Causal Inference Methods to Integrate Omics and Complex Traits

Abstract

Major biotechnological advances have facilitated a tremendous boost to the collection of (gen-/transcript-/prote-/methyl-/metabol-)omics data in very large sample sizes worldwide. Coordinated efforts have yielded a deluge of studies associating diseases with genetic markers (genome-wide association studies) or with molecular phenotypes. Whereas omics–disease associations have led to biologically meaningful and coherent mechanisms, the identified (non-germline) disease biomarkers may simply be correlates or consequences of the explored diseases. To move beyond this realm, Mendelian randomization provides a principled framework to integrate information on omics- and disease-associated genetic variants to pinpoint molecular traits causally driving disease development. In this review, we show the latest advances in this field, flag up key challenges for the future, and propose potential solutions.

Most common diseases have 30%–80% heritability (Ge et al. 2017) and the remaining causes are comprised of modifiable environmental, lifestyle, and molecular factors. The identification of all genetic and nongenetic factors remains elusive for multiple reasons.

The major hurdle in deciphering the genetic basis of complex traits is the necessity of very large sample sizes to identify contributors of a genetic architecture very much resembling an infinitesimal model (where a very large number of genetic variants have increasingly small effects). Nongenetic factors, on the other hand, tend to have larger correlations with diseases. However, accurate measurement of all relevant (molecular and high-level) phenotypes in human populations in a noninvasive fashion is difficult; hence, we often rely on noisy proxies, and dissecting true causes from mere disease correlates has proved to be particularly challenging.

The aim of this review is to provide an overview of how principled causal inference methods can be adapted to combine molecular phenotypes (i.e., omics data) and complex diseases to reveal robust causal biomarkers. First, we introduce genome-wide association studies (GWASs) of complex and molecular traits. We then explain how results generated from these two types of studies can be combined together, with strong emphasis on various forms of Mendelian randomization (MR). Next, we delve into the key findings of transcriptome-, proteome-, and methylome-wide MR studies. We also review key limitations of current approaches (heritable confounding, tissue-specificity, reconciling different omics). Finally, we point out future challenges: linking disease causes and consequences, teasing out sex-specific effects, deriving causal regulatory networks, and leveraging findings for drug repurposing. Because of the extremely advanced state of transcriptomics research and hence the dominating abundance of (publicly available) gene expression data sets (>1.8 million human samples in the Gene Expression Omnibus), our review is slanted toward this type of omics data. However, most concepts, difficulties, and challenges presented for RNA-based biomarkers are, in principle, transferrable to other omics data types.

GENOME-WIDE ASSOCIATION STUDIES

Whereas whole genome sequencing is not yet a viable option in millions of samples (current costs are ∼$1000/sample), measuring common genetic variations has become affordable (cost <$50/sample for genome-wide genotyping arrays, probing >700,000 variants). There is growing evidence that these measured common variants capture the majority of the heritability for several model traits (Yang et al. 2015). GWASs have been designed to identify single-nucleotide polymorphisms (SNPs) associated with traits/diseases (Visscher et al. 2017). Vast amounts of financial and human resources have been invested in the past decade into data collection for large population cohorts, including whole-genome-scale genotyping/sequencing and extensive characterization for dozens of clinically relevant phenotypes. Indeed, these efforts have led to thousands of association studies on hundreds of phenotypes and diseases. These traits include anthropometric traits (e.g., height and body mass index [BMI]; Yengo et al. 2018), blood chemistry variables (e.g., lipid levels; Global Lipids Genetics Consortium et al. 2013), cardiovascular traits (e.g., blood pressure; Evangelou et al. 2018), complex diseases (e.g., type 2 diabetes [T2D]; Mahajan et al. 2018), and cognitive traits (e.g., educational attainment; Lee et al. 2018).

These studies have shown that each identified variant alone has a minute impact (0.01%–0.5% explained variance) on the respective phenotype. However, as study sizes continue to grow larger, the cumulative effect of the increasing number of discovered DNA variants is steadily approaching their respective heritability (Maier et al. 2018) and the remaining gap could possibly be filled by additive effects of rare variants (Wainschtein et al. 2019). GWASs have also shed light on biologically meaningful pathways (Pers et al. 2015) and key tissues (Ongen et al. 2017; Finucane et al. 2018; Richardson et al. 2020). They have also enabled accurate estimation of narrow sense heritability due to common variants (Yang et al. 2010; Bulik-Sullivan et al. 2015b) and genetic correlation (Bulik-Sullivan et al. 2015a).

