Introduction

Tuberculosis (TB), caused by Mycobacterium tuberculosis, is a significant global public health problem, especially with the rise of the HIV pandemic. Among the one-third of the world infected by M. tuberculosis, almost 8 million incident cases of TB are diagnosed annually, and 2 million deaths are attributed to the disease each year.

Support for genetic susceptibility to TB in humans was first provided by twin studies, which have estimated that the concordance rate of TB is higher in monozygotic than in dizygotic twins (Comstock 1978; Kallmann and Reisner 1943). Differences in susceptibility between racial groups have been demonstrated and further support a genetic component; macrophages in blacks are more permissive to tuberculosis infection than those in whites (Crowle and Elkins 1990). A segregation analysis by Shaw et al. (1997) suggested that a general two-locus model was marginally favored over a single-gene codominant model for TB. Several animal models have also provided evidence for a role of genetic factors in TB susceptibility (Blackwell et al. 1995; Flynn et al. 1995; Kramnik et al. 2000; Lurie 1941). Furthermore, candidate gene studies have been conducted in a variety of world populations, though the results have been equivocal (Bellamy 2003).

As discussed in our previous work (Stein et al. 2003), one inherent complication in previous genetic studies of human TB has been the definition of the disease phenotype; as stated by Möller et al. (2010), studies of TB are “exquisitely sensitive to phenotype definition.” TB is heterogeneous in its presentation, severity, and duration. TB most often presents with pulmonary disease, but can also affect nearly all organ systems. Even for pulmonary TB, the presentation may differ among people with progressive primary disease and those with reactivation of a latent form of infection. Since TB is such a heterogeneous disease, it is difficult to define a reliable phenotype for genetic analysis. As we discuss below, diagnostic criteria for TB vary considerably based on available resources. Moreover, TB may develop at any time after infection, and active disease may not present during the data collection phase of the study but rather develop years later. The objective of this review is to examine closely the spectrum of criteria that have been used for clinical characterization of subjects in TB genetics studies and illustrate how these varying criteria may affect study results.

Furthermore, complex traits are characterized by polygenic effects, gene–gene and gene–environment interactions, unclear phenotype definitions, pleiotropy, and genetic heterogeneity. In addition to the aforementioned inconsistency in phenotype definition, many of these genetic epidemiological principles also apply to TB; throughout this review we illustrate examples in the current literature and emphasize the need for ongoing study.

Impact of phenotype definition

TB natural history and disease: factors affecting diagnosis

The pathogenesis of TB can be thought of as a two-stage process (Comstock 1982). In the first stage, exposed individuals can acquire latent M. tuberculosis infection (LTBI), in which M. tuberculosis establishes a productive infection but does not produce symptoms. LTBI is diagnosed by a positive tuberculin skin test (TST) and/or positive interferon-γ sresponse assay (IGRA) and the absence of clinical signs and symptoms of full-blown disease (ATS/CDC 2000; Nyendak et al. 2008). Approximately 10% of individuals with LTBI progress to develop active TB disease, which is characterized by growth of M. tuberculosis in sputum and cultivation in culture or positive acid-fast bacilli (AFB) smear, plus characteristic radiological signs on chest X-ray and hallmark symptoms, including persistent productive cough, fever, and weight loss (ATS/CDC 2000; Garay 2004).

Some studies have provided evidence that these different stages of TB natural history and disease have different genetic influences. Two whole-genome scans (Cobat et al. 2009; Stein et al. 2008) and a candidate gene study (Thye et al. 2009) examined traits related to LTBI and found distinct regions linked to these phenotypes compared to those linked to TB disease. Other studies (Flores-Villanueva et al. 2005; Motsinger-Reif et al. 2010; Stein et al. 2007) have contrasted TB cases to LTBI controls and concluded that observed genetic associations were related to progression from LTBI to active TB disease. However, these studies are in the minority. There are two issues that make the examination of TB genetics literature difficult: (1) characterization of controls and (2) diagnosis of TB. These two issues relate to the two stages of TB pathogenesis, respectively, in which we go into detail below.

