Tuberculosis: a quarter of humanity is infected, and almost none of them are ill

Start here: what tuberculosis is

Tuberculosis is caused by a single slow-growing bacterium, Mycobacterium tuberculosis, which travels between people through the air. Someone with active disease in their lungs coughs; the bacilli ride out on droplet nuclei small enough to stay aloft; someone else breathes them in. That is the whole transmission story, and it is why TB has followed crowding, poverty and confinement for as long as there have been cities.

What makes TB unlike almost any other infection is what happens next. In most people the immune system does not clear the bacterium and does not lose to it either. It walls it off. Macrophages and T cells assemble into an organised structure called a granuloma, the bacteria inside are contained but not necessarily killed, and the person feels entirely well — sometimes for the rest of their life, sometimes until decades later when something tips the balance and the infection breaks out. That contained state is latent tuberculosis infection, and it is the single most important fact about this disease ^{[1] [2] [3]}.

The scale of that containment is hard to take in. A 2016 re-estimation using mathematical modelling put the global burden of latent infection at 23.0% of the world population in 2014 — approximately 1.7 billion people ^[4]. Against that, roughly 10.8 million people developed active TB in 2023 ^[5]. Almost everybody infected with M. tuberculosis is not ill and never will be. The disease is what happens to the small remainder.

And yet it is one of the largest infectious killers on earth. Before COVID-19 displaced it, TB was the most common cause of death from a single infectious pathogen: an estimated 10.0 million people developed TB in 2019, with 1.2 million deaths among HIV-negative people and a further 208,000 among people living with HIV. Eight countries accounted for two thirds of the global total — India, Indonesia, China, the Philippines, Pakistan, Nigeria, Bangladesh and South Africa ^[6].

The binary is a simplification. "Latent" and "active" are useful clinical categories but poor biology. Some frequently exposed people appear resistant to infection altogether; some become infected and then eliminate the bacilli; active disease itself runs from mild to overwhelming, in the lungs or outside them. The field now describes a spectrum rather than two boxes ^{[7] [8]}.

Why it is in the news. Two reasons, developed below.

1. A vaccine that works in adults. The only licensed TB vaccine, BCG, is more than a century old and protects children against the most lethal forms while doing much less for adult pulmonary TB — the form that transmits. The candidate M72/AS01E showed 54.0% protection against progression to active pulmonary disease in infected adults in a phase 2b trial, and 49.7% at the three-year final analysis ^{[9] [10]}. That is the first adult efficacy signal of its kind, and it is now in phase 3. 2. Drug-resistant TB became curable in six months. In December 2022 the WHO recommended a 6-month all-oral BPaLM regimen — bedaquiline, pretomanid, linezolid and moxifloxacin — in place of the 9-to-18-month regimens previously used for multidrug- and rifampicin-resistant TB ^[11]. In the trial behind that change, unfavourable outcomes fell from 41% under standard care to 12% ^[12].

A note on the occupational connection. This review is a sibling to two others in this series. Breathing crystalline silica sharply raises the risk of tuberculosis — not only through silicosis but through the dust exposure itself, even in workers who never develop radiological silicosis ^{[13] [14]}. The relationship is quantified in Pillar 3, and it is the reason TB and the mineral-dust pneumoconioses have always been read together ^[15].

Three pillars follow — measurements, medicines, and progress — with a latent-reservoir model between the first two that explains why a disease almost nobody who carries it will ever develop is nevertheless so hard to eliminate.

Pillar 1: measurements and diagnosis

Two different questions, two different sets of tests

TB diagnostics answer two questions that are easy to confuse. Do you have active tuberculosis right now? is a microbiology question, answered by finding the organism. Have you ever been infected? is an immunology question, answered by finding the immune memory. No test answers both, and the failure to keep them apart causes real clinical harm.

Finding the organism: smear, culture, and the molecular shift

The oldest test is sputum smear microscopy — stain a sputum sample and look for acid-fast bacilli. It is cheap and fast and it misses a great deal, particularly in people with HIV and in children. Culture remains the reference standard, and it also produces the isolate needed for drug-susceptibility testing, but M. tuberculosis grows slowly enough that culture takes weeks.

