Mendelian Randomization with Molecular QTLs: Methods, Challenges, and Advances

Introduction

Mendelian randomization (MR) is an analytic framework that leverages genetic variants as instrumental variables to infer causal relationships between an exposure and an outcomenature.com. In MR, if a genetic variant (instrument) is associated with a modifiable exposure (such as a biomarker or molecular trait) and that variant also shows an association with an outcome (disease or phenotype), one can test whether the exposure causally influences the outcome. The appeal of MR lies in the random allocation of genetic alleles at conception, which mitigates confounding and reverse causation biases typical of observational studiesnature.com. However, MR relies on strict assumptions: (i) the instruments must be strongly associated with the exposure, (ii) the instruments must not be associated with any confounder of the exposure-outcome relationship, and (iii) the instruments affect the outcome only through the exposure (no horizontal pleiotropy)nature.com. Violations of these assumptions – for example, a genetic variant influencing multiple biological pathways (horizontal pleiotropy) – can bias causal effect estimates and produce false positivesnature.com.

Molecular quantitative trait loci (QTLs) are genomic variants that affect molecular traits such as gene expression (eQTLs), protein levels (pQTLs), or DNA methylation (meQTLs). Integrating QTL data into MR has become a powerful strategy to identify causal genes or molecular mechanisms underlying complex traits. In such analyses, the molecular trait (e.g. the expression level of a gene or the abundance of a protein) serves as the exposure, with QTL-associated variants as instruments, and a disease or trait is the outcomegenomemedicine.biomedcentral.com. This approach enables researchers to prioritize likely causal genes at GWAS loci and to propose mechanistic links between genetic variation and disease. For example, using eQTL-based MR, Zhu et al. identified over a hundred gene-trait associations where gene expression is putatively causal for the traitpubmed.ncbi.nlm.nih.gov. Likewise, MR with pQTL instruments can pinpoint proteins that mediate genetic risk, highlighting potential drug targetspubmed.ncbi.nlm.nih.gov.

Despite its promise, MR using molecular QTLs faces significant methodological challenges. Pleiotropy (variants affecting multiple pathways) can violate assumption (iii) and confound results. Instrument weakness is also common – many eQTLs or meQTLs explain only a small fraction of variance in the molecular trait, risking weak instrument bias toward null or inconsistent estimates. Additionally, molecular QTLs are often context-specific (e.g. tissue-specific eQTLs), which complicates the MR assumption that instrument effects are consistent in the exposure and outcome datasetspubmed.ncbi.nlm.nih.gov. Finally, establishing causal directionality (does the molecular trait cause the disease, or vice versa?) requires careful analysis, sometimes involving bidirectional MR. In the following sections, we review key MR methods developed to address these challenges, focusing on approaches leveraging eQTLs, pQTLs, meQTLs, and other QTLs. We discuss the specific issues each method targets – such as pleiotropy, weak instruments, and causal direction – and how recent advances improve inference. Throughout, we highlight foundational studies and recent developments (particularly from the past five years), drawing on peer-reviewed literature for examples and technical details.

MR Assumptions and Challenges in Practice

Instrument Strength and Weak QTLs: The first MR assumption requires that genetic instruments are strongly associated with the exposure. In practice, many molecular QTLs have modest effect sizes, raising concern about weak instrument bias. Weak instruments can produce attenuated estimates and inflate Type I error, especially in two-sample MR when sample overlap exists. A common metric is the F-statistic for instrument strength; low values indicate potential bias. When many instruments are used, as in genome-wide MR, there is also a “many weak instruments” problem where standard errors are underestimated. Methods like MR-RAPS (Robust Adjusted Profile Score) were introduced to address this by modeling the uncertainty in weak SNP effects and applying heavy-tailed distributions to down-weight outliersnature.comnature.com. Another approach is to perform stricter instrument selection (e.g. using only genome-wide significant QTLs or strong cis-QTLs) to ensure relevancegenomemedicine.biomedcentral.com. Recent MR frameworks also adopt probabilistic fine-mapping ideas: for instance, the MRAID method (MR with automated instrument determination) selects a subset of candidate SNPs in high linkage disequilibrium (LD) that have non-zero effects on the exposure, effectively filtering out weak or null instrumentspmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov. By improving instrument strength (and validity) through such selection, MRAID and similar approaches maintain power while controlling false positives. Nonetheless, a trade-off exists – very stringent instrument selection increases strength at the cost of potentially missing true causal signals. The limited availability of strong QTL instruments, especially for genes with context-specific regulation, remains a fundamental challengepubmed.ncbi.nlm.nih.gov.

Horizontal Pleiotropy: The most pervasive threat to MR validity is horizontal pleiotropy, where an instrument influences the outcome through pathways other than the intended exposure. Undetected pleiotropy violates assumption (iii) and biases causal estimatesnature.com. For example, a variant may be an eQTL for gene A but also independently affect another gene or trait that influences the outcome, producing a spurious association. Traditional MR analysis often uses inverse-variance weighted (IVW) regression to combine multiple instruments, but IVW yields a biased estimate if even one instrument has horizontal pleiotropic effectnature.com. A telltale sign of pleiotropy is heterogeneity in the SNP-specific causal estimates. Statistical heterogeneity tests (Cochran’s Q) can flag this, and one simple remedy is to remove outlier SNPs contributing to heterogeneitynature.com. If the majority of instruments are valid, eliminating a few suspected pleiotropic SNPs can restore consistencynature.com. This approach underlies methods like GSMR (Generalized Summary MR), which implements a "HEIDI-outlier" procedure to detect and discard pleiotropic outliers that deviate from the expected causal effect patternnature.comnature.com. Verbanck et al. (2018) similarly developed MR-PRESSO, which identifies outliers with residual patterns suggesting pleiotropy and removes them to provide an adjusted estimate.