In case–control studies, controls should be similar to cases in every way, only nondiseased. This is why case–control studies can be used to establish a relationship between a (genetic) risk factor and disease. For studies of TB, this means that controls should have had the opportunity to become cases and were therefore exposed to other infectious TB cases. The difficulty with many published reports is that there is no characterization of exposure in controls. Some studies assume that in TB-endemic communities, all individuals are exposed (Hoal et al. 2004; Taype et al. 2006), thereby assuming that all individuals who do not have clinical TB disease are latently infected (LTBI). However, other research has demonstrated that individuals may be persistently exposed to M. tuberculosis but never develop LTBI (ATS/CDC 2000; Stein et al. 2008). Some studies characterize exposure in controls by conducting TST on unaffected individuals (e.g., Dubaniewicz et al. 2005; Flores-Villanueva et al. 2005; Motsinger-Reif et al. 2010). Other studies utilize unaffected household members as controls (Cervino et al. 2000; El Baghdadi et al. 2003; Stein et al. 2007; Velez et al. 2009a). However, many other studies have used population controls, but this design runs the danger of possible misclassification bias (Edwards et al. 2005).

These issues have an impact on the interpretation of results, depending on study design. If a study used population controls and it is not known whether these individuals have LTBI, a positive association between a polymorphism and TB could mean that the polymorphism is associated with LTBI (since all TB cases must first develop LTBI) or progression to active TB disease. A second disadvantage is that individuals who are never exposed to infectious TB cases may actually carry the risk polymorphism, but because they are unexposed, they never develop TB. In this case, misclassification bias results in a bias toward the null hypothesis, and this could be problematic when the disease is common (McCarthy et al. 2008).

Second is the issue of TB diagnosis, which varies widely by study site. The gold standard for TB diagnosis is growth of M. tuberculosis in culture (ATS/CDC 2000). Many research sites lack facilities available for culture confirmation of TB and thus base diagnosis of TB on a positive AFB smear. Research has shown that an AFB smear is less sensitive than culture and that the AFB smear grade could reflect differences in disease severity (Garay 2004). Furthermore, smear-negative, culture-positive TB is a problem in developing countries because of HIV coinfection (Garay 2004). Because diagnostic criteria for TB vary by study, this could reflect underlying differences in disease severity. If host genes influence severity of TB, these differences in diagnostic criteria become very important. Furthermore, because sensitivity and specificity differ between culture and AFB smear, the problem of potential misclassification arises, which could introduce further bias into genetic studies.

Another factor related to differences in TB diagnosis is the different M. tuberculosis strains. Researchers have categorized M. tuberculosis into six main strain lineages that are associated with particular geographical regions (Gagneux and Small 2007) as well as clinical presentation (Thwaites et al. 2008) and rate of progression to active TB disease (de Jong et al. 2008). Therefore, not only do different diagnostic criteria potentially reflect differences in disease severity, as discussed above, but specific M. tuberculosis strains may also influence disease severity. The potential impact of strain lineage on genetic epidemiological studies was demonstrated by a recent study that observed an interaction between host genotype and M. tuberculosis genotype whereby the TLR2 genotype is associated with TB caused by the Beijing strain (Caws et al. 2008).

Lastly, another contributing factor to the differential diagnosis, clinical presentation, and pathogenesis of TB is age of onset. Diagnosis of TB in children is complicated by the absence of positive M. tuberculosis cultures in the vast majority of cases (Lewinsohn et al. 2004) and is often delayed (Iriso et al. 2005). Furthermore, children may be asymptomatic (Iriso et al. 2005). The immune response is different in young children versus adults (Lewinsohn et al. 2008). An association study of NRAMP1 in pediatric TB cases suggests that this gene may be associated with primary TB disease as opposed to reactivation disease (Malik et al. 2005). Leung et al. (2007) also found a more significant association between NRAMP1 and TB in younger individuals, with age cutoff of 65 years and youner. Besides these few studies, the confounding influence of age of onset on TB genetics studies has not been explored, but it may have an impact on the heterogeneity among published studies.

Studies of NRAMP1 as examples

So far we have discussed how phenotype definition matters in the analysis of genetic susceptibility to TB. Next we illustrate these issues using studies of the NRAMP1 gene, aka SLC11A1, as examples. Table 1 summarizes each publication by population, diagnosis method used, information provided about study controls, and the most significant p value observed for any polymorphism within the gene. Note that this table includes only those studies that focused on pulmonary TB in all age groups.