The decisive change was a cartridge-based molecular assay. Xpert MTB/RIF detects M. tuberculosis DNA and, in the same run, mutations in rpoB that confer rifampicin resistance — a result in hours rather than weeks, at a level of the health system where no laboratory previously existed ^{[16] [17]}. Its measured accuracy, pooled across 86 studies and 42,091 participants, is a sensitivity of 85% and specificity of 98% for pulmonary TB; for rifampicin resistance, sensitivity 96% and specificity 98% ^[18].

Two caveats in that same meta-analysis matter more than the headline. Sensitivity was 98% in smear-positive but only 67% in smear-negative, culture-positive participants, and 88% in HIV-negative against 81% in HIV-positive participants ^[18]. The test is weakest exactly where the disease is hardest to find. The newer Xpert Ultra trades a little specificity for sensitivity — 88% versus 83% in a head-to-head study, with specificity 96% versus 98% ^[18] — and has been evaluated in the settings where the older assay struggled most, including tuberculous meningitis in HIV-infected adults ^[19]. Xpert has also been validated for extrapulmonary disease, where obtaining any specimen at all is the limiting step ^{[20] [21] [22] [23] [24]}.

Finding the immune memory: TST and IGRA

For latent infection there are two instruments: the century-old tuberculin skin test (TST) and the interferon-γ release assay (IGRA), a blood test measuring T-cell interferon-γ production in response to M. tuberculosis-specific antigens.

Their limits should be stated plainly, because they are routinely oversold. Both are indirect markers of exposure and both are imperfect. Neither can distinguish latent infection from active disease, neither can separate reactivation from reinfection, neither resolves the stages within the infection spectrum, both lose sensitivity in immunocompromised patients, and — the important one — both have low predictive value for progression to active TB ^[25]. A positive IGRA in a person from a high-burden setting tells you they have joined the 1.7 billion; it tells you very little about whether they are among the small fraction who will fall ill. The practical consequence is that latent-infection screening is worth doing only in people whose risk of progressing is high enough to justify treating them ^[25], a principle the low-burden-country guidelines are built on ^{[26] [27]}.

This is also why biomarker-guided approaches are being tested: a randomised trial used a blood transcriptomic risk signature to select who received preventive therapy rather than treating everyone with a positive IGRA ^[28], and host RNA expression has been evaluated for the notoriously difficult problem of diagnosing childhood TB ^[29].

Imaging, and machines that read it

Chest radiography finds disease that symptoms have not yet declared, and it remains the mainstay of TB screening. Its weakness is that it needs a reader. Three deep-learning systems evaluated against Xpert as reference in 1,196 outpatients in Nepal and Cameroon achieved areas under the curve of 0.92–0.94, and when matched to radiologist sensitivity had significantly higher specificity; used as a triage step, they could reduce the number of Xpert tests needed by 66% while holding sensitivity at 95% or better ^[30]. The same paper's warning is worth carrying: a single universal cutoff performed differently at each site, so the operating point has to be chosen for the population being screened. PET/CT, at the other end of the cost spectrum, has been used to study what infection actually looks like inside a body over time ^[31].

Drug-susceptibility testing, and reading resistance from the genome

Once TB is confirmed, the next question is which drugs will work. Phenotypic drug-susceptibility testing grows the isolate against each drug and is slow. Molecular testing reads the resistance mutations directly: rpoB for rifampicin, katG and inhA for isoniazid ^{[32] [33] [34]}. Whole-genome sequencing extends this to population surveillance, and has been used to trace how multidrug resistance emerged and spread across globally diverse strains ^{[35] [36]}. It is worth being clear that non-tuberculous mycobacteria are a separate problem that these tests must also distinguish ^[37].

Centerpiece: a simple simulatable model of the latent reservoir

Everything above sets up one arithmetic question. If 1.7 billion people carry this organism and only 10.8 million develop disease in a year, what rate connects the two — and what does that rate imply for anyone trying to end the epidemic?

Model one infected person as facing two routes. A primary route, front-loaded: disease within a few years of infection, conventionally defined as within five ^[38]. And a reactivation route: a small constant annual hazard acting for the rest of life on everyone the primary route did not take.

R(t) = p_fast · (1 − e^(−k·t)) + (1 − p_fast) · (1 − e^(−h·t))

The total is pinned by a grounded number: an infected person carries a 5–10% lifetime risk of progression into TB disease ^[39]. Taking the upper end, and requiring the model to reach exactly that at a 50-year horizon, fixes the reactivation hazard h rather than leaving it free. An independent modelling estimate of lifetime risk, from a completely different approach, puts the age-weighted average at approximately 12% ^[40] — close enough to confirm the anchor is not eccentric.

