Jonathan Edwards
Senior Member (Voting Rights)
Fascinating. Keep at it.
Genetics: Chromosome 17 CA10
Fascinating. Keep at it.
It's using an older assembly/coordinate system called GRCh37. The more recent one that DecodeME uses is called GRCh38. Defining where exactly a SNP is on a chromosome isn't an exact science, so as they learn more, they make updates to the positions.
"Liftover" is converting the coordinates to a different assembly. I used liftover for the "Getting up in the morning" GWAS to make it match DecodeME.
On gnomAD, if you look up a SNP, it'll tell you where that SNP is in the other assembly. For example, here's the page for the lead SNP of the neck and shoulder pain GWAS in GRCh37 coordinates, as reported in Table 2: https://gnomad.broadinstitute.org/variant/17-50259142-A-C?dataset=gnomad_r2_1
Partway down the page, it says:
Partway down the page, it says:
ScoutB
Senior Member (Voting Rights)
Thank you for finding/sharing this. These results are so fascinating. I hope I get less foggy soon so I can read more of the details.
One question for you forestglip (and maybe @jnmaciuch if you're not too busy) -- do we have any idea how much of the similarity in 'shape' of the DecodeME stats and the ease-of-getting-up stats is due to variants at all those elevated positions being inherited together? i.e. would we expect that all the dots I circled in orange tend to be inherited together? (In that case I guess any condition where one of the dots is elevated you'd expect to see them all elevated?)
(Crop from post #57)
(Maybe this question translates to is 'is this what linkage disequilibrium looks like in summary statistics'?)
Jonathan Edwards
Senior Member (Voting Rights)
I think that is probably right at first pass.
It intrigued me that if you have a batch of variants tightly linked - so that they produce a flat line of linkages - you cannot simply ask 'did SNP 45 or SNP 52 cause the risk' because 'cause' in this situation is a complicated concept dependent on what options are biologically available. If these variants always go hand in hand, even if you can in theory trace a transcription factor binding being critical at a particular point it may not be legitimate to attribute cause to any one variant (which might be a variant not in your SNP library but closely linked to those in the line I guess).
The others still have a clearer understanding of this than I do so let's see what they say.
Yes, this is showing linkage disequilibrium. The following plot actually shows the strength of LD between each of the variants in the plot with the lead variant (purple diamond).
The variants in red have very strong LD with the lead variant, so would be expected to show up very often in people that have the lead variant. Hence, they are just about as significant as the lead variant. The yellow variants have less LD with the lead variant, so the significance might be further off, as we see here.
Yes, if the causal variants in the two studies were two different high LD "red" variants, the plots would probably look pretty similar. In theory, coloc helps to mathematically determine the probability that there is a shared variant based on the overall pattern, which may subtly change even if the other study's causal variant is a high LD "red" variant.
With a set of variants that have near perfect LD with each other (LD~1), I would think it's probably very difficult to determine which one is causal.
I guess it may be wise to leave open the possibility that the "getting up in the morning" study has a different causal variant that is just in high LD with ME/CFS's causal variant.
For the PheWAS analysis I did a few posts ago (looking up which other traits have significant associations at a variant), I used GWAS Atlas. They have 4,756 GWAS datasets, and they have not added any datasets since 2019.
I found out that IEU Open GWAS also has PheWAS functionality, and they have over 10 times as many GWAS datasets (50,069), so it seemed worth looking for associations there too.
Unfortunately, the Open GWAS website doesn't have a simple PheWAS lookup using the web browser like GWAS Atlas, and instead it requires using their API. But they provide Python and R packages for accessing the API, and at least the Python package is fairly straightforward. I haven't tried the R package.
Keep in mind that the variant being significant in multiple traits doesn't necessarily indicate a shared causal variant between ME/CFS and each of these traits. Though it at least shows that a causal variant is likely in the same general region near CA10 in the two traits.
The most significant in this case is a reaction time trait. The link in the table doesn't take you directly to the BioBank details for the trait, but here are the details for that first trait, including a little video of what the cognitive test is: https://biobank.ndph.ox.ac.uk/ukb/field.cgi?id=404
Basically, it's a "game" where two cards are shown on a screen, and the participant presses a button as quickly as possible if the two cards match. This trait is how long it takes to press the button (durations even if the participant got it wrong are counted here).
