Ion Channel SNP Paper

[user9876]

After further digging it appears that, unfortunately, the paper doesn't correct for multiple comparisons after all.

The study used the chi-squared test to compare the frequency of a SNP in patients vs controls, and helpfully it reports the chi-square statistic as well as the p value. You can calculate the p value directly from the this chi-square statistics. For example, for the first SNP in table 1, the Chi-squared statistic (using the appropriate one degree of freedom) gives an uncorrected p value of 0.003. That's exactly the p value reported in the same table, indicated the p value is uncorrected. Which suggests that false positives are likely to be an issue.

That said, it's not clear what the most appropriate correction would be. Although they tested 233 SNPs, that doesn't count as 233 independent tests due to linkage (two SNPs close together on a chromosome are often inhertited together, so if one SNP is significant in the test, its partner will almost certainly be too: effectively this is one test, not two).

I'm not convinced that significance testing makes sense here anyway unless there is more evidence that the control set is a good representative sample of the overall population. One very big danger is that they end up with control volunteers who have family members with ME which could skew results.

 

[Sidereal]

That said, it's not clear what the most appropriate correction would be. Although they tested 233 SNPs, that doesn't count as 233 independent tests due to linkage (two SNPs close together on a chromosome are often inhertited together, so if one SNP is significant in the test, its partner will almost certainly be too: effectively this is one test, not two).

Exactly. Personally, I almost always find genetic studies of "multifactorial" illnesses (where effect sizes are tiny and lots of SNPs pop up, frequently different ones in different studies) impossible to interpret. As I understand it, special techniques have been developed for genetics research to control for false discovery rate that are less severe than Bonferroni correction but it's pretty clear from reading this paper that nothing of this sort was done, they just ran 233 chi squares. Which isn't to say that the results are necessarily wrong, just that the p values reported are uninterpretable.

 

[user9876]

Exactly. Personally, I almost always find genetic studies of "multifactorial" illnesses (where effect sizes are tiny and lots of SNPs pop up, frequently different ones in different studies) impossible to interpret. As I understand it, special techniques have been developed for genetics research to control for false discovery rate that are less severe than Bonferroni correction but it's pretty clear from reading this paper that nothing of this sort was done, they just ran 233 chi squares. Which isn't to say that the results are necessarily wrong, just that the p values reported are uninterpretable.

I'm still feeling a little confuse by this paper and it relates to the significance testing. I guess the question in my mind is over the units of interaction. We seem to have a gene mentioned and then individual mutations also mentioned but one individual may have multiple mutations on a single gene. So is it the combination of them that are important in the way the gene works or does each singularly have an effect; I guess we may not know.

The question becomes important when doing significance testing because it may be that a given variant of a gene (with a number of mutations) is significant rather than each individual one. If this were the case I would look at significance by comparing the two sample discrete probability distributions (patient, control) over the set of combined variants on a gene rather than for each individual mutation. Something like the Kolmogorov-Smirnov test comes to mind in comparing whether two sample distributions are the same but I think it makes assumptions that the distributions are over continuous data.

But I guess if it is each difference that acts individually then the pairwise comparison would seem fine. If we are not sure about how the system works I guess I would like both.

 

[sdmcvicar]

TL;DR - Data suggests that PWCFS may be predisposed to having errors in expression and function of TRP ion channels. This may alter signaling, cell function, and homeostasis within the patient population, speculatively tying these channels to a host of symptoms common in PWCFS. We worry that since they didn't correct their p-values, this is all bunk.

Long version:

The biostatisticians at UTMB were fans of using q-value for corrections a few years ago when I took an introductory bioinformatics course. It's less stringent than Bonferroni, but it does take into account false discovery rates. Link to a document discussing some the various analysis options here: http://kmplot.com/multipletesting/@multiple_testing_OH.pdf

Page 14 has a great table of that might help us visualize what some of the commonly used statistical methods would give for significance based on uncorrected p-values.

This is discovery science, and with any -omics type study, they're trying to generate new hypotheses based on what has high correlation. Too many leads is problematic, because you can only invest so much time and money in each one, and may be chasing false positives. On the other hand, if you're too stringent, your study will generate no significant leads, even if there were true differences in the two populations. Since you have to publish, you err on the side of caution, which leads to the "looking for babies in the bathwater" phenomenon.

The key message they're trying to sell here is that commonality of alterations in these TRP genes suggests a mechanism by which CFS might be acquired, or how certain symptoms may affect patients with varying degrees of severity, based on their genetic background. Since we have no mechanism to test against, and don't know the effects each of these SNPs have on expression and function of the receptor, the only practical way to evaluate multiple SNPs on a single gene is to treat them as unlinked and do the pairwise comparison.

