Preprint Initial findings from the DecodeME genome-wide association study of myalgic encephalomyelitis/chronic fatigue syndrome, 2025, DecodeMe Collaboration

[obeat]

Curiosity genes. Very rare!

 

[ME/CFS Science Blog]

Coloc problem
I think there might be an issue with the colocalization analysis and Figure 4 in the DecodeME preprint. This is important because these results determined whether a candidate gene is considered Tier 1 or not. In brief, In think the paper used only the significant SNPs from the GTExv10 data to match it with DecodeME data while the coloc analysis probably requires the full dataset.

GTEx gives data on gene expression in 49 different tissues. It allows researchers to check if a certain DNA variant or SNP is associated with increased expression of a particular gene in each of the tissues. The preprint took the regions around the 8 DecodeME hits and looked for genes that have a significant association to a SNP in that same region in the GTEx dataset. The preprint then calculated if the GTEx gene expression signal matched the one found in DecodeME using an analysis called colocalization. But I suspect there is an issue here because it appears that the analysis only used SNPs that had reached the significance threshold in GTEx, rather than the full data.

Supplementary table 6 shows the datasets end in 'signif_pairs.parquet' which usually refers to this limited dataset and I was also able to largely replicate the coloc results in that table using the significant SNPs only retrieve using the GTEx API. The problem is that in several cases these are only a fraction of the SNPs that do not represent the data in this region.

Hopefully the plot below clarifies the problem. The gray dots are GTEx data* for SNPs that were also in DecodeME. The blue ones are the significant SNPs. The first graph shows that these were top results in our example namely the gene ARFGEF2 in visceral adipose tissue in the locus on chromosome 20. The second graph shows that these significant SNPs exclude the top SNPs in DecodeME. The preprint reports that these two signals 'colocalize' while this isn't actually the case if you use the full dataset.



*One caveat is that I didn't' actually manage to get the full GTEx dataset (the GTExwebsite say it's on a Google Cloud bucket but the link looks broken). So I used eQTL Catalogue version of the GTEx data (they used the same raw data but analysed it using their own standardised pipeline).

My guess is that many of the reported associations in Figure 4 do not hold up and are an artifact of using only the significant SNPs in the coloc analysis. Here's my reasoning about this. Coloc has 5 possible outcomes:

• H0 Neither trait has a genetic association in the region.
• H1 Only trait 1 is associated with the region.
• H2 Only trait 2 is associated with the region.
• H3 Both traits are associated, but they are driven by independent, distinct causal variants.
• H4 Both traits are associated and share a single, common causal variant.

Because the analysis is only done if there is a significant association in both traits (DecodeME and gene expression), H3 and H4 are the most likely ones. Coloc uses a bayesian analysis and when n, the number of SNPs is small, the prior probability becomes more important than the actual evidence.

• For H3 the prior is n * 1e-4 * (n-1) * 1* 1e-4
• For H4 the prior is n * 1e-5.

So when n is very small, H4 is more likely than H3. That might explain why there are many associations in Figure 4, while a colocalization normally occurs in < 50% of GWAS hits.

using the full (eQTL Catalog) dataset, I only found 4 eQTL associations in the chr20 locus (none involving ARFGEF2).

 

[mariovitali]

I have a rare PER2 variant (one of the genes in the circadian rythm machinery) and early onset ME/CFS (maybe). In my case it seems to have all started with sudden, extremely unrefreshing sleep on some days (for 1 year), with no apparent link to anything.

Does the fact that a symptom was the first to manifest have a special meaning in other diseases? Maybe it is the part of physiology from which the dysfunction began spreading out to other parts?

Looking further at PER2 function, I found a study (in mice) where Peroxisome Proliferator Gamma is mentioned and also triacylglycerols (which we have come across before in ME/CFS). More specifically :

Furthermore, mutant mice showed a drastic decrease in plasma levels of total triacylglycerol (TAG) and nonesterified fatty acid (NEFA) (Figures 1E and 1F). These alterations could not be attributed to differences in food intake, as food consumption was identical in Per2−/− and WT mice (Figure 1G). When analyzed by indirect calorimetry, Per2−/− mice showed a significant increase in oxygen consumption as compared to WT (Figure 1H). These results demonstrate a pronounced alteration of lipid metabolism in mice lacking the Per2 gene.

Also, regarding PPARγ and PER2 , it appears that there is an association with dopamine neuron survival via NURR1





Link of study : https://www.cell.com/cell-metabolism/fulltext/S1550-4131(10)00357-8

 

[forestglip]

Hopefully the plot below clarifies the problem. The gray dots are GTEx data* for SNPs that were also in DecodeME. The blue ones are the significant SNPs. The first graph shows that these were top results in our example namely the gene ARFGEF2 in visceral adipose tissue in the locus on chromosome 20. The second graph shows that these significant SNPs exclude the top SNPs in DecodeME. The preprint reports that these two signals 'colocalize' while this isn't actually the case if you use the full dataset.

