Molecule Property Analyses of Active Compounds for Mycobacterium Tuberculosis

Abstract

Tuberculosis (TB) continues to claim the lives of around 1.7 million people per year. Most concerning are the reports of multidrug drug resistance. Paradoxically, this global health pandemic is demanding new therapies when resources and interest are waning. However, continued tuberculosis drug discovery is critical to address the global health need and burgeoning multidrug resistance. Many diverse classes of antitubercular compounds have been identified with activity in vitro and in vivo. Our analyses of over 100 active leads are representative of thousands of active compounds generated over the past decade suggesting they come from few chemical classes or natural product sources. We are therefore repeatedly identifying compounds that are similar to those that preceded them. Our molecule-centred cheminformatics analyses points to the need to dramatically increase the diversity of chemical libraries tested and get outside of the historic Mtb property space if we are to generate novel improved antitubercular leads.

Graphical Abstract

INTRODUCTION

It is widely known that Mycobacterium tuberculosis (Mtb) infection is causative of tuberculosis (TB), and claims the lives of around 1.7 million people with 10.4 million infections reported per year.³ Hence TB continues to be the focus of many international efforts to develop new therapeutics to address resistance to the available first- and second-line drugs.⁴ Due to increasing TB morbidity in developed countries in the late 1990’s and the global expansion of drug-resistant strains, a high priority has since been given to the development of new TB drugs by the World Health Organization, Bill & Melinda Gates Foundation, National Institutes of Health as well as other international institutions and funds. While the widely used TB drugs were developed in the late 1950’s-1960’s (Table 1, compounds 1–4), it is only recently that the first new drugs (bedaquiline, delamanid and pretomanid) came on the market in over 40 years^{5, 6} (Table 1, compounds 5–7); these latter molecules are currently not widely used because of their cost (Table 1) and their use is limited to multidrug resistant Mtb (MDR-TB).^{7, 8} Interestingly, these three molecules also have the highest hydrophobicity (ALogP) out of the 7 approved TB drugs and no hydrogen bond donors which is unusual (Table S1).

Table 1.

First line and recently approved antituberculars.

Drug	Year approved	Pros (and target where known)	Cons	References
Isoniazid 1	1952	A prodrug activated by KatG. Inhibits synthesis of mycolic acids by targeting InhA. In the United States a month of treatment costs less than $25.	Extensive side effects including neuropathy and liver enzyme effects.	9
Pyrazinamide 2	1954	Can penetrate granulomas. Kills non-replicating Mtb. A prodrug that is converted by pyrazinamidase (PncA) to the active pyrazinoic acid.	PZA resistance exists.	10, 11
Ethambutol 3	1962	Inhibits the enzyme arabinosyl transferase. In the United States it costs $100 to 200 per month.	Ocular side effects.	12
Rifampicin 4	1967	Targets RNA polymerase. In the United States costs about $120 per month.	PXR inducer and leads to hepatotoxicity	13
Bedaquiline 5	2012	Blocks the proton pump for ATP synthase.14–18	Discovered by M. smegmatis whole-cell screening19, MIC99 − 0.03 μg/ml Used for multidrug resistant TB. induce phospholipidosis at high doses20 potent inhibition of hERG increasing clinical QTc interval prolongation21. Costs $30,000 in high income countries. Mutations in atpE can lead to resistance	22, 23
Delamanid 6	2014 (EU)	Blocks manufacture of mycolic acids in cell wall. Increased sputum culture conversion.	Used for multidrug resistant TB. QT prolongation. Costs $30,000 in high income countries	23
Pretomanid 7	2019 (USA)	Blocks manufacture of mycolic acids in cell wall. Costs $364 for a six-month treatment.	Used in combination regimen with bedaquiline and linezolid for people with XDR-TB or treatment-intolerant/non-responsive MDR-TB	24, 25

The appearance and increasing number of Mtb strains with multiple drug resistance (i.e, MDR is defined as resistance to at least rifampicin (RPN) and isoniazid (INH)) and extensive drug resistance (i.e., XDR means MDR and additional resistance to at least one fluoroquinolone and one injectable second-line drug) represent a potential serious public health crisis. These intractable TB strains are associated with low cure rate and high mortality^{3, 26} as well as significant added treatment costs. Even more concerning, totally drug-resistant strains (TDR-TB) have been reported in clinical practice.^{27, 28}

Widespread latent tuberculosis²⁹ and the low potency of current TB drugs against latent disease represent one more weakness with current TB therapy.^{30, 31} According to the WHO, Mtb now exists as latent disease in approximately one quarter of the global population. Latent infection remains asymptomatic for a long time. Over a lifetime, it becomes active in about 5% of infected persons³, and reactivation is associated with a weakened immune system (including pediatric and geriatric cases) and in patients with compromised immune status, such as individuals undergoing immunosuppressive therapy or infected with HIV. HIV-positive patients carrying latent tuberculosis also have a several-fold higher risk of developing active tuberculosis (8–10 % annually). Therefore, although millions of cases of reactivated tuberculosis are reported worldwide, reliable methods of timely diagnosis and treatment of latent TB have not been developed. Thus, latent TB and reactivation risk represent a significant health issue. There is no doubt that the development of advanced TB drugs will play a key role in the reduction of morbidity and mortality levels required for achievement of WHO’s global international goals.³² This state of affairs and the general importance of this issue were also emphasized during a recent WHO Ministerial Conference held in Russia³³, where TB represents a major healthcare challenge.

There are several other difficulties involved with Mtb drug discovery, namely the bacteria is slow growing, a BSL3 laboratory is required to work with the bacteria and studies are complicated by the need to assess MIC and intracellular killing in infected macrophages.^{34, 35} Ultimately in vivo assessment is also slow because of the length of infections required as well as the need for acute and chronic animal models. Subsequently, the pipeline for antituberculars, which is mostly driven by the TB Alliance, is by any standards very thin with approximately a dozen new tuberculosis drugs undergoing Phase 1–3 clinical studies currently as summarized by the working group on new TB drugs (Figure 1).

Figure 1.

Current TB drug pipeline^{1, 2}, * denotes new chemical class. Molecules in phase 3 have all been FDA (bedaquiline and pretomanid) or EU (delamanid) approved but also components of ongoing trials. Image courtesy of Dr. Barbara Laughon (NIAID) www.newtbdrugs.org.

Some of these drugs are expected to become viable products. In particular, six compounds to combat MDR-TB are being investigated within the framework of Phase II and III clinical trials. Bedaquiline (5, Sirturo®, Janssen Therapeutics, Titusville, NJ, USA), delamanid (6, Deltyba®, Otsuka Pharmaceuticals, Tokyo, Japan) and pretomanid (7, TB Alliance, USA) have been approved for MDR-TB, respectively and are still undergoing additional clinical trials as combinations with other drugs. Repurposing of the known drugs clofazimine and levofloxacin are in Phase II and III trials, as tuberculosis therapy.^{36, 37} Mtb strains resistant to some of these molecules have been already reported^38–40, which suggests the limited time window with a new drug before losing sensitivity in the field. This issue also points to the continued urgent need for rapid development of the most advanced TB drugs and continual feeding of the research pipeline. For all the decades and billions of dollars spent on tuberculosis drug discovery, this begs the question why is there not more to show for it? Let us begin by looking at the two of the most recent drugs to be approved for Mtb.

Nitroimidazoles as ‘new’ drugs for Mtb

Initially, nitroimidazoles were produced as bicyclic nitroimidazofuran derivatives developed for cancer chemotherapy as radiosensitizers.⁴¹ Once they were identified in 1989 to be potent against actively growing non-fissile tuberculosis, interest in this class of compounds took off and this work has been extensively reviewed.⁴¹ Delamanid (OPC-67683 (6)) and Pretomanid (PA-824 (7)) are examples of this class of compounds (Table 2) and there is continued interest in developing additional compounds in this series.⁴² This class of compounds is therefore far from new dating back 30 years. When you compare the two compounds it is apparent that they are very similar in terms of activity against drug sensitive and resistant Mtb.

Table 2.

Comparison of Delamanid and Pretomanid

	Delamanid 6	Pretomanid 7
MIC (μg/ml)	0.012 μg/ml43	(MIC90) of 0.039 μg/ml44
Drug resistant clinical isolates MIC (μg/ml)	~0.01 μg/ml45	0.031–0.531 μg/ml44
Compound class	nitro-dihydro-imidazooxazole	nitroimidazopyran
Mechanism	Non-mutagenic. Delamanid inhibited the synthesis of methoxy- and keto-mycolic acid.45	Protein synthesis and lipid synthesis are substantially inhibited. Pretomanid is activated by a process that requires the unique deazaflavin F420 and depends on electron transport. Active against latent TB.
Human dosing	Delamanid was well tolerated in initial studies.46 Delamanid was associated with an increase in sputum-culture conversion at 2 months among patients with multidrug-resistant tuberculosis.47 It was also shown to improve outcomes and survival in MDR TB.48	Maximum plasma levels in 4–5h, steady state after 5–6 days daily dosing, elimination half life 16–20 h49, causes creatinine to rise by inhibiting renal tubular creatinine secretion50, bactericidal from 200–1200 mg daily over 14 days51, a dose of 100 mg – 200 mg appears safe and effective52, pretomanid-moxifloxacin-pyrazinamide is potentially suitable for drug sensitive and MDR TB53, pretomanid does not inhibit CYP3A4 in a clinically relevant manner54, increase in exposure with high fat meal55, combination with bedaquiline and pyrazinamide indicates a new treatment approach56, moxifloxacin, pretomanid and pyrazinamide was safe and well tolerated and showed superior efficacy.57

Compound 6 is a nitroimidazoxazole that was developed⁴⁵ for the treatment of MDR-TB. Compound 6 was developed from a smaller nitroimidazole⁵⁸ and 6 inhibits biosynthesis of methoxy-mycolic and keto-mycolic acids (Table 2) which are essential components of the TB mycobacterial cellular wall. It is notable that, both 6 and 7 require metabolic activation by deazaflavin-dependent nitroreductase (Ddn) to provide anti-tubercular activity.⁵⁹ Compound 6-resistant cells were characterized by Rv3547 (ddn) showing the gene encoding Ddn mutations and lacking the ability to provide metabolic activation of this compound.⁴⁵ Mtb produces the inactive desnitro-imidazoxazole as the major metabolite of 6. We can infer that Rv3547 reduces the delamanid nitro group forming an intermediate between the 6 prodrug and the inactive terminal metabolite (i.e., desnitro-imidazoxazole).⁴⁵ Compound 6 has potent anti-tubercular activity including against drug-resistant strains and no cross resistance has been reported with any current tuberculosis drug.⁶⁰ Compound 7 is a bicyclic nitroimidazole, nitroimidazooxazine or nitroimidazopyran derivative. Studies have shown both bactericidal and sterilizing effects on drug-resistant, drug-susceptible Mtb strains and resting mycobacterial cells.⁶¹ In vitro studies using Mtb clinical multidrug resistant isolates and in vivo studies using mouse and guinea pig models demonstrated anti-tubercular activity.^{44, 62} According to mouse preclinical studies, threshold toxicity is more than 1000 mg/kg after single oral administration.⁶² Compound 7 does not demonstrate cross resistance with conventional tuberculosis drugs.⁴⁴ Compound 7 is transformed into three major metabolites causing a release of nitric oxide (NO, a natural antibiotic released by macrophages⁶³). Similar to isoniazid, 7 inhibits cell wall growth by affecting the biosynthesis of mycolic acids under aerobic conditions. Under anaerobic conditions it causes a cyanide-like action, such that pretomanid acts as a respiratory poison leading to NO production and respiratory inhibition in Mtb by binding to the heme/ cytochrome iron orbital that normally binds oxygen and acts like cyanide would.⁶⁴ NO is produced by imidazole ring two-electron reduction rather than extraction from the nitro group.⁶¹ Compound 7 effect on the mycobacterial respiratory chain is manifested at lower ATP intracellular levels.⁶⁴ According to a recent study, intracellular accumulation of methyl glyoxal (i.e., a toxic metabolite) may be another 7-induced mechanism of mycobacterial killing.⁶⁵ Curiously, 7 also “imitates” the natural immune response of the host organism such that only penetration into Mtb can induce NO production.⁶⁴ If we are to greatly improve upon these compounds and leverage the same approach of NO production, we would need to address the high frequency of resistance and overcome drug resistance seen in the clinic.⁶⁶