MOLECULAR TRAITS UNDERLYING DISEASES

Whereas GWASs have revealed many biological insights in the past decade, the underlying mechanisms as to how the effect of SNPs leads to disease are still poorly understood. Even if effect sizes are unbiasedly estimated, the fine-mapping of the association signals is very difficult and largely insufficient for pinpointing the implicated gene or understanding the underlying molecular mechanisms. One opportunity to fill the gap from genetic variants to complex traits is combining GWAS findings with other “omics” (e.g., transcriptomics, methylomics, metabolomics, proteomics) data, to functionally characterize the statistical associations (Passador-Gurgel et al. 2007; Petretto et al. 2008). SNPs influencing gene/protein expression level are referred to as gene/protein expression quantitative trait loci (eQTLs/pQTLs) and genes amenable to genomic regulation are called eGenes (GTEx Consortium 2015). Encouragingly, trait-associated SNPs are three times more likely to be associated with gene expression (Nica et al. 2010; Nicolae et al. 2010; Fehrmann et al. 2011; Hernandez et al. 2012), suggesting strongly shared genetic mechanisms between molecular and complex traits.

CAUSAL INFERENCE

The presence of a causal effect of an exposure on an outcome implies that a modification of the exposure (and the exposure only) would lead to a change in the outcome. Elucidating the existence and the strength of a causal effect is inherently hard because it requires an exposure-specific intervention. Estimating causal effect is primarily hampered by confounding and reverse causation.

In the special case where the exposure is a DNA variant, reverse causation can be excluded (because of temporal considerations), but confounding can arise because of linkage disequilibrium (being only correlated to the causal variant), population stratification (Kang et al. 2008), indirect effects (e.g., parental [Kong et al. 2018] and assortative mating [Morris et al. 2019] effects), and selection or participation bias (Taylor et al. 2018; Hughes et al. 2019). Despite careful efforts to control for population stratification (by using ancestry principal component correction and mixed effect models), large meta-analyses of GWASs have still been influenced by confounding, as suggested by an emerging body of evidence for human height and education level (Abdellaoui et al. 2019; Haworth et al. 2019; Sohail et al. 2019). Family-based designs can be a solution to eliminate most of the confounding effects, which are often shared among siblings (Belsky et al. 2018; Davies et al. 2019).

When the exposure is not a genetic variant, limitations and biases (including residual confounding and reverse causation) in causal effect inference are far more substantial. This means that many of the associations found in classical epidemiological studies are mere correlates of disease risk, rather than causal factors directly involved in disease development (Fewell et al. 2007; Pingault et al. 2018). As nongenetic molecular traits are also vulnerable to the effects of confounding and reverse causation, challenges of causal inference also extend to the investigation of molecular traits as causal risk factors underlying disease.

For example, transcriptome-wide association studies (TWASs) have aimed at identifying genes whose (genetically predicted) expression is strongly associated to complex traits (Nica et al. 2010; Gusev et al. 2016). Later developments were based on a Bayesian sparse linear mixed model (Mancuso et al. 2018), genetic best linear unbiased predictor (GBLUP) (Mancuso et al. 2017), or elastic net (Barbeira et al. 2018). However, these studies were not designed to estimate the strength of the causal effect nor to distinguish causation from horizontal pleiotropy (i.e., when a genetic variant influences multiple phenotypes independently). Nonetheless, when the same genetic variant is associated with both the expression level of a gene and a phenotype, there is some evidence that the implicated gene is causally involved for that trait. To detect such overlapping associations, several methods have been proposed. To quantify the colocalization signal of gene expression and phenotype associations one can use Bayesian model comparison methods, such as various versions of coloc (Wallace et al. 2012; Giambartolomei et al. 2014; Fortune et al. 2015; Guo et al. 2015) and eCAVIAR (Hormozdiari et al. 2016). The disadvantage of such methods is that they are blind to the directionality of effect and sensitive to the presence of heritable confounding factors. A more principled approach based on model comparison aiming to disentangle various causal models has been long proposed (Millstein et al. 2009), but it is based on a single genetic marker as instrument.

