Mendelian Randomization: potential use of genetics to enable causal inferences regarding HIV-associated biomarkers and outcomes

Abstract

Purpose of review

It is unknown whether biomarkers simply correlate with or are causal for HIV-associated outcomes. Mendelian randomization (MR), is a genetic epidemiologic approach used to disentangle causation from association. Here, we discuss the potential use of MR for differentiating whether biomarkers are correlating with or causal for HIV-associated outcomes.

Recent findings

MR refers to the random allocation of alleles at the time of gamete formation. In observational epidemiology, this refers to the use of genetic variants to estimate a causal effect between a modifiable risk factor and an outcome of interest. A formal MR study using a genetic marker as a proxy for the biomarker has not been conducted in the HIV field. However, in the post-genomic era this approach is being used increasingly. Examples are evidence for the causal role of body mass index on blood pressure and non-causal role of CRP in coronary heart disease. We discuss the conceptual framework, uses and limitations of MR in the context of HIV infection as well as specific biomarkers (IL-6, CRP) and genetic determinants (e.g., in CCR5, chemokine, and DARC genes) that associate with HIV-related outcomes.

Summary

Making the distinction between correlation and causality has particular relevance when a biomarker (e.g. IL-6) is potentially modifiable, in which case a biomarker-guided targeted treatment strategy may be feasible. Although the tenets of MR rest on strong assumptions, and conducting an MR study in HIV infection presents many challenges, it may offer the potential to identify causal biomarkers for HIV-associated outcomes.

Introduction

In contrast to the ‘one size fits all’ approach of contemporary medicine for common complex diseases such as cancer and coronary heart disease (CHD), the goals of personalized medicine are to use individual biological signals such as biomarkers to predict the risk of developing disease, improve disease diagnosis, risk stratification of patients with active disease, and selection of therapies. This patient-specific approach is also the hope for HIV-1 medicine.

Many challenging barriers prohibit the utility of biomarkers in personalized medicine. Foremost, it is difficult to differentiate whether a biomarker is diagnostic, prognostic, or etiologic. For prognostic research, all factors associated with the outcome are of interest, whether they are causal or not. However, in etiologic research, causality is of fundamental importance. Although significant progress has been made in the identification of biomarkers that associate with HIV-related outcomes (e.g., see articles in this series [1–4]), it is unclear whether these biomarkers simply correlate with or are causal for HIV-associated outcomes, independent of all other risk factors (i.e. ceteris paribus). Making this distinction and adopting strategies to determine whether a biomarker may be causally linked with an HIV-associated outcome may have particular relevance when the biomarker is potentially modifiable, in which case a biomarker-guided targeted treatment strategy may be feasible.

Inferring causality from observational data is difficult as it is not always clear which of the two associated variables is the cause and which the effect, or whether both are common effects of a third unobserved variable, or confounder. The direction of causality can sometimes be determined by temporal criteria (e.g., the cause must precede the effect) or from knowledge of the underlying biology. However, confounding is always hard to address fully as it is due mostly to factors that are difficult to measure and control for. Also, it is nearly impossible to be certain that all the relevant confounders have been identified and accounted for. Also, for ethical reasons, risk factors cannot be assessed by randomized controlled trials. Studies combining genetics and biomarkers may help establish causality of a biomarker with outcome, and a genetic approach has been proposed to tackle some of these tribulations of observational studies and is termed ‘Mendelian randomization’.

The term Mendelian randomization (MR) is used in observational epidemiology to refer to the use of genetic variants as an instrument to estimate a causal effect between a specific modifiable risk factor (e.g., biomarker) and a trait/disease of interest [5–13]. The goal of this approach is to overcome some of the confounding that is encountered in observational epidemiology, such as residual confounding and reverse causation, by taking advantage of the natural random allocation of alleles during meiosis. It is important to emphasize that the while the goal of the MR approach is to help identify causal biomarkers that may provide useful diagnostic and/or prognostic information, it does not ascertain factors that are prognostic/diagnostic only, i.e., with no causal element.

In principle, MR may also hold promise as a tool for evaluating whether a biomarker is causally associated with HIV-related outcomes. However, infection with HIV imposes special constraints and confounders that may limit the full utility of this approach. That is, HIV infection through diverse mechanisms may induce or depress levels of a biomarker that may or may not be in the causal pathway for a particular HIV-associated outcome. Moreover, following infection the biomarker level may be altered to such an extent that it may inflate or obscure the association of a genetic variant with biomarker concentrations observed in HIV-negative persons. Additionally, epidemiological factors (e.g. differences in cohort characteristics, HIV outcome definitions) that are discussed below may further confound the results and interpretation of genetic variant–biomarker or genetic variant–HIV outcome association studies. Hence, the goals of this review are two-fold: First, as this approach has been used in the non-HIV field with some success in making the distinction of whether a biomarker is causal with or correlated with an outcome [10,11,14], we sought to bring the principles, assumptions, strengths, limitations and potential promise of MR to the fore of HIV investigators who evaluate biomarkers or other risk factors associated with HIV-related outcomes, and second, to illustrate this approach within the context of specific biomarkers and genetic determinants that influence HIV-associated outcomes. For these reasons we do not provide a comprehensive listing/discussion of genetic variants that associate with HIV-specific biomarkers [e.g., CD4+ T cell counts, plasma HIV RNA levels (viral load)], or HIV-associated outcomes (e.g., AIDS, death, opportunistic diseases), for which the reader is referred to other outstanding reviews [15–19].

