Advances in the design of combination therapies for the treatment of tuberculosis

Abstract

Introduction:

Tuberculosis requires lengthy multi-drug therapy. Mycobacterium tuberculosis occupies different tissue compartments during infection, making drug access and susceptibility patterns variable. Antibiotic combinations are needed to ensure each compartment of infection is reached with effective drug treatment. Despite drug combinations’ role in treating tuberculosis, the design of such combinations has been tackled relatively late in the drug development process, limiting the number of drug combinations tested. In recent years, there has been significant progress using in vitro, in vivo, and computational methodologies to interrogate combination drug effects.

Areas covered:

This review discusses the advances in these methodologies and how they may be used in conjunction with new successful clinical trials of novel drug combinations to design optimized combination therapies for tuberculosis. Literature searches for approaches and experimental models used to evaluate drug combination effects were undertaken.

Expert opinion:

We are entering an era richer in combination drug effect and pharmacokinetic/pharmacodynamic data, genetic tools, and outcome measurement types. Application of computational modeling approaches that integrate these data and produce predictive models of clinical outcomes may enable the field to generate novel, effective multidrug therapies using existing and new drug combination backbones.

1. Introduction

Since the introduction of antibiotics, tuberculosis (TB) treatment has required months or years to achieve effective disease resolution. Culturable Mycobacterium tuberculosis (Mtb), the causative agent of TB, from patients decreases rapidly during the first few weeks of treatment [1]. However, historically, short-duration therapies (e.g., up to a few months) resulted in patients with disease at the end of treatment and a high percentage of disease relapse [1,2], indicating that there are Mtb that can survive drug treatment to regrow and cause disease once drug treatment ends. Some patients are successfully treated with short treatment times, while others are not. There are currently no diagnostic tests to indicate when enough therapy for a relapse-free cure has been given to an individual patient. Thus, treatment times must be lengthy for all patients to ensure the best possible outcome for most patients. Why are such prolonged treatment regimens required to achieve the durable curing of TB? Decades of research into this question suggest that the answer is partly due to the immune response to TB infection and the resulting disease pathology and partly due to the ability of Mtb to alter its metabolism and physiology to survive changing environmental cues [3].

Clinical trials have demonstrated that TB treatment must include multiple drugs given for several months to achieve an effective and durable cure. The current standard of care treatment for drug-sensitive (DS) TB is six months and consists of two phases of drug treatment: four drugs (isoniazid (H), rifampicin (R), pyrazinamide (Z), and ethambutol (E)) given for two months followed by two drugs (HR) for four months. This treatment has been used for over 40 years, with an estimated treatment effectiveness of 85% [2,4,5]. Furthermore, the estimated number of TB cases resistant to currently used drugs (including drug-resistant (DR) and multidrug-resistant (MDR)) remains high, at more than 500,000 annually [4]. Treatment for DR TB has historically required more than four drugs given for up to 24 months due to lower drug sterilizing abilities and toxicity, was without standardization and has left many patients without positive outcomes (<50% curing) [6]. Given the development of novel antibiotics, new clinical trial designs, and positive clinical trial results [7,8], there is the promise to develop more effective therapies for TB. Treatment guidelines now recommend a standardized and shorter treatment (HRZE) for most DR TB patients. Therefore, there is reason to be optimistic about the discovery of other effective treatment regimens for TB using new drugs and new design principles.

This review will discuss the lessons learned, the current state of primarily empirically determined drug treatment regimens, and developments in the systematic and rational exploration of combination drug space to identify the next generation of TB multidrug regimens. We will cover a brief history of how the current standard of TB drug treatment was empirically determined (Section 2), why multidrug therapies are needed to target Mtb during treatment (Section 3), and current and future approaches for identifying the best new candidates for shortening treatment time and increasing treatment efficacy (Section 4).

2. Empirical establishment of combination therapy for Mtb.

2.1. TB treatment from palliative care to combination antimicrobial therapy

Historically, TB treatment was necessarily empirical and iterative. Until the mid-twentieth century, TB treatment consisted of palliative care and the isolation of infected individuals to minimize the spread of disease [1,2,9]. Drug treatment for TB had been long sought after but not proven effective until the discovery of streptomycin (S) and para-aminosalicylic acid (PAS) in 1944 [2,9]. These antibiotics were quickly found to be efficacious for treating TB in the clinic [2]. Only a few years after the first uses, treatment with both S and PAS in a clinical trial demonstrated the increased efficacy of combination therapies while preventing the development of common drug resistance in monotherapies [10]. The discovery of isonicotinic acid hydrazide (H, isoniazid) in 1952, which also displays activity against Mtb, quickly led to a three-drug combination therapy H+S+PAS that had predictably high cure (>90%) in patients with as low as 4% disease relapse [2,10].

2.2. The modern short-course therapy

These advances in treatment were significant but required continuous antibiotic therapies for 24 months. In the 1960s, ethambutol (E) was introduced and led to the replacement of PAS in the three-drug therapy (HSE) because it shortened treatment time to 18 months while also being better tolerated than PAS. Many antibiotics currently used as second-line treatments for drug-resistant TB were also developed during the 1960s [2]. Rifamycins, antibiotics produced by soil bacteria, were discovered in 1957 and chemically modified to show more efficacy in vivo during the 1960s [2,9,11]. A semisynthetic antibiotic, rifampicin (R), was developed during the 1960s [2] and added to the TB combination regimen (HRSE) because of the dramatic improvement in treatment time from 18 to 9 months [1]. Pyrazinamide (Z) was rediscovered in 1972 and in clinical trials was shown to provide sterilizing activity in the first few months of treatment that further reduced treatment time to 6 months when combined with isoniazid and rifampicin (HRZ) [1]. Treatment outcome improvement was achieved through the incremental addition or substitution of drugs. The short-course therapy and treatment regimen was developed in the late 1970s based on these decades of clinical trials. Ethambutol was added shortly after the development of the short-course therapy regimen to combat the resistance to isoniazid and led to the standard of care today (HRZE). Further treatment improvement using substitutions was recently realized after more than 40 years of attempts with the shortening of drug therapy to four months by substituting isoniazid and rifampicin for moxifloxacin (M) and rifapentine (P) [7]. The design principles leading to our best TB treatment options have been primarily driven by the iterative addition or substitution of drugs to clinically effective combinations. To understand how to develop the next generation of effective TB drug combinations, it is helpful to consider what is achieved by multidrug therapy and what population of Mtb requires lengthy treatment so that it can be specifically targeted to achieve shorter-duration therapies.