Beyond outlier removal, more robust estimators have been developed to obtain consistent causal estimates even when many instruments are invalid. MR-Egger regression is a landmark method that introduces an intercept term to model any average pleiotropic effect across instrumentsnature.com. Under the InSIDE assumption (Instrument Strength Independent of Direct Effect), the Egger regression intercept provides a formal test for directional pleiotropy; a non-zero intercept indicates that on average the instruments have a pleiotropic biasnature.com. The slope from MR-Egger yields a causal effect estimate corrected for pleiotropy (assuming the pleiotropic effects are uncorrelated with instrument strength). The strength of MR-Egger is its ability to detect and adjust for unbalanced pleiotropy, but it has notable weaknesses: it is statistically less powerful than IVW and can be sensitive to instrument selection and allele codingnature.comnature.com. Moreover, Egger’s reliability hinges on the InSIDE assumption, which is often untestable and can be violated in the presence of heritable confoundersnature.comnature.com.

Other robust methods make different trade-offs. The weighted median estimator combines instruments by median weighting: if at least 50% of the total weight comes from valid instruments, the median-based estimate is consistentnature.com. This approach tolerates up to half the instruments being invalid (with arbitrary effects) and often improves precision over MR-Egger. Likewise, the weighted mode estimator groups instruments by similarity of their ratio estimates; the idea is that the largest cluster of concordant SNP effects corresponds to the true causal pathwaynature.com. As long as the largest group of instruments is valid and oriented toward the true effect, the mode-based MR gives an unbiased estimate. These median and mode methods sacrifice some efficiency but are more robust than IVW when pleiotropy is widespread. A drawback is that when the number of instruments is small (e.g. only a few strong QTLs are available), their assumptions (majority valid or largest cluster valid) are hard to verify, and estimates may become unstable.

Recent years have seen an explosion of advanced methods to handle pleiotropy in the presence of LD between instruments, which is particularly relevant for molecular QTL-based MR (where multiple variants in a cis region are often correlated). Standard MR methods assume independent instruments, but cis-eQTLs or cis-pQTLs are typically in LD blocks. Clumping to select independent variants can reduce instruments to one or two per gene, hampering power and robustnessnature.comnature.com. To address this, researchers have developed generalized MR models that incorporate the LD structure. Generalized IVW (G-IVW) and generalized MR-Egger extend the classical methods by using the covariance between instruments in a generalized linear model, allowing correlated instruments in the analysisnature.com. However, these still assume all or most instruments are valid (for G-IVW) or satisfy a generalized InSIDE condition (for Egger), which may not hold in cis-QTL scenariosnature.com. Methods like LDA MR-Egger (LD-aware Egger) further attempt to account for correlations when performing Egger regressionnature.com, but they share Egger’s limitations of low power and stringent assumptions.

More recent innovations explicitly model the distribution of pleiotropic effects to improve inference with correlated instruments. For instance, MR-LDP (MR with LD and Pleiotropy) uses a Bayesian framework to jointly model SNP effect sizes on exposure and outcome, accounting for LD and allowing for some SNPs to have direct effectsnature.comnature.com. MR-Corr2 is another approach that focuses on correlated horizontal pleiotropy, recognizing that if SNPs are in LD, their pleiotropic effects on the outcome may be correlated. MR-Corr2 uses a covariance term for pleiotropic effects in its model to correct biasnature.com. A similar idea is extended in MR-CUE, which stands for Mendelian Randomization with Correlated and Uncorrelated Effects (as implied by its descriptionnature.com). MR-CUE (Cheng et al., 2022) can simultaneously handle both correlated and uncorrelated pleiotropy by making flexible assumptions about the pleiotropic effect distributionnature.com. MRAID (described earlier) likewise explicitly accounts for two types of pleiotropy – one where invalid instruments have effects correlated with their instrument strength, and one where they are independent – in a joint likelihood modelpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov. Finally, RBMR (Robust Bayesian MR) is a Bayesian approach that assigns prior probabilities to each instrument being valid or invalid and then infers the causal effect while averaging over these possibilitiesnature.comnature.com. Each of these advanced methods seeks to relax the classical MR assumptions by modeling pleiotropic effects rather than assuming them absent. Their strength is in scenarios like cis-QTL MR or polygenic MR where LD and pleiotropy are inevitable; their weakness is added complexity and often heavy computational burden (e.g. MRAID uses Gibbs sampling, which is slower than simple IVW)pmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov. Moreover, each method’s performance depends on whether its statistical assumptions (e.g. specific prior on pleiotropy) match the true underlying biology. No single approach is universally best, and it’s now common to apply multiple robust MR methods in parallel to see if a causal inference is consistent across themnature.com.

Causal Directionality and Bidirectional MR: Even if MR identifies a robust association between a molecular trait and an outcome, one must consider the direction of causality. It is possible that the outcome influences the molecular trait (reverse causation) or that both are influenced by some unmodeled factor. To address this, researchers use bidirectional MR by swapping the roles of exposure and outcome. For example, an MR analysis might test if DNA methylation at a CpG site causes type 2 diabetes and also test if genetic liability to diabetes (using disease-associated SNPs) causes methylation changes at that CpGpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov. In a study of diabetes and blood methylation, investigators applied two-sample MR in both directions for dozens of CpG sites; they found one methylation site in the DHCR24 gene where methylation likely causally increased diabetes risk, whereas for most other sites the direction was consistent with diabetes causing the methylation changespmc.ncbi.nlm.nih.gov. Such bidirectional analyses help infer the causal arrow: if only the forward direction (methylation → disease) is significant, it supports the exposure causing the outcome, whereas if only the reverse is significant (disease → methylation), it suggests the molecular change is a consequence of disease or a parallel effect. Additionally, a statistical test called the Steiger test can be used in two-sample MR to formally test directionality: it evaluates whether the genetic instruments explain more variance in the exposure than in the outcome. If not, it questions the assumed direction of effect. Applying Steiger filtering can improve causal inference by excluding instruments that appear to affect the outcome more directly than the exposure, which could indicate reverse causation or violation of assumptions.