Mendelian randomization (MR): conceptual framework

MR has been used extensively in the cardiovascular field (reviewed in [11]). Recent findings pertaining to the associations among single nucleotide polymorphisms (SNP) in the gene encoding C-reactive protein (CRP), CRP levels and coronary heart disease (CHD) serve as a blueprint for evaluation of whether biomarkers are causal for disease outcome [14]. We use this case-study from the cardiology field to illustrate the concept of MR as it also has direct relevance to the HIV field with many studies suggesting that CRP serves as a biomarker of HIV-associated outcomes (see related articles in this series [2,3]).

To establish the causal relationship between biomarker and disease outcome, investigators have traditionally used observational data showing an association between biomarker and outcome (Fig. 1A). However, this association may be confounded. One way to overcome this confounding is to integrate genetic strategies wherein a genetic factor is used as a proxy for the biomarker (Fig. 1B–C). Conceptually then, if a genetic variant is related to the biomarker (e.g., CRP levels) but is not otherwise associated with the outcome (e.g., CHD or HIV-related outcomes), then the relationship of the genotype with the outcome can be exploited to assess the causal relationship between biomarker and outcome (Fig. 1B–C). In the setting where the relationship between the biomarker and outcome is causal, then the genotype should associate with outcome, and the magnitude of this association should approximate the association of the genotype with the biomarker.

Fig. 1.

Approaches for evaluation of biomarkers in observational studies. (A) Evaluation of the relationship of biomarker with disease outcome using a non-randomized conventional approach. (B) Basic framework of Mendelian randomization approach. In its most basic form the approach is to study the causal relationship between a putative risk factor (biomarker levels) and the outcome (association 1) using a genetic variation that is known to influence the biomarker levels under investigation (association 2) as an instrumental variable, by estimating the association between the genetic variant and the outcome (association 3). (C) Same association study in panel A, but conducted within the principles of Mendelian randomization. (D) Detailed framework and assumptions of Mendelian randomization for evaluation of the relationship between biomarker and disease outcome. Directed acyclic graph (DAG) represents the relationships of a modifiable risk factor or biomarker B with the disease D with measured confounding variables C, unmeasured confounding variables U, unmeasured processes I impacting on both B and D, a SNP in the candidate gene designated as G that influences levels of B and which is also related to other genetic variants in another gene G1 that is possibly related to G and B, and other individual characteristics C’ possibly related to G, B, and D. In a DAG, a node represents a variable and an arrow a direct causal effect. Because a cause must precede an effect, no cycle is allowed and this is why the graph is termed acyclic (there is no loop from one node back to itself following the arrows). The methodology of Mendelian randomization uses the relationship of G with D, which presumably is unaffected by known confounding variables, C, and unknown confounders, U, to evaluate the relationship of B with D. Thus, in this situation, for causal inference of the impact of B on D, a genetic variant, G, which strongly influences the biomarker, can serve as a valuable instrumental variable. The premise of the instrumental variable approach is that integrating the estimates of the relationships of G with B and G with D can yield inferences regarding the relationship of B with D. There are many challenges to this method: the assumption that a gene influences disease solely through B is strong and unverifiable, as a single gene can influence disease risk through multiple pathways other than B (the phenomenon of pleiotropy); other alleles, G1 may be linked to G (linkage disequilibrium) and influence D through other pathways, which may result in confounding; and other characteristics of individuals at birth, C', that independently predict the development of D can be associated with G (population stratification) or influence the expression of G (e.g., epigenetics), and may void the conditions for a valid instrumental variable. Additionally, other alleles and patient characteristics can modify the effect of G on B, the effect of G on D, or both. These and other limitations are further illustrated in Figure 3. This figure is adapted from [12,13,20].

MR, is based on the premise of random allocation of alleles at time of gamete formation, and hence is thought to be independent of confounding factors [5–9]. In MR, a specific genotype results from two randomized transmissions, one from the paternally inherited allele and the other from the maternally inherited allele. A consequence of these randomizations is that genotypes are not expected to be associated with known (measurable or not) or unknown confounders for any outcome of interest, except for those that lie on the causal pathway between the genotype and the outcome (e.g., CRP→CHD; Fig. 1C). In theory, this should allow for the elucidation of two associations in a non-confounded manner (Fig. 1B–D): (i) genotype (G)-risk factor/biomarker (B) and (ii) the genotype (G)–disease outcome (D). By integrating appropriately the results of these two associations, one can, in theory, obtain an estimate of the risk factor/biomarker (B)-disease outcome (D) association, which is itself not confounded. Figure 1D outlines the principle and assumptions of Mendelian randomization, and it is important to emphasize that this approach does not define the association of the path from disease to biomarker (i.e., no D→B in Fig. 1B or D), a situation that might be common in the setting of HIV infection.

In some respects, the MR approach is analogous to randomized controlled trials (of sufficient sample size), in which the random allocation of treatment (or preventive measure) is expected to lead to an even distribution of (known or unknown) confounding factors across each group (Fig. 2). Thus, this strategy may be particularly helpful when direct intervention studies are not feasible due to the absence of a specific inhibitor of the risk factor. An additional strength of this approach is that while the levels of a biomarker may vary throughout a person’s life in response to diverse and unmeasured lifestyle and age-related biologic processes, genetic constitution is invariant. Another advantage is the accuracy of genotyping, which contrasts with biomarker measurements that may vary with time (e.g., diurnal changes, age) and there may be laboratory-to-laboratory differences in biomarker quantification. Finally, a genetic variant can be used as a proxy for a exposure that is difficult to measure (e.g., alcohol intake; [21]) and to possibly even infer biological causal pathways [22].