3. Combination therapy as a means to treat TB in multiple physical compartments.

TB’s disease pathology helps us understand why Mtb infections can be challenging to treat and require multidrug therapies for extended periods. The pathology of pulmonary TB involves a coordinated immune response that includes multiple cell types that evolve during the progression and treatment of the disease. The complex and dynamic nature of the infection environment suggests that Mtb must survive a variety of stressors and possess the physiological flexibility to adjust to the changing surroundings. The pulmonary lesions of TB disease are organized, inflammatory, immune structures that consist of multiple cell types surrounding a core of either intracellular or extracellular Mtb [12,13]. Several lesion (granuloma) types have been described, each with characteristic immune cell composition, organization, degrees of necrosis of host cells, and eventually cavitation [14,15]. Necrotic and cavitating lesions are associated with high levels of extracellular Mtb and are more difficult to treat in humans [16] as well as animals[17].

3.1. Differential drug distribution

There is evidence in humans that lesion types have differential responses to drug treatment that may be influenced by the amount of drug that reaches Mtb. Recent positron emission tomography/computed tomography (PET/CT) studies monitoring lesion response to drug treatment within individual patients during 14 days of drug treatment observed that some lesions within a patient shrink while others enlarge [18] suggesting there is heterogeneity of lesion response to drugs within individual patients. Studies of drug distributions in patients with MDR TB demonstrated that linezolid (L) penetrates evenly through caseous lesions. In contrast, other drugs like clofazimine (C) do not penetrate the necrotic core of lesions, confirming what had been observed in animal models (Figure 1) [19–23]. Together these studies support the idea that Mtb in different lesions will respond to drugs differently in part because the concentrations of drugs where Mtb are located within the lesions are variable. These results provide a plausible explanation for why treatment may take a long time and why therapy with multiple drugs and/or higher doses provides the shortest and most effective cure. Furthermore, the design of new drug combination therapies should ensure that more than one drug is present in every lesion type at levels that will kill Mtb.

Figure 1.

Diagram of tissue drug distribution in Mtb infected lungs. Diagram of lungs with cellular (solid arrow) and caseous (dashed area) lesions (A). Pyrazinamide penetrates both lesion types (B) [103,199]. Rifampicin penetrates both lesion types and accumulates in the caseous core (C)[114,199]. Bedaquiline penetrates cellular lesions but is partially excluded from the core of caseous lesions (D) [103]. The relative drug concentration scale (from low to high) is shown in greyscale bar in the upper right.

3.2. Mtb dormancy and persistence

The lesion microenvironment (e.g., lipid carbon sources, iron deprivation) and stressors (e.g., acidic pH, oxygen levels) can induce a non-replicative or dormant state [17,24–27]. Mtb adapted to these conditions respond with changes in transcriptomic, proteomic, and metabolic activities [28,29]. Many antibiotics target actively replicating cells and thus may provide limited inhibition when treating infections containing non-replicating Mtb. Support for this idea comes from in vitro experiments [28,30,31], such as those demonstrating that some drugs in the standard of care regimen targeting the growth of new mycolic acid cell wall components (H and E) are more effective against actively growing Mtb [1,2,32–37]. The other drugs in the regimen (R and Z) can kill both actively growing and non-replicating Mtb and may partially explain why including these two drugs in regimens were associated with treatment shortening during the development of the current standard of care regimen [1,2,36,37]. More recently, drugs such as bedaquiline and pretomanid that display activity towards both replicating and non-replicating Mtb have been used effectively to treat Mtb in vitro [38,39] and clinically [8,40,41]. Furthermore, there are antibiotics that can selectively kill non-replicating Mtb [30,31,34,42–47]. Thus, there is hope that treatment-shortening regimens will be found by combining drugs that target both replicating and non-replicating Mtb.

As defined in the Consensus Statement from Balaban et al., tolerance describes cells that survive lethal or inhibitory concentrations of a drug without genetic resistance [48]. Tolerance of a subpopulation (persistence) allows the survival of a subgroup of cells for extended periods without changing the minimum inhibitory concentration (MIC) for the whole population [48]. This is particularly relevant for mycobacterial cells because, as a population, they are intrinsically heterogeneous in their drug response. Differential drug susceptibility at the single-cell level is observed in vitro in a drug-specific manner [49], suggesting that there is always a subpopulation of Mtb cells that may tolerate a single drug treatment and therefore could contribute to the challenge in rapidly clearing Mtb using antibiotic treatments. Persister cells may also be a stepping-stone to resistance [50], as has been observed in other bacterial species [51,52]. Moreover, in these studies, the tolerance to one drug preceded resistance development to a second drug. It may be that Mtb persisters in TB patients could present a risk for developing resistance to multiple drugs in the current treatment regimen. Care must also be taken to ensure drug combinations are designed so they are not prone to cross-resistance [53]. Targeting tolerant and dormant cells is a focus of new TB drug development [54].