Epidemiologists have sought to develop statistical methods to tease out causal effects based on observational data, which may lend themselves to evaluating causality of molecular phenotypes. A large family of causal inference methods is instrumental variable (IV) analysis (Lawlor et al. 2008). It requires that beyond the exposure (X) and the outcome (Y) we also measure an IV (G) in a population, which fulfils the following three assumptions: (1) Relevance—G is robustly associated with the exposure. (2) Exchangeability—G is not associated with any confounder of the exposure–outcome relationship. (3) Exclusion restriction—G is independent of the outcome conditional on the exposure (i.e., there is no path between the instrument and the outcome independent of the exposure). A special case of IV analysis is MR, where the instrument is a genetic variant.

In MR with risk factors that are not directly linked to the DNA sequence (e.g., not related to chromatin properties, gene expression, protein, or methylation levels), multiple genetic variants across the genome can be used as instruments to look for a homogenous effect across all instruments (Palmer et al. 2012). However, if the exposure of interest is the expression level of a gene or a protein, then the optimal instrument(s) would be genetic variants lying within or near the gene coding for the protein itself (i.e., in cis). In this way, the presumed effect of the instrument on the outcome can only possibly be through the exposure.

As listed above, many genetic studies now combine genetic information of diseases with one or multiple molecular traits such as gene expression (Zhu et al. 2016; Porcu et al. 2019), metabolomics (Bell et al. 2020), protein data (Al Awam et al. 2015), and DNA methylation (Wahl et al. 2017) to gain a better understanding of complex disease etiology. It has been shown that a large fraction of trait-associated SNPs are also associated with molecular traits (i.e., molecular quantitative trait loci [molQTLs]) (Nica et al. 2010; Hannon et al. 2017; Sun et al. 2018), suggesting their potential involvement in the molecular mechanisms. However, understanding how such variants also influence complex traits is challenging. The most efficient way to address this question is to combine summary-level data from molQTL and GWAS studies in a two-sample MR framework to evaluate whether a molecular trait has a causal influence on a complex trait (Richardson et al. 2018). Given the complexity of human phenotypes, no single molecular feature is expected to fully capture all key biological mechanisms. Furthermore, molecular markers form highly complex networks and disentangling their individual roles in isolation is challenging due to largely shared instruments leading to contamination of horizontal pleiotropic effects.

MR STUDIES WITH OMICS EXPOSURES

Transcriptome-Wide MR

Recently, we adapted a multivariable MR method (Sanderson et al. 2019) tailored to gene expression exposure, termed TWMR (transcriptome-wide Mendelian randomization) (Porcu et al. 2019). Key improved features include approximate conditional analysis to select potentially correlated instruments and iteratively including genes with shared instruments for joint causal analysis. TWMR integrates summary-level data from GWAS and eQTL studies in an MR framework to estimate the multivariate causal effect of gene expression on complex human traits (Fig. 1). The robustness, validity, and interpretation of such a multivariable approach is described in Sanderson (2020). Because proximal genes often have correlated expression levels, multivariate causal effects are key to disentangle direct, mediated, and shared effects.

Figure 1.

Multivariable Mendelian randomization (MR) applied to expression traits. (eQTL) expression quantitative trait loci.

Applying TWMR to complex traits revealed thousands of putative genes with a causal effect on at least one phenotype. Notably, about one-third of these gene–trait associations were not prioritized by previous GWASs (i.e., no SNP reached genome-wide significance level within the gene ±500 kb). For example, while educational attainment GWAS entirely missed the BSCL2 locus, TWMR highlighted this gene as potentially causal (P_TWMR = 1.89 × 10⁻⁶). BSCL2 has been previously shown to be involved in type 2 congenital generalized lipodystrophy (OMIM [Online Mendelian Inheritance in Man]:#615924) (Guillén-Navarro et al. 2013), which has been associated with some degree of intellectual impairment. It is generally accepted by now that the gene in closest proximity to the lead GWAS SNP may not necessarily be the key player. For instance, in the height-associated 2p21 region (Wood et al. 2014), TWMR results suggested that SOCS5 is not the causal gene, despite being closest to the lead SNP. Instead, TWMR revealed that high expression of CRIPT is causally linked to high stature. Loss-of-function mutations in CRIPT are known to be associated with short stature (OMIM:#615789) (Shaheen et al. 2014), making it a plausible candidate.