Fig. 2.

Similarities between Mendelian randomization and randomized clinical trials, illustrated by the example of the relationships between IL-6 gene variants, IL-6 levels and HIV outcomes. Figure adapted from [12].

Historically, the first description of the concept of MR in observational epidemiology is attributed to Katan, who suggested the use of ApoE genotypes that associate with cholesterol levels as a way to distinguish whether low cholesterol levels were a cause of cancer or a consequence of carcinogenesis [23]. The MR concept builds on what is known as an instrumental variable method in econometrics [24]. In the case of MR, the genetic variant acts as an instrumental variable. Three essential assumptions of MR must be met to allow for accurate application of MR and interpretation: (i) genotype is independent of confounding between biomarker and outcome [i.e., the graph has no arrow (in either direction) connecting CRP gene with the confounders; Fig. 3A], (ii) genotype is associated with the biomarker (i.e, there is an arrow connecting CRP genotype to serum CRP and this relationship can be accurately quantified, with a stronger association being most favorable; Fig. 3A), and (iii) genotype is independent of outcome except as mediated through the biomarker (i.e., no single arrow between CRP gene and CHD; Fig. 3A). Although these assumptions are strong and may have some untestable aspects, as for many modeling strategies, the approach may yield useful insights.

Fig. 3.

Limitations and considerations while applying Mendelian randomization. (A) The CRP genotype as an instrumental variable in Mendelian randomization. The arrows can be considered to represent causal relationships; it is important to note that there is no arrow between CRP gene and the confounders and there is no single arrow between CRP gene and CHD, i.e., there are no other routes in the pictorial between CRP gene and CHD. (B) The problem of pleiotropy. A single gene can influence disease risk through multiple pathways other than the biomaker. This violates the second assumption of Mendelian randomization. (C) Problem of linkage disequilibrium. In this instance, the genetic variant is in linkage disequilibrium with the variant associated with, or causal for, the disease. SNP in Gene 1 is in linkage disequilibrium with another SNP in Gene 2. Here, Gene 2 directly affects the disease level or risk, and hence Gene 1 is not an adequate genetic instrument for MR. The SNP in Gene 1 could be in linkage disequilibrium with a SNP in Gene 2 that is also causal for the biomarker (dashed line). In this instance, if Gene 2 SNP only affects disease via its effects on the same biomarker, there is no violation of the assumptions, and SNP in Gene 1 can still be considered as an instrument in a Mendelian randomization analysis. (D) Problem of gene-gene interaction. Example shown is for the interaction between genetic variants in IL-6 and CRP wherein a polymorphism in IL-6 gene correlates more closely with CRP levels than IL-6 levels. (E) Low explained variability. The Mendelian randomization approach is dependent on the proportion of the variance (explained variability depicted along the left of the upright triangle) of the biomarker explained by the genetic instrument (SNP) for the genes depicted on along the right of the inverted triangle. Abbreviations for genes: MC4R, melanocortin 4 receptor gene; FTO, fast mass and obesity associated gene; CRP, C-reactive protein gene; LCT, lactase gene; LPA, lipoprotein A gene; and ALDH2, aldehyde dehydrogenase gene. Data is adapted from [11].

CRP in CHD: a case-study of Mendelian randomization

Although CRP is a well-established biomarker for inflammation and CHD, studies have been inconclusive regarding a causal role of CRP in CHD. Using a powerful multi-staged study design, Elliott and colleagues recently identified SNPs in the CRP gene that associated with CRP levels [14]. However, in this MR analysis, these CRP SNPs did not associate with CHD [14]. These results suggest that CRP may not be causal for CHD.

Considerations for Mendelian randomization

MR studies represent a special case of conventional genetic association studies and hence many of the same genetic and non-genetic parameters/features considered for interpretation of such studies also apply to MR [11,20]. Many of the assumptions, challenges, and considerations necessary to account for in a MR study are outlined in Fig. 1D. We discuss some of these considerations and limiting factors pertinent to the conduct of an MR study within the context of the results of the CRP case-study described above as well as IL-6, as both CRP and IL-6 have relevance to HIV-associated outcomes [1–4].

(i) In an MR study it is important to consider that a gene may influence disease risk through multiple pathways other than the biomarker of interest (i.e., pleiotropy, which refers to the situation where a genetic variant has more than one specific phenotypic effect; Fig. 3B). E.g., the association of CRP SNPs may be indirect, influencing other biomarkers that cause or prevent CHD. (ii) A suitable functional genetic variant to study the biomarker of interest may not be identifiable for a MR study, such that most associations will be indirect, relying on linkage disequilibrium (e.g., SNPs in the CRP gene are in close proximity to SNPs in another gene that associates with CHD, Fig. 3C). Also, linkage disequilibrium patterns (i.e., the genomic architecture of the locus of interest) may vary from one population to the other and can potentially confound analyses. Copy number variation is a distinct polymorphism that is also thought to play a major role in gene expression levels [25] and failure to account for the genomic architectural complexity of the copy number variation region may also potentially confound MR [26–30].