Another complication in directionality is the possibility of mediators on the causal path. For instance, a genetic variant might affect gene expression which in turn affects a protein level that finally influences disease risk. In a naive MR of the gene expression on disease, the true exposure might actually be the protein, and the gene expression is just an intermediate step. Techniques like two-step MR or multivariable MR can help dissect such causal chains. Two-step MR involves first performing MR of the genetic variant on an intermediate phenotype (e.g. methylation to expression), then MR of that intermediate on the final outcomepmc.ncbi.nlm.nih.gov. If both steps show causality, it supports a chain (genotype → methylation → expression → outcome). Multivariable MR (MVMR), on the other hand, includes multiple exposures in the same MR modelnature.com. For example, one can include both gene expression and DNA methylation as exposures and a disease as outcome, using genetic instruments that affect either or both exposures. MVMR will partition the SNPs’ effects between the two exposures, providing an estimate of each exposure’s direct effect on the outcome conditional on the othernature.com. This is a powerful way to account for horizontal pleiotropy via measured mediators: if a variant has effects on two traits that both influence the outcome, a multivariable approach can jointly estimate those pathways. Burgess et al. (2015) formalized MVMR, and it has been applied in contexts like metabolic traits (e.g. partitioning the effects of BMI and lipid levels on coronary disease). The strength of MVMR is that it can control for known pleiotropic pathways (treating them as additional exposures)nature.com. Its weakness is the need for valid instruments for each exposure and sufficient power to estimate multiple coefficients. In the context of molecular QTLs, MVMR is particularly useful if one variant influences several molecular traits (say an eQTL for two genes); including both gene expression traits in a multivariable analysis can reveal which one (if any) drives the outcomenature.comnature.com. This strategy was explicitly used in recent transcriptome-wide MR frameworks as discussed next.

Integrating eQTLs: Linking Gene Expression to Disease

Expression QTLs provide natural instruments to test if gene expression changes have causal effects on complex traits. The foundational approach in this area is the Summary-data based Mendelian Randomization (SMR) method by Zhu et al. (2016)pubmed.ncbi.nlm.nih.gov. SMR was designed to integrate GWAS findings with eQTL data to prioritize candidate genes at loci. In SMR, for each gene, the top cis-eQTL (the variant most strongly associated with that gene’s expression) is used as the instrument to estimate the effect of that gene’s expression on the trait, using the ratio of the SNP’s effect on trait to its effect on expressiongenomemedicine.biomedcentral.com. An SMR association indicates that the gene’s expression level is associated with the trait due to a shared genetic variant, either because the expression change is causally influencing the trait or due to pleiotropy (the same variant independently affects expression and trait)pubmed.ncbi.nlm.nih.gov. To distinguish these scenarios, Zhu et al. introduced the HEIDI (Heterogeneity in Dependent Instruments) testgenomemedicine.biomedcentral.com. The HEIDI test leverages multiple SNPs in the gene’s cis-eQTL region: under the hypothesis of a single shared causal variant, all SNPs’ association patterns with expression and trait should be consistent with that one variant. If instead multiple independent variants are affecting the trait and expression (linkage or multiple pleiotropy), the SNP association ratios will be heterogeneousgenomemedicine.biomedcentral.comgenomemedicine.biomedcentral.com. HEIDI effectively tests for excess heterogeneity; a non-significant HEIDI test (no heterogeneity) suggests the association is likely driven by one variant, supporting a model of the gene being causal (or at least colocalized with the trait via the same variant)genomemedicine.biomedcentral.com. A significant HEIDI p-value indicates that more than one variant in that region influences the trait/expression, so the observed SMR association might be due to linkage (two separate causal variants) rather than true mediated causalitygenomemedicine.biomedcentral.comgenomemedicine.biomedcentral.com. In the initial SMR studies applying this method across complex traits, a subset of gene-trait associations passed the HEIDI test, highlighting novel candidate causal genes that were not necessarily the nearest genes to GWAS SNPspubmed.ncbi.nlm.nih.govgenomemedicine.biomedcentral.com. For example, SMR analyses identified genes like TRAF1 and ANKRD55 for rheumatoid arthritis where each gene’s expression, influenced by a local eQTL, was linked to disease riskpubmed.ncbi.nlm.nih.gov. These results provided leads for functional follow-up, illustrating how eQTL-MR can help move from association to mechanism.

The SMR approach addresses pleiotropy (to an extent) through the HEIDI test and by focusing on cis-eQTLs (which are more likely to have specific effects on the target gene). Its simplicity – using one top eQTL as instrument per gene – is a strength in terms of clear interpretation and low data requirements (only summary stats needed)genomemedicine.biomedcentral.com. However, it also has weaknesses. By using a single genetic instrument, SMR cannot apply the multi-instrument pleiotropy-robust methods described earlier (Egger, median, etc.), so it relies on HEIDI and the assumption that the top eQTL is the key variant. If the top eQTL itself has horizontal pleiotropic effects (for instance, it also regulates a nearby gene or has an independent effect on the trait), the SMR estimate will be biased and the power to detect that via HEIDI is limited when eQTL sample sizes are smallnature.comnature.com. Indeed, the developers noted that when only one instrument is used, the method is less powerful and cannot easily flag the case where that single instrument is an outliernature.comnature.com. Another limitation is that eQTLs are highly tissue-specific. SMR analyses are usually performed with eQTL data from a single tissue (e.g. blood or brain); if the trait is affected by gene expression in a different tissue, the available eQTL may not capture the causal expression change. This can lead to false negatives (missing a causal gene because the eQTL in the studied tissue is weak or absent).