What the model explains. Four things.

First, why a positive test is such weak news. The reactivation hazard the model derives is about 0.1% per year. Told that they carry a lifelong infection, a person naturally hears a sentence; the arithmetic says something closer to a lottery ticket that mostly does not pay out. This is the same fact the IGRA literature states from the diagnostic side — low predictive value for progression ^[25] — arrived at from the natural-history side instead.

Second, why that tiny hazard is nevertheless the engine. Divide 10.8 million cases by 1.7 billion carriers and you get 0.64% per year, an upper bound on the average annual rate at which the reservoir converts to disease. Under one percent. That is what a global epidemic of ten million cases looks like from the reservoir's point of view — and it is why the epidemic cannot be switched off by treating the sick. The people who will produce next decade's cases are already infected and currently well.

Third, why elimination is slow. The source of the reservoir estimate ran exactly this counterfactual: assuming no additional infections at all from 2015 onwards, existing latent infections alone would still generate around 16.5 cases per 100,000 per year in 2035 and 8.3 in 2050 ^[4]. Perfect transmission control today leaves a two-decade tail. This is the quantitative case for the two interventions that act on the reservoir rather than on transmission — preventive therapy, and a vaccine that works in the already-infected.

Fourth, why the same infection is a different disease in different people. The reactivation hazard is not a constant of nature; it is a number that other conditions multiply. HIV raises it 20-fold ^[41], and the figure shows what that does: within the five-year primary window a person with HIV has already exceeded the entire HIV-negative lifetime risk. Silica exposure, diabetes, malnutrition and TNF-blocking drugs move the same dial in the same direction ^{[13] [42] [43]}.

What the model deliberately does not do. It does not include reinfection, which is a real and sometimes dominant route: exogenous reinfection has been shown to cause recurrent TB after curative treatment ^[44], and prior infection confers only partial protection against disease following reinfection — estimated at 16–41% in adolescents and adults ^[38], with a separate analysis putting the protection against progression following reinfection at around 79% ^[45]. It also treats the hazard as constant, when risk is highest immediately after infection and depends on age. And it says nothing about who transmits. The claim is about magnitudes and the shape of the tail, which is the part that is grounded.

Pillar 2: medicines and prevention

Treating active TB: why it takes months

Drug-sensitive TB is curable, and the regimen has been stable for decades: two months of isoniazid, rifampicin, pyrazinamide and ethambutol, followed by four months of isoniazid and rifampicin — six months in total, four drugs, taken daily ^[46].

The obvious question is why an ordinary bacterial infection needs half a year of chemotherapy. The answer is that a substantial fraction of the bacilli are not growing. Inside granulomas, and inside the lipid-rich foamy macrophages that accumulate there, M. tuberculosis enters a slow or non-replicating state in which most antibiotics — which target processes of active growth — barely touch it ^{[47] [48] [3]}. The long tail of the regimen exists to sterilise that population, and shortening it without relapse has been the central problem of TB therapy for fifty years.

Multiple drugs are used for a second reason: resistance. Any large bacterial population contains rare pre-existing resistant mutants to any single drug; combining four makes simultaneous resistance vanishingly unlikely. That logic is also why interrupted or partial treatment is so dangerous, and why adherence is not a side issue but the mechanism by which resistance is manufactured ^[49].

The regimen finally got shorter. Study 31/A5349, a phase 3 trial across 13 countries, tested two four-month rifapentine-based regimens against the standard six months. The regimen replacing rifampicin with rifapentine and ethambutol with moxifloxacin was non-inferior — 15.5% versus 14.6% unfavourable outcomes in the microbiologically eligible population — while the rifapentine-only substitution was not ^[50]. Earlier four-month moxifloxacin-based attempts had failed this test ^[51], and a patient-level pooled analysis of treatment-shortening trials mapped out which patients can safely be treated for less time ^[52].

Drug-resistant TB, and the drugs that changed it

Multidrug-resistant TB (MDR-TB) means resistance to at least isoniazid and rifampicin, the two most powerful first-line drugs; rifampicin-resistant TB (RR-TB) is treated the same way. Around 500,000 people develop rifampicin-resistant TB each year ^[12]. Until recently the treatment was 18–24 months of injectable and oral drugs with serious toxicity and poor success rates ^{[53] [54]}.