(iii) A MR study will need to consider the growing evidence that complex diseases are influenced by numerous gene–gene and gene–environment interactions, which again may differ from one population to the other (Fig. 3D). In this instance, a false-negative association may be due to failure to account for a second gene that modifies the levels of the biomarker of interest, e.g., a SNP in the promoter of the IL-6 gene that associates with serum IL-6 levels [31] has been shown to correlate more strongly with CRP than IL-6 levels [32], and ApoE alleles can also influence CRP concentrations [33,34]. Illustrating gene-environmental interactions, a recent study showed that environmental factors modify the effects of the functional IL-6 promoter SNP [35]. (iv) An important limitation of MR is the weakness of the instrument, i.e., most single genetic variants only explain a small proportion of the trait variance (typically <5%; Fig. 3E), and hence very large sample sizes are needed to achieve a reasonable power. (v) The assumption of linearity between biomarker and outcome may not always be true.

(vi) The success of MR heavily rests on the existence of allelic homogeneity, i.e., a common causal allele is shared by many individuals. (vii) It is difficult to exclude canalization (e.g., genetic elevations in CRP levels are counterbalanced by compensatory changes in other systems) (viii) Epigenetics may confound the MR analyses. Given that levels of cytokines such as IL-6 and CRP have a heritable component [36–38] that may be further modified by environmental influences (e.g., gender, age), MR studies need to not only assume a random distribution of alleles in the offspring, but also a random distribution of epigenetic changes (e.g., gene expression) at conception, in order for the core assumptions of the MR strategy to remain valid [39].

Mendelian randomization and HIV-AIDS: special considerations

In general most association studies in the HIV field evaluate the association between a genetic variant with plasma HIV RNA levels (viral load), an intermediary phenotype, and/or with more distal outcomes such as AIDS, specific-AIDS defining illnesses or death (Fig. 4A). These studies using both candidate gene approaches and genome wide association studies have yielded many advances [15–19], including identification of genetic markers in the MHC locus that associate strongly with viral load [40], and the insight that HLA alleles such as HLA-B*57 [41] and variations in CCL3L/CCR5 genes may influence HIV pathogenesis by impacting on both viral load-dependent and –independent parameters such as cell-mediated immunity [42]. These viral load-independent associations are in accord with the notion that although highly contributory, viral load is not the sole determinant of CD4+ T cell loss and HIV-associated outcomes [43]. In some cohorts, the explained variability of viral load for rates of CD4 cell decline [44] and progression to AIDS [42] was <5% and <15%, respectively. Additionally, despite suppression of the viral load by highly active antiretroviral therapy, many parameters such as T cell activation that correlate highly with viral load persist, suggesting that mechanisms other than viral load may underlie HIV-associated outcomes [1,4,16]. These observations have prompted a search for and identification of host-centric rather than virus-centric intermediary phenotypes such as inflammatory factors (biomarkers) that associate HIV-associated outcomes (reviewed in [1]). (x) In HIV-1 infection, an MR study may need to account for the possibility of shared genetic influences. Although a significant part of the variation in inflammatory markers is due to genetic influences [36–38], almost 50% of the shared variance among biomarkers such as IL-6, CRP and fibrinogen may be due to a common genetic factor, suggesting that these biomarkers may lie within common causal pathways [37]. This has relevance to the results from the SMART and other observational studies that have found similar biomarkers to associate with HIV-associated outcomes [1].

Fig. 4.

Association study for HIV-associated outcomes, considerations for MR approach in HIV infection and candidate biomarkers for HIV-associated outcomes. (A) Schema depicting a conventional genetic association study for HIV-associated outcomes. The figure illustrates the association between a genetic variant with HIV-associated outcomes (CD4+ T cell counts, AIDS or death) via either viral load (VL)-dependent or –independent parameters such as cell-mediated immunity. (B) Considerations for MR study in HIV infection. In addition to the confounding factors that may influence any genetic association study, HIV may introduce additional variable that needs to be considered when conducting an MR study. a, HIV infection may alter biomarker levels (Δ biomarker), which in turn could impact on viral load. b, A biomarker may correlate with HIV-associated outcomes, and conversely, specific HIV-associated outcomes may influence biomarker levels. c, Specific HIV-associated outcomes (e.g., T cell activation) may influence HIV viremia, and conversely, HIV may influence outcomes through mechanisms that are independent of the biomarker of interest. d, HIV may theoretically impact on the gene locus of interest (e.g., by epigenetic reprogramming of gene locus that encodes the biomarker under investigation), and thus impact on gene expression and biomarker levels. (C) Candidate biomarkers that may influence HIV-associated outcomes and proportion of variance of the biomarker explained by the genetic instruments (SNP) depicted along the right of the inverted triangle.

In the ensuing sections of this review, we focus on the application of MR in HIV infection for investigating whether an intermediary phenotype such as a biomarker is causal with or correlating with HIV-associated outcomes. It should be noted that MR has, to date, not been discussed or conducted within the context of communicable or infectious diseases [11]. Although in theory the tenets of MR should also hold true in the context of HIV infection, both the evaluation and conduct of an MR study and meeting the assumptions of MR in HIV infection may be difficult for many reasons, some of which are summarized in Fig. 4B, and discussed below.