To overcome some of these limitations, subsequent methods have extended the eQTL-MR paradigm. A significant advance was the development of Transcriptome-Wide MR (TWMR) by Porcu et al. (2019)nature.com. TWMR adapts a multivariable MR framework tailored to gene expression, allowing multiple eQTL instruments per gene and even multiple genes per locus to be analyzed simultaneouslynature.com. In practice, TWMR takes summary statistics for all SNPs in a locus for both eQTL and GWAS, and it can incorporate many instruments (not just the top SNP) for a given gene’s expression. It also can include multiple gene expression traits as parallel exposures if their eQTL signals overlap (since in cis regions, one variant often influences several nearby genes’ expression)nature.com. By doing so, TWMR replaces the single-instrument assumption with a weaker assumption akin to InSIDE, enabling the use of robust MR estimators even in the cis-eQTL contextnature.comnature.com. Essentially, TWMR can perform something akin to a generalized IVW or Egger across a set of correlated instruments, while also distinguishing between effects of multiple genes. This helps address pleiotropy at two levels: (a) if an eQTL for gene A is actually affecting gene B which causes the trait, a multivariable analysis including both A and B’s expression can reveal the true source, and (b) if a variant has pleiotropic effect not through measured genes, robust methods like Egger (with multiple instruments) could detect it via interceptnature.com. Porcu et al. indeed reported that a previous single-instrument analysisnature.com struggled to distinguish causality from pleiotropy, whereas their multiple-instrument, multiple-gene approach improved power and identification of likely causal genesnature.comnature.com. They demonstrated, via simulations, better control of Type I error and lower bias compared to standard MR, and produced an “atlas” of putative causal genes for 43 traits using eQTL data from GTEx and eQTLGen consortianature.comnature.com.

Another challenge in eQTL-based MR is tissue specificity. A gene might be causal only when expressed in a specific tissue relevant to disease. Traditional MR would either miss it (if eQTL data from that tissue aren’t available) or potentially violate assumptions (if using an eQTL from a non-causal tissue where the variant’s effect differs). A recent development to tackle this is mintMR (multi-context integrative MR)pubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov. MintMR is a multivariable MR framework that explicitly models gene expression effects across multiple tissues and even across different molecular layers (e.g. expression and methylation) as joint exposurespubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov. One of mintMR’s strategies is to use eQTLs that show consistent effects in more than one tissue as instruments, thereby improving the reliability of the instrument across contextspubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov. It iterates between per-gene multi-tissue MR analyses and a global model that learns which tissues are likely disease-relevant for each gene, borrowing strength across genes to identify patternspubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov. By doing so, mintMR addresses two issues: limited instruments (it can pool instruments across tissues) and inconsistent effects (it downweights tissue-specific signals that don’t generalize)pubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov. In a 2024 study, Lu et al. applied mintMR to evaluate causal effects of gene expression (and also DNA methylation) on 35 complex traits, finding that it controlled false positives and offered new insights into tissue-specific disease mechanismspubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov. The strength of such an approach is in leveraging the growing multi-tissue QTL datasets and acknowledging that the classic MR assumption of a single causal effect can be violated if the genetic effect differs by context. A weakness is that the model becomes quite complex and reliant on accurate cross-tissue effect estimates; if data in some tissues are noisy, the model’s latent variable approach may or may not correctly identify the relevant tissue. Nonetheless, mintMR represents an important step towards integrative, context-aware causal inference.

In summary, eQTL-based MR methods have evolved from simple single-variant analyses (SMR) to sophisticated multi-variant, multi-gene, and multi-tissue frameworks (TWMR, mintMR). Each development is motivated by a specific methodological issue: SMR tackled the gene-mapping problem with a pleiotropy test for linkage, TWMR addressed pleiotropy and power by using multiple instruments and exposuresnature.comnature.com, and mintMR confronts the tissue heterogeneity and sparse instrument issues by integrative modelingpubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov. The net effect is a more reliable identification of causal genes, though challenges like distinguishing truly causal genes from correlated signals in the same locus still require careful interpretation and often complementary evidence (e.g. colocalization probabilities or experimental validation).

pQTL-Based MR: Proteins as Causal Mediators and Drug Targets

Proteins are the functional effectors of many genetic associations, and MR analysis using protein QTLs (pQTLs) has gained traction as a way to pinpoint causal proteins for diseases. The principle is analogous to eQTL MR: if a genetic variant influences the circulating level of a protein, and that variant is also associated with a disease outcome, MR can test whether the protein level has a causal effect on disease risk. A distinctive advantage of pQTL MR is its direct relevance to therapeutics – identifying a protein that causally influences disease highlights a potential drug targetpubmed.ncbi.nlm.nih.gov. Indeed, successful examples like PCSK9 have validated this approach: genetic variants in PCSK9 that lower LDL cholesterol were shown by MR to reduce coronary disease risk, presaging the development of PCSK9-inhibitor drugsnature.comnature.com.

Several large-scale studies have constructed a “genetic atlas” of the human plasma proteome and applied MR. Sun et al. (2018) measured thousands of proteins and identified hundreds of pQTLs, then performed MR analyses to test which proteins might causally mediate genetic influences on various diseasespubmed.ncbi.nlm.nih.gov. They found that linking genetic factors to disease via specific proteins can highlight therapeutic targets and even anticipate on-target side effects of drugspubmed.ncbi.nlm.nih.gov. For instance, MR revealed proteins whose increased levels protect against disease but might raise risk for another trait, which is crucial information if that protein were to be inhibited or supplemented by a drugpubmed.ncbi.nlm.nih.gov.