ANALYSIS OF ACTIVE ANTITUBERCULAR COMPOUNDS

In recent years, there has been an abundance of reviews written on the topic of tuberculosis drug discovery as well as advances in treatment regimens and newer approaches to therapy. While we do not intend to cover the same ground, it is worth noting that most of these prior publications tend to focus on the underlying biology and targets used.⁶⁷ They also hone in on a relatively small number of targets such as InhA, DNA gyrase, RNA polymerase combined with a few newer common targets like MmpL3, DprE1, CTP synthetase and QcrB.^{68, 69} Others have described emerging targets with some of the hits from high throughput screens that appear structurally similar across different groups.⁷⁰ A recent review⁷¹ covered more current research but makes no attempt to organize the compounds by structure class. Other reviews have also covered a larger period of time but without a distinction of any kind of bioactivity threshold.^{72, 73} While such reviews have tackled why TB drug discovery is so challenging, they focused heavily on the TB Alliance pipeline with little if any discussion of the chemistry.⁷⁴ More recent reviews have provided small selections of lead compounds for TB grouped by pathways targeted, while also commenting on the difficulty in developing fast acting drugs.⁷⁵ In contrast to these many reports, the objective of the current review is to provide a new perspective on the design and development of TB drugs. Perhaps importantly none of these prior reviews has attempted to analyze or understand recent antitubercular chemistry in the context of the large numbers of active compounds described over the past few decades.We will therefore focus our attention on issues with these compounds and specifically the gains made in the development of these molecules (including against drug-resistant strains and MDR-TB). In order to dramatically limit the number of compounds addressed over the last decade, we will focus on those affecting drug-susceptible or drug-resistant Mtb strains that have demonstrated predominantly minimum inhibitory concentrations (MICs) < 10 μM, low toxicity and in many cases have shown promising in vivo efficacy. While we have also assessed those publications that have taken a target-based approach and identified potent molecules, in many cases they do not have an acceptable MIC. For example, a fragment-based screening approach took a weak 4-phenyl-1H-imidazole compound hit for inosine-5’-monophosphate dehydrogenase (IMPDH) and improved it to IC50 of 0.27 μM, but did not show inhibition of bacterial growth up to 50 μM.⁷⁶ This molecule therefore does not fit our selection criteria.

Most of the compounds assessed are from relatively recent years with a few older compounds for comparison. Many of the scaffolds described have significant liabilities in selectivity or drug disposition/ toxicity which we will highlight. While several of the papers highlighted have been published in this journal, we are not limiting our scope to this alone. It is also unlikely we have captured every TB drug discovery paper to all readers satisfaction, but our intent is to provide a sampling, a “taste of TB drug discovery” if you will, and then place them in the bigger context. We have therefore divided the 118 molecules into separate tables (Tables 3–12). The selection of what are the actual cores of these molecules may vary by the differing perspectives of readers. We will demonstrate how the fruitful search for these new highly active anti-tuberculosis compounds performed by many groups globally has led to some chemical diversification of antitubercular drugs, but we question whether it is enough. Our goal is to structurally categorize the most promising leads rather than focus on biology, allowing an analysis from a molecular property perspective, which is fundamentally different compared with any of the preceding reviews. Thus, we can attempt to more clearly understand the chemical as well as mechanistic diversity of these potential innovative drugs that have been identified to date.

Table 3.

Pyrrole, indole-2-carboxamide, Pyrazole [1,5-a] pyridine-3 carboxamide, Pyrazolepyridone, imidazole [1,2-a] pyridine and oxazolidinone chemical classes. MmpL3 = Mycobacterial membrane protein Large 3, DprE1 = decaprenylphosphoryl-β-d-ribose oxidase. PAINS were calculated as described by Lagorce et al.¹⁰⁵ Compound assay promiscuity was also predicted with BadApple.¹⁰⁶

Molecule	ID	Chemical class	Mtb Activity	Target	Notes	References
	8	Pyrrole	MICs 0.7–6.2 μg/ml also active against drug-resistant bacilli (including ones with multiple resistance) and infected macrophages	MmpL3	PAINS pyrrole_B(29) BadApple: 2 Alerts Also known as BM212.	107, 108
	9	Pyrrole	MIC: 0.5	MmpL3	PAINS pyrrole_B(29) BadApple: 2 Alerts Thiomorpholine fragment improves antimycobacterial activity	40,
	10	Pyrrole	MIC90 – 0.5 μg/ml		PAINS pyrrole_B(29) BadApple: 2 Alerts Hybrid molecule.	109
	11	Pyrrole	MIC 0.025–0.12 μg/ml also activity against isoniazid- and rifampicin-resistant Mtb strains		BadApple: 2 Alerts Sudoterb, in vivo activity in infected mice as 12-week therapy (12.5 mg/kg) led to complete lung and spleen sterilization. Has undergone phase I clinical studies.	110, 111
	12	Indole-2-carboxamide	MIC90 was 0.006–0.047 μM against Mtb including MDR- and XDR-TB strains	MmpL3	BadApple: 1 Alert in vivo efficiency assessment using an infected mouse model demonstrated that the compound can reduce pulmonary bacterial load by 2.12 orders of magnitude in 4 weeks (dose - 100 mg/kg)	112
	13	Pyrazole [1,5-a] pyridine-3 carboxamide	MIC90: 11.1–223 nM		PAINS anil_di_alk_D(198) anil_di_alk_D(198) In vivo efficacy led to lower visceral bacterial load	113
	14	Pyrazolepyridone	MIC 1.6 μg/mL	DprE1	Suboptimal water solubility (1 μM, logD 3.9), free plasma protein binding (1 %) and clearance (27 ml/min/kg)	81, 114–116
	15	Imidazole [1,2-a] pyridine	MIC90: 0.09–0.13 μM active against two INH- and rifampicin resistant MDR clinical isolates		BadApple: 1 Alert In vivo studies in mice and rats showed satisfactory pharmacokinetic properties (peak serum concentration (Cmax) was 225 ng/ml, half-life (T1/2) was 1.5 h and clearance was 86.284 ml/h/kg)	117
	16	Oxazolidinone	MIC50 1 μg/ml and MIC90 0.5 μg/ml	Binding with mycobacterial 50S ribosomal subunit	PAINS anil_di_alk_A(478) Associated with peripheral and optic neuropathies which are the most common side effects. Toxicity detected during clinical trials resulted in the withdrawal	118–123
	17	Oxazolidinone	MIC: 0.03–0.50 μg/ml	Binding with mycobacterial 50S ribosomal subunit	PAINS anil_di_alk_A(478) The therapeutic regimen involves combination with moxifloxacin and pyrazinamide which is more potent than rifampicin + isoniazid + pyrazinamide	124–126
	18	Oxazolidinone	MIC 0.25 to 1 μg/ml against a panel of clinical isolates of M. tuberculosis, including single-drug-resistant strains	Binding with mycobacterial 50S ribosomal subunit	Multiple use resulted in several non-life-threatening side effects, the most common being gastrointestinal disorders and a slightly increased reticulocyte level observed during the course of drug treatment	127–129

Table 12.

More recent lead compounds of different structural classes. Ddn = Deazaflavin-dependent nitroreductase, InhA = enoyl-ACP reductase, PyrG = CTP synthetase, QcrB = The b subunit of cytochrome bcc complex, IMPDH = inosine-5’-monophosphate dehydrogenase, DprE1 = decaprenylphosphoryl-β-d-ribose oxidase, Pks13 = polyketide synthase 13, Mmpl3 = Mycobacterial membrane protein Large 3, Mptpb = protein tyrosine phosphatase B. PAINS were calculated as described by Lagorce et al.¹⁰⁵ Compound assay promiscuity was also predicted with BadApple.¹⁰⁶

