Systematic discovery of drug interaction mechanisms

Introduction

Drugs play a crucial role in elucidating fundamental relationships in cell physiology (Falconer et al, 2011). When drugs are combined, interactions like synergism and antagonism can occur (Loewe, 1928; Keith et al, 2005) (Fig 1A); such drug interactions are often critical for the success of multidrug treatments (Pillai et al, 2005) and can slow or accelerate antibiotic resistance evolution (Chait et al, 2007; Hegreness et al, 2008). A case in point is the synergistic combination of trimethoprim and sulfa drugs which has been applied successfully for decades (Pillai et al, 2005) even though the mechanism of synergism has long remained elusive (Nichols et al, 2011). Synergism and antagonism occur frequently between antimicrobials and are largely determined by the primary cellular target of the drugs that are combined (Yeh et al, 2006; Ocampo et al, 2014). However, synergistic drug interactions are rarely explained by the genetic interactions between the corresponding drug target genes (Cokol et al, 2011). To design combinations exploiting the full potential of existing drugs, a deeper understanding of the underlying mechanisms of drug interactions is urgently needed. Due to the vast number of possible drug combinations, general principles that are valid across diverse drug pairs could greatly facilitate the identification of drug interaction mechanisms. Recent work in this direction revealed scaling laws describing the effects of resistance mutations on drug interactions (Chait et al, 2007; Wood et al, 2014); further, the effects of three or more drugs appear largely predictable from the two-drug effects (Wood et al, 2012). However, the underlying mechanisms of most two-drug interactions remain unknown.

Drug interactions could be caused by physicochemical effects, for example when one drug simply enhances the permeability of the cell envelope for another (Jawetz & Gunnison, 1953); alternatively, they may have more complex causes, specifically if one drug triggers a regulatory response, which affects the action of another. While many genes affect the cell's sensitivity to individual drugs (Hillenmeyer et al, 2008; Nichols et al, 2011) and recent work suggested that certain drug resistance mutations may affect drug interactions (Munck et al, 2014; Wood et al, 2014; Rodriguez de Evgrafov et al, 2015), it is unclear to what extent genetic perturbations can alter drug interactions. Likewise, the cellular functions that control these interactions are largely unknown. To pinpoint their underlying causes, we developed a systematic approach for identifying genes that reshape drug interactions: using precise growth rate measurements of a genome-wide set of E. coli gene deletion mutants (Baba et al, 2006), we show that drug interactions are robust to most, but not all, genetic changes. We present a general principle encapsulating this robustness, which enables the quantitative prediction of mutant growth under drug combinations. Rare mutants violating this principle expose cellular functions governing each drug interaction. We establish that diverse drug interactions are recurrently controlled by central cellular functions, in particular polysaccharide synthesis and ATP synthesis.

Results and Discussion

We began by quantifying the growth rates of ~4,000 nonessential E. coli gene deletion mutants (Baba et al, 2006) under six different antibiotic pairs and their constituent individual drugs (Materials and Methods). We selected antibiotics with diverse modes of action (Table 1) and drug pairs covering all interaction types (Fig 1A and B). Drug concentrations were adjusted to inhibit wild-type growth under individual drugs by ~30%; the same concentrations were used when drugs were combined, leading to different levels of growth inhibition (Materials and Methods). In total, we measured over 50,000 growth curves (optical density increase over time) using a dedicated robotic system (Fig 1C–F; Materials and Methods). These growth rate measurements were highly reproducible (Supplementary Fig S1), and consistent with established changes in mutant sensitivity to individual antibiotics (Tamae et al, 2008; Liu et al, 2010; Nichols et al, 2011): for example, DNA repair mutants were sensitive to ciprofloxacin. Drugs with related mode of action had similar effects on growth of genome-wide mutants (e.g., Pearson correlation ρ = 0.68 for chloramphenicol and tetracycline), and the effects of drug combinations were usually most similar to those of the constituent drugs (Supplementary Fig S2; Materials and Methods). These observations confirmed our expectation that perturbations of cell physiology caused by drug combinations are mostly an overlay of effects caused by the constituent drugs.