A key methodological consideration for pQTL MR is instrument selection. Genetic influences on protein levels can be cis-acting (near the gene encoding the protein) or trans-acting (distant, often affecting regulation upstream or in pathways). Cis-pQTLs are generally preferred instruments because they are more specific to the protein of interest and less likely to have pleiotropic effects on other proteinsgenomemedicine.biomedcentral.comgenomemedicine.biomedcentral.com. In contrast, trans-pQTLs (especially variants in pleiotropic genes like those encoding master regulators) may affect multiple proteins and traits, violating MR assumptions. For this reason, many studies initially restrict to cis-pQTLs. For example, Jiang et al. (2022) performed MR of proteins on 211 phenotypes using only cis-pQTLs, noting that cis instruments are “known for being less prone to horizontal pleiotropy”genomemedicine.biomedcentral.com. In their first phase, they identified dozens of proteins in cerebrospinal fluid, plasma, and brain whose levels were putatively causal for various traitsgenomemedicine.biomedcentral.com. Interestingly, when they expanded the instrument set to include significant trans-pQTLs (with a safeguard of excluding variants that affected many proteins), the number of protein-trait associations roughly doubledgenomemedicine.biomedcentral.comgenomemedicine.biomedcentral.com. However, even then, most proteins had at most one or two instruments after clumping, meaning conventional pleiotropy-robust methods could not be applied for lack of multiple instrumentsgenomemedicine.biomedcentral.com. In such cases, results must be interpreted alongside evidence from colocalization or biology because a one-instrument MR per protein cannot fully distinguish a truly causal protein effect from a pleiotropic association.

To bolster confidence in pQTL MR findings, it has become standard to perform colocalization analysis in parallelgenomemedicine.biomedcentral.com. Colocalization approaches (like Bayesian coloc) evaluate whether the genetic association signals for the protein and the disease overlap, i.e. are likely driven by the same variantgenomemedicine.biomedcentral.com. This provides a complementary check: an MR finding that a protein affects a disease is far more convincing if the highest probability is that one common SNP drives both the protein and disease association (colocalization), rather than two different SNPs in the region. In practice, MR and colocalization are often used together to prioritize targetsgenomemedicine.biomedcentral.com. For example, in a proteome-wide MR on coronary artery disease (CAD), several proteins like PCSK9 and COLEC11 not only showed significant MR effects on CAD, but also had high colocalization probability (H4 > 0.7) indicating the same variant influenced both protein level and CAD risknature.comnature.com. These were considered strong causal candidates. On the other hand, another protein (FGFR1) had an MR signal but colocalization indicated distinct causal variants for the protein and CAD (suggesting the genetic signal for CAD in that locus was independent of the FGFR1 pQTL)nature.comnature.com. In such a scenario, the MR association could be a false positive due to LD hitchhiking, and indeed the authors noted insufficient evidence to implicate FGFR1 in CADnature.com.

The pQTL MR domain has also spurred specialized methods. One example is the cis-MR for drug target discovery framework by Qian et al. (2023), which introduced cisMR-cMLnature.comnature.com. This method extends the robust MR-cML approach (a contamination mixture model for MR) to handle correlated cis-pQTLs. CisMR-cML includes both variants associated with the protein (exposure) and those associated with the outcome in the cis region as candidate instrumentsnature.comnature.com, then uses a model selection (via Bayesian Information Criterion) to identify and downweight invalid instruments that show evidence of horizontal pleiotropynature.comnature.com. By modeling joint SNP effects rather than marginal and allowing some SNPs to have direct effects, it is robust to violations of the standard IV assumptions even with multiple correlated variantsnature.comnature.com. In an application to plasma proteins and CAD, this approach was able to detect known causal proteins (like PCSK9) and flag invalid instruments in regions where a cis-pQTL coincided with an independent CAD variantnature.comnature.com. The authors compared cisMR-cML against simpler methods (like G-IVW and Wald ratio per protein) and showed that naive methods often yielded inflated false positives when a pleiotropic SNP was present, whereas cisMR-cML controlled false discoveries by removing or correcting those influencesnature.comnature.com. The strength of such targeted methods is their focus on a single cis-locus at a time, where they can leverage the local LD structure and explicitly test each variant’s model fit (via likelihood or information criteria) to discern a causal mediation versus pleiotropy. The obvious trade-off is complexity and computational cost, but given the smaller number of cis-loci (relative to a genome-wide MR), this is tractable.

Overall, MR with pQTLs has rapidly advanced our ability to identify causal proteins. Its success stories are often in alignment with pharmaceutical interventions (providing a genetic rationale for drug targets), and its methodological toolkit now includes strategies to mitigate pleiotropy (using cis-instruments, colocalization filters, and robust cis-MR models). One lingering limitation is that many proteins lack a strong cis-pQTL, or have only trans-pQTLs which are harder to interpret causally. Large cohorts with proteomic data (employing technologies like SOMAscan or Olink) are increasing the yield of pQTLs, but the instruments identified can be context-dependent (e.g. present in plasma vs CSF) and sometimes overlap with eQTLs (since protein levels can correlate with mRNA expression). Multivariable MR could be applied in these cases too – for instance including both the mRNA and protein as exposures to see if the protein has an effect independent of transcript levels. Some studies have begun integrating multi-tissue pQTLs, as seen with Jiang et al. who compared plasma vs CSF vs brain protein causal effects and found largely consistent directionsgenomemedicine.biomedcentral.comgenomemedicine.biomedcentral.com. This hints that while tissue differences exist, many protein effects may generalize. Future work will likely refine pQTL MR by adding more tissues and considering longitudinal effects (since protein levels can be acute or chronic). But even with current methods, pQTL-based MR is a cornerstone of the burgeoning field of genetic proteomics, marrying human genetics with drug discovery.