Interestingly, TWMR-implicated genes show global trends for functional relevance, defined as being linked to a more severe version of the same trait in the OMIM database. For example, height-, total cholesterol-, and educational attainment–associated TWMR genes are 1.3/3.7/2.6-fold enriched for being linked by OMIM to abnormal skeletal growth syndrome, hypercholesterolemia, and cognitive impairment, respectively. These results provide a hint that both mild (e.g., modified expression level) and strong (e.g., protein truncating mutation) perturbation of the same genes may impact the same trait, but to a different extent.

Metabolome-Wide MR

As emerging technologies have made feasible the assessment of hundreds of metabolites, many studies explored the role of metabolites in several diseases (Sabatine et al. 2005; Shah et al. 2010; Wang et al. 2011). Additionally, many GWASs on metabolite concentrations (metabolite GWAS [mGWAS]) have been performed to investigate how genetic variants affect changes in metabolite levels in human urine (Raffler et al. 2015) and blood (Shin et al. 2014). Such large studies enable two-sample MR approaches to estimate the causal effect of metabolites on complex traits using metabolite QTLs (mQTLs) as genetic instruments.

MR studies have been particularly important in providing insights into the role of low-density lipoprotein cholesterol (LDL-C) on coronary artery disease (CAD) development. For example, many independent trials have shown statins to reduce LDL-C levels and risk of CAD, proportional to the dose of the statin, owing to the causal link between LDL-C and CAD (Cannon et al. 2004). Using all known genetic variants associated with LDL-C as instruments, the causal role of LDL-C has been substantiated by MR (Ference et al. 2012). Furthermore, MR models have been successful in predicting the effect of specific LDL-C-lowering drugs by restricting the analysis to the gene target (e.g., NPC1L1) of the drug (e.g., ezetimibe) in question (Holmes et al. 2013; Myocardial Infarction Genetics Consortium Investigators et al. 2014; Ference et al. 2015; Würtz et al. 2016). Moreover, MR studies have encouraged the development of novel drugs, such as PCSK9 inhibitors, which have been recently shown to reduce cardiovascular events in phase III trials (Giugliano and Sabatine 2015; Sabatine et al. 2017). Further details on these links and broader discussions on the role of MR in drug repositioning can be found in the section “Omics-Based MR to Boost Drug Development and Repositioning” below.

MR analysis also pointed to a potential causal effect of serum leucine concentration on T2D, which has been long known to be an important biomarker (Melnik 2012). High triglyceride levels have also been confirmed by MR studies to be a risk factor for T2D, while another longitudinally implicated biomarker (Vangipurapu et al. 2019), alanine, seems rather to be a consequence of diabetes (Liu et al. 2017). To our knowledge, systematic comparison between MR and longitudinal observational studies is scarce and whereas a rare attempt showed good concordance (Würtz et al. 2014), it still lacked directionality by comparing exposure change to outcome change.

Proteomics MR

Blood-based biomarker measurement is a mainstay for patient management and fundamental for diagnosis, prognosis, and treatment. Biomarkers that are known to be causally linked to a disease and can subsequently be modified are key to lowering a patient's risk. Because of the limitations of most alternative approaches, MR has become an increasingly common technique to identify and shed light on causal relations between protein biomarkers and disease.

We have applied MR to identify novel, causal mediators of CAD in the ORIGIN Trial by examining a comprehensive panel of 237 biomarkers (Sjaarda et al. 2018a). Using genetic variants residing at or near the gene for each biomarker under study, MR analysis revealed six biomarkers, in which there was evidence for a causal effect on CAD, including two novel markers: CSF1 and CXCL12. Both biomarkers had been previously linked to inflammatory processes characteristic of atherosclerosis in both animal models and human studies, consistent with MR results showing a causal link with CAD. These findings were also consistent with results from the CANTOS study (Ridker et al. 2017), which showed that an intervention aimed at decreasing inflammation through interleukin-1β inhibition can lead to lower rates of recurrent cardiovascular events. Together, these studies shed light on the role of inflammation in CAD and highlight the utility of MR in revealing specific protein markers directly involved in disease progression that could be targeted via pharmacological inhibition.

