Rational prioritization strategy allows the design of macrolide derivatives that overcome antibiotic resistance – pnas.org

Results

All computational results are derived from a consistent set of relative free energy perturbation (FEP) simulations. The ligands were simulated both in the unbound state in aqueous solution and in the bound state within the bacterial ribosome. The bound states of clarithromycin and the acylide derivatives were modeled with the structure of the 50S ribosome of Deinococcus radiodurans together with explicit water and with an ionic concentration of 0.1 M. The clarithromycin–ribosome complex was equilibrated with molecular dynamics simulations (SI Appendix, section A and Fig. S1). In parallel, relative FEP protocols for all ligands were set up and tested in aqueous solution. The simulations decouple the side chains of the acylides as shown in Fig. 1, leaving only the common backbone as the alchemical intermediate state. This also allows the use of enhanced sampling methods for exploring the conformational flexibility of the side chains (19).

First, we estimated the relative affinities for water, in the absence of the ribosome, using implicit solvent calculations (stage IIa) (SI Appendix, section C) (20). These calculations represent an upper bound for their relative solubilities and neglect self-solvation effects (21), as well as the free energy of sublimation (22). Stage IIa was motivated by the relatively poor solubility of clarithromycin and the assumption that hydrophilic molecules might be less affected by promiscuous adenosine triphosphate (ATP)–dependent efflux pumps (17, 23). Since all molecules were predicted to be more hydrophilic than clarithromycin (SI Appendix, Table S1), we used the five compounds with the highest affinity for water (6, 3, 8, 5, 4) for the next stage (stage III, where we calculated the binding free energy of the ligands toward the ribosome).

Likewise, we estimated membrane permeabilities using an implicit membrane model (stage IIb) (SI Appendix, section C) (24). The resulting transfer free energies gauge membrane permeability based on Fick’s First Law (25) but neglect self-solvation effects (26), as well as the nanostructure of the membrane. Since all candidates were considered to be membrane permeable (SI Appendix, Table S2), the five molecules with the highest affinity for the cell membrane (7, 2, 9, 18, 14) were also prioritized for the binding free energy calculations in stage III. Stage IIb was motivated by the assumption that membrane-permeable molecules are more likely to enter bacterial cells (17).

Accordingly, we selected molecules 2 to 9, 14, and 18 for FEP binding free calculations (stage III). The final frame from the molecular dynamics simulations of 1 bound to the ribosome served as the starting point for relative FEP calculations of the bound state (SI Appendix, section D). To reduce computational costs, the ribosome was truncated to a subsystem comprising 0.1 million atoms, and all atoms outside of the binding pocket were kept frozen. Thus, any slow structural transitions of the ribosome are neglected.

The equilibration phase of the FEP binding simulations was employed to filter out nonbinders by postprocessing the trajectories of compounds 2 to 9, 14, and 18 with molecular mechanics energies combined with generalized Born and surface area (MM-GBSA) continuum solvation calculations based on one trajectory (stage IIIa) (SI Appendix, section D) (27, 28). These types of calculations only provide very approximate estimations of the free energy changes since the structural relaxation of the unbound state is not accounted for and due to the implicit treatment of the solvent. In addition, any slow relaxation of the binding pocket is neglected due to the short simulation length. Nevertheless, this step allowed us to eliminate compounds 14 and 18 because they exhibited highly unfavorable relative binding free energies (SI Appendix, Table S3).

Next, we calculated the relative free energies of binding (ΔΔGbind) and relative free energies of hydration (ΔΔGhydr) of the remaining prioritized set of compounds (2 to 9) with respect to clarithromycin using a rigorous free energy estimator and explicit solvent (stage IIIb) (Table 1 and SI Appendix, section B). Based on similar FEP protocols in the blind Statistical Assessment of the Modeling of Proteins and Ligands (SAMPL) prediction challenges (29, 30), the expected error of these calculations is between 2 and 3 kcal mol^{– 1}. Thus, molecule 6 is a significantly better binder than clarithromycin, while molecules 3, 4, 5, and possibly 9 are likely to exhibit improved binding affinities toward the ribosome. The explicit solvent ΔΔGhydr values confirm that all tested compounds are more hydrophilic than clarithromycin, especially 6.

