#eqtl

[doctorveera]

Every MR study teaches something new

A Mendelian Randomization (MR) study published in Nature Medicine reports a causal association between blood levels of OAS1 and susceptibility to COVID-19 infection. I liked many things in this study, of which I’d like to highlight three.

First, I liked that the authors performed the MR analysis using not just pQTLs, but also eQTLs and sQTLs as genetic instruments. In doing so, the authors reveal the mechanism underlying the association between the genetic variant and the OAS1 protein level in the blood, which, in turn, offers protection against COVID-19 infection.

In a typical scenario, one would expect a genetic variant to increase the protein level by increasing the mRNA level, considering the central dogma of biology (i.e. we expect an increased translation as a consequence of an increased transcription). But often the scenario is not that simple. Many of the pQTL studies published so far have shown that often pQTLs are not eQTLs. That is, often genetic variants influence protein levels without influencing the mRNA levels. The current study’s finding is one such example.

Here, the genetic variant increases the OAS1 protein level not by increasing the OAS1 transcription, but by favoring the OAS1 splicing towards one particular isoform. This is a scenario where a pQTL is not an eQTL but an sQTL. So, this finding teaches us that genetic variants can influence protein levels without influencing the transcription, but instead by influencing the splicing.  The authors have provided a table of posterior probabilities for pairwise colocalizations across eQTLs, sQTLs, pQTLs, and COVID-19 genetic associations, which clearly shows that there is no colocalization between eQTLs and pQTLs or between eQTLs and COVID-19 genetic associations.

The second thing I liked about this study is it draws a clear causal pathway from genetic variant to mRNA to protein to disease risk. This means it is an opportunity to explore how the effect sizes change as we move from one end of a causal pathway to its other end. For e.g. I’d expect that the effect size of the genetic variant’s association with splicing should be larger than its association with protein levels, which in turn should be larger than its association with disease risk (assuming that the genetic variant’s effect on disease risk is mediated entirely via its effect on the protein levels). However, I wasn’t able to retrieve the eQTL’s and pQTL’s effect sizes, which were disappointingly not reported in the study, and the reference pQTL study has not released the summary statistics publicly.

Thirdly, the authors report that the phenotypic association between OAS1 blood level and COVID-19 risk reverses between infectious and non-infectious states. While the non-infectious state OAS1 levels decrease the risk, the infectious state OAS1 levels increase the risk for COVID-19. This is expected as the viral infection itself causally increases the OAS1 levels in the blood. So, here, by using MR, we are able to dissociate the effect of OAS1 levels on COVID-19 risk from the effect of COVID-19 infection on the blood OAS1 levels, which, I think, is a great example of the wonderful things that we can do using human genetics.

I like well-conducted studies like this because they teach us things that are beyond their original aims, things that are fundamental to all kinds of human genetic studies.

[doctorveera]

GWAS drug targets glitter only in hindsight

A GWAS of peptic ulcer disease (PUD) published in Nature Communications reports an interesting drug target locus. The index variant at the locus is located upstream of gene CCKBR that codes for cholecystokinin receptor B, which is a known drug target. Proglumide, an old peptic ulcer drug, is a cholecystokinin antagonist, and it acts by blocking cholecystokinin receptors A and B sub-types. This a nice example of a GWAS locus supporting existing drug targets.I have tweeted similar findings before (ex1, ex2)

A quick lookup in GTEx database revealed an eQTL association between the index variant rs10500661 and CCKBR, but only in the testis tissue. However, I couldn’t stop noticing that the direction of the eQTL association align with rs10500661’s direction of association with PUD risk and proglumide’s mechanism of action. The effect allele T, which decreases the CCKBR expression, has a protective effect on PUD (OR-0.90). And this aligns with the fact that proglumide treats PUD by blocking the receptor protein.

If you think about it, what we are doing here is reverse engineering. We are leveraging the knowledge about the drug proglumide and pinpointing the causal gene (CCKBR) and the disease variant’s functional consequence on the causal gene (decreasing the CCKBR expression). Though this has no translational benefit, it is kind of cool, isn’t it? Perhaps, all such loci with known drug targets could be curated and used as a training set for machine learning methods that are recently being developed to identify GWAS causal genes.

Learning about this CCKBR locus branched out two streams of thoughts in me—one, about eQTLs and the other, about GWAS loci leading to drug discoveries.