3.3. Multidrug therapies reduce treatment time and drug resistance development

Drug resistance is a major issue in treating diseases such as bacterial infections and cancer. The likelihood of developing resistance increases with the duration of drug exposure. Therefore, the months-long treatment time for TB is sufficient for resistance development to occur [55,56]. Drug resistance can occur when cells acquire a factor (for example, a mutation in the drug target or mutation that increases the expression of an efflux pump) that allows cells carrying this factor to survive in the presence of drug concentrations that will inhibit or kill cells without the factor [48,57,58]. These factors are heritable, and once resistance develops, a drug is ineffective at clinically useful concentrations; the resistant population can survive and expand into the space previously occupied by the now-killed susceptible bacteria.

The use of multiple drugs that target different cellular processes is a strategy that can prevent the development of resistance experimentally and clinically. By targeting various essential cell processes, the likelihood for a single cell to be simultaneously resistant to multiple drugs is low. Historically, multidrug therapies for TB showed improved curing with fewer patients relapsing with drug-resistant Mtb compared with single-drug therapies [1]. Using as a reference the clinically determined spontaneous streptomycin resistance rate of one in 10⁵, one in 10¹⁵ Mtb bacilli would be triple drug-resistant [59]. Thus multidrug therapies with more than three drugs should be effective because there should be fewer than one spontaneously triple drug-resistant Mtb cell in a severely infected patient, using the severe case of cavitary disease bacillary burden as a reference [2,60]. Based on these calculations, the likelihood of multidrug therapy success is supported by the success of three and four-drug regimens for treating TB [1,7,8].

Decreased resistance development was clinically demonstrated during the early trials using S and PAS. In one early study, S monotherapy resulted in 20% of patients developing resistance [2] and the randomized control trials from the Medical Research Council (MRC) in the United Kingdom found streptomycin resistance in 70% of cases [2,61,62]. In contrast, the MRC studies found that streptomycin resistance could be detected in 41% of cases receiving both S and PAS and was found in as few as 9% of cases when PAS was administered every three days [2,61,62]. These early studies demonstrated the necessity of combination therapy to prevent resistance development. Therefore, it is not surprising that the development of short-course therapy for TB treatment utilizes at least three drugs. The clinical studies have demonstrated that we require combination therapy to bet-hedge treatment against variation in access and susceptibility to Mtb that resides in different lesions. Combination therapies should therefore include drugs that coordinate to rapidly reduce bacterial burden and then sterilize persister cells.

4. New era of combination therapy design

The use of animal and in vitro models of TB has been key to the development and design of drug combinations, as well as to the optimization of combination therapies. In the following sections, we will discuss the models, their uses, and computational approaches that will aid in identifying new combinations and designing future TB treatment regimens.

4.1. Optimizing combinations based on backbones

Dozens of drugs are being considered for inclusion in new TB drug combinations [63–65]. These include drugs that have already been approved for use in humans, like moxifloxacin or linezolid and are being repurposed or considered for use in TB treatments. There is considerable optimism and excitement that treatment improvement can be achieved with combinations including drugs, such as bedaquiline, and pretomanid, from newly discovered drug classes. However, the number of possible drug combinations among TB antibiotics is immense. Therefore, one successful approach has been to develop new combination therapies by augmenting and modifying effective backbones. Isoniazid, rifampicin, and pyrazinamide (HRZ) form the backbone of the current four-drug standard of care (HRZE) [1]. Over the last 15 years, the three-drug combination of pretomanid, moxifloxacin, and pyrazinamide was developed based partly on preclinical models’ success [66–68] and evaluated in a clinical study [69]. Early optimism about the potential clinical efficacy of this three-drug combination backbone led to the initiation of the ongoing SimpliciTB clinical trial [70] to evaluate the addition of bedaquiline to pretomanid, moxifloxacin, and pyrazinamide for the treatment of both DS and DR TB. Modifying these backbones through the addition and substitution of drugs has identified regimens with proven or potential treatment improvement [7,70,71].

The Nix-TB [8] and ZeNix-TB [41] clinical trials recently demonstrated dramatic treatment shortening of DR TB with the three-drug combination of bedaquiline, pretomanid, and linezolid (BPaL). BPaL was developed, in part, with preclinical studies which evaluated the addition of various drugs to backbones such as BPa [67,72–74]. Despite the treatment shortening and efficacy of BPaL, toxicity is associated with high and prolonged dosing of linezolid [8,75]. Therefore, there is interest in identifying drugs to substitute for linezolid (e.g., less toxic oxazolidinones or other drug from other classes).

The process of addition and substitution of drugs has continued to be used to identify new treatment-improving drug combinations. However, clinical trial successes and failures highlight the importance of making the best choices for combinations to test in the clinic. The choice of drugs in the current standard-of-care treatment regimen was identified through decades of clinical evaluation of the addition or substitution of drugs [1]. Clinical trials evaluating the substitution of rifapentine for rifampicin [76,77] and the substitution of moxifloxacin for either ethambutol or isoniazid [78] were unsuccessful at identifying treatment-shortening regimens. However, the combination of rifapentine, moxifloxacin, pyrazinamide, and ethambutol was demonstrated to be treatment-shortening in a recent clinical trial [7] and shares two drugs with the current standard of care. The combination was in part tested because of the potential seen in the earlier unsuccessful clinical studies, and consideration of moxifloxacin as a potential treatment-shortening drug for inclusion was also strongly supported through testing in preclinical mouse models [67,68,79–87]. The ongoing TRUNCATE-TB trial [71] further explores the potential for treatment shortening using substitution and addition of drugs to the current standard of care. Another recent example comes from the early optimism in the STAND clinical trial [69] evaluating pretomanid, moxifloxacin, and pyrazinamide that led to the addition of bedaquiline and the initiation of the ongoing SimpliciTB clinical trial [70] to evaluated BPaMZ for treatment of both drug-sensitive (DS) TB and MDR TB. These examples of recent trials evaluated seven drug combinations and identified between one and two treatment-improving regimens and have taken more than 12 years to complete. This underscores that choosing combinations for evaluation is essential to maximize the possibility of identifying treatment-improving regimens.