Structure	ID	Chemical class	Activity	Target	Notes	references
	104	Triazine	MIC 0.078 μM	Ddn and InhA	BadApple: 1 Alert Original hit from a Bayesian model screen	66
	105	Triazine	MIC 0.63 μM	Ddn and InhA	BadApple: 1 Alert Optimised lead, did not show efficacy in mouse	66
	106	Thiophenecarboxamide	MIC 0.5 μg/ml	PyrG	BadApple: 1 Alert Compound activated by EthA	86
	107	Quinoxoline di-N-oxide	MIC 1.2 μg/ml		Lead optimization of hit discovered using a pharmacophore search.	87
	108	Pyrazolo[1,5-a]pyrimidine	MIC 1.1 μg/ml		BadApple: 1 Alert Compound identified by a Bayesian model screen	85
	109	Imidazole	MIC 4.8 μg/ml		Compound identified by a Bayesian model screen	225
	110	Thymidine	MIC 7.8 – 15.6 μM	thymidylate kinase	BadApple: 1 Alert IC50 of 0.95 μM versus Mtb TMPK	88
	111	Indolyl Azaspiroketal	MIC90 versus Mtb BCG of 0.8 μM	Selective membrane permeabilization of Mtb membranes	BadApple: 1 Alert Modest reduction in CFU counts in the lungs (−0.75 log10) and spleens (−0.84 log10) of Mtb H37Rv infected mice when dosed at 100 mg/kg daily over 4 weeks (6 days per week, total of 24 doses)	89
	112	Morpholino-Thiophenes	MIC of 0.24 μM	QcrB	BadApple: 1 Alert Activity in a murine intratracheal infection model at 100 mg/kg for four days led to a modest reduction of CFU in the lung of 0.8 log compared to 2.8 log for moxifloxacin
	113	Benzoxazole	MIC = 1.0 μM	IMPDH	BadApple: 2 Alerts Issues with mouse microsomal stability and high serum binding	91
	114	Bis-Substituted Cyclam	6.25 μg/mL		BadApple: 1 Alert Relatively large (MW = 967.31 amu), charged species which might impact oral bioavailability	92
	115	Phenyl Aminothiazole	MIC90 against several strains of Mtb in the range of 0.060 to 0.125 μg/mL		BadApple: 2 Alerts Compound is metabolically stable and apparently is not susceptible to drug efflux pumps	93
	116	6-Dialkylaminopyrimidine Carboxamide	MIC99 1.3 μM	BCG_3193 (Rv3169) and BCG_3827 (Rv3768)	BadApple: 2 Alerts The compound has a relatively high cLogP (5.0) and was tested in various assays for PK and demonstrated moderate uptake, stability, and showed low clearance	95
	117	Heterocyclic N-oxide	Growing Mtb H37Rv (MIC90 = 1.10 μM) and dormant (hypoxic) Mtb H37Rv (MIC90 = 6.62 μM)		BadApple: 2 Alerts Complete sterilization in the lungs of female BALB/c mice at 200 mg/kg given daily for 5/7 days over a three week period.	96
	118	Hydantoin	MIC 8.3 μM	DprE1	pIC50 = 7.0, no toxicity to HepG2 cells up to 100 μM, and good kinetic aqueous solubility (202 μM)	97
	119	Coumestan	MIC90 = 0.125−0.25 μg/mL	Pks13	PAINS mannich_A(296) (SIT) assay by dosing BALB/c mice at 100 mg/kg via oral gavage and collecting serum at 30, 60, and 120 minutes. After processing the serum, the 30 minute cohort gave good inhibition of Mtb H37Rv in the in vitro MABA assay suggesting that the lead is orally available and that metabolism likely intervenes past the 30 min time point.	99
	120	Coumestan	0.0039 to 0.0156 μg/mL	Pks13	PAINS mannich_A(296) Significantly reduced toxicity against the four human-derived cell lines	98
	121	Riminophenazine	MIC of 0.016 μg/mL		PAINS quinone_A_ter(370) BadApple: 1 Alert 3–5 logs of CFU reduction in the lungs after 20 days in comparison to untreated control animals	100
	122	Spirocyclic	MIC90 = 0.06 μM	MmpL3	BadApple: 1 Alert At a 50 mg/kg dosing regimen, 123 gave a cidal response and a 4.2 log reduction in CFU in the lungs compared to untreated controls	101
	123	Biarylpyrazole imidazole	MIC of 6.25 μg/mL	CYP121A1	BadApple: 1 Alert The MIC was found to correlate with calculated logP values	102
	124	Hydroxypyrimidinone	MIC 4.7 μM	DprE1	A single dose of 10 mg/kg indicated poor absorption	103
	125	Phenyl-isoxazole	Not reported	Mptpb	BadApple: 1 Alert IC50 Mptpb = 2.98 μM at least 1 log reduction in bacterial burden in lungs and spleens of guinea pig.	104

Our combined decades of experience in drug discovery applied to tuberculosis spans several areas such as high throughput screening (HTS) of hundreds of thousands of compounds^77–79, medicinal chemistry that has led to two compounds currently in clinical trials for TB^{80, 81} and machine learning computational approaches to categorize and identify new antitubercular leads as well as developing a structural understanding of crucial features for antituberculosis activity.^82–84 Therefore, we will focus on some of these key areas to highlight molecules of interest. Writing this perspective from our different unique viewpoints has indicated that there is scope for more diversity in chemistry and approaches to attack Mtb.

It is apparent from the recent antitubercular leads that there is a very limited palate of chemical classes. Some of the structural classes shown in this perspective represent common chemistries that have been applied elsewhere for anti-infectives or other types of therapeutics. For example, we have included pyrrole, indole-2-carboxamide, pyrazole [1,5-a] pyridine-3 carboxamide, pyrazolepyridone, imidazole [1,2-a] pyridine and oxazolidinone (Table 3), ethylenediamine and dipiperidine, hydrazone (Table 4), indole, 1,3,4-oxadiazole, 1,3,4-thiadiazole, oxadiazole, tetrazole-5-thiol, thiazole, benzothiazole, thiadiazole (Table 5), riminophenazine, pyridinone, phthalimide, quinoline, 4-aminoquinolone (Table 6), formamidopyrimidine, pyrimidine, diarylpyrazole, tetrahydropyrazolopyrimidine, pyrimidine-azaindole, indole (Table 7), quinazolinone, oxadiazole-pyranopyridine, chromeno[3,2-с]pyridine, methylcoumarin, triazoloquinolone and benzothiazinone (Table 8), dispiropyrrolothiazole, spiro-pyrrolothiazoles, octylberberine, capuramycins, caprazamycins, spectinomycins and galactose linked nitroimidazoles (Table 9), isooxazoline, oxazole, Benzo[d]oxazole and triazole (Table 10), benzimidazole, indazole, phthalazine, nitrofuran and 3-aracylphthalide and oxoborol (Table 11).

Table 4.

Ethylenediamine, dipiperidine, hydrazone chemical classes. MmpL3 = Mycobacterial membrane protein Large 3. Compound assay promiscuity was also predicted with BadApple¹⁰⁶.

Molecule	ID	Chemical class	Activity	Target	Notes	References
	19	Ethylenediamine	MIC 0.16–0.64 μg/ml	MmpL3	Low oral bioavailability, efficacy in mouse model when combined with isoniazid and rifampicin. has been in Phase I, IIA and IIB clinical studies	130–137
	20	Dipiperidine	MIC: 4 μg/ml		identified in a high throughput screen of over 10,000 compounds. in vivo Mtb models and favourable pharmacological safety	138, 139
	21	Hydrazone	MIC − 0.1 μg/ml	inhibition of mycolic acid cyclopropanation	Thioacetazone is used in combination with isoniazid	140
	22	Hydrazone	MIC 3.1 μg/ml	inhibition of mycolic acid cyclopropanation	BadApple: 1 Alert Greater potency in a mouse model	141
	23	Hydrazone	MIC < 0.05 μg/ml	inhibition of mycolic acid cyclopropanation	greater potency in a mouse model	141

Table 5.

Indole, 1,3,4-oxadiazole, 1,3,4-thiadiazole, oxadiazole, tetrazole-5-thiol, thiazole, benzothiazole, thiadiazole chemical classes. PanC = pantothenate synthetase, PrrB and PrrA are a two-component regulatory system, DprE1 = decaprenylphosphoryl-β-d-ribose oxidase. PAINS were calculated as described by Lagorce et al.¹⁰⁵. Compound assay promiscuity was also predicted with BadApple.¹⁰⁶

Molecule	ID	Chemical class	Activity	Target	Notes	References
	24	Indole	MIC90 is 6.16 μM against MDR Mtb clinical isolates	PanC	Docking in PanC	142, 143
	25	1,3,4-oxadiazole	MIC90 < 0.25 μM also active against non-replicating Mtb strains (streptomycin-starved Mtb18b)	Nucleic acid biosynthesis	BadApple: 1 Alert Low cytotoxicity and non mutagenic	144
	26	1,3,4-thiadiazole	MIC90 < 0.25 μM also active against non-replicating Mtb strains (streptomycin-starved Mtb18b)	Nucleic acid biosynthesis	Low cytotoxicity and non mutagenic	144
	27	Oxadiazole	Addition of 25 μM BDM31343 to 0.1 μg ml−1 ethionamide abolished the growth of Mtb whereas 1 μg ml−1 ethionamide was necessary to obtain a similar growth inhibition in the absence of BDM31343	Allosteric inactivation of DNA-binding EthR activity	BadApple: 1 Alert A booster for ethionamide co-administration of ethionamide and BDM31343 to M. tuberculosis–infected mice for 3 weeks reduced the bacterial load throughout the entire treatment as efficiently as a three times–higher dose of ethionamide administered alone	145–147
	28	Tetrazole-5-thiol	MIC90 < 0.5 μM active and selective agents to combat MDR and XDR Mtb strains		Low cytotoxicity, might utilize the FAD systems	147
	29	1,3,4-oxadiazole	MIC90 < 0.5 μM active and selective agents to combat MDR and XDR Mtb strains		BadApple: 1 Alert Low cytotoxicity, might utilize the FAD systems	147
	30	Thiazole	MIC90 < 1.68 μM	PrrB and PrrA	BadApple: 2 Alerts solubility is 31 μM, mouse plasma protein binding is 17 % and mouse clearance is 170.4 ul/min/mg	148–150
	31	Thiazole	MIC90 of 7.1–12.0 μM	PrrB and PrrA	BadApple: 4 Alerts PAINS thiaz_ene_D_bis(8)	151
	32	Benzothiazole	MIC90 < 4.62 μM	DprE1	BadApple: 1 Alert Low-toxicity on the A549 human cell line (IC50 > 100 μM), negative Ames test and moderate CYP inhibition	152
	33	Thiadiazole	MIC90: 0.08–0.66 μM	Isocitrate lyase	BadApple: 1 Alert Target plays a key role in Mtb persistence (chronic and latent infection)	153, 154

Table 6.

Riminophenazine, pyridinone, phthalimide, quinoline, 4-aminoquinolone chemical classes. QcrB = The b subunit of cytochrome bcc complex, DprE1 = decaprenylphosphoryl-β-d-ribose oxidase. PAINS were calculated as described by Lagorce et al.¹⁰⁵ Compound assay promiscuity was also predicted with BadApple.¹⁰⁶

Molecule	ID	Chemical class	Activity	Target	Notes	References
	34	Riminophenazine	MIC 1 μg/ml	DNA intercalation	PAINS quinone_A_ter(370) The riminophenazines lack interactions with CYPs, are generally active against various mycobacterial infections (such as leprosy), but they often have poor solubility and drug-induced side effects. Cross resistance with other classes of tuberculosis drugs has not been observed so they have the potential to treat MDR-TB	155–157
	35	Riminophenazine	MIC 0.25 μg/ml	DNA intercalation	PAINS quinone_A_ter(370) Milder skin pigmentation side effects	155–157
	36	Riminophenazine	MIC 0.12 μg/ml	DNA intercalation	PAINS quinone_A_ter(370) Milder skin pigmentation side effects	155–157
	37	Pyridinone	MIC90 1.56 μg/mL is active against the MDR-TB strain	RNA polymerase	PAINS anil_di_alk_A(478) BadApple: 2 Alerts Half-life in human hepatic microsomes (14.4 h) and a slight inhibition of CYP3A4 and CYP2D6	158
	38	Phthalimide	MIC90 0.132 μM	QcrB	Identified by phenotypic screening of 6000 molecules	159
	39	Phthalimide	MIC90 0.065 μM	QcrB	Mouse pharmacokinetic studies showed high clearance and low plasma levels	159
	40	Quinoline	MIC90 0.05 μM	cytochrome bc1	BadApple: 1 Alert In vitro intrinsic clearance (Clint) was 14.3 ml/min/kg and T1/2 was 21.8 min	160
	41	Quinoline	MIC90 1 nM against MDR-TB clinical isolate	cytochrome bc1	BadApple: 1 Alert Clint (14.8 ml/min/kg and T1/2 19.4 min)	161, 162
	42	4-aminoquinolone	MIC90 of 0.2–3.12 μM	DprE1	A non-covalent inhibitor. Cmax of 4.9 μM, plasma clearance of 34.4 ml/min/kg and a T1/2 of 0.5 h	163

Table 7.