To me, more than the question of accuracy & significance of the data set, is whether the lack of correction was deliberate or due to a lack of experience. If the former, future resources might be put into a weak lead. If the latter, it strengthens the argument for a central database for study data.

Not sure if this helps, or just muddies the waters further.

 

[Simon]

A final comment on this study

Page 14 has a great table of that might help us visualize what some of the commonly used statistical methods would give for significance based on uncorrected p-values.

Thanks, that's useful. False Discovery Rate (FDR) corrrections using an FDR of 5% would reduce the p-value for significance to p<0.0027, and none of the findings in this study would pass that test. That said, because of linkage between the SNPs (they tend to be inherited together), this is probably overly strict. Using q-values instead, the threshold would be p<0.017 and about half of the findings would still be significant at this level.

It is interesting that most of the significant SNPs are for TRPM3: 9 ex 13 at p<0.05, 3 ex 3 at p<0.01, considering they looked at 21 different ion channels This could mean that the study does provide evidence that TRPM3 is significantly associated with mecfs. Another possiblity is that the TRMP3 SNPs are in 'linkage disequilibrium' ie are inherited together so that if one TRMP3 SNP (pronounced 'snip', btw) is significant, the other TRMP3 SNPs are likely to be too, and it could effectively be a false positive amplified.

It's a shame the authors didn't discuss such issues in the paper as it could make the findings easier to intepret.

One simple analysis that might shed more light on the problem is to look at dosage effects: is the association with mecfs stronger for patients that have two copies of the SNP in question (ie both chromosomes carry the same SNP) than in those with single copies? The Dubbo study found exactly this when analysing the link between cytokine genes and mecfs: Cytokine Polymorphisms Have a Synergistic Effect on Severity of the Acute Sickness Response to Infection

 

[Myrtle]

I don't know if its the TRPM channels or not but in my 23andme raw data when I was looking at my double copy mutations in the past, I found I had double copy gene mutations in the ion channels to do with calcium. It was one of the things which had interested me in my results which I wondered how it was affecting me. I found those mutations when my raw data was put through through Valenjtns program picking out the more uncommon gene mutations.

I'll have to take another look at exactly what those mutations were so others can check these ion channel ones out too.

Could you tell me what this 'Valenjtns program' is please? Also, where could I get it if I am interested in it?
Thanks

 

[Sushi]

Could you tell me what this 'Valenjtns program' is please? Also, where could I get it if I am interested in it?
Thanks

@Valentijn

 

[Valentijn]

Could you tell me what this 'Valenjtns program' is please? Also, where could I get it if I am interested in it?
Thanks

http://sourceforge.net/projects/analyzemygenes/ . Though we're hoping to get an updated and expanded version up there in the next couple weeks (currently delayed by my inability to sit up for long).

 

[student]

We are greatfull for this insight of such detailed SNP knowledge from your data. Thank you dear Valentijn.

Is there a CFS project – for the larger 23andme Valentijn groop? Do PR Forum readers send more genetic SNPs results to you? What is todays procedure for such sendings. What addiditves are wanted? Do you plan more capacity to store (and process) such data of 2015, or in the near future?

Would it help to start with money support for such a task?

 

[Valentijn]

Is there a CFS project – for the larger 23andme Valentijn groop? Do PR Forum readers send more genetic SNPs results to you? What is todays procedure for such sendings. What addiditves are wanted? Do you plan more capacity to store (and process) such data of 2015, or in the near future?

Yes, ME patients (CCC/ICC/SEID with PEM) send me their files if they want to contribute. At the moment
I'm mostly looking at a compilation of the very rare results for patients, to see if there's a trend at specific SNPs, multiple SNPs on the same gene, or multiple SNPs on multiple genes which have a shared function. Myself and another forum member are then looking the possible implication of rare missense mutations using an online site for protein modelling predictions, to determine if those mutations might be causing problems.

More complicated processing will likely happen when I get a new computer, but that's delayed again due to upcoming moving expenses.

Would it help to start with money support for such a task?

Not really. At this point, I'm staying quite busy with just looking at the rare results, and I can do that well enough on my laptop.

 

[student]

Dear Valentijn. There are some minor details (in your post #21)– if you would find time: What are the A C TG findings for these TRPM3 SNPs, missing in the 23andme results. rs12682832, rs4454352, rs10115622. ??
What is now checked with todays running chip version - at 23andme (does anyone know)?

It seems that – for the TRPA1 (2) and TRPC4 (2) there was less specific Data for %-age of europeans available, no? TRPA1- rs2383844, rs4738202. TRPC4- rs6650469, rs655207.