Clinical trials have demonstrated that adjustments (addition or substitution) of the drugs to a backbone combination can lead to improvements in treatment outcomes. In addition, recent trials have established paths for new regimens based on novel backbones and significant shifts on old backbones. Parallel work on novel backbones and refinement of existing backbones should be pursued to maximize the potential to identify treatment-improving drug combinations.

4.2. Preclinical animal models of TB disease

The drug combination space to be explored with new drugs is large and cannot be surveyed in the clinic. Animal models are important for evaluating drugs and drug combination design in modern TB regimen design and have been extensively reviewed [88–90]. Consideration of the drugs highlighted in the clinical trials above (e.g., moxifloxacin, bedaquiline, pretomanid) as potentially treatment-shortening drugs for inclusion in regimens was strongly supported through testing in preclinical mouse models of TB disease. While systematic evaluation of drug combinations efficiencies in animal models are not feasible, evaluation in animal models is essential before progression to clinical trials. We will briefly cover animal models and their uses for drug combination design.

Animals have been used to test new TB drugs’ efficacy and regimens since the introduction of streptomycin [2]. Additionally, our understanding of disease pathology and the biology of Mtb and host response to infection have depended on animal studies [88–92]. Mice are the most used animal models, but most strains lack many of the pathological hallmarks of human TB disease [93]. Other animal models (e.g., guinea pigs, rabbits, non-human primates) have granulomas more like those in human disease, but they cannot be practically used for large-scale drug treatment efficacy studies.

Mouse models of TB are the most practical and economical for drug efficacy testing due to their ubiquity in basic research and cost, ease of genetic modification, ease of handling, and the need for relatively small amounts of drugs [85,94]. The most commonly used mouse strains (e.g., BALB/c, C57BL/6, and Swiss) form lesions where Mtb is intracellular and lack the caseous, necrotic lesions or cavitations that occur in human disease. Despite these and other limitations, drug efficacy studies using these strains have produced a large amount of data, and these strains have demonstrated utility in evaluating new drugs and combinations. The treatment-shortening effects of pyrazinamide and rifampicin [95] were shown in mice. The newly approved multi-drug resistant treatment of bedaquiline, pretomanid, and linezolid was also effective in mouse studies [67,68,73,96]. Because of these successful examples and the practical aspects outlined above, these mouse strains will continue to play a prominent role in preclinical drug efficacy testing.

More recently, the C3HeB/FeJ mouse strain has been used in drug efficacy studies. These animals form heterogeneous lesion types: intracellular (similar to those observed in BALB/c mice); ones with caseous, necrotic cores that are similar in pathology to human lesions; and intracellular lesions enriched with neutrophils. C3Heb/FeJ mice also have heterogeneous drug responses. Treatment-responsive and non-responsive mice can be observed with pyrazinamide, which is thought to be influenced by heterogeneous lesion types. Drug combination efficacy studies have also shown the utility of the C3HeB/FeJ strain for evaluating drug combinations that show clinical efficacy [83,96–100].

Rats are commonly used in toxicology and pharmacology studies to evaluate the pharmacokinetics of drugs but were considered too resistant to Mtb infection to be useful. More recent studies have shown that Wistar rats can be infected and show similar pathology to mice [88,90]. Furthermore, a DprE inhibitor was tested in a rat infection model because optimal drug exposure was not attainable in mice [101]. The common use of rats in toxicology makes future development of this animal model for studying drug combination efficacy and pharmacokinetics promising.

Guinea pigs, rabbits, and non-human primates (NHP) form lesions that show caseation necrosis, while cavitation can also be observed in rabbits and NHP. These human-like pathologies make these animal models important for evaluating drug efficacy. Guinea pigs were the preferred animal model in the mid-1900s and were influential in developing early TB drug regimens. Recent studies of combination drug efficacy showed similar outcomes in mice and the caseous necrotic lesion contains abundant extracellular Mtb during drug treatments [88,90]. Even with these advantages, guinea pigs are likely to remain models for follow-up studies because they are larger, more expensive, and require more drugs during treatment because they clear drugs faster than mice. Rabbits form caseating necrotic lesions as well as lung cavities after infection with Mtb. In humans, longer TB treatment, disease relapse, and development of resistance are all associated with lung cavity formation. Additionally, drug distribution studies in the caseous lesions and pharmacokinetic studies have shown similarities with humans [17,20,21,94,102–104]. The rabbit models of TB disease will continue to provide important insight into the drug treatment of refractory lesions, but due to their size, cost, and other limitations, they will not take the place of mice in evaluating drug combination efficacy early in the development process. Non-human primates infected with Mtb have the pathological hallmarks, including cavitation, as well as the symptoms and range of drug treatment outcomes that occur in human TB disease [88,90]. Cynomolgus macaques, and more recently marmosets, have been used to study drug TB disease and treatment efficacy [88,90,105–109]. A study in marmosets showed improved treatment of cavitary lesions with the current four-drug standard of care compared to an older streptomycin-containing drug combination [105]. Despite the necessary scarcity of drug combinations that can be tested in NHPs, experiments in the NHPs provide rich information that helps us understand drug combination efficacies in the context of immune effects that are similar to those in human disease.