DNA Methylation QTLs (meQTLs) and MR

DNA methylation is an epigenetic mark that can be influenced by genetic variation (meQTLs) and has been associated with numerous diseases and risk factors. MR analyses using meQTLs as instruments aim to determine if methylation changes at specific CpG sites play a causal role in disease or if they are consequences of disease processes. The use case for meQTL-MR often arises from epigenome-wide association studies (EWAS) that find CpG sites differentially methylated in cases vs controls. MR can help sort out causality: is the differential methylation a causal contributor to disease, or is it simply reflecting downstream effects (for example, inflammation from disease causing methylation changes)?

An important distinction in methylation MR is the high likelihood of reverse causation or confounding by environmental exposures. Unlike germline DNA, methylation can change over the life course and in response to disease or lifestyle. However, if a CpG site has a known cis-meQTL (a genetic variant affecting methylation at that site), that variant can serve as an instrument for lifelong predisposition to a certain level of methylation. Using genetic proxies for methylation levels, MR studies have investigated causal links. For instance, experiments with bidirectional MR in the context of type 2 diabetes (T2D) have been done for methylation sites. In one study, investigators took 58 CpGs correlated with T2D from observational data and found genetic instruments for both those CpGs and for T2D itselfpmc.ncbi.nlm.nih.gov. They then performed MR in both directions: T2D liability (genetic risk score) on CpG methylation, and CpG methylation on T2D riskpmc.ncbi.nlm.nih.gov. After correcting for multiple tests, they detected one methylation site where higher methylation causally increased T2D risk (cg25536676 in the DHCR24 gene, with an odds ratio ~1.43 per unit increase in methylation)pmc.ncbi.nlm.nih.gov. For the majority of other sites, they did not find evidence that methylation caused T2D; if anything, some showed evidence consistent with T2D influencing methylationpmc.ncbi.nlm.nih.gov. This illustrates a general finding in methylation MR: many disease-associated CpGs are likely reactive (consequences or markers of disease) rather than drivers. Identifying the exceptions – truly causal epigenetic changes – is valuable, as those could be therapeutic targets (e.g. drugs that modify methylation at that locus).

Methodologically, meQTL MR faces several hurdles. First, the availability of strong meQTLs is limited. Many methylation sites have only one or few genetic variants that explain a small fraction of variance in methylationgenomebiology.biomedcentral.com. This exacerbates the weak instrument problem and often necessitates large sample sizes or consortium data for meQTL discovery. Second, meQTLs can themselves be in LD with eQTLs for nearby genes, making it difficult to ascribe an effect purely to methylation vs gene expression. If a genetic variant influences both methylation at site M and expression of gene G, and disease risk, then a univariable MR of methylation on disease might be confounded by the gene expression pathway. An elegant approach to resolve this is to use multivariable MR including both the methylation and gene expression as exposures. If the effect of the instrument on the outcome is fully accounted for by one of the two, it suggests that pathway is causal. This was done, for example, in studies where methylation and expression from the same locus were tested together: often the gene expression carries the signal, and the methylation effect vanishes in multivariable MR, suggesting methylation was not the driver but rather correlated with the gene’s expression which truly affected the outcomepubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov. However, there are cases where the opposite might hold, or both matter.

Another consideration is the tissue and timing: blood meQTLs are commonly used (because large cohorts like ARIES have measured them), but disease-relevant methylation might occur in specific tissues (e.g. brain meQTLs for neurological diseases). If the genetic control of methylation is context-specific, using meQTLs from an accessible tissue could lead to incorrect conclusions. Some MR analyses leverage mQTLs identified across multiple tissues to mitigate this issue, analogous to the multi-context approach for eQTLs.

Despite these challenges, there have been successes in identifying methylation sites with probable causal roles. For example, Gaunt et al. (2019) developed a systematic framework to integrate meQTL and GWAS data, identifying instances where a genetic variant appears to influence disease via methylation changespmc.ncbi.nlm.nih.gov. They outlined scenarios: (1) genetic variant → methylation → disease, (2) genetic variant → disease → methylation, or (3) genetic variant influences both (pleiotropy). By combining MR with colocalization and mediation analyses, they found evidence for some loci conforming to scenario (1). One such finding was in the PAOF1 gene region for lupus, where a meQTL-mediated effect was implicated (the meQTL allele that increased methylation also increased disease risk, and colocalization supported a single variant explanation).

On the whole, MR with meQTLs is a promising but still developing area. Its strengths lie in providing a causal lens to interpret EWAS hits and uncovering mechanisms (for instance, if methylation mediates the effect of an exposure like smoking on lung function, as some studies have exploredsciencedirect.com). Its weaknesses are that methylation is often an intermediary in complex pathways and can be influenced by environmental confounders not accounted for by genotype. The genetic instrument approach circumvents confounding by non-genetic factors, but if the genetic variant affects multiple molecular phenotypes (like both methylation and expression), one must be careful. The use of two-step MR (genotype → methylation → expression → outcome) can help map out the chain of causality, and emerging methods are combining genetic and epigenetic data in comprehensive models. As data grows (e.g. brain meQTL catalogs, single-cell methylation QTLs), MR analyses will refine our understanding of epigenetic causality. So far, the takeaway from the literature is that while many methylation changes are likely effects, there are select cases where methylation is on the causal path, and identifying those can shed light on disease etiology and potential epigenetic therapies.

