed between GSK1007102B treatments. Statistical significance is given by p < 0.05 (*). Statistics were calculated using PRISM (GraphPad). Compound colors correspond to those in Figure 1. The FIC value for GSK1007102B with and without meropenem and clavulonate is given above the bar and was calculated based on the ratio of the MIC of the combination treatment to the MIC of GSK1007102B alone. Conservative upper limits were used for the MICs. MICS for each culture are given in Table S2.

Collectively our microbiological data point to the IPAs as authentic mycobacterial and nocardial inhibitors with the ability to specifically potentiate β-lactam antibiotics.

Discussion

Our data demonstrate that for well-defined targets, such as Hank’s-like kinases, in silico screening offers a cost-effective way to prioritize compounds for pharmacologic testing. We identified a family of inhibitors, the IPAs, that are biochemically active and bind the target, PknB, as our model predicted. Traditionally β-lactam antibiotics are almost never used for mycobacterial infections due to at least perceived inadequate potency. Here we show our inhibitors are microbiologically active and when coadministered with β-lactams, both types of drugs work at lower concentrations.

The computational model was highly predictive of the actual crystal structure (Fig. 5). Having a high resolution structure and well-characterized kinase likely helped increase this accuracy, as did the relative uniqueness of IPA binding mode. The core scaffold, aminopyrazine ring, and dimethylpropargyl alcohol were in almost identical positions with similar hydrogen bonding patterns. The dimethylpropargyl alcohol at R₁ was necessary for both biochemical and microbiological activity (Fig. 1C) suggesting that R₁ position hydrogen-bonding in the back pocket is necessary for activity. Our structure shows two critical hydrogen bonds: to the backbone and to E59, highlighting this importance. Although similar to the binding of GSK690693 in Akt2, multiple differences could be exploited to increase Akt-vs-PknB selectivity. The entrance to the back pocket is wider for PknB, and deeper on the E59 side than Akt (Fig. 5D). Previous SAR with Akt shows that wider R₁ groups (such as a phenyl) substantially inhibit an IPA’s activity¹⁸, suggesting that careful tuning at this position may be useful for increasing selectivity against Akt. Although most interactions between GSK690693 and PknB involve backbone hydrogen bonds or conserved residues (Fig. 4C), our data suggest that point mutations that collapse the back pocket may be a possible resistance mechanism (Fig. 6). The back pocket is well conserved (Fig. S9), but we cannot rule out the development of resistance by mutations in this area. Although the computational model was less predictive of the R₂ position, the basic orientation was correct, and based on the crystal structure, opportunities to improve the biochemical potency and selectivity exist. Interestingly, after deposition of our structure, a new PknB structure with a semi-selective inhibitor was made available (6B2P)³⁹. This inhibitor did not bind the back pocket but instead adopted an extended conformation along the hinge region and occupied similar hydrophobic interactions as GSK690693, although the molecular overlap was mostly confined to the aminofurazan ring along the hinge region. The structure varied more significantly in the DFG region than did ours or 1O6Y; however, the back pocket architecture was still mostly preserved suggesting that targeting these areas or even hybridizing these compounds could yield to even more selectivity and potency.