4.3. Integrative computational modeling

Preclinical animal models will continue to serve an important role in drug efficacy testing. However, none of these models are amenable to evaluating the thousands of candidate drug combinations; therefore, another path to realizing the potential treatment-improving drug combinations is to prioritize drug combinations for animal testing and enable exploration of novel drug combinations. In vivo efficacy is an integration of drug dosing and the time resolution of drug distribution (pharmacokinetics, PK) and susceptibility of Mtb populations in different metabolic states (pharmacodynamics, PD). One scalable and economical way to integrate these data is through computational modeling. Simulations that employ mathematical and computational modeling are being used to determine drug dosing and predict clinical outcomes of tuberculosis mono- and combination therapies [21,110–118]. Modeling approaches utilizing drug PK and PD are predictive of drug distribution in both plasma and lungs as well as predicting potential drug-drug-interactions, both of which may aid in choosing drugs to combine that minimize adverse side effects (reviewed in [119,120]). Utilization of PK/PD data to model drug combination effects and develop improved treatment regimens is a highly active area and has been reviewed elsewhere [119,121,122]. These models can include patient clinical and demographic data, infection burden, blood, and serum measurements for biomarkers, as well as tissue and serum drug concentrations. Experimental measurements from human, non-human primate, small animal studies, and the Mtb response in in vitro hollow fiber models (see below) can also be used as variables in mechanistic models that are being used to design improved multidrug regimens. Treatment regimens with established clinical efficacy (HRZE and BPaL) or that are being evaluated (BPaMZ) contain drugs with differential drug distributions and effects on Mtb that are actively growing or non-replicating [19–23] (Figure 2). Therefore, by using information about Mtb sensitivity combined with the penetration of these drugs into compartments containing Mtb, we may be able to explain why these regimens are clinically effective. By incorporating PK/PD with models of granuloma structure and immune effects, computational simulations predict drug and combination drug response in vivo and the clinic [123–128]. These simulations of granuloma formation use in vitro drug exposure measurements in addition to tissue pharmacokinetic/pharmacodynamic drug distribution and immune cell recruitment measurements from animal studies to model the dynamism of the infection environment. In recent years, integrated PK/PD models have been developed for many TB antibiotics, including isoniazid, rifampicin, rifapentine, pyrazinamide, fluoroquinolones, pretomanid, and drug combinations [21,111–117,129]. The use of clinical modeling for predicting treatment outcomes of new drug combinations has been limited because the small number of drug combinations that have been trialed in human studies limits the ability to evaluate the predictive ability of the models. Clinical data are also complicated by other factors, such as other conditions and compliance, which may make the outcomes harder to interpret in these mathematical models. Yet, the predictive abilities of PK/PD and integrative models of single- and combination-drug responses in vivo make these models, especially those that incorporate the immune state, important tools moving forward because they allow large-scale exploration of combination regimens and optimization of dosing.

Figure 2.

Diagram of tuberculosis lesion with drug penetration and Mtb cell type susceptibility. Mtb-infected lungs are shown on the left with different lesion types. Higher magnification of the lesions, highlighting the difference in lesion structure and Mtb cell growth within these lesions. These examples are high-level and generalized descriptions of major lesion types; there is a wide spectrum of lesion types that are not fully represented here but are reviewed in-depth [14,15]. The standard of care (HRZE) and two drug combinations showing efficacy in clinical trials (BPaMZ and BPaL) are listed on the right. *hypoxia refers to the area around the periphery of the caseum; hypoxia levels are variable in many lesion types. Antibiotics in each combination are listed below, and the relative penetration into lesion is shown with arrows (the upward arrow shows that drug penetrates readily into lesion, and the downward arrow shows drug penetration into a lesion may be uneven or restricted). The Mtb cell type susceptibly for each drug are displayed next to each drug. BPaMZ = bedaquiline + pretomanid + moxifloxacin + pyrazinamide. BPaL = bedaquiline + pretomanid + linezolid

4.4. in vitro models of drug combination effects

In vitro models are important for understanding the response of Mtb to its environment and to drug challenges. These models also play a crucial role in prioritizing new drugs for development and testing in animal infection models (see [88,90,130,131] for comprehensive reviews of models and their uses). However, in vitro models cannot capture the entire complexity of the infection environment. Until recently, there have been sparse data to articulate a clear path on how to best use Mtb drug response in vitro to evaluate in vivo drug efficacy. The increase in data availability of in vitro drug combination data has emphasized the need to develop in vitro models of microenvironments where persisters reside and to balance the simplicity of these models for systematic measurement that can be used for the prediction of in vivo outcomes.

Mtb encounters and must adapt to numerous factors and stressors during infection, including changes in nutrient availability, pH, oxygen availability, and other host-mediated stressors (Figure 2). In vitro growth conditions have been developed that mimic many of these important microenvironmental factors and allow for their study in laboratory settings [88,90,130,131]. These include liquid media conditions that model single as well as multiple factors and include in vitro immune cell-based infection assays (see Table 1 for selected models and the environmental factors they mimic).

Table 1.

Selected microenvironmental factors encountered by Mtb during infection and in vitro models developed to mimic these factors.