Three genuinely new drugs broke that. Bedaquiline is a diarylquinoline that inhibits mycobacterial ATP synthase — a target no antibiotic had used before, and one that works against dormant as well as replicating bacilli precisely because it collapses ATP homeostasis rather than blocking growth ^{[55] [56]}. Pretomanid and delamanid are nitroimidazoles ^{[57] [58]}. Together with linezolid they made an all-oral regimen possible for the first time — the first anti-TB drugs with novel mechanisms in more than fifty years ^[46].

The trial that settled it was TB-PRACTECAL, an open-label randomised non-inferiority trial in Uzbekistan, Belarus and South Africa. Comparing 24-week all-oral BPaLM against 36–80 weeks of standard care, unfavourable outcomes occurred in 12% of the BPaLM group versus 41% of the standard-care group — superior, not merely non-inferior — with grade 3 or higher adverse events in 23% versus 48% ^[12]. WHO adopted BPaLM in December 2022, replacing the 9-month and 18-month regimens for most MDR/RR-TB ^[11], and implementation experience has followed ^[59]. Cost-effectiveness analyses across South Africa, Georgia and the Philippines supported the switch ^{[60] [61] [62]}.

The obvious risk is resistance to the new drugs, and it is already appearing. Acquired bedaquiline resistance was described soon after introduction ^[63], mutations in pepQ confer low-level bedaquiline and clofazimine resistance ^[64], and an analysis across four clinical trials examined baseline and acquired resistance to all three of bedaquiline, linezolid and pretomanid and its effect on outcomes ^[65]. Balancing access against resistance is now an explicit policy problem ^[66]. Linezolid dosing in particular has to be managed against its toxicity ^{[67] [68]}.

Treating latent infection: acting on the reservoir

If the model's argument holds, the highest-leverage intervention is treating people who are not ill. Tuberculosis preventive treatment does exactly that.

The options have shortened in the same direction as active treatment. Six or nine months of daily isoniazid works but is completed by too few people. A network meta-analysis of 16 randomised trials covering 44,149 patients found that all the active regimens except nine months of isoniazid significantly reduced active TB relative to placebo, that differences between the active regimens were not significant, and — decisively — that regimens of 3–4 months were more likely to be completed ^[69]. Once-weekly isoniazid plus rifapentine for 12 weeks (3HP) is now a standard option ^{[70] [71]}, with its own tolerability profile ^[72] and favourable cost-effectiveness in high-burden settings ^{[73] [74]}. Current guidance is set out in the WHO low-burden-country guidelines ^[26] and the US recommendations ^{[27] [75]}, and preventive treatment in high-burden settings has been reviewed as its own strategic question ^{[76] [77] [78]}.

BCG, and its honest limits

BCG is a live attenuated Mycobacterium bovis strain, first used in 1921 and still the most widely given vaccine in the world. What it does and does not do is well quantified and frequently misstated.

A landmark meta-analysis of 14 prospective trials and 12 case-control studies found BCG reduced TB risk by about 50% overall — but protection against tuberculous death was 71% and against meningitis 64%, both markedly higher than against total TB cases ^[79]. Meta-analyses in newborns and infants confirmed strong protection against the disseminated childhood forms ^[80]. In children with known recent household exposure, BCG conferred 19% protection against becoming infected at all, 71% against active disease, and, among those infected, 58% protection against progression from infection to disease ^[81].

So: BCG is a good vaccine against a child dying of miliary TB or TB meningitis, and a poor one against the adult pulmonary disease that sustains transmission. Efficacy against adult pulmonary TB varies enormously by latitude — geographic latitude and study validity together explained 66% of the between-trial heterogeneity ^[79] — and the reasons are still argued over ^[82]. This gap is the entire rationale for the vaccine work in Pillar 3.

Pillar 3: progress

M72/AS01E: the first adult efficacy signal in a century

The M72/AS01E candidate is a recombinant fusion protein of two M. tuberculosis antigens combined with the AS01E adjuvant. Its development ran through dose-finding and safety studies in PPD-positive adults, adolescents in an endemic setting, and people with active TB ^{[83] [84] [85]}, with immunogenicity confirmed in meta-analysis ^[86].