In stage IIIc, we carried out quantum mechanics/molecular mechanics (QM/MM) calculations to further explore the relative stability of the ribosome–ligand complexes in explicit solvent (SI Appendix, section E). The relative average interaction energies of the ligands with their surroundings were determined in the ribosome (ΔΔEribo) and in aqueous solution (ΔΔEaq). The use of QM/MM approaches previously improved the accuracy of computational free energy predictions by about 1 kcal mol^{– 1} (30, 31). The QM/MM calculations indicate that compounds 6, 5, and 4 exhibit the most favorable interaction energies with the ribosome. They were, therefore, prioritized for the in vitro testing (stage IIId). Molecules 7 and 8 were eliminated because of the predicted instability of their complexes with the ribosome (Table 1). The results for the other molecules are less clear, as stronger interactions with the ribosome are often offset by a simultaneous high affinity for the unbound state in water. Overall, the QM/MM interaction energies confirm the FEP results, indicating that 6 is a very favorable binder to the ribosome. In addition, ligands 5 and 4 emerge as potential effective inhibitors. A detailed discussion of the simulations of the ligand–ribosome complexes is provided in SI Appendix, section E.

To evaluate our rational design strategy, the inhibitory activities of all 19 candidates were determined by the half maximal inhibitory concentrations (IC50) from an in vitro transcription–translation assay against the Escherichia coli ribosome (stage IIId) (SI Appendix, part 2). As evidenced in Fig. 2, compounds 6 and 4 are indeed the most active in the whole set, and 5 is more active than clarithromycin. Molecule 7, which was predicted to be the worst binder in the prioritized set, also turns out the be the least active in this subset. Ligands 11, 13, 15, and 16 also show high affinity toward the ribosome but were not part of the prioritized set because of their solvation properties. Membrane permeability and hydrophobicity are expected to play a prominent role for biological activity.

Fig. 2.

IC50 data for the inhibition of the E. coli ribosome by telithromycin (TEL), clarithromycin (CLY), the precursor molecule (21), and all 19 candidates. The three best candidates from the computational preselection are highlighted in red.

In stage IV, we tested the acylides for in vitro antimicrobial activity by assessing minimal inhibitory concentrations (MICs) against gram-positive bacteria (Bacillus subtilis and Staphylococcus aureus) as well as against gram-negative bacteria (E. coli, Acinetobacter baumanii, Pseudomonas aeruginosa, and Klebsiellapneumoniae) and yeast strains (Candida albicans and Cryptococcus neoformans). The MIC results are listed in Fig. 3. We also tested the cytotoxicity and hemolytic activity by assessing cytotoxic concentration (CC50) against the human kidney cell line HEK293 and hemolytic concentration (HC10) against human red blood cells (stage V) (the last two rows of Fig. 3). For MIC values up to 32 μg mL–1, therapeutic indices are calculated (CC50/MIC) and shown in parentheses below the respective MIC value in Fig. 3. As the ester bond in acylides might be prone to enzymatic breakdown, all acylides were tested to be stable in human blood serum (32).

Fig. 3.

Activity against a variety of gram-negative strains and yeast, the gram-positive bacteria B. subtilis, and the high-priority pathogen S. aureus (32). MICs are given for erythromycin (ERY), clarithromycin (CTY), telithromycin (TEL), and all candidates. The three best candidates from the computational stage (4 to 6) are highlighted in red. Additionally, CC50 (Homo sapiens HEK293) and HC10 (H. sapiens red blood cells) values are provided. Selectivity indices (CC50 value divided by MIC) are shown in parentheses below the respective MIC value. The selectivity indices are color coded together with the respective MIC value to facilitate readability. Not determined values are indicated by a dash.