Microenvironment mimicked	in vitro model	References
general replicating Mtb	standard laboratory growth media	[88,90,94,130,131]
lipid-rich infection compartment	cholesterol	[17,33,88,90,94,130–138]
	fatty acids
intracellular occupancy	in vitro infection of cell lines	[88,90,94,130,131,140,144,145]
	in vitro infection of primary cells
	acidified media
oxygen limited infection compartments	hypoxic environmental chamber	[17,88,90,94,98,130,131,153]
	fatty acids + nitrate
caseous infection compartment	ex vivo caseum	[17,88,90,94,130,131]
	caseum mimic
	neutral pH
general non-replicating Mtb	nutrient limitation	[88,90,94,130,131,156,160,161]
	multistress models
dynamic nutrient and drug concentrations	hollow fiber systems	[88,90,94,130,131]

Broth-based in vitro conditions have been important for understanding the environmental factors influencing Mtb growth and survival, including the effects of different lipids as carbon sources [17,33,132,133]. In vitro studies have helped us understand the key role of cholesterol for Mtb survival during infection [134–138], including its use as a carbon source via beta-oxidation through its secondary metabolites as important precursors to virulence [133,137–139]. Mtb utilization of other lipids for growth and virulence, as well as the metabolic flux of carbon and important cofactors (e.g., Acetyl-CoA and propionyl-CoA), were described in in vitro conditions that used single lipid carbon sources such as short-chain fatty acids.

Mtb must also respond to changes in pH and oxygen levels during infection, and the details of how Mtb adapts to these changes have been studied using in vitro models. Mtb response to low pH has been well studied using growth media buffered to different pH levels that mimic the acidification of the phagosome [140] as well as the more neutral pH of portions of the extracellular caseum [17]. The importance of studying Mtb during pH adaptation for understanding drug response has been well documented for pyrazinamide. Pyrazinamide is generally inactive in standard laboratory media at neutral pH and is more potent at more acidic pH [141–143]. Mtb response to intracellular residency has also been studied using in vitro infection in both primary cells and cell culture systems [130,144,145]. These studies have identified important roles for the host cells in affecting Mtb response to antibiotics and in the availability of drugs at the sites where Mtb bacilli are found [146–149]. For example, both clofazimine and bedaquiline can be taken up by host cells which then serve a role in drug delivery to the intracellular Mtb. Furthermore, in vitro host cell infection models identified several potential treatments-improving drug combinations [80,150] that were confirmed in mouse studies [80,99,150–152].

Oxygen availability for Mtb varies during infection, including in the caseating lesions, which have hypoxic compartments [17,98,153]. Mtb can respond to low oxygen levels by slowing growth or entering a dormant or non-replicative state where many drugs that require active replication are not as effective [17,24,34,154–157]. Hypoxic growth conditions can be induced using environmental chambers that regulate gas concentrations but make monitoring drug response more challenging. When oxygen is low, Mtb can conduct anaerobic respiration and generate ATP by utilizing alternative electron acceptors, such as nitrate [158]. This process has been shown experimentally during intracellular infection [159] as well as modeled in vitro using fatty acids combined with nitrate. The reduction of nitrate during anaerobic respiration produces nitrite, which can be toxic for Mtb [159]. Consequently, anaerobic respiration of nitrate in vitro can lead to an accumulation of nitrite that also induces a hypoxic non-replicative/dormant state that may be relevant to the in vivo infection environment. As the field has started to focus on developing treatment regimens that target persisters, models where multiple factors are combined in a single in vitro system to study Mtb cellular and drug response have been described [130,156,160–162]. These multi-stress models are thought to more closely mimic the complex environments that Mtb encounters during infection.

Traditional in vitro methods for culturing and treating Mtb using broth and solid media represent controlled and consistent but static environments. Hollow fiber systems are an alternative in vitro approach in which Mtb are housed within chambers and can be exposed to dynamic media states [163]. This can include changes to the culture media for monitoring growth dynamics to changing environmental cues as well as changes in drug concentration over time as well as multiple drug treatments [118,129,164–166]. Hollow fiber systems allow for evaluating the PK/PD profile of a drug or combination more consistent with in vivo PK studies [118,129,164,165]. To date, this approach cannot be scaled to accommodate large-scale systematic studies of drugs or combinations. However, it is the only system that can monitor sequential drug treatments and dynamic changes in drug concentration more similar to the treatment environment during infection.

4.5. Measures of drug combination effects: dose responses and interactions

Systematic high-order (three or more) drug combination studies are not routinely conducted because the scale of the experiments and the size of the combination space make them logistically challenging. The checkerboard assay is the traditional method for measuring a single pairwise drug-drug interaction, wherein every drug-dose combination is measured. Checkerboard assays are resource-intensive and require, for example, one microtiter plate to measure all possible dose combinations of two drugs at ten doses (10×10 = 100 drug dose combinations). Three-dimensional checkerboard assays have been developed to study high-order drug combinations [167] but require thousands of microtiter plates to measure a few high-order drug combinations. New methods have been developed to approximate essential elements of the checkerboard to enable efficient, systematic measurements of drug interactions (pairs or high-order) using in vitro models [150,168,169]. Combinations that decrease relapse and shorten treatment time in mouse models of tuberculosis have been developed based on data from in vitro experiments [80,150].

The development of new methods to measure drug interactions has also highlighted the different outcome measurements from these assays. Traditionally, there is a tradeoff between the ease and the interpretation of a measurement, and these data may not always be directly comparable. Clinically, standardization in MIC determination using solid and liquid media has been set by the European Committee on Antimicrobial Susceptibility Testing (EUCAST) [170,171] and can be used to measure drug combinations. However, the method has not yet been adapted to scaling for systematic drug combination studies. Colony-forming units (CFUs) are time-consuming (assayed ~4 weeks after drug treatment) and are not amenable to large-scale systematic studies. However, CFUs are considered the best measure of the ability of a drug to reduce Mtb viability. In about two weeks, cell viability can be estimated by adding resazurin to charcoal agar outgrowth plates instead of waiting until visible colonies are formed [50,155]. This method can be scaled in multiwell plates but cannot be performed in drug treatment plates for very high throughput studies. Optical density (OD) is one of the easiest measurements of an in vitro assay because it can be made directly and repeatedly on drug-treated broth-based cultures and can easily scale with the use of multiwell plates. CFU and OD are used to determine minimal inhibitory concentrations (MIC) for drugs and therefore are highly interpretable. However, its interpretation is limited because it cannot necessarily differentiate between treatment cidality and growth inhibition. OD is further limited to optically clear media and apparatus. Liquid resazurin assays report on cellular metabolic activity and can be scaled in multiwell plates while also providing a quantitative readout of cell viability [172–174]. Resazurin assays typically require separate assay plates increasing the difficulty and scalability for systematic measurements. Genetically modified Mtb that can produce light (either autoluminescent or with the addition of substrate) or fluorescence provide other alternative measurements to OD for estimating cell growth [133,175–182] with luminescence, including the advantage of potentially providing a readout of cell viability. The genetic modifications have the limitation of potentially disrupting the Mtb physiology and requiring antibiotic selection, which may confound drug studies. Recently, drug sterilization of Mtb was estimated by measuring ribosomal RNA (rRNA) synthesis [183] as a sensitive and orthogonal metric to CFUs. Whichever measurement type is made, identifying candidate drug combinations requires both an efficient approach and an interpretable measurement.

