Skip to main content
Loading Explore guides…
Mendelian Randomization · Acedra
Home
Explore
Biology
Mendelian Randomization
Sign in to save
Chat
If YouTube asks you to sign in, open the video on YouTube.
Watch on YouTube
Reload player
Transcript
57 segments
Copy
Hide video
0:08
Can you see that. Yes, Hi there, my name is Joey Ward, thanks for tuning in today for this talk on Mendelian Randomization it's overview and uses. So, in this talk we're going to cover a bit of the background of why mendelian randomization is necessary and why it has an advantage over more basic association analyses like Linear Logistic Regression it will cover the issues of interpretation and association analysis
0:40
things like spurious correlation reverse causality and confounding and what mendelian randomization does to overcome these issues. Then we will look at how mendelian randomization can act as like a natural randomized control trial and we're also going to find out how to select genetic variants to us and our analyses that then we're going to look at what's actually going on in mendelian randomization analysis,
1:05
we’ll cover the assumptions of the instrumental variable and issues that can arise will selecting these genetic variants for the instrumental variable things like weak instrument bias and play atopy and we'll look at some of the simpler types of mendelian randomization techniques of which, there are now quite a few just so you can get a feel for what's actually being measured and then finally we'll look at a few real-world examples of how these techniques can be used.
1:33
So we all know that the basic experimental design is to do a bit of background meeting to form a hypothesis you then gather some data and run some association analysis again like a linear logistic model and then you hope and pray that your P is less than 0.05 after you've collected from all protesting but interpretation of these models can run into difficulty and, as we all know that correlation isn't necessarily causation So what are the kind of things that can go wrong.
2:01
So firstly we can find spurious associations there, and you can see here there's an unbelievably high R squared for per capita for cheese consumption and death by bedsheet entanglement and I’m sure if any of us found an R squared have over 90% and our analysis would be absolutely over the moon, the chances of this being causative ae pretty slim and similarly, the number of people drowning in pools correlates very well with the number of films that Nicolas Cage
2:26
appears in each year, and even when you consider the dive in the quality of his outputs since about 2005 it's still be unlikely that will lead to more people throwing themselves into a pool and drowning. We've also got the issue of reverse causation so using the example here of BMI over time, it can be assumed that weight loss will eventually lead to dementia however it's more likely that weight loss, maybe an early indicator of dementia and the declining cognitive faculties
2:53
may lead to weight loss before the condition get serious enough to warrant a clinical diagnosis and then finally we've got the issue of confounding where another variable known as the confounder affects both the predictor and the outcome variable and it could be concluded that yellow teeth causes lung cancer, if the confounder was smoking, which both yellow’s teeth and causes lung cancer isn't considered. So, how does MR help with all this, so instead of just looking at the association between
3:25
two phenotypes allows us to make a causal estimate of an exposure X on an outcome come Y and it does this by the use of an instrumental variable or IV, which is called Z in those in the form of genetic mutations So here we assume that the confounders you act both on the outcome X and on the exposure Y and another way to visualize what's going on yours compare another mendelian randomization analysis to a randomized control trial so, with a randomized
3:56
control trial you take your sample and you randomly allocate them into two groups. So one group gets to the treatment, for example, a drug, like a statin in the control group will either get nothing or a placebo we then see if the treatment group have lower LDL levels than the control group and overtime can see whether the cardiovascular event rate is lower in the statin group than in the control group and this would allow us to conclude
4:21
that statins lower the rate of CV events by lowering LDL levels in the blood and mendelian randomization works in a similar way, but instead of us choosing who goes in which group nature, does that for us. So we know the LDL levels will vary naturally from person to person and there'll be some people with lower LDL levels than others, which will be based in part on their genetics so if we can find mutations in the genome that can lower LDL levels, we can then compare
4:50
those with the Lord genetic propensity for the LDL against those with a higher genetic propensity and, as with the randomized control trial will get then conclude that lower LDL lab was the lead to lower CV events but due to the genetic makeup of the individuals but then, how do we know exactly what genetic mutations it is we want to use so in the genome there are positions where there's a difference in the base per sequence so, for example,