(i) Null or false-positive associations may arise because infection with HIV-1 may grossly alter levels of a biomarker. Fig. 4B illustrates the possible ways by which infection with HIV may distort many of the assumptions that need to be met for an MR study. We use the example of the recent associations of elevated IL-6 levels with adverse HIV-associated outcomes [1,4] to illustrate this important point. Although these associations raise the possibility that an IL-6 level-guided targeted treatment strategy could become an attractive therapeutic option, caution may need to be exercised as IL-6 levels may simply be a reflection of a more proximal and causal unmeasured biological pathway (e.g., T cell activation). Another possibility is that the impact of IL-6 on HIV-associated outcomes is via an additional biological factor. One can also envisage that the association of IL-6 or any other biomarker with disease outcome may be highly sensitive to the timing of IL-6 measurements, as associations may differ depending on whether biomarker concentrations were assessed during early or late stages of HIV disease. There is also the concern of reverse causation as CD4 cell depletion or specific forms of opportunistic infections may result in higher IL-6 levels and the pathways that mediate these raised IL-6 levels may remain activated despite suppression of viral replication.

If one were to apply the principles of MR to IL-6 in HIV infection, the same tenets described for the integration of CRP gene variant→CRP levels→CHD would have to apply, i.e., IL-6 gene variant→IL-6 levels→HIV-associated outcome. However, one will have to be very mindful of the fact that HIV infection may perturb IL-6 levels in a manner that the association of an IL-6 SNP with IL-6 levels may differ (inflated or obscured assocation) from that observed in HIV-negative persons. Consider then the following hypothetical scenario: assume for a moment that the association between IL-6 levels and HIV-associated outcomes is strong (as it appears to be the case). However, also consider that for the caveats and reasons mentioned in the preceding sections, future studies with large numbers of HIV-positive persons show that, at least based on the MR criteria, IL-6 is not causal for HIV-associated outcomes. These null genetic findings would limit our ability to distinguish whether IL-6 a risk marker or a risk factor.

Several additional types of confounding could have influenced the results of this hypothetical MR study with IL-6 in HIV (e.g., gene-gene interaction, pleiotropy). For example, the effects of IL-6 are pleiotropic, and interestingly, as noted above, a functional IL-6 SNP is known to correlate more strongly with CRP than IL-6 levels [32], and hence it is conceivable that the associations of the IL-6 SNP might not relate to IL-6 but CRP levels. Also, the interpretation of null MR studies is challenging, especially because in most instances a genetic variant explains only a small variance in the level of the biomarker (e.g., Fig. 3E and Fig. 4C). Also, the IL-6 SNP that may serve as a proxy for IL-6 is highly European-specific [45,46] and this will limit cross-comparisons of the same instrumental variable (i.e. SNP) across ethnicities. These limitations and considerations that we describe for a hypothetical MR study designed for IL-6 will also apply to any biomarker evaluated within the context of HIV infection.

(ii) Associations for gene or biomarker may vary because of differences in the criteria used to define an HIV-associated outcome. Although investigators oft-times do not differentiate between ‘soft’ non-clinical (e.g. CD4+ T cell counts or viral load) versus ‘hard’ clinical (AIDS or death, or specific AIDS-defining illnesses) endpoints, this differentiation has important consequences for any genetic or biomarker association study. Differences in the definition of the HIV-related phenotypic endpoint may greatly impact on the interpretation of association studies because genotypes have been identified (e.g. HLA alleles [47] and CCR5 haplotypes [48]) that impact on specific stages of HIV disease course (early vs late). Hence, it would not be surprising if genetic associations made for an endpoint classified as ‘progression phenotype’ and defined on the basis of the time to treatment initiation, or the time to the drop of CD4 cell counts below 350 cells/mm³ yields contrasting results when the associations for the same genotype are made with the 1987 CDC criteria of AIDS or death. Similarly, a biomarker because of its biological characteristics may causally associate with a specific AIDS-defining illnesses (e.g. cancer or HIV-associated dementia) but not with clinical AIDS (CD4 count <200 cells/mm³) and vice versa. For the same reasons, the association of a biomarker with overall AIDS (which is an amalgamation of several AIDS-defining illnesses) may be dependent on the frequency of the AIDS-defining illness present in the cohort, which may differ according to the risk characteristics of the cohort subjects (e.g. intravenous drug use versus mucosal transmission).

(iii) Associations for gene or biomarker may vary across cohorts because of differences in the epidemiologic characteristics of the cohort. Lead-time bias may confound association studies conducted in a seroprevalent cohort versus a seroconverting cohort. Other epidemiological features may also impact on the results of biomarker or genetic association studies. First, patients enrolled in distinct cohorts may have contrasting access to health care (e.g., differential antiretroviral use across cohorts in the pre-HAART era may confound associations of biomarker levels measured on samples during this time-period). Second, risk characteristics (e.g. intravenous drug use versus mucosal transmission) will impact on rates of coinfections with other viruses (e.g. Hepatitis C virus or GB virus-C) that may modify biomarker responses. Third, gender may also impact on biomarker status. For example, CCR5 levels are higher in women compared to men [49], and differential expression of biomarkers in men versus women have been observed [50]. Fourth, continent-of-origin may influence HIV outcomes. Because of environmental factors or co-infections specific to distinct geographic regions (e.g. malaria, TB), the gene-biomarker associations may differ. For example, studies have shown that some African populations have higher levels of T cell activation [51,52] and CCR5 [53], which in turn may influence biomarker levels. Also, pertinent to the evaluation of gene-biomarker associations across ethnicities, studies have suggested that compared to persons of European descent, persons of African ancestry might be over-represented and under-represented, respectively for genetic variants that associate with upregulation and downregulation of inflammation [45,54]. Finally, HIV subtypes may also associate with contrasting HIV outcomes and biomarker responses [55,56].