Other Molecular QTLs and Emerging Applications

The paradigm of using genetic instruments for molecular traits in MR extends beyond mRNA, proteins, and DNA methylation. Any molecular phenotype with genetic predictors can potentially be analyzed. For example, metabolite QTLs have been used to study causal roles of metabolites (like lipids, amino acids) in disease. Large metabolomics GWAS have identified loci controlling levels of metabolites; MR can then test if those metabolites mediate disease associations. This has been fruitful in cardiometabolic research (e.g., demonstrating causal effects of LDL cholesterol and urate levels on various outcomes, which in fact validated known risk factors)nature.comnature.com.

Another area is microbiome QTLs (sometimes called mbQTLs): human genetic variants that influence the composition of the gut microbiome. A few studies have used host genetic instruments for microbiome features to test causality on diseases like inflammatory bowel disease. However, microbiome data add extra complexity due to horizontal pleiotropy (diet-related genes, for instance, might affect microbiome and disease independently). This field is still nascent in MR.

Chromatin accessibility QTLs (caQTLs) and histone modification QTLs are yet another frontier. If a genetic variant affects chromatin state (e.g. an QTL for open chromatin at an enhancer), one could hypothesize that the chromatin change is causal for gene expression changes and hence disease risk. MR could be applied in a two-step manner (variant → chromatin → disease) or jointly with gene expression. Some integrative methods consider such hierarchical models, effectively extending MR into mediation analysis territory.

One particularly interesting extension is using summary MR as a tool for fine-mapping gene networks. Methods like CAUSE (Causal Analysis Using Summary Effect estimates) blend MR with polygenic modeling to infer whether the shared signal between an exposure and outcome is causal or just pleiotropy. CAUSE was applied, for example, to distinguish whether genetically predicted HDL cholesterol has a truly protective effect on coronary disease or whether genetic variants affecting HDL also independently affect CAD (the conclusion was largely pleiotropy, not causality, for HDL in one analysis). While CAUSE is not QTL-specific, it has been used with molecular traits too. It represents a class of methods aiming to improve causal inference when both exposure and outcome have polygenic architectures with potentially many overlapping variants.

Phenome-wide MR (pheWAS-MR) is an approach where one takes a particular exposure (say a protein or gene expression) and evaluates its causal effects across a wide array of outcomes. This is often done using two-sample MR with large biobank data. For example, picking a protein like CRP (C-reactive protein) and performing MR on dozens of disease outcomes to see where it has an effect. This can reveal unanticipated impacts or repurpose known biology to new indicationspubmed.ncbi.nlm.nih.gov. When the exposure is a molecular QTL-predicted trait, this approach effectively scans for all the phenotypes that trait might causally influence, highlighting potential pleiotropy as well. In one study, over 600 exposure-outcome pairs (many being protein-disease combinations) were screened in UK Biobank using MRAID, identifying several lifestyle factors and biomarkers with broad causal rolespmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov.

As these examples show, MR in the genomics era is expanding into a framework for causal inference across the “ome”. The main caution is that as we include more exotic molecular traits, the assumptions may become harder to satisfy – genetic architecture might be more complex, instruments fewer, and measurement error in exposures (when predicted from genotype) higher. Each scenario benefits from tailored methods (as we saw for cis-MR, mintMR, etc.). Nonetheless, the unifying principle remains: if we can find genetic predictors for a factor, we can probe its causal relationship with outcomes. The continuous development of biobank-scale QTL studies (for transcriptomes, proteomes, methylomes, metabolomes) and sophisticated MR methods suggests that in the next few years, our ability to map out causal pathways from genome to phenome will substantially improve.

Strengths and Limitations of QTL-based MR Approaches

In reviewing these methods, certain common strengths and weaknesses emerge. MR using molecular QTLs is powerful for mechanistic insight: it bridges GWAS findings with specific genes or molecules, thus moving a step closer to biological causality than a genetic association alone. It also exploits the naturally randomized experiment provided by meiosis, which is a strength in avoiding environmental confounding that often plagues molecular correlations (for instance, both an inflammatory protein and disease severity might increase with age, but a pQTL-MR would not be confounded by age). Moreover, the development of two-sample MR has been crucial – it allows us to use large consortia data for exposures and outcomes separatelynature.comnature.com, dramatically increasing power by leveraging public GWAS results. Many of the modern methods (SMR, TWMR, etc.) operate entirely on summary statistics, reflecting how MR has adapted to the massive scale of genomic data availability.

The key strength of newer QTL-MR methods is their focus on robustness to pleiotropy and LD. Techniques like MR-Egger, MR-PRESSO, and HEIDI test introduced the ability to detect pleiotropynature.comnature.com, while methods like MR-Corr2, MRAID, and cisMR-cML go further to correct for it in complex scenariosnature.comnature.com. By modeling pleiotropic effects or using multi-variable approaches, these methods aim to salvage causal inference even when classical MR assumptions are partly violated. This is especially pertinent for molecular traits where genetic pleiotropy is common (e.g. a variant in a transcription factor might influence many genes’ expression and multiple phenotypes). The strength here is in not throwing away information – using all SNPs in a cis-region (via joint modeling) rather than a single clumped SNP can improve power and make full use of the genetic signalnature.comnature.com.

However, with complexity comes limitations. Many advanced methods make their own assumptions (e.g. a distribution of pleiotropic effects) that if unmet can bias results in new ways. For example, MR-Egger’s InSIDE assumption or MRAID’s assumption that pleiotropic effects are partly uncorrelated with instrument strength need to hold reasonably wellnature.compmc.ncbi.nlm.nih.gov. If not, one might get a false sense of security – a non-significant Egger intercept might occur even if pleiotropy is present but just happens to correlate with instrument strength (violating InSIDE). Thus, results often need to be consistent across different methods to be credible.