Table 1 Summary of TB association genetic studies of NRAMP1/SLC11A1 a

First, we can see how different studies diagnosed TB. Some studies used the gold-standard definition of growth of M. tuberculosis in culture, while other studies used only positive AFB smear. Other studies had more heterogeneous criteria; for example, many studies diagnosed TB based on growth in culture or smear, whereas others also included individuals with response to TB treatment or evidence of TB disease on chest X-ray. Although they did not conduct an association analysis of NRAMP1, another approach to classifying TB cases is illustrated by the first genome-wide linkage scan of TB (Bellamy et al. 2000), where patients diagnosed based on positive AFB smear were categorized together with individuals with a history of TB treatment. One problem in using TB treatment as a diagnostic criterion for TB disease is that some research sites may be unable to distinguish TB from disease due to nontuberculous mycobacteria, but standard TB therapy cures both diseases. It is also possible that some research sites “overtreat” cases to make sure that patients recover from whatever ails them. Furthermore, some studies included both pulmonary and extrapulmonary TB and analyzed them together in the analysis (Ma et al. 2003; Motsinger-Reif et al. 2010; Rossouw et al. 2003; Taype et al. 2006; Velez et al. 2009a). Thus, the classification of TB cases is not comparable across studies.

Second, we can see the wide variety of designs used to categorize individuals as controls. Some studies utilized unaffected individuals within households who were clearly exposed to TB and therefore had the opportunity to develop disease but did not. Other studies conducted TST in unaffected individuals. Not only does a positive TST indicate past exposure to a TB case, it also allows the authors to deduce that the genetic association observed is with progression to active disease. However, the clinical status of other controls is unclear. For example, many studies utilized blood donors, so nothing is known about the past exposure or LTBI status of these individuals. Other studies recruited individuals presenting at clinics for diseases other than TB, so again, both past exposure and LTBI classification are unknown.

Notice that of the 12 studies that demonstrated an association between NRAMP1 and TB, only four used culture positivity as their diagnosis method. Also note that only one of the NRAMP1 associations has been observed in studies in which exposure has been characterized in some way, by either utilizing household members as controls or evaluating TST in nondiseased individuals. To investigate differences in a more quantitative fashion, we used Fisher’s method for combining p values (Fisher 1950). When we combined p values, we looked for the most significant p value for any marker within the gene. This answers the question of whether the gene is associated with TB at all; as we point out later, with the variation seen in polymorphisms analyzed in this gene, this may actually be the most appropriate question. Although not nearly as rigorous as a full meta-analysis, this method provides a quick look at these different groupings of studies. We do not present a proper meta-analysis here because of the various stratifications we are interested in; we refer the interested reader to published meta-analysis of this gene (Li et al. 2006) for additional insight.

First, to assess the impact of the TB diagnosis method used, we grouped studies with culture positivity as the only diagnostic criterion versus all other studies. There was a significant association with TB in both categories (p = 0.0008 and p < 10−5, respectively), so it does not appear that the TB diagnosis method has a significant impact on the overall conclusion of association between TB and NRAMP1.

Second, we analyzed the impact of characterization of controls on association results. Here, we grouped together studies in which controls had been clinically characterized, in terms of either exposure to an infectious TB case or TST (Cervino et al. 2000; Dubaniewicz et al. 2005; El Baghdadi et al. 2003; Motsinger-Reif et al. 2010; Velez et al. 2009a). When p values across these studies were combined, the result was still statistically significant (p = 0.012), although it must be interpreted with caution. Because a p value was not reported by Motsinger-Reif et al. (2010), we conservatively inferred a nonsignificant p value of 0.055. However, if we assumed a higher p value of, say, 0.3, the pooled p value was less significant (p = 0.04). If the studies were restricted to those with family-based controls, the result is significant, but not overwhelmingly so (p = 0.014). When we consider that these gene-centric analyses utilized the smallest p value across the polymorphisms analyzed for NRAMP1, one might argue that the more appropriate significant cutoff would account for multiple comparisons. In this case, if we consider that most studies analyzed four polymorphisms within NRAMP1, a conservative Bonferroni-like significance threshold would be α* = 0.05/4 = 0.0125; then, the latter analysis that restricted studies to those with family-based controls would not show a statistically significant association between TB and NRAMP1. By contrast, the combined p value in the remaining studies was highly significant (p < 10−10). Does this contrast suggest an influence of misclassification of controls because of undocumented exposure? There are not enough studies to firmly conclude this. In addition, there are not enough studies with documented TST to conclude whether associations with NRAMP1 are with progression from LTBI to active TB disease or simply susceptibility to LTBI.