We hypothesized that selecting small and hydrophilic inhibitors (Lipinski’s rules followers) would facilitate both entry into bacterial cells and future drug development efforts⁵³. Although effective antibiotics vary greatly in these parameters, most bacterial protein targeting drugs obey these guidelines⁷⁶. Furthermore, the class of inhibitors we have identified, while having some human kinase activity, are particularly drug-like. GSK690693 has already been used in clinical trials while, to our knowledge, other known PknB inhibitors have unknown effects on human biology. Our IPAs substantially inhibit growth of a non-pathogenic mycobacteria (Msmeg), four pathogenic mycobacteria (Mtb, Mche, Mab, and Mm), and a pathogenic non-mycobacterial actinobacteria (Nast) at micromolar levels when potentiated by a β-lactam, and they can inhibit BCG sufficiently on their own as well (Fig. 1C). While microbiologic activity of the IPA obviously requires biochemical activity, certain IPAs such as GSK949675A and GSK614526, are only active against certain bacteria, and likely factors other than enzymatic inhibition influence which drugs penetrate the different bacterial cell walls. While other PknB inhibitors have been optimized with medicinal chemistry to achieve microbiological efficacy comparable to our IPAs^40–42, our inhibitors have not, suggesting that sub-micromolar level inhibition may be attainable. Combination therapy is traditionally used for mycobacterial infections, and the limited microbiologic activity of previously optimized PknB inhibitors in combination with traditional first line drugs such as isoniazid and ethambutol did not show any potentiation of activity⁴⁰. In contrast, our data suggest that β-lactams improve of PknB inhibitor utility (Figs.1C, 8 & 9). Although the mechanism for this potentiation is unknown, it is not likely that β-lactams are altering general cellular access, as IPAs which cannot enter the cell but still have comparable biochemical activity, such as GSK614526, fail to potentiate β-lactams (Fig. 1C). In agreement with previous results, isoniazid also failed to potentiate IPAs (Fig. 7), as does kanamycin, further suggesting this pathway is specific for β-lactam potentiation. Additionally, although Msmeg and Nast did not require a β-lactamase inhibitor to see this effect, despite having a wild type β-lactamase profile (Bla+), Mtb and Mche did, and our Mab strain was a β-lactamase knockout; therefore, a tripartite therapy with a β-lactamase inhibitor may be required to protect the β-lactam in organisms with a different β-lactamase profile. Interestingly, IPAs were effective on BCG in isolation and potentiation was not observed (Fig. 1C). This could be due to an increased susceptibility to IPAs or this potentiation pathway being less rate-limiting in BCG. We also found that GSK690693 does not potentiate β-lactams in M. marinum (Fig. S7C); however, similar to BCG, GSK690693 is effective on its own at reducing mycobacterial growth (Fig. S7D). Additionally, the β-lactam target which cooperates in this pathway is unknown, and penicillin binding proteins vary across mycobacterial species⁷⁷. It is possible that different species have different β-lactam binding preferences, leading to an additional interaction to be optimized.

While mycobacteria have up to 10 other serine/threonine kinases (such as PknA and PknG) that potentially are also inhibited by IPAs, Nocardia only have three other kinases, suggestive of growth inhibition being inhibited at least in part by PknB. Notably PknA is involved in the cell wall stress pathway of mycobacteria⁷⁸, but a BLAST analysis using the NCBI and UniProt databases did not reveal any close homolog in Nocardia. Additionally, IPAs definitively bind PknB (Fig. 4) with high affinity (Fig. 1C), and our SAR and mutagenesis suggests the importance of back pocket binding. The available structures of Mtb PknA (6B2Q), PknE (2H34), PknI (5XKA), and PknG (2PZI) all show a reduced or absent back pocket^{39, 79–81}. Docking GSK690693 into these structures gives no consistent binding pose and estimated maximum binding energies ranging from −7.70 to −9.05 kcal/mol, which are well below the most active IPAs and below the known inactive SB-747651, suggesting some preference for PknB over the structurally known Mtb kinases. Still it is possible that IPAs may act through a combination of PknB + some other targets. PknB is the only mycobacterial serine/threonine kinase to have PASTA domains⁸², and these domains are predicted to bind β-lactams, suggesting a route for β-lactam potentiation. Furthermore, as the other mycobacterial kinases are also involved in growth and virulence⁷⁸, off-target inhibition of these structurally similar kinases, if present, should only serve to enhance its drug potential. Therefore several lines of evidence from this work (crystallographic, biochemical, microbiologic (presence of PknB in various actinomycetes) and our past work (IPAs have no activity against Lmo with the PASTA kinase deleted) suggest PknB is at least the primary target of this class of compounds in vivo.

Taken together, our data suggest that computational modelling accurately predicted the binding modes and biochemical activity of the IPA family of drugs. These un-optimized PknB inhibitors have mycobacterial activity similar to other discovered inhibitors, but less of a discrepancy between biochemical potency and microbiologic potency than known inhibitors, potentially suggesting better bacterial access/retention. Importantly, we found that combining a β-lactam with hits which are not yet optimized to be PknB inhibitors can create microbiological inhibition near levels seen with other biochemically optimized inhibitors, opening the field for further improvements. This approach allowed us to leverage a large body of industrial and academic work in a cost-efficient and timely manner to repurpose a failed drug as a new lead for developing new mycobacterial antibiotics.

Non-standard Abbreviations

IPA

imidazopyridine aminofurazan

NTM

non-tubercular mycobacteria

Msmeg

Mycobacterium smegmatis

Nast

Nocardia asteroides

BCG

Mycobacterium bovis, Bacillus Calmette-Guerin

Mab

Mycobacterium absessus

Mmar

Mycobacterium marinum

Mtb

Mycobacterium tuberculosis