Another MR study revealed new mediators of chronic kidney disease (CKD) whereby human epidermal growth factor receptor 2 (HER2) and uromodulin (UMOD) were both identified as causal mediators of CKD (Sjaarda et al. 2018b). Further MR exploration of the HER2 pathway also revealed ACE as a regulator of HER2 levels. These results implicate HER2 not only as a mediator of ACE inhibitors’ protective effect on CKD, but also as a marker to identify patients who would benefit from ACE/RAAS inhibition. Furthermore, these findings suggest HER2 inhibitors (e.g., gefitinib) as a potential novel treatment for CKD.

Recent proteome-wide work (Zheng et al. 2019) has probed the causal effect of over 1000 proteins on 225 phenotypes and identified 105 putatively causal effects of 64 proteins on 51 phenotypes. Importantly, they have demonstrated that protein–disease links supported by MR and colocalization are far more likely to indicate potentially successful therapeutic targets.

Methylome-Wide MR

As epigenetic processes are putative intermediate mechanisms between socioenvironmental exposures and health outcomes, epidemiologic studies of epigenetic marks may shed light on the biological pathways embodying exposures or provide biomarkers of exposures alternative to self-reported questionnaires (Relton and Davey Smith 2015; Sharp and Relton 2017). Applications of MR in epigenetics have been limited to DNA methylation as, so far, it is the most common and simplest epigenetic process measurable in epidemiologic studies. To perform MR, cis genetic variations related to DNA methylation levels (i.e., mQTLs) are typically used as instruments. Traditional bidirectional MR has been applied to interrogate exposure–methylation associations, such as maternal hyperglycemia (Allard et al. 2015), adiposity (Richmond et al. 2016), or methylation–complex traits associations (Richardson et al. 2018). Other studies have applied a two-step MR framework (Relton and Davey Smith 2012) to elucidate the mediating role of methylation between the exposure–outcome association of interest, such as obesity and cardiometabolic diseases (Mendelson et al. 2017). A recent systematic review of MR studies suggests that DNA methylation may be mediating the causal effect of pre- and postnatal exposures to tobacco smoke on birth weight and inflammation markers, respectively, and prenatal exposure to vitamin B12 on cognitive outcomes (Grau-Perez et al. 2019).

MR has mostly been performed only on a selected set of methylation loci and a limited number of exposures and health outcomes. The reason for this is that currently available platforms (e.g., Illumina Infinium HumanMethylation450 or MethylationEPIC BeadChips) can tag only a small subset of CpGs (∼3% genome-wide). In addition, samples with genome-wide methylation and genomic data are still smaller than those with gene expression measurements, but the gap is closing rapidly (Huan et al. 2019), greatly catalyzing methylation MR studies.

CURRENT LIMITATIONS

The Curse of Unmeasured Heritable Confounders

In the case of most MR methods (using multiple instruments), the exclusion restriction assumption can be replaced by the weaker InSIDE assumption, which requires that instrument strength is uncorrelated to the direct (horizontal pleiotropic) effect on outcome. Despite pleiotropy being pervasive, the InSIDE assumption is reasonable when the (alternative) pleiotropic pathway from the instrument to the outcome does not involve the exposure. However, in the presence of a confounder (V) of the exposure–outcome relationship (Fig. 2), all SNPs associated with V would serve as instruments for the exposure (X) and their effect on X will be proportional (q_y/q_x) to their effects on the outcome (Y). In classical MR, all instruments G for X would be included in the analysis, both direct (G_x) and indirect (G_v). As the genetic basis of the confounder becomes more prominent (i.e., the confounder is heritable), increasing numbers of SNPs will yield such biased estimates. Therefore, the estimated causal effect, which would be a weighted combination of the true causal effect (α) and (q_y/q_x), will become more biased. To evaluate this bias, more sophisticated methods (Darrous et al. 2020; Morrison et al. 2020) or functional validation is necessary (Lepik et al. 2017).

Figure 2.

A key violation of the InSIDE assumption of Mendelian randomization. An unmeasured heritable confounding factor V and separate direct (π_x) and indirect (q_xπ_v) genetic effect for X for a set of single-nucleotide polymorphisms (SNPs) G_x and G_v. The causal effects of V on X, V on Y, and X on Y are denoted by q_x, q_y, and α, respectively.