Formamidopyrimidine, pyrimidine, diarylpyrazole, tetrahydropyrazolopyrimidine, pyrimidine-azaindole, indole chemical classes. DprE1= decaprenylphosphoryl-β-d-ribose oxidase. Compound assay promiscuity was also predicted with BadApple,¹⁰⁶

Molecule	ID	Chemical class	Activity	Target	Notes	References
	43	Formamidopyrimidine	MIC90 ≤ 0.2 μg/ml		BadApple: 2 Alerts No toxicity in Vero cells	164, 165
	44	Quinolinyl pyrimidine	MIC 0.87 μg/ml	type II NADH-menaquinone oxidoreductase (Ndh)	BadApple: 4 Alerts Analogs with improved ADME properties were generated	166, 167
	45	Pyrimidine	MIC90 − 0.7 μg/ml and an identical activity versus an MDR strain		BadApple: 2 Alerts LD50 in mouse of 315 mg/kg	168
	46	Pyrimidine	MIC90 – 0.7 μg/ml and an identical activity versus an MDR strain		BadApple: 2 Alerts LD50 600 mg/kg	169
	47	Pyrimidine	MIC90 of 0.37 μM and activity against isoniazid- and rifampicin-resistant Mtb strains		BadApple: 2 Alerts LD50 45 mg/kg	170
	48	Diarylpyrazole	MIC90 0.125 μg/ml		BadApple: 3 Alerts	171
	49	Diarylpyrazole	MIC90 0.25 μg/ml		BadApple: 3 Alerts	171
	50	Tetrahydropyrazolopyrimidine	MIC 0.15 μM		High potency in animal studies after oral administration (100 mg/kg; once a day; therapeutic course - 28 days) leading to a decrease in lung CFU by 3.5 orders of magnitude	172
	51	Pyrimidine-azaindole	MIC90: 0.78–1.56 μM	DprE1	BadApple: 1 Alert (dose − 300 mg/kg) was also tested in vivo and found to reduce lung bacterial load by 1 log	173
	52	Indole	MIC90 14.3 μM And inhibited INH-resistant strains	DprE1	BadApple: 2 Alerts	174

Table 8.

quinazolinone, oxadiazole-pyranopyridine, chromeno[3,2-с]pyridine, methylcoumarin, triazoloquinolone, benzothiazinone chemical classes. AHAS = Acetohydroxyacid Synthases, DprE1 = decaprenylphosphoryl-β-d-ribose oxidase. Compound assay promiscuity was also predicted with BadApple.¹⁰⁶

Molecule	ID	Chemical class	Activity	Target		Notes	References
	53	Quinazolinone	MIC 2.5–10 mg/L Also had intracellular activity against XDR clinical isolates.	AHAS	Compound Identified by a virtual screen		175
	54	Oxadiazole-pyranopyridine	MIC 0.31 μg/ml				176
	55	Oxadiazole-pyranopyridine	MIC 0.73 μg/ml				176
	56	Oxadiazole-pyranopyridine	MIC 0.07 and 0.14 μg/ml against drug-susceptible M. tb and MDR-TB				177
	57	Chromeno[3,2-с]pyridine	MICs 0.22 and 0.07 μg/ml against drug-resistant Mtb and MDR-TB		BadApple: 1 Alert Reduced lung and spleen bacterial load by 1.11 and 2.94 orders of magnitude, respectively in vivo (dose − 25 mg/kg)		178
	58	Methylcoumarin	MIC 0.78 μg/ml		BadApple: 1 Alert		179
	59	Methylcoumarin	MIC 1.56 μg/ml				179
	60	Triazoloquinolone	MIC 0.5 μg/ml		Also active in macrophages		180
	61	Benzothiazinone	MIC 1ng/ml	DprE1	High activity against all the clinical isolates of Mtb (including MDR and XDR strains). Efficacy in the chronic mouse model at 37.5mg/kg		81, 181
	62	Benzothiazinone	MIC 1 ng/ml (2.3 nM) and 4 ng/ml (9.2 nM), respectively against Mtb H37Rv and M. smegmatis	DprE1	Has completed an open-labelled, dose-escalation phase I study in healthy male volunteers followed by a multiple ascending dose trial in 2016 (in Russia) and in 2018 (in Switzerland). This most recent study investigated the safety, tolerability and pharmacokinetics at doses up to 640 mg once daily for 14 days and revealed a good safety profile. In 2017, a phase IIa EBA study (monotherapy during 14 days) was initiated in drug sensitive-TB patients in Russia and Belarus. It also confirmed safety in drug sensitive-TB patients and statistically significant EBA after 14 days monotherapy in the group of patients treated with 640 mg of		182–184

Table 9.

Dispiropyrrolothiazole, spiro-pyrrolothiazoles, octylberberine, capuramycins, caprazamycins, spectinomycins, galactose linked nitroimidazoles chemical classes. MraY = phospho-MurNAc-pentapeptide translocase, WecA = phosphoglycosyltransferase, InhA – enoyl-ACP reductase. Compound assay promiscuity was also predicted with BadApple.¹⁰⁶

Molecule	ID	Chemical class	Activity	Target	Notes	References
	63	Dispiropyrrolothiazole	MIC 0.21 μM Also active against the isoniazid-resistant Mtb strain MIC 8.31 μM		Compound was not cytotoxic in Vero cells	185
	64	Spiro-pyrrolothiazoles	MIC − 0.07 μM		Reduced bacterial count in lungs and spleen by 1.30–3.73 orders of magnitude	186
	65	Octylberberine	MIC 0.125 μg/ml		Significant activity against RIF- and isoniazid-resistant Mtb strains. These quaternary lipophilic agents may not be ideal and show FtsZ activity which may be related to membrane disruption	187, 188
	66	Capuramycins	MIC 1.0 μg/ml	MraY	Selected in a library including more than 7000 analogs -natural nucleoside compound extracted from Streptomyces griseus - poor water solubility and fast elimination as well as P-glycoprotein-mediated efflux are limiting factors, poor in vivo activity in the mouse	189–193
	67	Caprazamycins	MIC 1.56 μg/ml Also active agains an MDR strain MIC 6.25 μg/ml	WecA	Extracted from Streptomyces sp. affects both non-fissile and fissile tuberculosis in vitro, so, it may be an efficient drug to treat latent TB	194–197
	68	Pyridomycin	MIC 0.4 μg/L	InhA	BadApple: 1 Alert From Dactylosporangium fulvum or Streptomyces pyridomyceticus enables killing of isoniazid-resistant Mtb clinical isolates that have katG mutations	198–200
	69	Spectinomycins	MIC 0.8 μg/L	Ribosome and ability to overcome intrinsic efflux mediated by the Mtb Rv1258c efflux pump	BadApple: 1 Alert Synergized with 11 drugs from six antibiotic classes (including clarithromycin, doxycycline and clindamycin) clarithromycin provided additional bacterial killing in a mouse model of acute tuberculosis infection, but not in a chronic infection model due to mismatched drug exposure	201–203
	70	Spectinomycins	MIC 1.6 μg/L	ribosome and ability to overcome intrinsic efflux mediated by the Mtb Rv1258c efflux pump	BadApple: 1 Alert Synergized with 11 drugs from six antibiotic classes (including clarithromycin, doxycycline and clindamycin) clarithromycin provided additional bacterial killing in a mouse model of acute tuberculosis infection, but not in a chronic infection model due to mismatched drug exposure	201, 202
	71	Glycosylated beta-amino acids	MIC 3.12 μg/ml	inhibit biosynthesis of mycobacterial cellular wall peptidoglycan and arabinogalactan	In vivo mouse studies showed toxicity at 50 mg/kg/day	204
	72	Galactose linked nitroimidazoles	MIC 1.56 μg/ml		Active under aerobic conditions	205
	73	Galactose linked nitroimidazoles	MIC 1.56 μg/ml		Active under aerobic conditions	205

Table 10.

Oxadiazole, isooxazoline, oxazole, Benzo[d]oxazole, triazole chemical classes. InhA = enoyl-ACP reductase. PAINS were calculated as described by Lagorce et al.¹⁰⁵ Compound assay promiscuity was also predicted with BadApple.¹⁰⁶

Molecule	ID	Chemical class	Activity	Target	Notes	References
	74	Oxadiazole	MIC 1.6 μg/ml		BadApple: 2 Alerts	206
	75	Oxadiazole	MIC 1.5 μg/ml		BadApple: 1 Alert	206
	76	Isoxazoline	MIC 0.0001 μg/ml		PAINS anil_di_alk_E(186) BadApple: 1 Alert	207
	77	Isoxazoline	MIC 0.00156 μg/ml		PAINS anil_di_alk_E(186) BadApple: 1 Alert	207
	78	Isoxazoline	MIC 0.0002 μg/ml		PAINS anil_di_alk_E(186) BadApple: 1 Alert 0.83 Log10 reduction in Mtb CFU in lung versus untreated controls	207
	79	Isoxazoline	MIC90 1.56 μg/ml		PAINS anil_di_alk_E(186)	207
	80	Oxazole	MIC90 − 8 μg/ml 1 μg/mL against XDR-TB and 2 μg/mL MDR-TB		BadApple: 1 Alert Original hit from a screen of 45,000 compounds tested against M. smegmatis	208
	81	Benzo[d]oxazole	MIC90 − 11.47 μM	InhA	BadApple: 1 Alert Active against a XDR-TB clinical isolate	209
	82	Triazole	MIC 2.50 μg/ml	InhA		210
	83	Triazole	MIC 2.50 μg/ml	InhA		210
	84	Triazole	MIC 0.50 μg/ml			211
	85	Triazole	MIC 0.25 μg/ml			211
	86	Triazole	MIC 1 μg/ml			212
	87	Triazole	MIC 1 μg/ml		BadApple: 1 Alert	212
	88	Triazole	MIC 2.5 μg/ml Similarly active against clinical isolates		PAINS imine_one_A(321) No cytotoxicity to human HCT116 and GM637 cells	212, 213
	89	Triazole	MIC 2.5 μg/ml Similarly active against clinical isolates		PAINS imine_one_A(321) No cytotoxicity to human HCT116 and GM637 cells	213

Table 11.

