Preprint Dissecting the genetic complexity of myalgic encephalomyelitis/chronic fatigue syndrome via deep learning-powered genome analysis, 2025, Zhang+

Depends on the buns and lemonade a bit.

 

OK, I am not in a position to judge. @jnmaciuch seems to think the associations must be real. More opinions welcome.

 

Like @Hutan I don’t think it’s definitive.

But given the rarity of their variants and the small sample size, the independent cohort validation actually carries a lot of weight for me.

If it was all (or even partly) a fluke, you’d expect that test cohort AUC to be barely above 0.5. The validation shows that despite looking at very rare variants in a small group of people, the same pattern of rare variants was replicable in another small group.

But I’d wait to see the overlap with DecodeME before getting too excited. My concern is not that the results are unreliable, but rather that they’re only a small part of the story.

[Edit: I honestly don’t know what to make of the lack of replicability in UK BioBank, but like [edit: others] have already alluded to, the strength of replicability in an association such as this relies on whether the endpoint you’re associating with is actually similar between the two studies. It’s more likely for things to get drowned out in BioBank without stringent diagnostic criteria].

 

But if I get Hutan's argument right, by the same token there should have been replication in a cohort of maybe 2000 in the UK Biobank?

Reliability is the issue here. There are bound to be caveats. The UK, Biobank might have been a bad sample, as might the others for various reasons. But if there is a discrepancy that doesn't seem to make statistical sense that is a worry. Or is it that they didn't actually test for replication for ME/CFS? As Hutan implies that would be a bit odd.

 

I may be misinterpreting the methods (they're a bit vague on details), but I suspect it would be because they are actually comparing p-values between their own cohort and BioBank. Depending on how loosely BioBank defined ME/CFS, it would have a profound difference on how strong the association is with any given gene.

What (I think) they're doing in the UK BioBank analysis is simply checking whether the highly associated genes in their data set are similarly highly associated (above background) [edit: in published studies for] other conditions. So if the UK BioBank GWAS are providing genes that are more broadly associated with general fatigue due to insufficiently stringent diagnostic criteria, the top genes in Zhang et al. may or may not come up among the top genes in BioBank.

However, the validation with the Cornell cohort in this paper is not a comparison of p-values after-the-fact, it's actually looking at the participant-gene-level data for the new group and seeing if the same combination of genes is similarly predictive of outcome.

[Edit: @Hutan, to your point, I think Zhang et al. would be substantially limited by whatever labels were already applied by UK BioBank studies, including that vague 'Covid-19 controls' label. If it was that vague, though, I would've just left it out of the paper.]

 

I think this paper is using their fancy non-linear HEAL2 algorithm to predict disease risk in the first two cohorts, while they had to rely on the traditional statistical tests for the Biobank, so I don't think it should be too concerning that it didn't replicate.

Apart from that, I'd consider depression being the top disease hit a semi-replication. There's a good chance many people in the ME/CFS cohort of this study and the depression cohort of the UK Biobank have a similar condition. For 15 years I was diagnosed with depression before getting an ME/CFS diagnosis, so that's how I would have signed up, and I'm guessing there are many similar cases of people who think they have depression but actually have ME/CFS. And even if the cohorts are perfectly diagnosed, depression is probably one of the conditions I'd rank in the top three if asked for conditions similar to ME/CFS, so seeing it be the top hit from the same genes as ME/CFS is interesting.

 

I think they're specifically saying that their highly associated genes have a greater-than-expected-by-chance overlap with the set of genes that another GWA study had already found to be associated with XYZ condition. As in, I don't think they actually did any direct analysis of the UK BioBank data, which is an important distinction.

I saw "I think" here because I am truly guessing. I'm having a hard time figuring out exactly what they meant.

I agree it's extremely messy, though I think part of that is because they could only use the same terms that were defined in other papers. They should have done a much better job clarifying whatever they referenced, though.

[Edit: cross posted the same thought with @forestglip]

 

I think we're on the same page, some signals are just getting lost in transmission!

 

I am struggling to get my head around this, while trying to write something else, but I wonder if the apparent failure to replicate with the Leeds UK Biobank reflects different methodologies. Do Leeds have a cohort that has had whole genome sequencing and documentation for ME/CFS status. There was the preliminary GWAS study of people who reported having ME/CFS but is that usable? I don't know, but I wonder if there is a reason why they do not say there is a negative result per se.

 

I think there is a question mark about UK biobank diagnostic reliability. People were asked either if they had ever had a diagnosis of chronic fatigue syndrome, or Myalgia and celery -itis/ Chronic fatigue syndrome. It’s too easy for people with a diagnosis for chronic fatigue – which is pretty common, to answer to these questions. And Louis Nacul did work in Canada showing this is what happened in a large BC cohort Identified in a general population with abroad question: a more detailed follow-up questionnaire established that many positive answers didn’t have ME.

DecodeME Is different because not only was that a detailed follow-up questionnaire, but most people were recruited from the ME Community, rather than from the general public. I think it’s unlikely that substantial numbers of people with chronic fatigue are in the ME community. But again, we should soon be able to compare results from UK biobank with DecodeME.

But the lack of replicability Versus UK biobank doesn’t concern at this stage

 

I agree but I am confused because this was a cohort that has been trawled for SNPs as per usual GWAS rather than completely sequenced was it not? And if the Zhang findings depend on rare genes from whole sequencing should we expect a replication to even be possible?

I would like to think that it is almost certain that Zhang et al. have picked up some genetic signals, even if their approach giving 115 genes Amy be much more difficult to interpret than what is likely to come from DecodeME by tracking commoner SNP variants across big numbers.

Hutan was concerned that Zhang failed to find they should have found if their data were reliable. I am still unclear whether this is so. Even if the UK Biobank 'ME/CFS' cohort was dilute I would expect with 2000 cases for there to be some degree of agreement.

I must say that I find the way the Zhang paper is written much less transparent than the way Edinburgh do things.

 

Will the underlying dataset from DecodeME have rare variants and other data used here? Or is that not coming until SequenceME?

Presumably if it is present replication would be fairly straightforward with access to the model (a shame it hasn’t been made available)?

Or even a wider attempt to take the DecodeME data and rerun the recipe from this paper to train or just finetune the model and validate it on a larger and more consistent dataset?

 

No, DecodeME won't have data on rare variants. Genome wide analysis studies such as DecodeME only look at common variants in specific locations on the genome. Whole genome analysis studies, such as this one and the proposed SequenceME, 'reads' the whole genome of each sample, and this is then used to look for rare variants.

 

That’s my understanding. It’s not a replication but looking for genetic correlations between ME/CFS and other diseases. They’re saying “okay here are our identified 115 genes, do these pop up for other diseases too”.

However if the genes are relevant and the people in the (Leeds) UK Biobank actually have ME/CFS you’d assume the genes would also show up in some way no?

Good questions. Anyone feel comfortable asking them?

 

Thanks Andy.