5:22
base position on 1 million on chromosome one does either are cytosine Cleo or Adilene Leal and then we'll look whether there's a change in the phenotype between those who have 2 Cleo’s those who have one C one Leal and those who have 2 Leal’s but we don't want to take just one of these genetic variations which we call single nucleotide polymorphous items or snips there's there are millions flowing through the genome, so what we do
5:48
is what's called a genome wide association study or GWAS and we look at all the snips from the top of chromosome one to the bottom of chromosome 22 and due to the number of tests that were performing, we need to set or significance threshold a lot lower than naught point naught five and to be considered significant, we need the p value to be more than 5 times 10 to the minus 8. You can also use mutations that have been found from candidate gene studies, there was
6:15
usually family studies and fame mapping of genes different mutations that correspond with a phenotype as well and so, this is an example of the output of Gwas which is called Manhattan plot this particular one is mood instability phenotype that I ran so along the bottom you've got chromosomes numbered one to 22 and on the y axis, the actual was the negative log 10 P value. So the higher up the Y axis the smaller the p value in any of the points above the red
6:47
dotted line, we consider that to be genome wide significant but you'll notice that the point cluster into spikes, and this is due to a phenomenon called linkage disequilibrium which I’m not going to go into detail here but what's important is that for mendelian randomization we don't need all the snips that make it above that line, we only want the snips at the very top of the spike so, know we know which snips we want to use and
7:12
we can move on to our MR analysis, it was not completely necessary but as preferable that this is done in an independent cohort of the same ancestry is our discovery Gwas. Once we know which snips, we want to use as our instrumental variable need to consider three assumptions. So the first assumption is that the genetic variants Z associates with the exposure X, now, given to these snips usually come from a Gwas where their P value is less than five
7:42
times 10 to the minus eight we should be okay but since then we're now looking at an independent sample some of these snips might not show the same strength of association and this can be for a number of reasons, like due to having a smaller sample size and if both the snips don’t strongly associate with exposure, then this can lead to what's called weak instrument bias, which will cause the results to tend towards the null and so to avoid weak instrument
8:07
bias the statistic of the snips exposure regression should be greater than 10. Then the second assumption is that the instrument is independent of confounders and again, this should be OK, as the confounders shouldn't affect genome type of the sample unless they are acting as some sort of selection pressure of E on Z which would skew the balance of Leal’s I mean we can just assume that U is independent of Z. Then the third assumption is that there's no independent pathway between our instruments
8:39
Z and our outcome Y, but there's two possibilities that could violate this assumption, so the first is known as vertical pleiotropy and that's whereby snip Z acts via some sort of secondary pathway, which is shown as the question mark on the left and diagram. The second way is a wee bit more complicated and there's moves horizontally, pleiotropy so there may be no direct link between Z and Y via secondary pathway but there is another
9:09
nearby snip which is denoted Z 2 which is in strong linkage disequilibrium with our instrumental variable and this secondary snip does impact the outcome via some unknown pathway, so in that instance we would see that the third assumption as being violated and the snip Z1 wouldn't be suitable as an instrumental variable. So now most of our analysis will use more than a single snip which we can group together to create prologetic risk scores are PRS.
9:38
So, the more snips are used in the score, the higher the likelihood that some of them will be violating this third assumption, either by vertical or horizontal pleiotropy so there is a balance to be struck between having a sufficient power in the score to be useful and not having so many snips that you are violating the third assumption to any significant degree. Okay, so now we've got our instrumental variable in the form of our snip PRS and we're ready
10:07
to find some causative associations, now, remember the thing we want to know what is the causative effect of X on Y the image on this diagram is donating Beta of X hat on Y so the X has got the wee hat symbol above it and if it's just the coefficient of this association, then we would say beta of X and Y. So, one of the simplest methods is called the two stage least squares estimator, and this is only allowed in cases with a continuous exposure and the continuous outcome.