Third, we examined the potential impact of the M. tuberculosis strain. Using the global phylogeography of M. tuberculosis strain lineage presented by Gagneux et al. (2006), we grouped studies by the M. tuberculosis lineage found in the study population (Table 1). Malawi was not studied by Gagneux et al., so we conservatively assumed that the Indo-Oceanic (present in Tanzania), East African (present in Tanzania and South Africa), and Euro-American (in South Africa) lineages were prevalent there. Note that there was a great ecological simplification made here; we assumed that all of the cases analyzed in articles from in a given geographic region were caused by the same strain(s). However, it presents an interesting hypothesis that genetic association studies of NRAMP1 may differ based on the endemic M. tuberculosis strain. It is not trivial to analyze these data because many populations were infected by more than one lineage, and populations that have unique combinations (e.g., Tanzania has both Indo-Oceanic and East-African-Indian lineages) cannot be pooled with others. Studies conducted in populations with the East-Asian strain were highly statistically significant (p < 10−5). When populations infected by Euro-American and West-African 2 strains were analyzed together, there also was a statistically significant association (p < 10−4). When populations with only the Euro-American strain were analyzed, the result was less significant (p = 0.033), although strong conclusions cannot be drawn about this. Again, if we apply the adjustment for multiple comparisons described above (α* = 0.0125), then this analysis of populations with Euro-American lineage were not statistically significant. Thus, we do not have evidence that there is a strain lineage by NRAMP1 interaction, although specific studies investigating this hypothesis are warranted. These results must be interpreted with caution because our assumption that the results of Gagneux et al. (2006) adequately capture the variation within these study populations is likely an oversimplification. Indeed, recent research has suggested that the distribution of the bacterial population structure has changed significantly over time (van der Spuy et al. 2009), so it is possible that strain diversity in study populations has changed.

Other complicating factors

Population genetics

Another explanation for differing results by population includes genetic heterogeneity or inestimable polygenic effects (Deng 2001; Möller et al. 2010). Polygenic effects may be caused by the combined effects of several rare variants (Bodmer and Bonilla 2008). Rare variants in TB have been much understudied. One study has done extensive resequencing of Toll-like receptor (TLR) genes and found association with TB (Ma et al. 2007). We have also conducted full-exon resequencing of TLR genes and identified novel polymorphisms in Ugandan and South African populations (Baker et al. 2009). Copy number variants, which have been associated with HIV acquisition and progression as well as autoimmune diseases (McCarroll and Altshuler 2007), may also prove to be important to TB susceptibility. Because these variants have low frequency, they must be treated as rare genetic variants in analysis, and thus they require either large sample sizes or novel analytical approaches to detect trait associations. Analysis of large extended pedigrees would be ideal for the detection and analysis of rare variants (Manolio et al. 2009).

Linkage disequilibrium (LD) is another population genetic factor that may explain differences between studies. LD differs widely by population (Jakobsson et al. 2008), and since trait-associated polymorphisms may actually be untyped and in LD with genotyped markers, variations in LD may confound the ability to detect and replicate marker–trait associations. This is illustrated by the NRAMP1 association study by Velez et al. (2009a, 2009b); this study did not find association between TB and the four polymorphisms in NRAMP1 that were examined in previous studies, but it did find association between TB and other single nucleotide polymorphisms (SNPs) in the gene. If these additional SNPs had not been analyzed, the association between TB and NRAMP1 in that population likely would have been missed. In terms of understanding how specific genes convey risk for developing disease, it may be more important to first understand which gene is associated with disease risk and then identify the specific polymorphism(s). Since TB is a complex trait, there is likely genetic heterogeneity, implying that different variants influence disease risk in different populations. Thus, it is important to conduct thorough genotyping when conducting genetic epidemiological studies rather than focus on a few promoter or nonsynonymous SNPs.

Complex genetic effects

Complex traits are often characterized by gene–gene and gene–environment interactions, but such interaction effects in TB have been understudied. One study suggested interaction between the NOS2A gene and IFNGR1 and TLR4 (Velez et al. 2009b), though neither IFNGR1 nor TLR4 had significant main effects in this analysis. A second study by the same research group found interaction between NRAMP1 and TLR2, but TLR2 did not have a significant main effect (Velez et al. 2009a). This suggests that many important genes may influence TB in combination with other genes, but these interactions may be overlooked because their individual effects did not meet criteria for statistical significance. Alternative statistical approaches may be used to discover epistatic effects; such an analysis was conducted by Motsinger-Reif et al. (2010) who used multifactor dimensionality reduction to identify a potential gene–gene interaction between the TLR4 and TNFα genes.