Regards student

 

[Valentijn]

Dear Valentijn. There are some minor details (in your post #21)– if you would find time: What are the A C TG findings for these TRPM3 SNPs, missing in the 23andme results. rs12682832, rs4454352, rs10115622. ??

Not tested on 23andMe's V3 chip.

What is now checked with todays running chip version - at 23andme (does anyone know)?

V4.

 

[Jo-rel]

This seems to be an old thread. I'm curious if any new information has come to light with this TRPM3 SNP association? I find it interesting that it seems to be located near a confluence of calcium signaling, glucose homeostasis and insulin release, as well as cholesterol synthesis, vitamin D synthesis (both via the DHCR7 gene, a few steps before pregnenolone synthesis) locations. I happen to share a DHCR7 SNP with a rarer form of Smith-Lemli-Opitz Syndrome, and although I don't have the condition I do have very low cholesterol (around 105). Thus I probably don't make pregnenolone efficiently. I am also nearly all red with the TRPM3 SNPs from the study. Sugars throw me over the edge. Steroid hormones are all low. I do also have chronic infections which drive down cholesterol and vitamin D levels, hence pregnenolone (probably), which was already compromised by TRPM3. Seems to answer a lot. Wondering if anyone else has put some pieces together via this study? I've researched cholesterol deficiency in chronic illness and there seems to be a continuum. The lowest being SLOS patients, then autism, some forms of depression and bipolar, and then CFS. Methylation defects might exacerbate things through the Kreb's Cycle and its effects on pyruvate and other cholesterol precursors. Thanks!

 

[montseniana]

Hi all!

I know this thread is little bit old, but I was looking this gene (TRPM3) in my Enlis program, and I'm shocked because I have 313 mutations in this gene!!!!! All are unknown clinical significance, but... 313!!!

Curiosity killed the cat...

What do you think about it?

 

[A.B.]

Hi all!

I know this thread is little bit old, but I was looking this gene (TRPM3) in my Enlis program, and I'm shocked because I have 313 mutations in this gene!!!!! All are unknown clinical significance, but... 313!!!

Curiosity killed the cat...

What do you think about it?

Ask someone from Griffit uni who has worked on TRPM3.

 

[Valentijn]

What do you think about it?

They're probably just 313 normal alleles, not significant mutations.

 

[Snow Leopard]

What do you think about it?

I think it's a null result.

Candidate gene association studies, unless there is 100% specificity (example: rare diseases caused by singular SNPs) are almost always false positives, because they fail to apply appropriate statistical correction (and have sufficient sample size).

 

[montseniana]

Thanks to all!

At first I was shocked to see too many SNP... it seems that the SNP of this gene are dancing "la conga"...

 

[pattismith]

They used Fukuda, with no apparent requirement for participants to have PEM.

23andMe tested for rs11142508, rs1160742, rs1328153, rs3763619, rs7865858, rs1504401, rs2383844, rs4738202, rs6650469, and rs655207 on the V3 (previous) chip.

Is it this scientific study you are talking about?
http://insights.sagepub.com/examina...rphisms-snps-in-transient-recep-article-a4824

"Results: Thirteen SNPs were significantly associated with CFS patients compared with the controls. Nine of these SNPs were associated with TRPM3 (rs12682832; P ≤ 0.003, rs11142508; P < 0.004, rs1160742; P < 0.08, rs4454352; P ≤ 0.013, rs1328153; P ≤ 0.013, rs3763619; P ≤ 0.014, rs7865858; P ≤ 0.021, rs1504401
; P ≤ 0041, rs10115622; P ≤ 0.050), while the remainder were associated with TRPA1 (rs2383844; P ≤ 0.040, rs4738202; P ≤ 0.018) and TRPC4 (rs6650469; P ≤ 0.016, rs655207; P ≤ 0.018)."

here what I can read on my data from 23andMe (I found only three of them, but all three have polymorphism, either hetero or homozygote)

TRPM3
rs12682832 (NA)
rs11142508 TT (ancestral C)
rs1160742 (NA)
rs4454352 (NA)
rs1328153 (NA)
rs3763619 (NA)
rs7865858 AG (ancestral A)
rs1504401 CC (ancestral T)
rs10115622 (NA)

TRPA1

rs2383844 (NA)
rs4738202 (NA)
rs6650469 (NA)
rs655207 (NA)

it is strange that you had datas from 23andMe that I can't have now, do they change the snp they analyze from one year to another?

 

[Valentijn]

it is strange that you had datas from 23andMe that I can't have now, do they change the snp they analyze from one year to another?

Yes, they've used different chips over the years.