10:34
So, first we need to calculate the association between our instrument PRS and the exposure and then, using the residuals from that analysis we plug that in as the productive for the outcome Y and this coefficient of the regression is the beta X hat Y that we're looking for, and this analysis is really easy to perform, and there are packages like SEM that have the functions allow you to do both steps in one go and other simple MR technique is the
11:07
Wald ratio, so you can only use us when the instrumental variables are single snip so you can't use pathogenic risk scores with this one and this one we measure the association between Z and X like we did before and, but we also look at the association between Z and Y so as the instrument and the outcome. So, once we have the two coefficients, we just divide the beta ZY by beta ZX and then that gives us the beta X hat Y that we’re looking for, but the problem of that method
11:37
is that you are limited to just a single snip so unless that's a very powerful snip, then the analysis is likely to be underpowered but what you can do is expand out on the method and by calculating the Wald ratio for several snips one at a time and then doing an inverse weighted, sorry an inverse variants weighted Meta analysis, you can then calculate the combined effect of all the snips you want to use as your instrumental variable this
12:02
will in turn increase the power of the analysis allowing you to use several snips as your instrument. So, a more sophisticated technique for MR is called MR Egger year we get to relax the third assumption of the IV and this method is only valid as long as the effects between your instrumental variable and your outcome are independent of the effects between your instrumental variable and your exposure. So here, what we need to do is we perform weighted linear regressions of our gene exposure,
12:38
so that’s Z on X and our gene outcome, and that’s Z on Y, and then the best fit line through that and the gap between the origin is in the Y intercept is a measure of the horizontal pleiotropy and the slope of the line is the beta X hat Y that we're looking for so to explain that, in the form of a graph which makes it easier to understand and so along the horizontal axis, we have the measure of the effect of the instruments on the exposure
13:08
and on the vertical axis, we have the effect of the instruments on the outcome, which is denoted by that gamma symbol. The dashed line is the true causal slope, the thing that we're trying to estimate, and then the solid red line shows the inverse variance weighted slope, notice that parts of it is constructed through the origin is there's an assumption of no pleiotropy. So then you've got the solid black line, which is called beta E and that shows the Egger
13:33
slope and as you can see, it runs parallel to the true slope and there's less steep than the inverse variance weighted slope is it's not constricted through the origin so it could be the inverse variance weighted slope is actually going to overestimate your causal association and then you've got on the right hand side sorry on the left hand side the beta not E, which is the measurement of the horizontal pleiotropy and, as I mentioned
13:57
earlier, the more snips you have in your IV the more likely as the third assumption is going to get violated, so it makes sense to relax assumption once your IV has a large number of snips. Okay so that was just a brief overview of some of the simpler techniques, but this is a rapidly expanding field and there are no lots of different variations of MR they've all got their own strengths and weaknesses and often what papers will do is use several
14:24
types of MR and to give a range of estimate, to show the robustness of the results kinda like sensitivity analyses. So now you've got an understanding of how to perform a medallion randomization here are some of the examples of how it can be used in the new world. So those with high levels of cholesterol in the blood or more, the risk of cardiovascular illnesses such as heart attacks and strokes and as a protein in the body copy PSK9 it
14:54
controls, the number of low density lipoprotein receptors which are proteins on the surface of cells in these receptors play a critical role in regulating blood cholesterol levels and there's a drug Evolcumab which is a PSK9 main inhibitor and it's a monoclonal antibody which has been shown to reduce LDLC levels by about 60%, which in turn is lead to reduction and risk of cardiovascular deaths, heart attacks and strokes over there was potential, but
15:23
non-significant adverse side effects like they diabetes and impacts on cognition. So now LDLC level as well, really naturally from person to person which is based on part and the mutations and their PSK9 gene. So what Schmidt et al did was they used snip to confer variations and PSK9 expression, so the act like a natural PSK9 inhibitor and to gauge the likely effective PSK9 in the additional diabetes risk so they used four PSK9 snips both individually and combined
16:01
as a score later on to show that they significantly do reduce circulating LDLC concentration, and then they looked at the fourth snip separately and also combined as a score to see the association between the instrument and the outcome and they found that there was a significant causal effect of lower DLC on type two diabetes status, with an estimated odds ratio of 1.29 the also found that there's a significant causal effect on higher fasting glucose level body we and waist to hip ratio.