Difficulty of Gene Prioritization

Shifting from the classical single-gene view of complex traits biology to a pathway/network perspective can help us to better understand the etiology of the variation in human phenotypes. Pathway enrichment analyses can pinpoint biological mechanisms based on gene lists usually created based on physical proximity to GWAS hits (Pers et al. 2015; Lamparter et al. 2016). Combining GWAS-implicated genes with causal genes could improve the prioritization and subsequently boost the power to detect enrichment in relevant pathways and regulatory networks. However, to do so, several hurdles need to be overcome.

First, TWAS/TWMR approaches are currently limited to only the 17 K established eGenes (Võsa et al. 2018), which substantially decreases power to detect enrichment of the prioritized gene set in relevant pathways and regulatory networks. We expect that larger eQTLs studies will allow for the identification of additional eGenes resulting in stronger enrichment when using causally associated gene sets, rather than selecting genes based on physical proximity to GWAS hits. Despite these promising developments, many genes lead to disease not through change in their gene expression, but via other mechanisms (e.g., modification of the RNA or protein sequence [Marouli et al. 2017]) to which TWAS/TWMR are blind. This problem can be mitigated by the simultaneous application of multiple omics MR. Whereas omics data provide quantifiable readouts for the impact of SNPs, it becomes less straightforward to quantify the impact of coding variants. The suitability of different measures of coding variant severity (based on conservation, pathogenicity, protein folding property, etc.) could be tested in the MR framework.

Second, as opposed to standard GWAS where magnitude of SNP-trait correlation and evidence (P-value) have a monotonic relationship, MR P-values and causal effect estimates across genes are not directly comparable. The reason for this is that the evidence for causal effect depends on the number and strength of instruments and the magnitude of the effect. Therefore, it is not clear whether genes should be prioritized based on estimated causal effect size or P-value. Moreover, the threshold above which a causal effect size is deemed of clinical interest may vary from gene to gene.

Third, a further limitation of TWAS/TWMR is the presence of high correlation between coregulated genes, which in turn results in shared eQTLs. Many of these shared eQTLs may represent direct effects on a common transcription factor and indirectly relate to the gene of interest. Larger eQTL studies may be able to reveal unique instruments for such genes and hence disentangle their multivariate causal effects.

Tissue Specificity

Typically, the methods integrating data from GWASs and omics data are focused on eQTLs/mQTLs/pQTLs in whole blood, because it is the easiest tissue to collect and summary statistics from large studies are available (e.g., Westra et al. 2013; Võsa et al. 2018). However, because gene regulation is tissue-specific and many diseases manifest themselves only in certain tissues, the possibility to interrogate more tissues could unravel causative genes whose whole blood expression is not disease-relevant.

For example, the causal effect of SORT1 on LDL levels is only detectable when eQTLs derived from liver are used (Porcu et al. 2019; Richardson et al. 2020). Furthermore, FBN2 expression is driving systolic blood pressure in heart tissues, but is associated to forced vital capacity when using lung eQTLs (Richardson et al. 2020). Finally, for CAD, TWMR pointed to MRAS and PHACTR1 as causal genes exclusively in arterial tissues (Porcu et al. 2019).

The gene expression data collected in 54 tissues by the GTEx Consortium (Aguet et al. 2019), ranging in sample size up to 706 (muscle) genotyped individuals, provide an extremely valuable resource for tissue-specific analyses. With the increased sample size, increased cis allelic heterogeneity has been observed (providing more independent instruments) and more robust allele-specific expression QTLs have been identified. The new data have also revealed fundamental differences in the genetic architecture of gene expression and splicing. Importantly, this work identified cell-type composition as a key driver for tissue-specific eQTLs. Unfortunately, the sample size for any given tissue in GTEx is still >30 times lower than meta-analyzed data from whole blood by the eQTLGen Consortium (Võsa et al. 2018). This represents a considerable limitation to identifying tissue-specific causal genes given the limited number of eGenes shared between the tissues. However, whereas blood may not necessarily be the causal tissue, gene expression in blood may be a sufficient proxy for expression levels in other tissues. For example, cis eQTL effects are highly correlated (r = 0.7) between brain and whole blood for genes expressed in both tissues (Qi et al. 2018). Until larger tissue-specific samples become available, maximizing the available resources in whole blood studies can thus be a viable strategy.