All compounds are highly active against the methicillin-susceptible strain of S. aureus (MSSA, DSM 20231, ATCC 12600) (33) with MIC values up to 0.1  μg mL–1 while displaying no cytotoxicity or hemolytic activity at the same concentrations, as well as no activity against either of the yeast strains. Exceptions are only compounds 19 and 20, the two benzoic acid variants, which have no antibacterial activity. The high antibiotic activity of the other compounds is retained against B. subtilis (DSM 402, NCIB 10106) and mostly against two mutant strains of MSSA: S. aureus BAA-976 and BA-977. S. aureus BAA 977 features inducible macrolide–lincosamide–streptogramin B resistance due to the erythromycin ribosomal methylase erm(C) gene that alters the ribosomal structure (34), while S. aureus BAA 976 is macrolide resistant due to the ATP-dependent efflux pump associated with the msr(A) gene. Both strains display resistance against the reference antibiotics erythromycin, clarithromycin, and telithromycin. Most compounds retain similar activity in the resistant strains compared with the MSSA strains. The only exceptions are compound 5, which displays a 10-fold reduction of activity (MIC value shift from 0.1 to 1  μg mL–1); compound 6 with a 5-fold reduction (MIC value shift from 0.1 to 0.5 μg mL–1); and compound 8 with a 4-fold reduction (MIC value shift from 4 to 16 μg mL–1). Compounds 3, 6, and 16 still retain a high antibiotic activity with MIC values <1  μg mL–1 in the resistant strains.

None of the compounds show high activity against the MRSA strain S. aureus ATCC 43300, which harbors, among other resistant genes, the erm(A) gene. The only notable exception is compound 7, for which an MIC value of 32 µg mL^{– 1} was measured. The same compound showed only moderate activity compared to the other strains tested. This may be related to structural rearrangements in the binding pocket, caused by the bulky side chain of 7, that mitigate the effect of RNA methylation on the binding affinity.

Next, antibacterial MIC data were determined against several gram-negative pathogens, including type strains from A. baumanii (DSM 30007, ATCC 19606), P. aeruginosa (DSM 50071, and ATCC 27853), E. coli (DSM 30083, ATCC 11775, and ATCC 25922), and K. pneumoniae (ATCC 700603) (Fig. 3 and Table 2). Those four species are the top-ranked bacteria from the World Health Organization (WHO) priority pathogens list for the development of new antibiotics (33). Similar to established macrolide antibiotics, most compounds display no activity against the panel of gram-negative bacteria. A notable exception is once more compound 6, showing activity against A. baumannii (MIC between 8 and 32 μg mL–1). To a lesser degree, compounds 11 (MIC of 64  μg mL–1) and 16 (MIC between 16 and 32  μg mL–1) are also active.

In addition, the compounds were tested against gram-negative strains that harbor mutations, which affect the drug permeability through the bacterial membrane, including E. coli Δ lpxC (variation of the lipopolysaccharide structure) and Δ tolC (without the major efflux pump) and P. aeruginosa Δ mexAB/CD/EF/JKL/XY (without five of the major efflux pumps). Interestingly, all compounds regain significant activity (up to MIC values of 2  μg mL–1) if the major drug efflux pump is eliminated in E. coli. This also includes compound 20, which is otherwise not even active against gram-positive bacteria. Variation of the outer membrane by Δ lpxC had a lesser effect, with only compounds 3, 5, 6, 11, 12, and 16 regaining activity to MIC values of 16 to 32  μg mL–1. A similarly small effect was displayed for the P. aeruginosa strain that lacks five of the major drug efflux pumps. Only five compounds regain activity to MIC values of 32 μg mL–1 (Fig. 3).

A similar outcome can be achieved by the coadministration of sub-MIC levels of colistin, an antibiotic that interacts with the outer membrane component Lipid A, thereby increasing the permeability of the membrane for other drugs. Colistin showed synergistic activity with each of the tested acylides 6, 11, and 16 against E. coli and A. baumanii, increasing the activity to MIC values of 0.25 to 1  μg mL–1, at colistin concentrations of 10% of its MIC value (Table 2). Like the efflux pump mutants, activity with colistin against P. aeruginosa could not be reestablished to the same degree as for E. coli or A. baumanii, suggesting that P. aeruginosa still has effective efflux mechanisms. The activity against the efflux-deficient E. coli Δ tolC mutant suggests that the main cause of the resistance in gram-negative bacteria is the efficient elimination of the compounds via efflux pumps. In addition, the activities against E. coli Δ tolC agree with the IC50 values, with most MIC values around 2 μg mL–1 and IC50 values <0.5  μM. Exceptions are the weaker binders 7, 17, and 20, which also display IC50 values above 1  μM and MIC values at 16 to 32  μg mL–1. This suggests that the activity against S. aureus, which exhibits a wider range of MIC values, might be governed by permeability as well.