16:37
So, using that same methodology we then looked at the potential impacts on cognition using the UK biobank and we looked at several cognitive assessments all taken it base line like reaction time fluid reason and the trail making test and although we had large sample sizes, we weren't able to provide any definitive evidence of the relationship with PSK9 genetic variants with cognitive traits and the best we got was in logged reaction time, which was only normally significant.
17:06
So, this analysis is all well and good, when you've got access to funding large data sets and ample storage space and processing power, but what can you do if you only have access to a laptop and an Internet connection. 2 sample mendelian randomization allows us to overcome these problems and estimate a causal effect by using only Gwas summary statistics for both the exposure and the outcome. The first step is to select the snips at the top of the peaks and the exposure Gwas and
17:39
then you find them in the outcome Gwas and it doesn't matter if they're significant in the outcome, Gwas, but they definitely do need to be in the exposure Gwas then you just need to plot, the effects of this snip on the exposure18mins on the horizontal axis and the effect of the snips on the outcome on the vertical axis and the best fit line through these points is the causal estimate, you have the exposure on the outcome, so in
18:06
that example it makes look pretty clear been the real world examples and looked up that messier. So, it has been shown that the those with higher educational attainment are less likely to develop rheumatoid arthritis and what Hang et al showed us an instrument of 373 snips from an educational attainment Gwas was indeed a negative causal effect of educational attainment on developing arthritis. So, as you can see, on that plot that are actually three causal estimates of the association
18:42
and we've got the red line, which is the inverse variance weighted slope we've got a green line, which is the weighted median which have not actually covered in these slides and then you've got the blue line, which is the MR Egger slope. So it's kind of hard to see in this plot, but the inverse variance weighted slope is steeper than there, MR Egger slope and again that suggests that the inverse variance weighted
19:04
slope does give an over estimate of the causal association compared to the Egger method and that's due to the presence of pleiotropy because they are using like you know, several hundred snips in this score and it will also be possible to do that same thing in reverse, so if you wanted to see if there's some sort of bi directional effective going on, you could get the top hits from the Gwas and treat that as the exposure and then look for those same snips in the
19:31
exposure Gwas and treat that as the outcome. Okay, so nobody can do, MR whether we've got a data set or not and, but I think is important is quickly cover experimental design of MR and for one sample MR I think the best design here is C which is to determine the snips to us as the instrument from an independent cohort than the cohort you're going to calculate your beta ZX and your beta Y’s in, this situation is going to minimize overfitting of the data and the models can be weighted
20:08
by the coefficients from the discovery Gwas. The downside is that it can lead to weak instrument bias if there isn't the same strength of association between the snips used an instrument in the exposure and but as possible to calculate both the beta ZX and the beta ZY’s from the same sample as shown in A, then it's also possible to calculate beta ZY in the whole sample but beta ZX in a random sub-sample of the cohort.
20:35
So this might be done due to phenotyping in the cohort and you may find that you've got lots of outcome data like CBD events only a sub-sample of the cohort would have MRI or biomarker data for to sample MR the best situation would be to have the beta ZX calculated in one cohort and then the beta ZY is calculated in a completely independent cohort and this is to ensure that the calculated beta’s are truly independent of each other and there's
21:04
no collider bias being introduced but then, as Gwas samples are getting larger and Meta analyses become more common, the chances of being an overlap between the two samples like shown in E may become apparent so that that can lead to overfitting and weak instrument bias somewhat seen in one example, MR and then the final example F, in that way occur when you're using MR for testing developmental origins so, for example when beta ZX is estimated
21:33
in pregnant women and beta ZY is estimated in their offspring or when one of the outcomes can only occur in one sex so things like cervical or prostate cancer. So mendelian randomization is by no means the be all and end all of analysis is just another tool we can use to show how two factors are related MR are limited by pleiotropy and the discovery Gwas and to performing a Gwas of the same type phenotype and different cohorts
22:05
may find different topics to use this instrument or variable and that in turn would lead to different estimates and possibly conflicting causal estimates and because the effect sizes are very small yours need very, very large sample sizes to find good signals often studies are meta analysed and can include several different cohorts and these can be from different ancestors of background, which again could introduce noise and error and MRs by no means