These observations, in conjunction with those factors noted above that require consideration during the conduct of any genetic association study (pleiotropy, complexity of genetic architecture of the locus, gene-gene and gene-environment interactions) indicate that results of gene-biomarker, or gene-HIV disease, associations, may not necessarily be invariant across cohorts with heterogeneous epidemiological characteristics. Such a viewpoint is dogmatic and requires caution.

Mendelian randomization and HIV-associated outcomes: case studies

Many genetic variants have been identified that influence biomarker levels, and some of these may have a relationship with HIV-associated outcomes. Table 1 provides a brief listing of some potential biomarker candidates to consider for further evaluation by MR. Because of the strong in vitro, ex vivo and in vivo data linking the chemokine—chemokine receptor system to HIV-AIDS pathogenesis, we discuss CCR5 and its ligands (primarily CCL3L/CCL4L) as examples. We also discuss another gene, Duffy Antigen Receptor for Chemokines (DARC) as it is a chemokine binding receptor and variations in this gene may potentially impact on levels of chemokines and other biomarkers relevant to HIV-associated outcomes.

Table 1.

Selected biomarkers (B) and genetic instruments (G) that may have utility for evaluation in an MR study for HIV-associated outcomes.

Category	Biomarker (B)	Genetic instruments (G)	Genetic consequence on biomarkers (G→B)	Genetic Effect on HIV disease outcome (G→D)	References
Chemokine receptor	CCR5	CCR5 −Δ32 (rs333) CCR5 haplotypes (e.g. P1/HHE)	Altered CCR5 surface expression	Associations with HIV-AIDS susceptibility [including surrogates of disease (CD4, VL)] and immune reconstitution during HAART.	[15–19,48,57,58]
Chemokines	CCL3L, CCL4L	CCL3L and CCL4L various gene copy number.	Increased gene dose of CCL3L1-containing segmental duplication associates with increased protein levels (the relationship is not linear, reaching a plateau at high gene dose)	Associations of CCL3L1-containing segmental duplications with HIV-AIDS susceptibility and immune reconstitution. Specific combinatorial content of CCL3L and CCL4L genes associate with HIV-AIDS outcomes. CCL4L gene copy number associates with HIV susceptibility.	[28–30,42,59–61]
	CCL5	CCL5 −471G>A (rs2107538) CCL5 In1.1T>C (rs2280789)	Altered transcription and possibly protein levels	Associations with HIV-AIDS susceptibility	[62–66]
	CCL2	CCL2 −2578 A>G (rs4795895)	Altered protein expression	Associations with HIV susceptibility and disease progression including HIV-associated dementia	[67,68]
	CCL2, CCL5	DARC Asp42Gly (rs12075)	Altered serum level of CCL2 (MCP-1), CCL5 (RANTES) and IL-8; SNP accounts for ~20% of the variability in serum CCL2 levels	Not studied	[69]
	CCL4	CCL4L2 rs4796217 C>T	Minor allele significantly associates with a low plasma level of CCL4 (MIPβ)	Not studied	[70]
Peripheral blood cell counts	CD4+ T cell counts	MHC locus rs2524054 C>A	Altered CD4 levels and CD4/CD8 ratio in HIV negative individuals. Explains ~5.7% variation of CD4/CD8 ratio	The A allele of rs2524054 strongly associates with viremia when HIV controllers were compared to a group of HIV-1 progressors	[71]
	WBC, neutrophils	DARC −46T>C (rs2814778)	Low WBC and neutrophil counts, and at the population level, explains ~27% and ~20% of the variation in neutrophil and WBC counts, respectively.	African-specific DARC −46C/C associates with an increased risk of acquiring HIV but slow disease progression. The latter association occurs mainly in those African Americans who are also leukopenic.	[72–75]
	Platelet counts	ATXN2 rs11065987 A>G	Minor allele associates with altered platelet counts	Not studied	[76]
Cytokines/inflammatory markers	IL-6	IL-6 −174 G>C (rs1800795)	Associates with altered IL-6 and CRP levels	Altered risk of KS development and variable recovery of CD4 cells during HAART	[77–80]
	IL-6sR	IL-6R rs4129267 C>T	Associates with plasma levels of IL-6sR	Not studied	[70]
	IL-10	IL-10 −592C>A (rs1800872)	Decreased IL-10 levels	Increased HIV-AIDS susceptibility	[81–84]
	IL-18	IL-18 rs2250417 A>G	Minor allele associates with low plasma IL-18	Not studied	[70]
	TNFα	ABO rs505922 T>C	Minor allele associates with low levels of TNFα	Not studied	[70]
	CRP	CRP rs7553007 G>A	Minor allele associates lower CRP levels	Not studied	[14]
	Fibrinogen	FGB rs1800789 G>A	Minor allele associates with high fibrinogen and explains less than 2% of variations in levels	Not studied	[85]

Abbreviations: CCR5, CC chemokine receptor 5; CCL, CC ligand; DARC, Duffy Antigen Receptor for Chemokines; WBC, white blood cells; VL, viral load; IL, interleukin; KS, Kaposi sarcoma, and CRP, C-reactive protein. CCL3L denotes CCL3L-1, -2, and -3 genes; CCL4L denotes CCL4L-1 and CCL4L-2 genes.