4.6. Systematic studies of drug combination effects

The expansion of methods for high-throughput measurement of drug combinations has ushered in a new era of measuring drug treatment response across a variety of in vitro growth conditions. A major obstacle in drug combination studies performed in vitro is understanding how different growth environments model in vivo niches. Decades of studies of pyrazinamide demonstrate increased efficacy in acid pH media. Other drugs with growth condition-specific activities support the continued use of multiple models for understanding Mtb drug response [54,143,184,185]. A systematic study of drug combination effects on Mtb in eight growth conditions found drug potency and drug interaction depended on the conditions in which Mtb were treated [160]. The authors found that the outcome of mouse drug combination treatment could be predicted using in vitro data from subsets of growth conditions and metrics (e.g., interactions and potencies). These model-dependent results indicate that some in vitro formulations provide orthogonal information while others provide parallel information about drug combination effects on Mtb. These findings suggest that Mtb in these models are in different physiological states, perhaps analogous to Mtb in different lesion compartments, and that computational models that include combination drug response data from different relatively simple growth environments were predictive of relapse. A separate study using three growth conditions (log-phase, acid pH, and non-replicating) and testing multiple drug pairs monitored drug interaction and predicted bacterial load. Pretomanid and either moxifloxacin or bedaquiline were among the best pairs in multiple conditions showing a large bacterial burden decline with little correlation to drug interaction [186]. Both studies concluded that in vitro metrics of inhibition (e.g., decreased growth and bacterial load) were more informative than drug interactions for prioritizing drug combinations. Future studies are required to quantify the utility of any in vitro models for directly understanding the in vivo combination drug response and to continue developing in vitro models for combination drug studies that may reflect Mtb in various dormancy states that enhance persistence [23,154–156,180,187]. Additionally, computational methods have been developed to aid drug discovery [188] and predict drug combination effects [189,190]. These computational methods can complement and assist the methods for systematic, high-throughput measurements across multiple growth conditions to accelerate the identification of new effective drug combinations.

We have also entered an era where functional genomic approaches utilizing genetic tools, such as CRISPRi, are determining genetic pathways that, when perturbed, render Mtb vulnerable to drug treatments [191,192]. Genes in the target pathways of isoniazid and rifampicin were identified as vulnerable (inhA and rpoB, respectively), highlighting the ability of this approach to identify druggable targets and pathways. One advantage of the CRISPRi approach is the identification of new druggable pathways and the intersection of these pathways with environmental cues without having chemical compounds in hand. Intriguingly, the authors identified many genes that lack chemical inhibitors but that are more vulnerable than inhA and rpoB, indicating that future compound screens focusing on these pathways may be successful. Because gene expression can be attenuated, investigation of both essential and non-essential pathways and determination of conditional essentiality under differing growth conditions is possible. Identifying pathways vulnerable to perturbation using CRISPRi can identify new druggable pathways while also helping understand Mtb physiology and survival during infection. Chemical genetic approaches using genetic hypomorphs and pooled parallel sequencing have successfully paired effective compounds with their putative pathways of action [193]. Many more hits were identified using this approach than when wild-type Mtb was used, and the target pathways identified included the synthesis of essential cell components: DNA, cell wall, tryptophan, folate, and RNA. The authors identified a mycobacterial efflux pump, EfpA [194], and optimized compounds to inhibit wild-type Mtb. A follow-up study identified a structurally distinct compound that works synergistically and inhibits resistance development [195]. Furthermore, a recent study determined that EfpA can induce drug tolerance to many antitubercular drugs [196], suggesting that expression levels may play a role in how Mtb responds to drugs during treatment. Screens to identify mediators and inhibitors of multidrug tolerance, such as CinA [35459278], which cleaves NAD adducts from isoniazid, ethionamide, pretomanid, and delamanid, may extend the use of current antibiotics. Both CRISPRi and chemical genetics approaches have the potential to identify targets and compounds important for understanding drug treatment and improving future combination therapies.

The increasing availability of in vitro models and ease of measurement for drug and drug combination response to TB therapies has also created new opportunities to interpret these data. The development of experimental and computational approaches shows promise for predicting in vitro drug interactions using fewer experiments. A basis for these predictions is to take advantage of signatures of gene expression changes. These signatures predict drug interactions by using in vitro measurement in response to single-drug treatment. This enables predictions of thousands of drug combination responses from RNA sequencing data from tens of treatment groups [190,197]. These approaches are being used in interpreting why drugs increase efficacy while others antagonize the activity of each other and which Mtb pathways mediate these drug interactions in vitro. Additionally, experimental and computational approaches have been developed that predict drug combination treatment outcomes in preclinical and clinical studies using in vitro measurements [80,150–152,160,162,183]. The insights from these treatment prediction modeling approaches can help us understand the physiological and environmental factors most relevant during drug treatment while using in vivo data to refine computational models.