Epidemiological factors also may modify the effects of genes. M. tuberculosis gene by human gene interaction, as discussed above, is one example of this. Another important effect modifier is HIV seropositivity. The effect of HIV on TB genetics has been understudied because most research studies exclude HIV-infected individuals from both their case and control populations, and HIV is a strong confounding factor of TB immune response. However, our previous work has shown an interaction between HIV serostatus and the TNF receptor 1 gene (Stein et al. 2007). Future studies may suggest that certain genetic effects are important only in HIV-uninfected individuals, while other genes modify TB susceptibility in HIV-infected individuals. On the other hand, studies examining interaction effects require large sample sizes (Velez et al. 2009b).

Summary and conclusions

Although numerous studies examining genetic risk factors for TB have been conducted, we are only beginning to scratch the surface of understanding the role of host genetics in TB susceptibility. Here we have described several principles of genetic epidemiology—phenotype definition, population genetics, and complex genetic effects—and we have illustrated how these factors cloud the current TB genetics literature. These factors should all be considered when synthesizing the literature. Although many of our combined p values attained statistical significance when considering NRAMP1 studies as examples, we do not wish to imply that this is true of studies of other genes. Many of these p values are highly significant due to the large number of p values that were combined. Moreover, these p values are all significant because we analyzed the most significant published p value. If we had conducted these analyses separately for each polymorphism, we likely would have had different results. For genes that have been studied less extensively (see Möller and Hoal 2010; Möller et al. 2010 for recent reviews), the impact of study design on study results is unknown. Although it is impossible to change these study design components retrospectively, future studies should consider these issues when developing the study’s design.

Seven genome-wide linkage scans for TB have been published. Recently, the first genome-wide association study (GWAS) of TB was published (Thye et al. 2010). Interestingly, this GWAS detected statistically significant association between TB and a novel genomic region on chromosome 18 (p < 10−8), but this region appears to be a gene desert, so no new candidate genes were immediately proposed by this study. Because of the increased power of association analysis over linkage analysis for detection of common variants with smaller effect sizes (Ardlie et al. 2002; Risch and Merikangas 1996), a GWAS analysis of TB should certainly provide new clues to the genetic underpinnings of TB risk. However, in order to provide sufficient statistical power for a GWAS, thousands of study subjects are needed. In the Thye et al. (2010) GWAS, populations from Ghana, The Gambia, and Malawi were analyzed using meta-analysis. As we show in this review, merging data across studies should be done with extreme caution because of differences in study design. Other GWAS analyses have merged data across studies, using meta-analysis techniques, and have successfully identified loci associated with Crohn’s disease and type II diabetes (Barrett et al. 2008; Zeggini et al. 2008); however, heterogeneity in study design may mask underlying associations when unaccounted for (Heid et al. 2009). However, if we are interested in detecting rare variants underlying TB risk, linkage analysis may be more powerful (Ardlie et al. 2002), so family studies will be advantageous, as stated above.

The ultimate goal in understanding TB genetics is to understand which factors make individuals more susceptible to developing disease in order to facilitate the development of better vaccines and other therapeutics. Another field of study involved in reaching this goal is immunology. Very few studies have examined genetic influences on the TB immune response (Hawn et al. 2007; Shey et al. 2010; Stein et al. 2007, 2008; Wheeler et al. 2006), and more studies are needed. Furthermore, the observation that some genes are associated with more than one phenotype demonstrates pleiotropic effects. For example, studies have shown association with both immunological traits and TB disease outcomes (Shey et al. 2010; Stein et al. 2007). These results could reflect the fact that these traits are on the same pathway (immune response influencing disease outcome) or that the traits themselves are correlated. In addition, some researchers have advocated the study of the “genetics of health” for a different perspective (Nadeau and Topol 2006), which would be useful in vaccine development. The few studies on TB that have focused on resistance to M. tuberculosis infection (Stein et al. 2008), or on LTBI and not TB disease (Cobat et al. 2009; Stein et al. 2007; Thye et al. 2009), are beginning to offer insight into the genetics of individuals who infected with M. tuberculosis and remain healthy.

Finally, unlike most genetic epidemiological studies that are conducted in developed countries, studies of infectious diseases like TB must focus on developing countries in Africa and Asia where the disease is endemic. Studies in Africa offer a variety of challenges (Sirugo et al. 2008), including limited resources for diagnosis of TB (Gustafson et al. 2001), geographic variations in M. tuberculosis lineage, and population-level variations in LD. Future studies must be mindful of these issues. From a public health perspective, these populations will gain the most from genetic epidemiological studies of TB, though the challenges are great.