I have been learning about the role of eQTLs in human diseases for quite some time now, and my understanding so far is this: disease relevant eQTLs are a drop of true signals mixed in a big bowl of noise. The noise doesn’t necessarily mean that the eQTLs themselves are false positive signals. Most of the eQTLs in fact represent true associations between the genetic variant and the mRNA level. However, how many of those genetic effects are propagated to protein, to protein’s molecular function, and finally, to the end phenotype is not clear. Hence, I am often inclined to trust eQTLs when they are presented not as the lead evidence, but as a supportive evidence such as the one I discuss in this post. I am sure I have seen many such examples before, and have tweeted about it. But for now, I am able to recollect one, which remains fresh in my memory because of a peculiar characteristic of the finding. In this paper, the authors found a significant association between a loss of function mutation in gene GIGYF1 and type 2 diabetes (OR - 4.15) in the UK biobank, and interestingly, they also found a GWAS locus containing GIGYF1 close to the index variant. And it turns out that the index variant is an eQTL for GIGYF1, and the effect allele increases the GIGYF1 expression, and has a protective OR (0.96). This aligns with the rare variant association, where loss of function of GIGYF1 increases the diabetes risk. So, what we are seeing here is that human genetics is offering evidence that if you knockout the gene, the risk of T2D increases, and if you increase the gene expression, the risk decreases. That’s something we don’t get to witness often. So, it stayed in my memory. The point is, in such scenarios, it’s hard to discard the eQTL evidence as irrelevant, isn’t it ? And that’s why I tend to trust eQTLs when it is presented as a supportive evidence, but not when when it is presented as the lead evidence. This is also the reason why my enthusiasm for transcriptome wide association studies (TWASs) that was once in the peak has started declining lately. And recently, my enthusiasm for proteome wide association studies (PWASs) is in the rise. I have tweeted about one recently. But it looks like PWASs too might follow the same trajectory as the TWASs.

The second train of thought that I had while reading this paper is about turning GWAS discoveries into drug discoveries. One question that lingers in my mind whenever I stumble upon a GWAS locus containing a known drug target is this: if this drug has not been discovered already, would this GWAS finding lead to a successful drug discovery ? Finding the answer for this question is very important. People who firmly believe that GWAS discoveries will lead to drug discoveries often base their beliefs on examples such as the ones that I mentioned earlier. Particularly, these examples are used to justify the importance of minuscule effect sized loci. Often, a comparison is made between the GWAS locus’s effects size and the existing drug’s effect size. The GWAS locus represents a trivial disruption of the gene hence the small effect size. The same gene can be disrupted strongly using drugs, which will lead to a large effect size. And, one particular paper is often being referenced in such scenarios--a Plos Genetics paper that retrospectively shows that, compared to drugs without genetic evidence, drugs with genetic evidence are more likely to succeed in clinical trials and get FDA approval. All these are convincing points, but my only worry is all these evidences make sense only in the hindsight.

The situation will likely turn out entirely different if the starting point of the drug pipeline is a GWAS locus. One of the important point that many miss is that the GWASs spew out hundreds or even thousands of minuscule effect size loci. One of these loci might harbour a drug target. But how would one isolate that out? This is like pulling a needle from a haystack. The needle becomes visible when the drug has already been developed. It is easy to trace back from the drug to the gene target in the GWAS locus. To do the reverse is, however, extremely challenging, unless the target gene is located in a core pathway of the disease, or if one can find supportive evidence from rare variants. And that is one main reason why the drug companies are keen on sequencing large biobanks. A large effect loss of function variant association in combination with an extremely small effect GWAS association is a powerful evidence that the gene might be a potential drug target. These large effect rare variants act like a large magnet that pulls the needle out of the haystack. Without such magnets, the GWAS loci with potential drug targets will remain deeply buried inside the haystack, and will surface only occasionally when someone realizes that a drug that targets a gene located within the GWAS locus already exists.

[doctorveera]

eQTLs--Europeans vs East Asians

A new preprint reports a ~84% correlation in the brain eQTL effect sizes between Europeans and East Asians. It’s amazing that the authors were able to demonstrate such a high correlation even with a small sample size (EAS N=145; EUR N=397). Imagine performing a GWAS in such small sample sizes in East Asians and Europeans and comparing the effect sizes between the two. It’ll be impossible to get such high correlations (because of the huge standard errors). This is something I often tweet about. Molecular phenotypes such as gene expressions, protein levels etc are highly proximal to the genetic variants . So the effect sizes of the genetic variants tend to be extremely large that just a few hundred samples will provide sufficient statistical power to identify associations. Also, the measurement errors tend to be small, hence minimal phenotypic heterogeneity compared to usual GWAS phenotypes. This study is a nice demonstration.