Benzimidazole, indazole, phthalazine, nitrofuran, 3-aracylphthalide, oxoborol chemical classes. FtsZ = Filamenting temperature-sensitive mutant Z, KasA = β-Ketoacyl ACP synthase I, LeuRS = leucyl-tRNA synthetase. PAINS were calculated as described by Lagorce et al.¹⁰⁵ Compound assay promiscuity was also predicted with BadApple.¹⁰⁶

Molecule	ID	Chemical class	Activity	Target	Notes	References
	90	Benzimidazole	MIC 4.3 μM	FtsZ	BadApple: 1 Alert	214
	91	Benzimidazole	MIC99 0.39 μg/mL	FtsZ	BadApple: 1 Alert in vivo equivalent to isoniazid in a murine tuberculosis model with a reported log10 reduction of 2.10 (i.p.) and 1.72 (p.o.) in the lungs and log10 reduction of 3.36 (i.p.) and 2.49 (p.o.) in the spleen. Computational modelling and QSAR optimization led to other analogs in this class and the suggestion that the trisubstituted benzimidazoles bind in the GTP site of FtsZ	215–217
	92	Benzimidazole	MIC 1.5 μg/ml			218
	93	Benzimidazole	MIC 3.1 μg/ml		PAINS quinone_A_ter(370)	218
	94	Benzimidazole	MIC90 − 0.75 μM		BadApple: 1 Alert	118
	95	Benzimidazole	MIC90 5 μM against MDR Mtb		BadApple: 1 Alert IC50 of 58 and 7.8 μM (newborn skin fibroblast cell lines (HDFn) and human epidermal keratinocyte precursors (HEK), respectively)	124
	96	Indazole	MIC 0.8 μM	KasA	Active in vivo in mice showing 2.4 −3.5 log reduction in chronic and acute mouse models	219, 220
	97	Phthalazine	MIC 1.4 μM		100–1800 times higher activity against isoniazid-resistant Mtb strains than isoniazid - mechanism of action related to mycolic acid biosynthesis inhibition	221
	98	Nitrofuran	MIC 0.19 μg/ml		BadApple: 2 Alerts Excellent selectivity index > 1800	222
	99	3-aracylphthalide	MIC 0.97 μg/ml		Lactones and ketones represent chemical alerts	223
	100	3-aracylphthalide	MIC 0.93 μg/ml		Lactones and ketones represent chemical alerts	223
	101	3-aracylphthalide	MIC 0.81 μg/ml		BadApple: 1 Alert Lactones and ketones represent chemical alerts	223
	102	3-aracylphthalide	MIC 1.24 μg/ml		BadApple: 1 Alert Lactones and ketones represent chemical alerts	223
	103	Oxoborol	MIC 0.08 μM	LeuRS	Highly specific (IC50 = 0.216 μM), as the IC50 for human cytoplasmic LeuRS is 140 μM	224

In terms of multiple structural classes inhibiting specific targets, this outcome is clearly revealed for pyroles, indole-2-carboxamide (Table 3), ethylenediamine (Table 4), which have been shown to inhibit MmpL3, pyrazolepyridone (Table 3) benzothiazole (Table 5), 4-aminoquinlone (Table 6), pyrimidine-azaindole, indole (Table 7), benzothiazinone (Table 8) which inhibit DprE1. While several other structural classes have evidence on their targets, such information is typically absent (Table 3–Table 11, compounds 8–125).

As examples using this perspective, we will specifically highlight some relatively recent publications or those that take different approaches to identify lead compounds (Table 12). For instance, a study described the triazine antitubercular 104⁶⁶ which was initially discovered using a machine learning model.⁸⁵ This compound was recently found to be activated by F₄₂₀H₂ and one or more nitroreductases in addition to Ddn. Formation of NO^⋅ and a des-nitro metabolite were identified.⁶⁶ An analog 105 was generated in an attempt to improve the in vivo pharmacokinetics (PK). Both 104 and 105 achieved their activity via release of intrabacterial NO^⋅ along with inhibition of InhA.⁶⁶ The thiophenecarboxamide 106 was found to be activated by EthA and then inhibited PyrG identifying the importance of metabolic activation.⁸⁶ A computational pharmacophore identified quinoxoline di N-oxides and a lead optimisation program produced several compounds such as 107.⁸⁷ An additional study used a Bayesian model to identify compounds, one of which 108 was pyrazolo[1,5-a]pyrimidine⁸⁵, a class of compounds that had also been identified by several others through a concurrent high throughput screening program. A third compound identified in a separate Bayesian model screen was an imidazole 109 which demonstrated reasonable in vitro activity. These computationally-derived hits are vastly outnumbered by those discovered by other approaches such as HTS.

For example, Song et al. have reported the structure activity relationship (SAR) of about 30 new analogs of an earlier thymidine-based hit against the Mtb thymidylate kinase (TMPK) ostensibly using ligand efficiency (LE) metrics.⁸⁸ LE ranged from 0.18 – 0.29 for all compounds. Compound 110 emerged as the lead active giving an IC50 of 0.95 μM versus Mtb TMPK and activity against whole cell Mtb H37Rv (MIC = 7.8 – 15.6 μM). Screening against Mtb H37Rv in the presence of verapamil, an efflux pump inhibitor, modestly improved activity (MIC with verapamil (40 μg/mL) = 3.9 – 7.8 μM) suggesting that efflux may play a role in the differential activity against Mtb TMPK and the whole cell MIC. A crystal structure of the lead compound with Mtb TMPK was also reported.

A series of indolyl azaspiroketal Mannich bases has been reported to show selective membrane permeabilization of Mtb membranes.⁸⁹ The lead agent 111 gave an MIC90 versus Mtb BCG of 0.8 μM and an IC50 against VERO cells of 29 μM for a modest selectivity index (SI) of 36 (SI commonly presented as IC50 VERO/MIC90 Mtb). Not surprisingly, the compound has a relatively high logP by medicinal chemistry standards (5.12) and demonstrates 50% hemolysis of human red blood cells at 64 μM. The lead demonstrated acceptable PK in mice and gave a modest reduction in CFU counts in the lungs (−0.75 log₁₀) and spleens (−0.84 log₁₀) of Mtb H37Rv infected mice when dosed at 100 mg/kg daily over 4 weeks (6 days per week, total of 24 doses).

Whole cell aerobic Mtb screening of the corporate Lilly screening deck produced a sub micromolar active in the morpholino-thiophene class.⁹⁰ Further assessment validated the active hit leading to structure-activity relationship (SAR) optimization and advanced screening. The minimal pharmacophore was the morpholino-thiophene and produced a new lead compound 112 with an MIC of 0.24 μM against Mtb H37Rv, VERO cell toxicity of 37 μM, and a resulting Selectivity Index (SI = VERO IC₅₀/Mtb MIC) of 154. Microsomal stability was improved over the initial hit but intrinsic clearance in mice using mouse liver microsomes remained significant at 8 mL/min/gram liver. Activity in a murine intratracheal infection model at 100 mg/kg for four days led to a modest reduction of CFU in the lung of 0.8 log compared to 2.8 log for moxifloxacin. Initial mechanism of action (MOA) studies suggested that QcrB, a subunit of the menaquinol cytochrome c oxidoreductase and part of the Mtb bc1-aa3-type cytochrome c oxidase complex was a potential target.

Mtb IMPDH has been the target of numerous drug discovery programs. Hedstrom et al. have pursued the SAR of the benzoxazole scaffold to produce a new lead inhibitor 113.⁹¹ This compound had showed potent inhibition of Mtb IMPDH (K_i,app = 18 nM) and good activity versus Mtb H37Rv (MIC = 1.0 μM). Additionally, the compound is active against conditional and knockdown strains easing current concerns that guanine salvage can protect Mtb bacilli in the host. The compound showed reasonable pharmacokinetics in mice, but there are issues with mouse microsomal stability and high serum binding suggesting that optimization of the class for targeting Mtb in vivo will be required.

The group of Spain et al. recently reported continuing studies developing the SAR of bis-substituted cyclams as antimycobacterial agents.⁹² One compound in the series 114 demonstrated modest activity against Mtb H37Rv (6.25 μg/mL) and showed improved thermodynamic solubility as the ZnCl₂ salt relative to other metal salt forms. The compound showed approximately a 3-fold reduction in M. marinum fluorescence/infection in the zebrafish embryo model using tdTomato-fluorescent M. marinum at a dose of 10 μM, refreshed every second day of infection. It remains unclear how effective the molecule is as the metal salt or whether it is just a delivery mechanism for Zn²⁺, a metal known to have antibacterial activity. Beyond the issue of stability of the metal salt to gut pH, these are relatively large (MW = 967.31 amu), charged species which might impact oral bioavailability and utility as drugs.

Several substituted phenyl-aminothiazoles were reported through the TAACF screens in 2009 as potent and relatively selective inhibitors versus M. tuberculosis. Other groups have explored the SAR of this class since that time, but no definitive target(s) have emerged. Azzali et al. recently published over 30 analogs designed to further develop SAR of the class leading to compound 115.⁹³ This lead had similar activity to highly related structures reported out of the TAACF screens, and the reported compound showed MIC90 values against several strains of Mtb in the range of 0.060 to 0.125 μg/mL with retention of activity against several established drug resistant Mtb strains. Furthermore, 115 was relatively selective for Mtb versus a mammalian cell line, HMDM, (IC50 = 53 μg/mL) verifying earlier toxicity reports for the class by the TAACF (VERO cells). The paper also includes metabolism (T_1/2 and intrinsic clearance) from human liver microsomes (T_1/2 = 16.1 ± 0.2 min and CL’_int = 38.8 mL/min/kg) and information relating to efflux in drug resistant Mtb strains suggesting that the compound is metabolically stable and apparently is not susceptible to drug efflux pumps. The paper does not address concerns regarding chelating ability of the class. This group has also described a similar compound that is antitubercular and inhibits efflux activity.⁹⁴

Wilson et al. report the SAR development of an HTS hit from the Biofocus DPI Softfocus Library that is in the 6‑dialkylaminopyrimidine carboxamide class.⁹⁵ A lead compound 116 was identified from some 50 analogs that demonstrated high activity against Mtb H37Rv (MIC99 = 1.3 μM) and comparable activity against several clinical strains. The compound has a relatively high cLogP (5.0) and was tested in various assays for PK and demonstrated moderate uptake, stability, and showed low clearance via intravenous administration despite poor microsomal stability, most likely resulting from extensive tissue distribution as indicated by a high volume of distribution (Vd = 21.8 L/kg). Early MOA studies using an affinity anchored analog and sepharose beads in a chemoproteomic approach suggested two potential novel targets of unknown function, BCG_3193 (Rv3169) and BCG_3827 (Rv3768).

A SAR was recently reported for approximately 20 heterocyclic N-oxides leading to advanced screening of a novel hybrid of INH and a benxo-oxadiazole-N-oxide, (E)-6-((2-isonicotinoylhydrazono)methyl) benzo[c][1,2,5]-oxadiazole 1-oxide.⁹⁶ Compound 117 shows good antitubercular activity against actively growing Mtb H37Rv (MIC90 = 1.1 μM) and dormant (hypoxic) Mtb H37Rv (MIC90 = 6.6 μM) including a small number of drug resistant strains (e.g. INH-R MIC90 = 8.6 μM, RMP-R MIC90 = 3.8 μM, BDQ-R MIC90 = 1.2 μM). While the agent was relatively nontoxic against MRC-5 cells with IC50 of 519.2 μM, it showed some toxicity versus HepG2 (IC50 = 16.0 μM). The active lead was further evaluated for mutagenicity using the micronucleus test in peripheral mouse blood reticulocytes, and there was no indication of genotoxicity up to a single dose of 500 mg/kg after 30 hours. Compound 117 was bactericidal in culture against Mtb H37Rv infection in the J774A.1, a macrophage cell line, at 23.86 μg/mL (4× MIC), 5.84 μg/mL (MIC), and 1.46 μg/mL (MIC/4), essentially sterilizing the culture at all doses. Extensive absorption, distribution, metabolism, excretion and toxicity (ADMET) properties including liver toxicity were determined for the lead resulting in screening in vivo against Mtb Erdman in an acute murine model of tuberculosis. Compound 117 also showed complete sterilization in the lungs of female BALB/c mice at 200 mg/kg given daily for 5/7 days over a three-week period.