We next identified a general principle characterizing drug interactions in mutants. We hypothesized that the shape of the two-drug growth response surface g(a,b), which defines the interaction (Fig 1A), does not change qualitatively in most mutants (here, g denotes growth rate and a, b the drug concentrations). To test this hypothesis, we measured 108 mutant response surfaces in two-dimensional drug concentration matrices covering different drug pairs; this selection included mutants with strongly altered sensitivity to the individual drugs (Materials and Methods). We found that the vast majority of mutant response surfaces g^mut(a,b) were well approximated by a linearly rescaled wild-type surface: g^mut(a,b) =γg^WT(αa,βb) with scaling factors for maximum growth rate γ and for drug concentrations α, β (Fig 2A and B; Supplementary Figs S3 and S4). The sensitivity to one or both of the drugs often changed considerably, yet the response surface shape was generally preserved (Fig 2A); these observations held for all drug pairs and for mutants affecting diverse cellular functions (Fig 2B; Supplementary Fig S4), suggesting that most genetic perturbations do not affect drug interactions.

To test whether this conservation of drug interactions holds generally, we devised a strategy for predicting genome-wide mutant growth responses to drug combinations. For all ~4,000 mutants, we calculated the expected response to the drug combination by first rescaling the wild-type response surface according to the individual drug responses measured at fixed concentrations. At the rescaled drug concentrations, we then used the interaction coefficient of the wild-type, which quantifies the response to the drug combination relative to the Bliss additive expectation (Yeh et al, 2006), to calculate each mutant's expected growth rate under the drug combination (Fig 2C; Materials and Methods). The central assumption of this procedure is that the drug interaction is universally invariant, that is it is the same in mutants as in the wild-type upon rescaling of the drug concentrations. For all drug pairs and the vast majority of mutants, the growth rates measured at fixed concentration of the drug combination (Fig 1C–F) faithfully followed this prediction (Fig 2D; Supplementary Fig S5). These observations thus revealed a general principle of bacterial growth under drug combinations, which encapsulates the high robustness of drug interactions to genetic perturbations and enables the quantitative prediction of mutant growth rates under drug combinations.

The identification of this general principle empowered us to pinpoint ‘outlier’ mutants with unexpected growth response to drug combinations. These outliers are of key interest as they could have altered drug interaction; together with functional information on the mutated gene, they can thus point at the underlying drug interaction mechanism. Clear outliers for which the observed growth rate under the drug combination (Fig 1F) deviated significantly from the expected growth rate were rare (typically < 1% of mutants; Fig 2D; Supplementary Fig S5), facilitating this investigation. We measured the response surface of the strongest outliers for each drug combination in fine resolution 12 × 8 concentration matrices (Materials and Methods). For each drug pair, we thus identified several mutants with clearly reshaped drug interaction (Fig 3, Supplementary Fig S6). Drug interactions were often weakened or removed in these mutants; they were also amplified in certain mutants and, in some cases, entirely new ‘synthetic’ drug interactions appeared (Fig 3C). In a thiamin synthesis hypomorph, chloramphenicol–trimethoprim even became reciprocally suppressive; that is, addition of chloramphenicol on top of trimethoprim increased growth and vice versa (Supplementary Fig S7). We further observed clear biases in interaction changes: chloramphenicol–nitrofurantoin suppression was weakened or entirely removed in most mutants affecting this interaction (Fig 3A); in contrast, chloramphenicol–trimethoprim antagonism was often amplified to suppression (Fig 3B and C; Supplementary Fig S6L–Q), while other drug combinations showed more balanced interaction changes in both directions (Supplementary Fig S5). These data show that different drug combinations have varying potential for changing their drug interaction type, even if the wild-type interaction is similar. Interestingly, for the additive chloramphenicol–tetracycline combination, none of the outlier mutants showed any change in drug interaction: while the sensitivity to the constituent drugs often changed, additivity was generally preserved (Supplementary Fig S3). This observation is consistent with previous results that this interaction changed little in strains that had evolved resistance to chloramphenicol–tetracycline (Munck et al, 2014). Thus, the chloramphenicol–tetracycline drug interaction appears robust to genetic perturbations. Together, these data show that most drug interactions can be removed, amplified, and even qualitatively changed to a different interaction type by rare genetic perturbations, while other interactions are robust.