CCR5 levels

Surface levels of CCR5, the HIV coreceptor, represent an important biomarker as it impacts on several outcomes. Variable human CCR5 expression is associated with variable HIV cell entry in vitro, HIV acquisition, AIDS progression rates, viral load and CD4 immune reconstitution during HAART [15–19,86–90] Also, CCR5 surface expression influences efficacy of CCR5 blockers and entry inhibitors [91–93], and pre-infection CCR5 levels correlate with rates of disease progression post-infection [94]. The association of low CCR5 levels with reduced AIDS susceptibility also has an evolutionary correlate: high vs. low CCR5 surface expression distinguishes pathogenic (e.g. human) vs non-pathogenic lentiviral (SIV) infection in some non-human primates, with low CCR5 levels associating with non-pathogenic infection as seen in Sooty Mangabey [95,96]. Studies have identified polymorphisms in the CCR5 coding and promoter regions that influence both CCR5 expression and HIV-AIDS susceptibility as well as CD4 recovery during HAART [15–19,42,48,57,58,61,86–90,97–105]. Thus, at many levels, the assumptions of Mendelian randomization are met in the case of CCR5. However, evaluation of CCR5 variations in MR studies requires special considerations: CCR5 SNPs display complex linkage disequilibrium patterns [97]; the associations of CCR5 haplotypes are race-specific and there is a strong interactive effect between CCR5 haplotypes with different evolutionary histories [57]; and, there is evidence for gene gene interaction. For example, a recent study demonstrated that the interaction between the CCR5 Δ32 mutation and variations in the IL-12B gene influence HIV disease progression [106]

CCR5 ligand levels

Supporting the strong in vitro results that CCR5 ligands have HIV suppressive properties, extensive data demonstrates that ex vivo expression of CCR5 ligands such as CCL3, CCL4 and CCL5 serve as a biomarker for HIV-AIDS susceptibility, with higher levels associating with beneficial effects [107–113]. However, there is a cautionary note: the biomarker potential of CCR5 ligands requires consideration of the fact that although antiviral immunity is afforded by release of HIV-suppressive CCR5 ligands from activated leukocytes, serum levels of these biomarkers may correlate poorly with HIV-AIDS outcomes [114] and high levels of chemokines by imparting pro-inflammatory effects [115] may accelerate disease progression.

Among the CCR5 ligands, CCL3L1 displays the maximal HIV-suppressive effects in vitro [116,117], suggesting that it might play a key role in HIV outcomes in vivo. Notably, CCL3L genes (CCL3L-1, -2, and -3) are subject to copy number variation (reviewed in [30,118]), with higher copies associating with increased CCL3L1 expression [59,119], and enhanced HIV-specific T cell responses [120]. These in vitro biological data along with the in vivo observations demonstrating a strong association between higher CCR5 ligand expression and reduced HIV-AIDS susceptibility, led investigators to determine the associations of the CCL3L1-containing segmental duplications with HIV-AIDS susceptibility. In several cohorts, a lower copy number of the CCL3L1-containing segmental duplication associated with lower CD4 counts, higher viral load, increased risk of acquiring HIV and faster HIV disease course (reviewed in [30]). A similar association of increased AIDS progression rates following experimental infection with SIV was observed in rhesus macaques with low CCL3L gene copies [121]. In some cohorts these associations were not observed, and some of the reasons for this discrepancy have been discussed [28,30]. Analogous to the associations of the copy number of the CCL3L1-containing segmental duplication, an association between the copy number of CCL4L genes and HIV susceptibility has also been reported [30]. Thus, multiple lines of evidence suggest that variations in CCR5 chemokine genes that influence expression levels may serve as a useful instrument to test the role for chemokines in HIV-AIDS susceptibility using the MR approach. However, a special consideration for evaluation of CCL3L/CCL4L gene dose in MR studies is the complex genomic architecture of this locus because it is likely that it is the combinatorial content of these genes that associate with HIV-AIDS outcomes[30].

DARC expression: relevance for biomarker research relevant to HIV

Duffy Antigen Receptor for Chemokines (DARC) is a CC and CXC chemokine binding protein that is highly expressed on red blood cells and endothelial cells [122]. The functions of DARC have relevance for evaluation of chemokines and other biomarkers with HIV outcome. Foremost, many chemokine-related functions have been proposed for erythroid DARC, ranging from its serving as a sink or reservoir for several chemokines (e.g., IL-8 and RANTES), as well as maintenance of chemokine plasma concentrations [reviewed in [122]]. Two polymorphisms in the DARC gene have relevance to HIV biomarker research. A close relationship between erythrocyte DARC and chemokine levels is underscored by a genome-wide association study showing that a polymorphism (Asp42Gly) in the DARC coding region is a major regulator of the DARC-mediated erythrocyte binding of CCL2, IL-8 and RANTES and in turn, the circulating serum levels of these proteins [69]. Second, is an African-specific genotype in the DARC promoter (−46C/C) that results in selective loss of expression of DARC on red blood cells (Duffy/DARC-null trait on erythroid cells) and resistance to Plasmodium vivax malaria [122]. A comparison of healthy subjects with and without DARC −46C/C revealed differences in circulating levels of chemokines, as well as coagulation responses to intravenously administered endotoxin, with Duffy-null subjects having, in general, reduced responses [123,124]. In light of these findings that DARC expression may modify the chemokine and coagulation responses to experimental endotoxemia it will be important to evaluate whether the Duffy-null state influences LPS levels during HIV disease.