22:32
the finished article, the new methods are being constantly device and updated and the methodology isn't free from assumptions and if these assumptions are being violated, then our estimates are going to be inaccurate. So, in conclusion mendelian randomization allows us to move beyond associates, and for a causal association between two phenol types it can act as around natural randomized control trial, and that is going to save time and money and drug development and improve pharmacal
23:03
vigilance and you don't even need access to large data sets storage space and processing power how's it going I’m going to be done on a desktop computer. So that’s me at the end of my talk, I’m going to open up to questions, and if you could try and phrase them is yes, no questions because I’m trying not to cough my lungs out. Thank you very much Joey in the absence of a chair or I’ll you know I’ll chair the questions, thank you very much, does anyone have anything they'd like to ask Joey, as
23:42
he said, ideally, yes, no. So what one question I have Joey, which is unfortunately not very yes, no is to some extent does, so your first all these different methods and how the instruments will develop as project scores get better and so on, I guess it's the case, then that with more infrastructure you could just have this catalogue of mendelian randomization instruments that you just run through a single pipeline and it's just it's it becomes almost automated you think is a
24:24
realistic thing or? Yeah, but I think again, it all depends on the context, because as like because of a frequency theory from population that may need to become quite specialized so you can just pull a new snips for diabetes, it would need to be specific to the kind of cohort that you're looking at and, and again you're like this go back to the whole thing you can do the Gwas or you can find your instruments from the same sample you're looking at your
24:51
exposure in your outcome with then again that could lead to kind of more issues yeah I mean you could try and automated but you would also need to do a lot of kind of pre-screening and double check your work as you go and not just rely on the computer to do all. Well, I’m not interested, if a paper good automatically write itself that would be ideal. Yeah, that's called R markdown. Yeah, any other questions for Joey. There isn't it in the chat or that you know the record. Yeah.
25:44
I am, this is less of a question more of a comment I think we're into a really interesting period in epidemiology and specifically public specifically genetic epidemiology or MR has kind of become this additional source of evidence kind of there’s a nice paper in the BMJ I think Neil Davies all about where does MR fit in the kind of hierarchy of evidence and they can get it some not quite on the tier of randomized control trials in terms of causality
26:18
but they put it in your certainly higher up than your basic observational associations so we're at this really interesting time where we really do have methods which are developing in front of our eyes. Yes, I think that okay. Yeah just these methods are you are advancing at a very rapid rate and they are very clearly, enhancing our understanding of epidemiology and causality really from one year to the next, I mean like a one MR paper, where one of the co-author said or not, you shouldn’t
26:49
use that method, I was at a conference, where he said, everyone will be laughing at this method in a year to do this way it's really it's incredibly quick how its advancing. Yeah, I mean I would be interested to see like yeah just given that you're starting with different Gwas summary statistics you do pick different instrumental variables, for the same phenotype to see like how robust estimates are you looking at different discovery samples.
27:12
Good one I heard was them whenever you're coming with a method, apparently, the one they like, to use urate and gout. So, if you're if you're if you're urate method doesn't link urate levels to gout it’s a bit of a duffer. Part of me always wants to kind of do an MR of something we knew is not causal at all and, and you know, like something which, which we know is not true and just see what an MR method would.
27:39
Or brainsizig something it's something you've absurd like I would start the study which got Gwas estimates of participation and biobank I always wanted to turn that back on itself and see is participating and biobanks good or bad for your health, but I don’t know if folk would see the funny side. Then everyone would drop out. Yeah, well exactly. Anyway, and yes, I don't have any additional questions for Joey, grand well in the context of your sore throat I think you're off the hook.
28:18
So, no, thank you very much I’m sure people will look up the recording some point in the future, even if they weren't here today. So, thank you very much Joey, I don't know I’m talking like I’m not chair them yeah so, thank you very much Joey and I’m sure if anyone has any specific queries they’ll get in touch. Just. okay everyone see you in a bit, bye.
Here to help you learn
Sign in to keep a private conversation with your course tutor.
Ask me anything about the material…
Sign in to chat