We next identified cellular functions controlling drug interactions; among these, the synthesis of secreted polysaccharides (capsular and lipopolysaccharides, LPS) and ATP synthesis stood out in that they affected virtually all drug interactions. Firstly, manual inspection of the functional annotation of the outlier mutants (Keseler et al, 2005) for each drug pair, and systematic gene ontology enrichment analysis congruently exposed specific functions controlling the respective interaction (Fig 3D, Supplementary Table S1, Materials and Methods): for example, perturbing tRNA processing consistently removed chloramphenicol–nitrofurantoin suppression (Fig 3A and C; Supplementary Fig S6X and Y); similarly, ribosome production and assembly altered ciprofloxacin–tetracycline suppression (dksA, rsgA, ksgA in Supplementary Fig S5C; Supplementary Table S1), confirming previous results (Bollenbach et al, 2009). Secondly, we noticed that certain functions recurrently control multiple drug interactions. Besides polysaccharide and ATP synthesis (discussed below), chaperone deletions (ppiD, dnaK) consistently caused amplified or synthetic suppression for distinct drug pairs and perturbing amino acid synthesis (metL) amplified chloramphenicol–nitrofurantoin suppression but removed trimethoprim–chloramphenicol antagonism (Fig 3C, Supplementary Fig S6). Thus, multiple drug interactions are controlled by few recurring cellular functions.

Polysaccharide synthesis affects the majority of drug interactions (Fig 3C and D; Supplementary Fig S6): perturbing this function removed chloramphenicol–nitrofurantoin suppression and trimethoprim–mecillinam synergy (gmhB, rcsD in Fig 3C; lpcA in Supplementary Fig S6); in contrast, it led to synthetic suppression between trimethoprim and chloramphenicol (lpxM in Fig 3C; Supplementary Fig S6M and N). Polysaccharide synthesis mutants have modified outer membrane composition, which affects the uptake of molecules dependent on their chemical properties (Nikaido & Vaara, 1985). Hence, a plausible cause of these drug interactions is that bacteria regulate polysaccharide synthesis in response to certain antibiotics, which then affects the uptake of other drugs. Consistent with this mechanism, antibiotics are known to affect polysaccharide synthesis (Rothfield & Pearlman-Kothencz, 1969) and LPS synthesis mutants have increased sensitivity to chloramphenicol (Fig 3C; Supplementary Fig S6V); thus, stimulation of LPS synthesis by nitrofurantoin could explain chloramphenicol–nitrofurantoin suppression. To directly test this scenario, we removed outer membrane LPS using ethylenediaminetetraacetic acid (EDTA) (Nikaido & Vaara, 1985). Strikingly, LPS removal increased sensitivity to chloramphenicol, abolished chloramphenicol–nitrofurantoin suppression, and rendered this drug interaction purely additive (Fig 4A and B). Together, these data support that regulated changes in cell envelope composition, which affect drug uptake, are a recurring mechanism underlying chloramphenicol–nitrofurantoin suppression and other antibiotic interactions.