In addition to impacting on chemokine homeostasis, DARC expression may also influence cellular biomarkers relevant to HIV-AIDS pathogenesis. There are striking inter-racial differences in white blood cell (WBC) counts, a cellular biomarker. Compared to persons of European ancestry, individuals of African descent have lower WBC counts, attributable mainly to reduced neutrophil counts [125–128]. This hematological condition is referred to as “ethnic leukopenia or neutropenia” [125–129]. Although generally thought to be without consequence, inter-racial differences in WBC counts may influence treatment outcomes [129] and infectious disease susceptibility [130]. Conclusive data now demonstrates that the Duffy-null mutation is causal for the leukopenia/neutropenia found commonly in persons of African ancestry [72,73]. Whether the Duffy null-associated neutropenia influences HIV-AIDS pathogenesis merits further investigation for several reasons. HIV-negative healthy subjects with low neutrophil counts have elevated CRP levels [131], and whether this association is a correlate of HIV acquisition is unknown. Also, Detels et al (1) demonstrated that resistance to acquiring HIV in men followed prospectively in the Multicenter AIDS Cohort Study (MACS) was associated with high neutrophil and CD8+ T cell counts (1) and He et al showed that the Duffy-null state associated with an increased risk of acquiring HIV infection [74]. In light of the aforementioned findings, one possibility is that the association of Duffy-null state with increased HIV risk may reflect the association of this trait with low neutrophil counts, or that the DARC-null state associates with increased HIV risk specifically in those persons with this trait who also have low neutrophil counts. However, additional mechanisms may be operative as suggested by the binding of HIV-suppressive and inflammatory chemokines [74,122] and HIV-1 viral particles to DARC on RBC [74].

The recently described association of the DARC-null state with low WBC counts in persons of African descent [73] prompted investigators to evaluate further the relationships among the DARC-null state, leukopenia and HIV disease course in HIV-positive European Americans and African Americans [74,75]. They found that although leukopenia occurs frequently during HIV infection, its influence on disease outcome differs depending on both race and DARC genotype [75]. In HIV-infected European Americans, leukopenia was associated with an accelerated disease course. By contrast, the association of Duffy-null state with HIV disease course differed according to WBC but not CD4+ T-cell counts, such that leukopenic but not nonleukopenic HIV-positive African Americans with the Duffy-null trait had a survival advantage compared with all Duffy-positive subjects. This survival advantage became increasingly pronounced in those with progressively lower WBC counts. [75]. These findings suggested that in African Americans, the DARC genotype may influence HIV disease course within the context of a specific cellular milieu, as reflected by leukocyte cell counts.

It is important to note that others did not identify an association of DARC genotype with HIV disease course [132–135]. However, as posited by He et al in their riposte to these correspondences, these disparate findings may in part be explained by differences in cohort characteristics and study design, including study endpoints but not population stratification as suggested by some [19,132–135]. Further it is possible that the recently described relationships between DARC genotype, WBC counts and HIV disease progression [75] may have played a role in these null associations and may help account for the different findings reported in these cohorts [74,132–135].

Taken together, much more needs to be learnt about DARC function and biology in health and disease. Nonetheless, these observations collectively highlight the importance of being mindful that DARC expression may impact on evaluation of cellular (WBC and neutrophil counts) and inflammatory (chemokine) biomarkers relevant to HIV-associated outcomes as well as influence the chemokine and coagulation response to endotoxemia.

Conclusions

The biomarker field in HIV medicine is clearly at a very exciting point. The hope is that some of these biomarkers would be identified as risk factors for HIV-associated outcomes and could be used for both risk assessment and to individualize specific therapies. We review briefly the conceptual framework, strengths, and limitations of MR, an application that has been employed in other fields with some success to evaluate the possible causal role of biomarkers.

However, MR is not, and should not be viewed as a panacea, and the limitations of MR should be clearly acknowledged. Although MR can make a distinctive contribution to the epidemiology of biomarkers and HIV-associated outcomes, it is important to emphasize what MR is not. (a) MR is not fundamentally about discovering how genetic variation influences HIV-associated outcomes. In the context in which we have discussed MR, genotype is used strictly as a proxy for a biomarker. (b) MR is not a strategy for detecting genes that confer a higher risk of HIV-associated outcomes. (c) MR is an epidemiologic tool, not a molecular or physiologic inquiry into underlying mechanism of HIV-associated outcomes, albeit it may indirectly provide information regarding potentially critical pathways underlying HIV-associated outcomes. HIV infection may further constrain the ability to meet the assumptions of MR.

Nonetheless, with full acknowledgement of the limitations and caveats related to MR, and under the appropriate clinical settings where many confounding factors can be accounted for (e.g. SMART study), and including especially the availability of a genetic instrument that is both strongly associated with the biomarker and independent of HIV-associated outcome, MR can complement the findings from traditional observational and functional studies. We envisage that large scale genetic studies and MR as highlighted by the studies of Elliott et al on CRP and CHD [14] may be necessary for definitive evaluation of the causal role of biomarkers in HIV-associated outcomes. Together these complementary strategies offer the potential of identifying which biomarkers may be useful as specific drug targets and have a role in risk assessment and therapeutic selection. Undoubtedly, better biomarkers and innovative strategies for combining them will be needed, along with statistical and functional evaluation of causality, to fulfill the promise of biomarkers for personalized medicine.