The direction of effect is as expected: the T allele is associated with increased risk of ME/CFS, and it is also associated with longer duration to press the button.
Here are all associations with p<1e-6, sorted starting from most significant:
I also did the PheWAS with the other 7 DecodeME variants, and will post the results on their respective threads. [Edit: Not all of them had other significant traits.]
Edit: Note, if you want to look up the details page of a UK BioBank trait (ID starts with "ukb"):
- Go to the Open GWAS page for the trait linked in the table
- In the "note" row of the table, there's a number. For example, for the most significant trait above, it says "404" in the "note" field.
- Replace the number at the end of the following URL with the number for the trait: https://biobank.ndph.ox.ac.uk/ukb/field.cgi?id=404
I found out that IEU Open GWAS also has PheWAS functionality, and they have over 10 times as many GWAS datasets (50,069), so it seemed worth looking for associations there too.
Unfortunately, the Open GWAS website doesn't have a simple PheWAS lookup using the web browser like GWAS Atlas, and instead it requires using their API. But they provide Python and R packages for accessing the API, and at least the Python package is fairly straightforward. I haven't tried the R package.
Keep in mind that the variant being significant in multiple traits doesn't necessarily indicate a shared causal variant between ME/CFS and each of these traits. Though it at least shows that a causal variant is likely in the same general region near CA10 in the two traits.
The most significant in this case is a reaction time trait. The link in the table doesn't take you directly to the BioBank details for the trait, but here are the details for that first trait, including a little video of what the cognitive test is: https://biobank.ndph.ox.ac.uk/ukb/field.cgi?id=404
Basically, it's a "game" where two cards are shown on a screen, and the participant presses a button as quickly as possible if the two cards match. This trait is how long it takes to press the button (durations even if the participant got it wrong are counted here).
The direction of effect is as expected: the T allele is associated with increased risk of ME/CFS, and it is also associated with longer duration to press the button.
Here are all associations with p<1e-6, sorted starting from most significant:
I also did the PheWAS with the other 7 DecodeME variants, and will post the results on their respective threads. [Edit: Not all of them had other significant traits.]
Edit: Note, if you want to look up the details page of a UK BioBank trait (ID starts with "ukb"):
- Go to the Open GWAS page for the trait linked in the table
- In the "note" row of the table, there's a number. For example, for the most significant trait above, it says "404" in the "note" field.
- Replace the number at the end of the following URL with the number for the trait: https://biobank.ndph.ox.ac.uk/ukb/field.cgi?id=404
Last edited:
ME/CFS Science Blog
Senior Member (Voting Rights)
Thanks for doing this.
One issue I see is that we don't have a standardized measure of effect size to filter these. I suspect that in a lot of the things that come up like smoking, overall health and cognitive tests, the DNA has only very minor effects and that they only come up because these traits could be tested with enormous sample sizes. So a large proportion of all possible DNA variants and genes are likely associated with them.
Yes good point. We can at least see the total sample size for studies in the table above, and they're all around 350,000 to 450,000 for this group of datasets. Though unbalanced group sizes in binary traits, for example, could make a study's effective sample size smaller, making the findings less directly comparable to each other.
(It's also possible the beta's might all be on a standardized scale making them more easily comparable as effect sizes, since they're all around the same order of magnitude, though I'm not sure.)
One thing is that the Manhattan plots can be viewed for some of these traits to see how significant a given locus is compared to others. For example, for "Duration to first press of snap-button in each round", there's a link on the side to a plot on the Genotype-Phenotype Map website. I copied the plot and made an arrow pointing to the area near the CA10 locus (labeled as 17:52176967 A/G):
Yes, there are a lot of genome-wide significant loci thanks to the large sample size. But it looks like this CA10 locus is in maybe the top 10 or 20 most significant, so not nothing, I think.
Edit: Converted image to thumbnail because the huge text for the title was a bit overwhelming.
Last edited:
Open GWAS has another tool called the Genotype-Phenotype Map where it can do actual colocalization testing against many GWAS datasets using uploaded summary stats.