Another limitation is that MR can only test linear causal effects (in general) and one exposure at a time (in univariable MR). Real biological systems are non-linear and interactive. While MVMR allows multiple exposures, it still assumes linear additivity of their effects. If, say, a gene’s effect on disease only manifests when a certain metabolite level is high, standard MR won’t capture that conditional relationship. Similarly, MR typically doesn’t handle time-varying exposures well – a genetic variant represents a lifelong difference, so MR estimates are like the effect of a lifelong exposure difference. For something like methylation which may change dynamically, the interpretation of MR is the effect of having that CpG more methylated across the life course (weighted towards the tissue/time where the meQTL is active). This might differ from an acute change’s effect.

Furthermore, phenotypic heterogeneity and measurement error in molecular traits can pose problems. QTL studies measure expression or protein in specific contexts; if the trait is measured with error or not representative of the true effector (e.g. measuring mRNA when the key is protein function), MR results could mislead. TWAS (transcriptome-wide association studies) and MR both grapple with this – a TWAS might find a gene’s expression (predicted by genetics) correlates with disease, but if the actual effector is the protein or a post-transcriptional modification, focusing on expression might misdirect effort. Combining QTL levels (like doing MR from genotype to both mRNA and protein, and protein to outcome) is one way to sort this out, but requires rich data.

Lastly, a practical limitation is the need for large sample sizes. Many molecular QTL effect sizes are tiny; detecting causal effects via MR needs well-powered GWAS for outcomes and QTL studies. The most robust findings often come from traits with GWAS sample sizes in the hundreds of thousands and QTL consortia with tens of thousands. For less common molecular phenotypes (e.g. a phospho-protein QTL or a rare cell-type-specific eQTL), power will be limited and MR results may be underpowered or null even if a causal effect exists. Negative MR results (no evidence of causality) should therefore be interpreted with caution – “absence of evidence is not evidence of absence,” especially when instruments are weak.

Conclusion

Mendelian randomization has become an indispensable tool for causal inference in human genetics, and its extension to molecular QTL data has enriched our understanding of the pathways from genotype to phenotype. By leveraging eQTLs, pQTLs, meQTLs, and other QTLs as instruments, researchers can test specific biological hypotheses: Does altered gene expression of X cause disease Y? Is protein Z a mediator of genetic risk for condition W? We have seen that foundational methods like SMR opened the door to integrating genomics with MRgenomemedicine.biomedcentral.com, addressing gene-trait mapping with tests for pleiotropygenomemedicine.biomedcentral.com. Building on this, a suite of robust techniques (MR-Egger, weighted median/mode, MR-PRESSO) tackled the generic problem of pleiotropic bias in MRnature.com. More recent innovations have targeted the intricate issues of correlated instruments and context-specific effects, with multivariable and multi-context MR frameworks (TWMR, mintMR) improving causal gene discoverynature.compubmed.ncbi.nlm.nih.gov, and advanced models (MR-Corr2, MRAID, cisMR-cML) offering pleiotropy-robust inference even in challenging cis-QTL regionsnature.comnature.com. Each approach comes with strengths tuned to specific scenarios and with assumptions that must be considered when interpreting results.

The landscape of MR methods in the past five years reflects a drive toward greater rigor and realism – accounting for LD, multiple mediators, and complex pleiotropic architectures that early MR methods treated as nuisances. Empirically, these methods have led to important discoveries: identification of drug targets like PCSK9 and IL6R through pQTL MR and colocalization, clarification that some biomarkers (e.g. HDL cholesterol, C-reactive protein) are likely not causal despite observational links, and the pinpointing of causal genes at GWAS loci which has energized functional genomics experiments. They have also sometimes delivered cautionary tales, such as the realization that naïve MR can be misled by subtle pleiotropy (e.g. the initial MR implicating HDL in heart disease later questioned by more robust analysis).

In applying these methods, best practice is to use a triangulation of evidence: if multiple MR methods (with different assumptions) and complementary analyses like colocalization all support a causal link, the inference is much strongergenomemedicine.biomedcentral.comgenomemedicine.biomedcentral.com. Conversely, if methods disagree, one must dig into why – perhaps certain instruments are driving the difference, pointing to pleiotropy or invalid instruments that need resolution. The continued improvement of MR methods – including Bayesian and machine-learning approaches to instrument selection, and integration with experimental perturbation data – will further enhance our ability to draw causal connections. As datasets grow (single-cell QTLs, large multi-omics panels), MR will likely incorporate those, requiring yet more complex models but promising an ever more detailed causal map of human biology.

In conclusion, Mendelian randomization leveraging molecular QTLs stands at the forefront of causal genomics. It synergizes statistical innovation with biological insight, each new method aimed at a particular methodological hurdle that once limited our confidence in causal claims. By understanding the purpose and limitations of each approach – whether it’s combating pleiotropy (Egger, MR-CUE), boosting power with multiple instruments (IVW, RAPS, TWMR), or ensuring the right gene/tissue is targeted (colocalization, mintMR) – researchers can design analyses that extract credible causal stories from the noisy complexity of genomic data. The result is a more coherent translation from genotype to phenotype: from associations that tell us “where” to look, to MR inferences that tell us “what” might happen if we could intervene on a gene, protein, or epigenetic mark. This knowledge is a critical stepping stone toward the ultimate goal of genetics research – improving human health through targeted intervention on the true causal pathways.

References: The references are embedded as inline citations in the format【source†lines】 corresponding to the key literature supporting each point, including seminal and recent peer-reviewed studies on MR methods and applications.