ATP synthesis also controls multiple drug interactions (Fig 3C and D). Specifically, the ATP synthase mutant atpF was more sensitive to trimethoprim and this sensitivity was reduced by chloramphenicol or mecillinam, leading to suppression (Fig 3C); a similar effect occurred for ciprofloxacin–tetracycline (Supplementary Fig S6S). A thiamin synthesis hypomorph, which also has perturbed energy metabolism, behaved similarly (Supplementary Fig S7). Further, ATP synthase expression increased two-fold in response to trimethoprim (Supplementary Fig S8), suggesting that cells respond homeostatically to ATP deficiency. To test independently whether ATP synthesis affects drug interactions, we specifically blocked the proton pore of the ATP synthase F_O subunit using drugs (oligomycin, venturicidin; Materials and Methods). Indeed, inhibiting ATP synthase led to suppression between trimethoprim and chloramphenicol (Fig 4C and D; Supplementary Fig S9), supporting that imbalances in energy metabolism cause this synthetic drug interaction. Together, these data suggest a mechanistic scenario in which impaired ATP synthase function leads to decreased intracellular ATP levels, which may become growth-limiting in the presence of trimethoprim or ciprofloxacin; concurrent translation inhibition by chloramphenicol would reduce global ATP turnover, replenish the intracellular ATP pool (Schneider et al, 2002), and thus lead to increased growth, which would explain the observed suppressive interaction. No suppression would occur in the wild-type where ATP is likely in excess and not growth-limiting. At a molecular level, decreased ATP concentrations might limit growth by aggravating DNA repair and synthesis (Waldstein et al, 1974), which is likely the growth-limiting process in the presence of drugs targeting DNA synthesis; however, other ATP-dependent processes could also contribute to suppression. Overall, our results show that perturbations of central cellular functions, unrelated to the common antibiotic targets, can reshape diverse drug interactions.

We established a general principle of bacterial growth, which enables the prediction of mutant growth rates under drug combinations from their growth under the individual drugs alone (Fig 2, Supplementary Fig S5). This principle may hold more generally and should be tested for combinations of other challenges such as osmotic, temperature, or pH stress in future work. While conceptually similar empirical laws are an integral part of physics, they are still scarce in biology (Scott & Hwa, 2011). Even without understanding their molecular origins, such principles are powerful since they enable the prediction of quantitative phenotypes. Here, such a principle was crucial for systematically revealing antibiotic interaction mechanisms.

The identification of cellular functions controlling drug interactions offers new strategies for the rational design of multidrug combinations. Specifically, we identified targets for potential adjuvants, which could reshape antibiotic interactions: thiamin synthesis inhibitors could render the chloramphenicol–trimethoprim combination reciprocally suppressive (Supplementary Fig S7); such reciprocal suppression may slow down resistance evolution but is extremely scarce among natural antibiotic interactions (Chait et al, 2007). LPS synthesis inhibitors could remove chloramphenicol–nitrofurantoin suppression (Fig 4A and B), thus preserving advantages of an untapped drug combination while increasing its potency. Drugs inhibiting cellular functions that control antibiotic interactions (LPS synthesis, ATP synthesis, and chaperoning; Fig 3C and D) are in development (Moreau et al, 2008; Evans et al, 2010; Du et al, 2011; Balemans et al, 2012). These inhibitors could reshape drug interactions even if they have poor antimicrobial activity alone since most mutants we identified have only mild growth defects. Finally, our approach revealed that certain drug combinations are robust to mutations (Supplementary Fig S3) or change primarily toward weakened antagonism (Fig 3A). The origins of such robustness and biases in drug interaction changes are unknown. Still, such insights can be used to avoid loss of synergism due to mutations occurring in treatments, which is a serious concern (Pena-Miller et al, 2013; Munck et al, 2014). It will be exciting to extend the systematic approach presented here to drug interactions in other systems including the most worrisome pathogenic microbes and cancer.

Materials and Methods

Author contributions

GC and TB designed the study. GC performed experiments shown in Figs 1-3; TB performed experiments shown in Fig 4. GC and TB analyzed the data and wrote the manuscript.

Conflict of interest

The authors declare that they have no conflict of interest.