The phase 2b trial enrolled adults aged 18–50 with M. tuberculosis infection defined by a positive interferon-γ release assay and no evidence of active disease, at centres in Kenya, South Africa and Zambia — 3,575 participants randomised to two doses of vaccine or placebo a month apart. The interim analysis reported 54.0% protection against bacteriologically confirmed pulmonary TB in infected adults, without evident safety concerns ^[9]. At the three-year final analysis, 13 of 1,626 vaccine recipients versus 26 of 1,663 placebo recipients had cases meeting the primary definition — incidence 0.3 versus 0.6 per 100 person-years — giving vaccine efficacy of 49.7% at month 36, with M72-specific antibody concentrations and CD4+ T-cell frequencies sustained throughout, and serious adverse events at similar frequencies in both groups ^[10].

Two things about that result matter more than the percentage. It was measured in people already infected — that is, in the reservoir — which is precisely the population the model above identifies as the target. And a roughly 50% reduction in progression, applied to a pool of 1.7 billion, is a very different quantity from a 50% reduction in an ordinary incident infection. The candidate has since moved into phase 3.

It is not the only route being tried. BCG revaccination in a phase 2 trial reduced sustained IGRA conversion, an infection endpoint rather than a disease endpoint ^{[87] [88]}, and subunit candidates including H4:IC31 and H56:IC31 have been compared head-to-head ^{[89] [90]}. The broader pipeline and the immunological questions underneath it have been reviewed repeatedly ^{[91] [92] [93]}, and the sobering non-human-primate result that boosting BCG with proteins or rAd5 did not enhance protection is part of the same record ^[94].

The occupational multiplier: silica and TB

This is where the review joins its siblings. Silica does not merely coexist with TB; it multiplies it.

A cohort of 2,255 white South African gold miners was followed from 1968–71 to the end of 1995, accumulating 39,319 person-years and 115 cases of pulmonary TB. Radiologically diagnosed silicosis carried an adjusted rate ratio for pulmonary TB of 3.96 (95% CI 2.59–6.06). But the more consequential finding is what happened in the miners without radiological silicosis: their TB risk still rose with cumulative dust exposure, with an adjusted rate ratio of 1.10 per mg/m³-year and a fourfold gradient across exposure quartiles (1.00, 1.46, 2.67, 4.01). Even at necropsy, fewer than five silicotic nodules — a burden invisible on radiography — was associated with increased TB risk. TB was diagnosed on average 7.6 years after dust exposure ended ^[13].

That is a dose-response for a hazard that surveillance based on chest radiographs will systematically miss, persisting after the job ends. Silica-associated TB remains high on the list of occupational health priorities in low-income countries, and the HIV epidemic amplifies it ^[14]; in South African miners, silicosis and HIV together produced far more mycobacterial disease than either alone ^[95]. The same relationship sits inside the broader picture of silica-related disease ^{[96] [97] [98]}.

The mechanistic reading is consistent with the model: silica-damaged macrophages are the very cells that maintain the granuloma. An exposure that degrades containment raises the reactivation hazard — the multiplier in the figure — rather than raising the chance of being infected in the first place.

TB and HIV

HIV is the most powerful known risk factor for progression, increasing the risk of latent TB reactivation 20-fold, and TB is the most common cause of AIDS-related death; the two pathogens act in synergy, accelerating immune decline ^{[41] [99]}. This is why TB diagnostics that lose sensitivity in HIV-positive patients ^[18] are a compounding failure, and why the 208,000 TB deaths among people living with HIV in 2019 ^[6] are counted separately.

What is still missing

The drug pipeline continues, with new molecules and regimens against drug-resistant disease in development ^{[100] [101] [102] [103] [104] [105]}, and dormancy and resuscitation biology being mapped for exactly the non-replicating population that makes treatment long ^{[106] [107]}. Host-directed approaches, including vitamin D supplementation for prevention of infection and disease, have been trialled with mixed results ^[108].

Three structural gaps remain. There is still no test that predicts progression — the biomarker problem the IGRA literature names explicitly ^[25], and the reason risk-signature-guided preventive therapy is being tested ^[28]. Children remain under-diagnosed and under-treated: only 30% of a 3.5 million five-year target for children treated for TB was met ^{[6] [109]}. And cure is not the end of the illness — initial and recurrent pulmonary TB leave chronic lung-function impairment behind ^[110], which is why post-TB wellbeing is now argued to belong inside national programmes ^[6]. Elimination frameworks for low-incidence countries ^[111] and the effect of migration on TB epidemiology in them ^[112] are the other half of a picture whose centre of gravity remains the reservoir.