Interest in the target DprE1 has led to a HTS screening campaign at GSK and the discovery of a novel hit in the hydantoin class 118.⁹⁷ In a recent report by Rogacki et al, the activity of the initial hit and an extensive SAR program are described.⁹⁷ The initial hit 118 showed potent inhibition of DprE1 (pIC50 = 7.0), good activity against Mtb H37Rv (8.3 μM) using the resazurin reduction assay with fluorescent readout, no toxicity to HepG2 cells up to 100 μM, and good kinetic aqueous solubility (202 μM) relative to the MIC. A relatively thorough SAR exploration of the scaffold resulting in over 100 analogs gave only modest improvements in activity and other medicinal chemistry parameters. Mechanistically, the compound is a reversible inhibitor of DprE1. The authors concluded that the compound class has few or no major liabilities, is generally characterized by good metabolic stability, has no appreciable cytotoxicity, and possesses a physicochemical profile acceptable for further development. Further optimization was reported as underway including in vivo experiments in a respiratory rodent model of Mtb.⁹⁷

Two recent publications present the discovery and further natural product-inspired SAR optimization of the coumestan scaffold. The first publication reports nearly 50 analogs resulting in the lead 119, demonstrating significant activity against Mtb H37Rv in the MABA assay (MIC90 = 0.125−0.25 μg/mL).⁹⁸ The compound showed relatively significant toxicity to VERO cells (IC50 = 4 μg/mL.), yielding a modest selectivity index - SI (IC50/MIC90) = 16–32. Interestingly, four other human derived cell lines [MRC-5 and HFL1 (human lung fibroblast cells), QSG-7701 (human liver cell), and HEK-293 (human embryonic kidney cell)] showed no toxicity up to 50 μg/mL. The lead was further evaluated in the Serum Inhibition Titration (SIT) assay by dosing BALB/c mice at 100 mg/kg via oral gavage and collecting serum at 30, 60, and 120 minutes. After processing the serum, the 30 min cohort gave good inhibition of Mtb H37Rv in the in vitro MABA assay suggesting that the lead is orally available, and that metabolism likely intervenes past the 30 min time point. Selection of resistant Mtb mutants and sequencing identified Polyketide Synthase 13 as a putative target. In a second paper⁹⁸, Zhang et al. continued optimization of the scaffold identifying 120 as a second generation lead that shows improved activity (MIC90 = 0.0039 μg/mL) against Mtb H37Rv and MIC90s in the range of 0.0039 to 0.0156 μg/mL against several drug resistant strains.⁹⁹ Compound 120 also demonstrated toxicity against VERO cells (IC50 = 4 μg/mL.), but, had significantly reduced toxicity against the four human-derived cell lines discussed previously. Compound 120 gave a good half-life (T_1/2 = 120.1 min) in human liver microsomes and showed 8-fold higher activity than INH in the SIT assay dosed at 100 mg/kg with serum taken at 30 min.⁹⁹

Clofazimine (CFZ) 34 is a highly active antitubercular agent, although the drug suffers from liabilities including high lipophilicity and fat solubility causing discoloration of fatty tissues in the patient. Zhang et al. describe optimization of the CFZ riminophenazine scaffold with the goal of reducing lipohilicity to manage these drug liabilities.¹⁰⁰ Several CFZ analogs gave comparable or better activity as compared to CFZ with significantly reduced lipophilicity (ClogP). Six compounds with potent activity against M. tuberculosis in vitro, low acute toxicity in mice, and excellent PK profiles were further screened in an acute mouse model of MDR-TB infection. All compounds were highly efficacious, yielding 3–5 logs of CFU reduction in the lungs after 20 days in comparison to untreated control animals.¹⁰⁰ Two compounds exhibited equal or better in vivo efficacy against MDR-TB compared to CFZ with the potential for reduced coloration in fatty tissues, and these analogs are being further evaluated for advanced development. For comparison, 121 gave an MIC of 0.016 μg/mL (CFZ 0.12 μg/mL), VERO cell toxicity of >64 μg/mL (CFZ 68.6 μg/mL), and a measured logP of 3.74 (CFZ 5.34).¹⁰⁰

Screening of the GSK compound library yielded a novel spirocyclic hit that was further optimized in a recent SAR program and made available through their open source program.¹⁰¹ This effort yielded an improved compound 122 with potent antitubercular activity (Mtb H37Rv MIC90 = 0.06 μM). Cell health for the SAR set was measured in HepG2 cells and is reported as a triple-readout (membrane permeability, mitochondrial potential, and nuclear morphology) phenotypic assay measuring changes in cell structure as a preliminary consequence of cytotoxic injury. The assay used automated imaging to measure IC50 values for the effect of compounds on human liver-derived, HepG2 cells (48 h incubation) as a surrogate of hepatotoxicity. For high doses (>100 mg/kg/day), numbers <80 μM raises alarms for hepatotoxicity. Compound 122 gave results 12.6 μM, 10.7 μM, and 11.0 μM respectively. Compound 122 also had high lipophilicity and poor solubility as well as activity in the hERG assay leading to cardiotoxicity concerns. PK results for 122 showed acceptable bioavailability with lower intrinsic clearance, a longer half-life, and oral AUC values suitable for progression into efficacy studies relative to other analogs. Compound 122 was further assayed in an acute murine model of Mtb – female CB57BL/6J mice infected with Mtb H37Rv intratracheally at ~10⁵ CFU. In a quick assay where the compound is administered for 8 days, the ED99 to give a 2-log reduction in lung CFU was reported to be 11.6 mg/kg. At a 50 mg/kg dosing regimen, 122 gave a cidal response and a 4.2 log reduction in CFU in the lungs compared to untreated controls.¹⁰¹ Further analoging showed modest effects on some of the liabilities but confirmed potential cardiotoxicity and tolerability issues leading to sidelining of this compound series. Mtb MmpL3 was proposed as a possible target either directly or through a related perturbation in the transmembrane proton gradient.¹⁰¹

Mtb cytochrome P450 CYP121A1 is essential for bacterial viability, and a recent paper by Taban et al. describes a small SAR study of three series of biarylpyrazole imidazoles and triazoles including antitubercular and Mtb CYP121A1 inhibition data.¹⁰² About 20 compounds were reported in the three series with MIC values ranging from 6.25 to >100 μg/mL. The −CH₂-imidazole series was the most active in terms of K_D for CYP121A1 and inhibition of Mtb H7Rv in the Resazurin Microtiter Assay. For example, 123 gave a K_D of 13.6 μM against CYP121A1 and a MIC of 6.25 μg/mL.¹⁰² In general, the MIC was found to correlate with calculated logP values. It is not clear regarding whether logP or logD is more relevant as the imidazole moiety is protonatable at physiological pH. Furthermore, there was no discussion relating to effects on human CYP enzymes or selectivity versus mammalian cells (e.g. VERO, HepG2 etc.). Co-crystal structures are reported for certain analogs with CYP121A1.¹⁰²

Oh et al. have reported a novel hydroxypyrimidinone hit 124 from the chemical library of the St. Jude Children’s Research Hospital that appears to target decaprenylphosphoryl-β-D-ribose 2-oxidase (DprE1, Rv3790).¹⁰³ The compound shows good activity against Mtb H37Rv in two media 7H9/ADC (MIC 4.7 μM) and GAST (MIC 0.4 μM) and no toxicity to HepG2 cells up to 100 μM. The hit was bactericidal to actively replicating Mtb in vitro, with a 1–2 log reduction in CFU units observed over 7 days with weak dose dependence over a wide range of concentrations.¹⁰³ Compound 124 did not show bactericidal activity against anaerobic, nonreplicating Mtb cells generated in the Wayne hypoxic Mtb model. PK profiling of 124 in C57BL/6 mice administered an oral suspension at a single dose of 10 mg/kg indicated poor absorption and in vivo efficacy against Mtb was not undertaken. Ex vivo activity against Mtb H37Rv in J774 murine macrophages gave apparent bactericidal activity. Nearly 20 analogs of the initial hit were reported with very little improvement in activity results. Furthermore, 124 is considerably less active than other reported DprE1 inhibitors that are progressing towards clinical use.¹⁰³

A structure-based program targeting Mycobacterium tuberculosis protein-tyrosine-phosphatase B (MptpB) was recently reported by Vickers et al. that resulted in a lead compound 125 (IC50 Mptpb = 2.98 μM).¹⁰⁴ Certain analogs showed high selectivity for the mycobacterial phosphatase relative to human counterparts and low toxicity up to 150 μM in the mouse J774 macrophage cell line. Compound 125 showed reasonable kinetic solubility (200 μM) and acceptable results in the PAMPA screen (78.1 nM/s), suggesting the potential for good cell penetration. Compound 125 also showed improved pharmacological properties relative to other analogs with reasonable orally bioavailability and a good PK profile. Since the target is only essential in vivo, it is not surprising that it has a very modest impact on the growth of H37Rv in vitro. However, at both 20 and 100 μM, the compound significantly impacts bacterial burden in the human THP-1 macrophage model against both MDR strain (Beijing-W) and the drug-susceptible M. tuberculosis H37Rv strains.¹⁰⁴ Furthermore, the compound at the low dose of 5 μM demonstrates compatibility in combination with rifampin and isoniazid (0.1 μg/mL for INH and 0.3 μg/mL for RIF) reducing bacterial burden by 93% in BCG infected J774 mouse macrophages. In vivo profiling of 125 in guinea pigs showed high exposure (C_max 112 μg/mL, AUC 230 μg·h/mL), long half-life (t_1/2 5 h), good oral availability, and appropriate tissue distribution in guinea pigs under either parental or oral dosing. Compound 125 was further tested for efficacy as monotherapy in the acute and chronic guinea pig models of TB (female Hartley Duncan). In the acute infection, animals were infected with approximately 96 CFU and, after 24 h, orally dosed once daily at 100 mg/kg for 4 weeks. Treatment resulted in a 0.9 log reduction of bacterial burden in the lungs relative to vehicle.¹⁰⁴ For the chronic infection, guinea pigs were infected with ca. 63 CFU, and 125 was orally administered daily at 100 mg/kg for 4 weeks starting at 28 days post infection. In this model, treatment with 125 resulted in at least 1 log reduction in bacterial burden in lungs and spleens.¹⁰⁴

MOLECULAR DIVERSITY OF RECENT COMPOUNDS FOR TB

Interestingly, only a relatively small number (17) of the 118 molecules in Tables 3–12 (representing 14%) are Pan Assay Interference compounds (PAINS) as assessed using the FAF-3 online software¹⁰⁵. However more compounds (62) are flagged with the BadApple software (representing 52% in Tables 3–12), which attempts to identify compounds with assay promiscuity.¹⁰⁶ There is generally good agreement with these two approaches. Previous comparisons of TB in vitro and in vivo datasets have determined the mean of molecular descriptors/properties across actives and inactives.^{226, 227} We have determined 8 simple molecular properties for the molecules in Table 2 (Table S1) and Tables 3–12 (Table S2). These data are remarkably similar to our previous analyses suggesting a mean molecular weight of 390.96 and ALogP for 2.84 for approved TB drugs and mean molecular weight 399.57 and ALogP of 3.66 for TB leads, which are comparable to in vivo actives described earlier with a mean molecular weight of 411 and ALogP of 3.11.²²⁶ These may represent property ranges to target for TB activity. A more extensive cheminformatics analysis is now described in order to assess the chemical property space coverage of these Mtb leads.