5. Conclusion

The push to empirically develop treatment for TB in the 20th century was a success and heralded randomized control trial design, which is now the gold standard for clinical trials [61,62]. We are entering a new era to leverage the incredible advancements in drug discovery for tuberculosis that have led to many new antibiotics and targets to be identified by focusing on how to design optimized combination therapies [198]. So far, a thoughtful empirical approach to determining effective drug combinations has led to the establishment of several combination backbones with successful clinical trials. Yet, a significant number of novel combinations remain to be considered experimentally, especially involving new antibiotics. In recent years, the field has made key advances in developing animal models, in vitro models, and outcome metrics. These models were designed to mimic better the microenvironments that house the bacteria that are hardest to reach and kill with antibiotics. In the next phase of TB drug combination optimization, we expect that computational models will be used to integrate data of different types to improve methodology and our understanding of how to rationally design optimized combination therapies.

6. Expert opinion

The TB field has successfully developed new antimycobacterial drugs over the last twenty years (newtbdrugs.org, [63,64]). Though novel multidrug regimens have been found to be superior to the standard of care and a new regimen was approved for the treatment of MDR TB, the development of optimized combinations is not at the pace at which new compounds are developed. Nonetheless, the recent success of two clinical trials with new combinations, including those incorporating antibiotics already in use, points to the potential of the combination space to yield treatment-shortening regimens. On the other hand, the failure of some clinical trials, despite the supporting in vitro and preclinical animal data, highlights that we must improve these models and refine the metrics used to evaluate treatment outcomes.

To shorten the duration of treatment, therapies must rapidly sterilize the entire Mtb population. The underlying heterogeneity of tuberculosis lesions and the bacterial population creates distinct subpopulations of Mtb that must each be penetrated by antibiotics and then must be susceptible to those drugs at the levels reached in the tissue. To achieve this, we must continue to foster a collaborative and innovative research environment to integrate higher-quality measurement across a broad range of novel combinations. These data types should take advantage of in vivo and in vitro models with predictive value against treatment-shortening outcomes in preclinical animal models and the clinic. The response of Mtb in vivo to drug treatment depends on the microenvironment to which the population has adapted; therefore, Mtb in one niche may be susceptible to a particular drug combination while Mtb in another niche are tolerant. Likewise, the response of Mtb to drug combination treatment in in vitro experiments depends on the growth environment. How should drug combinations be evaluated when shown to be effective or synergistic in one condition but not in another? We must move forward by taking advantage of the scale of in vitro studies that allow for efficient and systematic measurement of the combination space and use computational modeling to understand which of those data are predictive of the in vivo response. Furthermore, as we refine which outcomes in preclinical animal models are most predictive of treatment shortening (and relapse-less cure) in clinical trials, we should update the standards of which in vitro experimental data are most practically useful for drug combination design, e.g., which conditions and outcome metrics can be used to model treatment efficacy of Mtb in niches that are most refractory to antibiotic treatment (such as persisters).

Our vision of an effective workflow for the design of combination therapy earlier in the drug development process is dependent on computational modeling to integrate disparate data on PD and PK in a lesion-specific manner (Figure 3). We anticipate that computational analysis will also enable us to realize the full potential of data in a way that balances resources and experimental scale. For example, clinical and NHP data are the scarcest and are therefore only available for a few regimens. Smaller animal studies should be used the prioritize combinations for testing in NHP and in the clinic, and outcomes should be quantitatively compared so that we learn how to make these animal studies more predictive (e.g., using different PK measurements, lesion type, and outcome measures). On the other end of this scale are in vitro measures that can be acquired rapidly and are relatively high throughput. We should continue to improve the conditions and metrics used for in vitro measures to balance throughput and predictive value against in vivo efficacy. In turn, these optimized in vitro conditions may be used with genetic tools to identify vulnerable molecular pathways of Mtb in growth conditions that model where current therapies are least effective. Measuring and modeling the drug degradation in both in vitro and in vivo models, as well as understanding the tissue distribution of active or unbound forms of drugs, will play an important role in better understanding the most effective drugs to combine. In evaluating combinations, integrating data from in vivo (preclinical model) experiments and clinical trials on both treatment successes and failures will play an important role in learning what distinguishes an effective treatment combination from one that does not show efficacy. Finally, as we focus on evaluating different drug combinations for treatment efficacy, we must also consider the long-term potential of novel combinations with regard to their potential to promote phenotypic and genetic resistance.

Figure 3.

Diagram of data flow for development of treatment improving candidate drug combinations. In vitro data of drug and combination effects can be directly used to inform candidate drug combination choice or can be used as input to design in vivo experiments and modeling efforts to predict optimal candidate combinations. Likewise, in vivo drug combination data (e.g., from preclinical animal models and human PK/PD studies) can be used directly to inform candidate combination choice or as input for modeling. Computational modeling that integrates multiple data sources, including effective in vitro combinations, drug tissue PK/PD, and preclinical drug combination efficacy can be leveraged to select candidate combination choice; as more combination data become available, the models can be refined and improved.

Article Highlights.

• Tuberculosis combination therapy has primarily been empirically determined and based on three-drug combination backbones

• Combination therapies for long durations are required to sterilize multiple and diverse infection microenvironments

• Experimental advances allow systematic evaluation of new drug target pathways and drug combinations

• To shorten the duration of treatment, therapies must rapidly sterilize the entire Mycobacterium tuberculosis population.

• Advances in modeling of drug therapies and treatment outcomes show potential to predict treatment outcomes and prioritize combinations for testing