I haven't independently checked any of that, but in case the analysis did only use significant SNPs, I just wanted to quote the coloc documentation which states that all SNPs should be used, as you say:

Please do include all SNPs you have in the region: do not trim by significance or MAF or other means. Coloc enumerates all possible configureations of shared and non-shared causal variants and evaluates their relative support; it cannot do this if it does not have all the SNPs available.

 

[Hutan]

I think there might be an issue with the colocalization analysis and Figure 4 in the DecodeME preprint.

What you say makes sense to me. Here's Figure 4 from the DecodeME Preprint with the chromosome 20 genes at the bottom of the column.


@Chris Ponting, could you or a team member respond to ME/CFS Science Blog's post about the SNP set used for the colocalisation analysis?

 

[Chris Ponting]

Coloc problem
I think there might be an issue with the colocalization analysis and Figure 4 in the DecodeME preprint. This is important because these results determined whether a candidate gene is considered Tier 1 or not. In brief, In think the paper used only the significant SNPs from the GTExv10 data to match it with DecodeME data while the coloc analysis probably requires the full dataset.

GTEx gives data on gene expression in 49 different tissues. It allows researchers to check if a certain DNA variant or SNP is associated with increased expression of a particular gene in each of the tissues. The preprint took the regions around the 8 DecodeME hits and looked for genes that have a significant association to a SNP in that same region in the GTEx dataset. The preprint then calculated if the GTEx gene expression signal matched the one found in DecodeME using an analysis called colocalization. But I suspect there is an issue here because it appears that the analysis only used SNPs that had reached the significance threshold in GTEx, rather than the full data.

Supplementary table 6 shows the datasets end in 'signif_pairs.parquet' which usually refers to this limited dataset and I was also able to largely replicate the coloc results in that table using the significant SNPs only retrieve using the GTEx API. The problem is that in several cases these are only a fraction of the SNPs that do not represent the data in this region.

Hopefully the plot below clarifies the problem. The gray dots are GTEx data* for SNPs that were also in DecodeME. The blue ones are the significant SNPs. The first graph shows that these were top results in our example namely the gene ARFGEF2 in visceral adipose tissue in the locus on chromosome 20. The second graph shows that these significant SNPs exclude the top SNPs in DecodeME. The preprint reports that these two signals 'colocalize' while this isn't actually the case if you use the full dataset.

View attachment 34492

*One caveat is that I didn't' actually manage to get the full GTEx dataset (the GTExwebsite say it's on a Google Cloud bucket but the link looks broken). So I used eQTL Catalogue version of the GTEx data (they used the same raw data but analysed it using their own standardised pipeline).

My guess is that many of the reported associations in Figure 4 do not hold up and are an artifact of using only the significant SNPs in the coloc analysis. Here's my reasoning about this. Coloc has 5 possible outcomes:

• H0 Neither trait has a genetic association in the region.
• H1 Only trait 1 is associated with the region.
• H2 Only trait 2 is associated with the region.
• H3 Both traits are associated, but they are driven by independent, distinct causal variants.
• H4 Both traits are associated and share a single, common causal variant.

Because the analysis is only done if there is a significant association in both traits (DecodeME and gene expression), H3 and H4 are the most likely ones. Coloc uses a bayesian analysis and when n, the number of SNPs is small, the prior probability becomes more important than the actual evidence.

• For H3 the prior is n * 1e-4 * (n-1) * 1* 1e-4
• For H4 the prior is n * 1e-5.

So when n is very small, H4 is more likely than H3. That might explain why there are many associations in Figure 4, while a colocalization normally occurs in < 50% of GWAS hits.

using the full (eQTL Catalog) dataset, I only found 4 eQTL associations in the chr20 locus (none involving ARFGEF2).

View attachment 34493

Thank you. Good spot: yes, we agree. The DecodeME project deprecated this analysis quite soon after the preprint was posted. This meant that we hadn't since gone back to double- (or triple-) check its findings. Rather, the team was working on a completely different Figure 4, using very different methods, to prioritise variants & genes. The new Figure 4 analysis needs access to the UK Biobank compute facility ("the RAP") so - as with all analyses - we're waiting for UK Biobank to reopen the RAP.
Again, thanks @ME/CFS Science Blog for spotting this.

 

[ME/CFS Science Blog]

Thank you. Good spot: yes, we agree. The DecodeME project deprecated this analysis quite soon after the preprint was posted. This meant that we hadn't since gone back to double- (or triple-) check its findings. Rather, the team was working on a completely different Figure 4, using very different methods, to prioritise variants & genes. The new Figure 4 analysis needs access to the UK Biobank compute facility ("the RAP") so - as with all analyses - we're waiting for UK Biobank to reopen the RAP.
Again, thanks @ME/CFS Science Blog for spotting this.

Thanks, look forward to this analysis! Hope the UK biobank access is reopened soon (can imagine it must be a very frustrating situation for your team).

In some sense, it might be good news that fewer genes show PPH4 coloc as it might lower the number of potential causal genes.

 

[ChronicallyOverIt]

I’m not following this well enough at the moment. Does this mean some of the tier one genes could be reprioritized? Especially the dense ones, thinking of the HLA region with the BTN genes