Cis versus Trans QTLs as Instruments

It is estimated that ∼70% of gene expression heritability is via trans effects, which is probably induced by the modulation of upstream genes (Liu et al. 2019). This observation is compatible with an underlying model whereby only a relatively small number of (partly correlated) core genes have direct effects on gene expression, while the bulk of eQTLs exert their effect through regulating these core genes. Thus, using cis eQTLs would be the most appropriate variants to estimate the impact of gene expression, because trans eQTLs are more likely to represent indirect effects and hence are subject to pleiotropy, thus violating MR assumptions. Still, a potential weakness of using cis eQTLs as instruments is that the signals they represent are all derived from the same region. While large eQTL data sets provide evidence that most genes have several statistically independent eQTLs, they may still be influenced by shared haplotype effects (i.e., a common confounder). In addition, cis eQTLs are sometimes shared among neighboring genes, hindering the distinction between causal effects of coregulated biomarkers. Identifying independent and not shared secondary associations in a multivariable MR setting can circumvent such issues, but this requires larger sample sizes (Porcu et al. 2019).

FUTURE DIRECTIONS

Sex-Specific Analysis

The extant literature shows that sex could modify the effect of causal variants (Ober et al. 2008; Randall et al. 2013; Winkler et al. 2015). Such sex-specific associations could arise as a result of sexual dimorphism in gene/protein expression. To explore this hypothesis and to better understand the genetic basis of sexual dimorphism, applying a sex-specific omics-MR, combining sex-specific summary statistics for both omics-QTLs and complex human traits would be necessary. Such analysis would reveal whether sexual dimorphism observed for complex traits (such as waist-to-hip ratio) is already present at the transcriptome/methylome/proteome level or appears only downstream.

Diseases Modify Gene Expression

Most TWMR efforts have focused on using cis eQTLs as instruments to tease out the causal effect of gene expression on a complex trait. The (causal) impact of diseases on the transcriptome program has only very recently been investigated in a large eQTL study (Võsa et al. 2018). It was found that disease-associated genetic variants affect expression levels more often in trans than in cis. Interestingly, only 4% of trans eQTL effects could be explained by mediation of a cis eQTL effect, indicating that in addition to gene–gene regulation, many other nontranscriptional stimuli play a role in gene expression modification. These facts imply that diseases may have more pronounced impact on gene expression than the reverse, thus comparing gene expression levels of diseased and healthy subjects may reveal the transcriptomic fingerprint of a disease. In fact, the identification of such differentially expressed genes (DEGs) has long been the prevailing approach to tease out disease biology (e.g., Rodriguez-Esteban and Jiang 2017). However, DEGs may be causes, consequences, or mere correlates of the disease under scrutiny.

As a proof-of-concept, we asked how highly TWMR-identified causal genes would rank in a DEG analysis. To address this question for LDL cholesterol, we computed the correlation between the expression levels of causally implicated genes for LDL (by TWMR) and the actual LDL level in an Estonian population (N = 490), part of the EGCUT biobank. We found that TWMR-implicated genes have only marginally stronger correlation with LDL than a random gene set of the same size (Fig. 3).

Figure 3.

QQ plot of the correlation P-values. Observational correlations were calculated between low-density lipoprotein (LDL) levels and whole blood expression levels of causal genes for LDL (by transcriptome-wide Mendelian randomization [TWMR]) in the EGCUT study (in red). Correlations were also calculated between LDL levels and a set of random genes (in blue).

Although the gene expression study was very small and many more traits need to be studied, the results indicate that analyzing DEGs might reveal disease-induced changes in the transcriptome rather than disease-causing genes.

Extension to (Gene Regulatory) Network of Causal Effects

Despite the fact that the basic network MR concept has been introduced (Burgess et al. 2015), its application to real biological/clinical data remains scarce. MR could be extended to estimate bidirectional causal effects for every pair of eGenes and then iteratively identify and eliminate indirect edges through a generalized version of summary statistics-based mediation analysis (Burgess et al. 2017). The resulting gene–gene regulatory causal network should be comparable to tissue-specific regulatory circuits built based on promoter-enhancer activity (Marbach et al. 2016). Networks should be drawn for each layer of omics data and their relative causal relationship to complex traits needs to be examined via mediation analysis, revealing direct and indirect effects.