To assess the molecular diversity of the compounds active against Mtb described in this perspective, we have generated a principal component analysis (PCA) plot based on eight 2D descriptors which included molecular weight, topological polar surface area, number of aromatic rings, number of hydrogen bond acceptors, number of hydrogen bond donors, number of rotatable bonds, ring count and MolLogP. All these descriptor calculations were generated with manual python scripts using RDKit²²⁸, Pandas²²⁹, NumPy²³⁰ libraries and Jupyter Notebook.²³¹ Before conducting PCA, the explicit descriptors were imputed “NAN” and ensured there were no missing values; and scaled from zero to one for each feature using MinMaxScaler from scikit-learn.²³² We then conducted PCA using three sets of descriptors on (1) Mtb approved drugs (Table 1) and 118 Mtb leads (Tables 3–12), (2) 7762 active compounds (MIC < 10 μM) from a set of 18,886 compounds with Mtb data²³³ (Mtb actives) and 118 Mtb leads datasets, and (3) Mtb actives, 118 Mtb compound dataset, approved drugs and Microsource spectrum datasets. The PCA figures were generated using Matplotlib²³⁴ library and Jupyter Notebook.

PCA analysis of Mtb approved drugs (Table 1) and 118 Mtb leads (Tables 3–12) indicates coverage of the same property space (Figure 2A). When visualized using the chemical network visualization this may represent a more useful approach to grouping the molecules than the Tables 3–12. We have summarized several “outlier” molecules away from the main cluster shown on Figure 2A. These molecules are also shown in Table S2 and S3. What is apparent is that many of these “natural product like” such as 66, 67, 69, 70, 71 or very large 114 with generally low ALogP (Table S2). Many of these molecules cluster together (Figure 2B). Several of outliers are very lipophilic 19, 41, 84, 85, 94. Interestingly, the high ALogP outliers are to the left and the low ALogP molecules are to the right. As one would expect similar molecules group together such as 1 and 2. (Figure 2B).

Figure 2.

Data visualization of Mtb approved drugs and Mtb lead compounds in this review. A. The PCA generated with eight descriptors showed that Mtb approved drugs are overlapped with the 118 Mtb leads dataset; with the three components of PCA representing 82.8% of the variance (x = 41.7%, y = 28.5%, z = 12.6%). Molecule numbers represent those outside the main cluster with structures shown (see also Table S3). B. Chemical network visualization containing compounds from Table 1 (green) and Tables 3–12 (red). This was generated using the same eight descriptors, with unsupervised hierarchical clustering using the Euclidean distance method, Wards linkage. Numbering corresponds to molecules in Tables 1, 3–12.

We have compared these TB lead molecules to a recently curated set of 7762 active compounds (MIC < 10 μM) from a set of 18,886 compounds with Mtb data²³³ (Figure 3). The only TB lead that appears separate from the main cluster of TB Actives is 114 which is a Bis-Substituted Cyclam (Table S2 and S3) with a high molecular weight and low ALogP. This analysis shows that the compounds in this review are representative of the chemical property space of the thousands of Mtb “active” compounds published by many labs to date and when compared to approved drugs (Figure 4). Digging deeper into the outlier molecules as described herein suggests that the extremes of ALogP and molecular weight (and perhaps other simple descriptors) need to be further explored in detail, as most of the TB leads seem to be in the “sweet spot” of mean molecule properties that we have seen covered over the years with our various analyses.^{226, 227}

Figure 3.

Principal Component Analysis of compounds in this review (blue) alongside a curated set of Mtb actives²²⁸ (red, MIC < 10 μM threshold). The PCA generated with eight descriptors showed that Mtb actives are overlapped with the 118 Mtb compound dataset; with the three components of PCA representing 85.78% of the variance (x = 55.82%, y = 18.65%, z = 11.31%). Lead molecule 114 is clearly outside of the main cluster of TB actives (for structure see Table S3).

Figure 4.

Principal Component Analysis of 118 lead compounds in this review (blue) alongside a curated set of Mtb actives²²⁸ (red, MIC < 10 μM threshold), approved drugs (green) and Microsource spectrum (yellow) and we then generated a PCA fitted with eight descriptors of the merged frame including all four datasets. The three components explained a variation of 89.57% (x = 52.89%, y = 24.39%, z = 12.29%). A majority of the four datasets were found overlapped in the same PCA space.

While we can attempt to classify the lead compounds structurally, we can also organize them by each property (Figure 5). This type of circular dendrogram approach while showing a relatively close proximity in log P property space alone, clearly indicates very few low logP molecules amongst these Mtb leads, for example structurally similar natural products or analogs 66 and 67, 69 and 70. This would suggest the need to further explore natural product or natural product derivative property space. These various molecule visualization methods are far more realistic in grouping the molecules than by classes based on structure types.

Figure 5.

Circular dendrogram obtained from the hierarchical clustering of the 118 lead compounds for Mtb using eight descriptors, Ward linkage, and Euclidian distance. Compound nodes and names are colored according to their MolLogP.

MACHINE LEARNING FOR TB DRUG DISCOVERY

Alternative approaches to identify new leads for TB can learn from past data. We and others²³⁵ have published widely using the naïve Bayesian approach for tuberculosis drug discovery.^{227, 236}. Models that combined bioactivity and cytotoxicity data were used to rank compounds such as the GSK antimalarial dataset.²³⁷ From the top 46 molecules, seven were chosen and five had MIC ≤ 2 μg/mL, the most active 104 being 0.0625 μg/mL.²³⁸ A second example used two different Mtb whole cell models to score three vendor libraries from which 550 compounds were tested and 124 actives identified.²³⁹ A third example filtered a library of >150,000 molecules and tested 48 compounds of which 11 were active²²⁵ including 109. The models achieved screening hit rates of 15–71% for suggested compounds, far higher than the 0.6 – 1.5% typical for random library HTS screening.^{85, 239, 240} Fusion of three dual activity models gave an excellent ROC value with a fourth external dataset from the same laboratory.²⁴¹ These models have also been used individually with a test set of 1,924 molecules for which cytotoxicity was determined in three cell lines and enrichments of 11.8 fold were observed in the best case.²⁴² Fusing single point data (bioactivity only) with dual activity data ultimately led to models with 345,011 molecules in them but these were no more predictive that the smaller dual activity datasets when tested with external data.²⁴³ We have also used Bayesian and additional machine learning approaches to model data from the mouse Mtb efficacy model that have been published over the last 70 years. Models were initially constructed with 773 compounds and used to predict 11 molecules from the literature (eight were correctly predicted).²⁴⁴ More recently these models have been updated and used with a test set of 60 molecules. The best Mtb in vivo models in this case now possess 5-fold ROC values > 0.7, sensitivity > 80% and concordance > 60%. These results indicated Bayesian models using literature in vivo Mtb data generated by different laboratories in many different mouse models can be predictive and also be used alongside other models to select antitubercular compounds with in vivo efficacy.²⁴⁵ We have also recently curated the Mtb data and developed newer machine learning models with 18,886 molecules (with activity cut offs of 10 μM, 1 μM and 100 nM).²³³ These datasets were used to evaluate many different machine learning methods (including deep learning and support vector machines) and were assessed with additional molecules published in 2017. The 100 nM cutoff model was tested with an evaluation set (n = 153 compounds) and showed comparable statistics to those seen with 5-fold cross validation (Accuracy = 0.83, Precision = 0.27, Recall = 1.00, Specificity = 0.81, Kappa = 0.36, and MCC = 0.47). Bayesian machine learning models constructed with literature Mtb data generated by different laboratories generally were equivalent to or outperformed Deep Neural Networks with external test sets.²³³ There are also individual efforts that have developed large numbers of promising leads such as the HTS of two million compounds that resulted in the GSK set of 177 lead compounds.²⁴⁶ We have previously shown that our Bayesian models identified some of these compounds before they were even published when they were a subset of an antimalarial library (described earlier²³⁷) providing an excellent example of prospective prediction.²⁴⁷ This raises the larger question of whether these models are able to identify the GSK leads because the models generated with large numbers of compounds from the SRI/TAACF screens were of a broad diversity that they essentially cover Mtb property space. Alternatively, we may be interpolating from these compounds and did the GSK library just sample close analogs of the SRI/TAACF library? Based on analysis of the literature it is clear from this work that we are quite adept at rediscovering compounds from the past, finding close analogs or structural scaffolds from years previously. Finding truly novel original molecules that are active against Mtb (and for that matter any bacteria) from machine learning models is a challenge we need to explore further in future because it is still a work in progress.

CONCLUSION

The recent development of tuberculosis research and development programs has been fruitful in terms of generating thousands of preclinical leads. Over the past decade alone, TB drug discovery has advanced with the screening of large drug-like and pharmaceutical company libraries starting with the TAACF in 2009 screening over 300,000 compounds.^77–79 Several companies such as Novartis and GSK have screened even more compounds, in the many millions.^{172, 246} The GSK library phenotypic screens alone covered 2 million molecules.²⁴⁶ It is likely that in total well over five million compounds have been tested against Mtb by academia and big pharmaceutical companies. These chemical libraries likely have some overlap but nevertheless many chemical classes have been examined, and more unique lead compounds were discovered in this way. Therefore, there is now a very low chance of finding new compounds and we will have to switch to looking for compounds against known targets or explore other sources of molecules. It may also be a challenge to come up with designs of new compounds which can address drug resistance. An alternative approach is the design and development of compounds which do not kill bacteria but instead block their virulence.^{248, 249} These compounds likely will not have an MIC in in vitro testing but will have in vivo activity. This points to further testing of compounds in in vivo animal models which is certainly expensive or optimizing combinations of drugs using other biochemical approaches. These alternative approaches will also have issues in how one goes after them and the need for expensive downstream animal/combination studies.