I uploaded the DecodeME GWAS-1 summary stats. The results are on the following page, indicating which traits colocalize (share a causal variant) with ME/CFS at various DecodeME loci: https://gpmap.opengwas.io/trait.html?id=a3069ff7-f52c-0c62-fde3-c38e99f0cd91
For example, here are the traits that appear to colocalize with ME/CFS at the CA10 locus (note the ME/CFS row is for the uploaded data):
I uploaded the DecodeME GWAS-1 summary stats. The results are on the following page, indicating which traits colocalize (share a causal variant) with ME/CFS at various DecodeME loci: https://gpmap.opengwas.io/trait.html?id=a3069ff7-f52c-0c62-fde3-c38e99f0cd91
For example, here are the traits that appear to colocalize with ME/CFS at the CA10 locus (note the ME/CFS row is for the uploaded data):
Note: They ask for sample size, and I gave a sample size of 58792, which is the effective sample size of DecodeME calculated using a common formula for unbalanced studies. For example, described on this page:
For illustration, a study with 10 cases and 1,000,000 controls doesn't have anywhere near the statistical power of a study with 500,000 cases and 500,000 controls, so n=1,000,010 wouldn't be useful for the first study. The effective sample size calculated with the above formula in this hypothetical study would be about 40.
The formula for effective sample size is sometimes written in a different but mathematically equivalent form, such as on the following page:
Edit: Fixed small error in second formula.
The formula is
Neff = 4 * [total sample size] * [proportion of cases] * [proportion of controls]. For DecodeME, that's Neff = 4 * 275488 * 0.056550558 * 0.943449442 ≈ 58792For illustration, a study with 10 cases and 1,000,000 controls doesn't have anywhere near the statistical power of a study with 500,000 cases and 500,000 controls, so n=1,000,010 wouldn't be useful for the first study. The effective sample size calculated with the above formula in this hypothetical study would be about 40.
The formula for effective sample size is sometimes written in a different but mathematically equivalent form, such as on the following page:
Neff = 4 / (1/[num cases] + 1/[num controls])Edit: Fixed small error in second formula.
Last edited:
Just seeing what's in this paper about CA10.
The most significant SNP associated with neck or shoulder pain in this UK BioBank cohort was at the CA10 locus. This SNP did not replicate in Finland's FinnGen biobank (p=.20), though they say the phenotype they used in FinnGen is "shoulder issues", which could represent quite a different phenotype, as they acknowledge:
I don't see a "shoulder issues" trait on FinnGen's website, so I'm not sure if it's called something else.
The CA10 SNP was genome-wide significant for neck/shoulder pain in the female-only GWAS (n=55783 cases), but not in the male-only GWAS (n=42869 cases). This is similar to DecodeME (females: p=9.5e-9, males: p=.01).
Discussion about CA10:
The above study said:
CA10 (and CA8, CA11) is unlike most members of the CA gene family, in that it does not catalyze the reaction of CO2 to HCO3-. It is called carbonic anhydrase-related protein 10, in contrast to, for example, carbonic anhydrase 9.
Carbonic anhydrase-related protein CA10 is an evolutionarily conserved pan-neurexin ligand (PNAS, 2017)
Therefore, I'm not sure if the carbonic anhydrase inhibitor acetazolamide (ACZ) would have any effect on CA10. The cited study about neuropathic pain, da Vitoria Lobo 2022, used ACZ because they saw increased levels of CA7, a protein which does have catalytic activity.
Maybe ACZ could still bind to CA10 and interfere with whatever it does, though I don't know the details of ACZ's mechanism of action.
CA10 (and CA8, CA11) is unlike most members of the CA gene family, in that it does not catalyze the reaction of CO2 to HCO3-. It is called carbonic anhydrase-related protein 10, in contrast to, for example, carbonic anhydrase 9.
Carbonic anhydrase-related protein CA10 is an evolutionarily conserved pan-neurexin ligand (PNAS, 2017)
Therefore, I'm not sure if the carbonic anhydrase inhibitor acetazolamide (ACZ) would have any effect on CA10. The cited study about neuropathic pain, da Vitoria Lobo 2022, used ACZ because they saw increased levels of CA7, a protein which does have catalytic activity.
Maybe ACZ could still bind to CA10 and interfere with whatever it does, though I don't know the details of ACZ's mechanism of action.