Causal networks would allow the dissection of direct and indirect effects in GWAS by adjusting the observed SNP-trait associations by the sum of the total SNP effects acting through all the incoming causal edges to that focal trait. In a previous work (McDaid et al. 2017), we used incoming causal edges to create prior effect estimates, but these priors could as well be subtracted from the observed effects to classify SNPs into core and peripheral genes—such a distinction has been promoted by the omnigenic model (Boyle et al. 2017).

More recent work on the omnigenic model proposed trans eQTLs as an explanation for the observed complex trait architecture (Liu et al. 2019). More research is needed toward this direction as their proposed model ignores reverse causality from trait to gene expression and assumes that most trans eQTLs are downstream effects of cis eQTLs, which has not been convincingly evidenced (Võsa et al. 2018). Linking eQTLs and traitQTL in a causal model setting has gained attention. An elegant extension of the LD score regression allowed the estimation of the fraction of trait heritability propagated through cis gene expression regulation (Yao et al. 2020). They estimated that ∼11% of trait heritability could be explained by cis eQTL regulation on average across 42 traits, with up to 30% for CAD.

Omics-Based MR to Boost Drug Development and Repositioning

It has been tempting to exploit disease-associated loci identified by GWAS to aid drug discovery (Yin et al. 2018). For this, not only disease onset, but disease progression genetics need to be exploited (Paternoster et al. 2017). Under ideal settings, the effect of an SNP might share mechanisms with the impact of a drug. Therefore, if an allele of a disease-associated SNP is predisposing to disease through increased gene/protein expression, a drug suppressing the level of the same gene/protein may be beneficial for that disease. Several examples exist where GWASs identified disease-relevant genes that are targets of efficient drugs. These include the statin-targeted HMGCR gene, which is associated with LDL levels (Swerdlow et al. 2015), psoriasis, and inflammatory bowel disease–associated IL12, T2D-associated ATM, and the LDL-associated PCSK9 gene (Robinson et al. 2018). In addition, phenome-wide association studies (pheWASs) may support the elucidation of side effect profiles (e.g., Neuraz et al. 2013).

Recent genome-wide approaches proposed to match a pharmagenic enrichment score of a drug with the polygenic risk score for a disease (Reay et al. 2020). Others (So et al. 2017) used MetaXcan (Barbeira et al. 2018) to impute genetically determined expression in disease cases versus controls. The differential gene expression is then contrasted to the impact of a drug on the transcriptome profile (obtained from the Connectivity Map [Subramanian et al. 2017]).

Notably, many of these approaches can lend themselves to predict off-target effects of drugs. For example, a recent MR study (Richardson et al. 2020) has shown that high HMGCR expression also lowers T2D risk, explaining why HMGCR-lowering drugs (e.g., statins) present diabetes-related side effects. The same study also demonstrated that ACHE—a target for several Alzheimer's medications—expression is positively correlated with blood pressure.

CONCLUDING REMARKS

With the increasing sample size of GWAS and sequenced reference panels, fine-mapping frequently yields credible sets containing only a handful of causal genetic variants. Still the functional characterization and the molecular consequences of those genetic variants mostly remain unclear. Although mapping leads SNPs to genes based on physical distance has already revealed meaningful insights (especially for coding variants), it is an oversimplification of the underlying biology. The emergence of large omics data sets (transcriptomics, proteomics, methylomics, metabolomics, chromatin states, etc.) with genomic information have led to the discovery of tens of thousands of omics QTLs (eQTL, splicing QTLs [Aguet et al. 2019], single-cell eQTL [Van der Wijst et al. 2019], pQTL [Suhre et al. 2017], mQTL [Bonder et al. 2017], meQTL [Shin et al. 2014], and cQTL [Delaneau et al. 2019]). The central challenge is to combine these QTLs with GWAS results of complex traits in a tissue-specific manner (Aguet et al. 2019). Clearly, competing risk factors need to be simultaneously included in a multivariable MR setting (Sanderson et al. 2019; Sanderson 2020). The MR framework lends itself to extensions integrating various sources of information and generate hypotheses on a massive scale. The novel insights gained from such an approach has tremendous potential to reveal the basis of the underlying genetic network, which in turn can be leveraged to boost drug repositioning and discovery.