It should be noted that academics and industry are not always truly using and acknowledging available public resources such as the earlier NIH screens, but are often times rediscovering known actives and only poorly improving on them. That includes the data on their toxicity, the drug-like properties of the scaffolds and general oral availability of these Mtb active molecules. Our earlier efforts with the Collaborative Drug Discovery database to make these Mtb data and machine learning models available represent some of the earliest attempts to help other research programs learn from this accumulated data and support open source drug discovery.^{83, 84} Mtb drug discovery needs a truly field-wide effort to succeed (it takes a village rather than isolated academics), and more needs to be done faster even still with all the issues of cost, compliance, drug resistance and many thousands of people still dying annually. The Global Alliance’s recent advances (their approval of pretomanid and general ownership of the TB pipeline) are worthy of highlighting especially as it has likely involved a massive investment in collaborative drug discovery and development. The GSK Open Source efforts (at a fraction of the Global Alliance’s budget) are also notable and should be lauded for their making hit and lead compounds accessible to the public as well as highlighting true modern pharmaceutical drug discovery research, particularly paradigms for advancement of compounds for in vivo and preclinical toxicity and pharmacokinetics assessment. Unfortunately, these many efforts have not yet led to new drugs and most of these resulting publications (some of which are included herein) represent molecules that did not perform well when they reached in vivo. One of our observations that derived from the efforts at collating leads for this publication was that many papers still appear on the same poor, toxic insoluble scaffolds without all the necessary minimal information to truly evaluate candidates (i.e. solubility, toxicity in mammalian cells etc.) that are now standards in antibacterial drug discovery. Still, the increasing academic and industry collaborations has tended to at least anecdotally improve the drug-likeness of the compounds that resulted. Unfortunately, our cheminformatics analysis suggests that still, the recent lead compounds occupy the same property space as known drugs (Figure 2) as well as the larger set of TB in vitro or in vivo actives (Figure 3 and 4). There are however, several “natural product like” or large molecules 66, 67, 69, 70, 71, 114 or others 19, 41, 84, 85, 94 that are at the extremes of the ALogP scale (Table S2) compared to the majority of the TB leads (Figure 2A) and they stand out from the rest. It may be worth exploring these outlier areas of property space further. Some of the TB lead molecules also appear to contain PAINS¹⁰⁵ and/or structural features that represent increased compound assay promiscuity.¹⁰⁶ These molecules may need to be avoided or deprioritized. It is likely that limited structural diversity within screening libraries is also limiting the target diversity covered, which might explain why we see so many compounds for the promiscuous targets such as Mmpl3 and DprE1. Alternatively, and a common perception, it is possible that Mtb drug discovery has unique challenges due to the cellular barriers and druggability of targets requiring extremely small molecules or larger lipophilic compounds to access and bind targets within the cell. While there are many issues with Mtb drug discovery still to be answered from the chemistry/target perspective, the question of why lipophilic sulfur-containing compounds are hits in many Mtb screens is also intriguing. Is this just a case of metabolic activation or is there something else involved?

In this perspective, we have reviewed many different classes of compounds representing respectable coverage of Mtb chemical property space that might suggest reasonable target diversity and a promising foundation for the development of new drugs that will address burgeoning Mtb drug-resistance in the clinic. This abundance of preclinical leads, many of which are active in mice and / or active versus MDR strains of Mtb provides some hope, although the current clinical trial pipeline remains limited (Figure 1). Other reviews have suggested there is a need for novel targets.⁷⁵ While the current perspective certainly confirms this finding, it is notable that there are also many compounds where the MOAs are unknown, and it is possible that unique and new targets may be hiding in plain sight. Perhaps a more critical observation is the generally limited variety of structural classes utilized and the general homogeneity of chemistry space. These current successes in generating new leads for Mtb have been driven primarily by funding from NIH/NIAID and BMGF as well as to some extent EU funding and to a lesser-extent, investments from pharmaceutical companies.⁵ While results from these programs are promising, there is still a dire need to develop more new leads in the hope that some of them will advance to the clinic for TB. More commitment to developing shorter treatment regimens are dependent on strong synergistic interactions between drugs and more screens for synergy are also needed. However, rarely does a compound with in vitro synergy translate to in vivo synergy²⁵⁰, more often than not the differences in ADME and PK properties of the two or more combined drugs will result in antagonism or no benefit²⁵⁰. Additionally, efforts to identify compounds affecting latent TB are of crucial importance to better target an essentially untreated population of bacilli in both actively diseased as well as latently infected patients²⁵¹. Certainly, the compatibility of new TB agents and antiretroviral drugs to treat HIV-infected TB patients is another key issue that has to be addressed.^{252, 253} The complexity of developing treatments for TB has led to the exploration of various chemical cores described herein that are outside traditional classes of antibacterials or antituberculars as shown in some of the nearly 118 lead compounds selected (Table 3–12). Furthermore, many perfectly good drugs are available that could be revisited for delivery via different routes. For example, a recent review has summarized prospects for using inhaled drug delivery to reinvigorate TB treatment.²⁵⁴ In this vein, there are a large number of published examples of leads that possess poor PK and efficacy when dosed orally (including those described in this review) and this set might serve as an excellent reservoir of new drugs using an alternative delivery route such as inhalation. After decades of relatively modest outcomes in Mtb drug discovery and development, we think it is time for a ‘breath of fresh air’ in this area.

While there have been some moderate successes in the field over the past decade, as researchers we are having to run faster just to stay in place. Clearly the investments to date have under-delivered but perhaps our expectations may be set unrealistically high. For example: should we move beyond relying on oral drug delivery, and do we always need to know the target for a new hit or lead? Hundreds of high-quality papers over the past decade that delve into the detailed mechanistic underpinnings of molecules for Mtb have not delivered more than a small fraction that are viable clinical candidates. How can we shift from having generations of TB scientists producing the same types of closely related molecules? We hope that this goes some way towards recognizing the need for fresh perspectives on Mtb drug discovery in order to answer some of these questions and lead us to another golden era in antibacterial drug discovery. The current pipeline is dominated by one organization, the TB Alliance. Maybe it is time for them and the community in general to rethink expectations of Mtb preclinical candidates and how we could perform drug discovery differently. We have presented evidence for how we can learn from the published Mtb data on lead compounds using computational models such as machine learning and cheminformatics analyses which may help us to explore new structural classes and property space. These types of approaches have been under-utilized in Mtb drug discovery likely due to a lack of funding and also acceptance, this may change with the many recent articles on machine learning. It is abundantly clear there is still a bias towards small molecules when there may be very large natural products (e.g. macrolactones etc.) that could have promising antitubercular activity. We certainly have not explored all of these large molecules to date. It is well overdue to do this and, as importantly, we need to better learn from the molecules and data that have accumulated over the years. It is not in our best interests to ignore the collective preceding chemistry for antituberculars, but we do need to shift to new areas of chemistry because we are continuously covering the same ground. This perspective represents our latest efforts in nearly a decade of analysis of small molecules for Mtb in order to provide fresh insights to improve our productivity and likelihood of success in future.

ABBREVIATIONS USED

ADMET

absorption, distribution, metabolism, excretion and toxicity

AHAS

Acetohydroxyacid Synthases

Ddn

deazaflavin-dependent nitroreductase

DprE1

decaprenylphosphoryl-β-d-ribose oxidase

FtsZ

Filamenting temperature-sensitive mutant Z

GSK

GlaxoSmithKline

hERG

human ether-a-gogo-related gene

HIV

human immunodeficiency virus

HTS

high throughput screening

IMPDH

inosine-5’-monophosphate dehydrogenase

InhA

enoyl-ACP reductase

KasA

β-Ketoacyl ACP synthase I

LeuRS

leucyl-tRNA synthetase

LE

Ligand Efficiency

MDR

multidrug resistant

MmpL3

Mycobacterial membrane protein Large 3

MOA

mechanism of action

Mptpb

protein tyrosine phosphatase

MraY

phospho-MurNAc-pentapeptide translocase

Mtb

Mycobacterium tuberculosis

NIAID

National Institute of Allergy and Infectious Diseases

NO

nitric oxide

PAINS

Pan Assay Interference compounds

PanC

pantothenate synthetase

PCA

principal component analysis

PK

pharmacokinetics

Pks13

polyketide synthase 13

PyrG

CTP synthetase

QcrB

The b subunit of cytochrome bcc complex

ROC

receiver operator characteristic

SAR

Structure Activity Relationship

SI

Selectivity Index, SRI/TAACF Southern Research International/Tuberculosis Antimicrobial Acquisition and Coordinating Facility

TB

Tuberculosis

TDR

totally drug resistant

WecA

phosphoglycosyltransferase

WHO

World Health Organisation

XDR

extensively drug resistant

Biographies

Biographies

Vadim Makarov is the Chief of Lab in the Dept. of Stresses of Microorganisms, A.N. Bach Institute of Biochemistry, Moscow, Russian Federation. He is an expert in medicinal chemistry and rational drug design with special interest in developing antimicrobial and antiviral agents. His recent work in drug development has focused on antituberculosis drugs and as result of his efforts in cooperation with several European scientific groups (NM4TB; LSHP-CT-2005–018923) a new class of synthetic organic compounds having high antituberculosis activity in vivo and in vitro was discovered (1,3-benzothiazinones). Two compounds from this group are under in-depth preclinical investigation, with PBTZ169 being the lead candidate. Dr Makarov’s team makes extensive use of advanced methods of medicinal chemistry, rational drug design and study of metabolic transformations.

Elena G. Salina received a M.Sc. degree from the Mendeleev University of Chemical Technology of Russia in 2002. Then she began graduate studies at the Bach Institute of Biochemistry of the Russian Academy of Sciences where she received a Ph.D. in Biochemistry under the direction of Professor A. Kaprelyants in 2006. She conducted postdoctoral studies at the same Institute working on M. tuberculosis persistence and modelling mycobacterial dormancy. Later she worked in the field of finding new drugs for latent tuberculosis at the Federal Research Center of Biotechnology of the Russian Academy of Sciences where she is currently a Senior Scientist. Dr. Elena G. Salina’s research interests concern the discovery of new drugs for the treatment of latent tuberculosis infection and multi-targeted non-selective inhibition.

Robert C. Reynolds received a B.S. in Chemistry from the University of Virginia in 1977 and a PhD. from Duke University in 1985. While a student at Duke University under Dr. Barbara Shaw in Chemistry and Dr. David Sedwick in the Department of Medicine. After postdoctoral fellowship at NIEHS in the Research Triangle Park with Dr. Robert London in the area of probe development for in vivo NMR metabolism, Dr. Reynolds began a 25-year career in drug design and synthesis at Southern Research Institute in Birmingham, Alabama reaching the position of Director of Drug Discovery Technology. His major areas of emphasis were new target and drug discovery against cancer and infectious diseases, particularly tuberculosis. In 2012, Dr. Reynolds joined the University of Alabama at Birmingham.

Phyo Phyo Kyaw Zin received her B.A. in Chemistry and Computer Science from Berea College. She is now a Ph.D. candidate in the Fourches lab at North Carolina State University and a recipient of the AAUW International Doctoral Fellowship and Olive Ruth Russell Scholarships. Her research areas involve molecular informatics, computer-aided drug design, development of cheminformatics approaches/software to generate virtual chemical libraries, analysis and visualization of big chemical databases, and implementation of AI-powered (ML such as DLNN) cheminformatics methods to build QSAR models.

Sean Ekins graduated from the University of Aberdeen; receiving his M.Sc., Ph.D. and D.Sc in Clinical Pharmacology. He was a postdoctoral fellow at Lilly Research Laboratories then worked as a senior scientist at Pfizer and Lilly Research Laboratories, Associate Director of Computational Drug Discovery at Concurrent Pharmaceuticals, Inc., Vice President of Computational Biology at GeneGo, Then worked for a number of companies, as CSO at Collaborative Drug Discovery and is currently CEO and Founder at Collaborations Pharmaceuticals, Inc. He is an Entrepreneur-in-residence at the Eshelman School of Pharmacy, UNC Chapel Hill. Sean has co-authored >300 peer reviewed scientific papers, book chapters and patents as well as edited/ co-edited five books. He has worked for over a decade on tuberculosis drug discovery using machine learning.