Overview of ME/CFS genes in DecodeME

[ME/CFS Science Blog]

The list of genes that the DecodeME preprint suggested based on ‘Tier 1’ and ‘Tier 2’ labels is incorrect. There was an error in the data analysis as confirmed by Prof Chris Ponting here:
https://s4me.info/threads/initial-f...2025-decodeme-collaboration.45490/post-726398

Over the past couple of weeks, I’ve assembled a list of genes that are likely associated with ME/CFS. As DecodeME only found 6 hits above the significance threshold, signals just below this threshold (p < 10⁻⁶) were also included, resulting in 23 loci.

Findings and comments about individual genes are best posted in the thread of that locus. This thread is aimed at finding patterns and commonalities among these genes, to better understand the genetic risk for ME/CFS.

Top hit (GRCh38)	P-value	Risk allele frequency and odds ratio	Most likely gene(s)	FLAMES prediction	Certainty of evidence
20: 48,914,387 T/TA rs11482246	9.5e-12	cases: 65.4% controls: 63.3% OR 1.095	ARFGEF2	ARFGEF2	WEAK The evidence is mostly based on position as the core SNPs lie just before the start of ARFGEF2, inside its antisense RNA gene. STAU1 and PREX1 are also plausible candidates.
17: 52,183,006 C/T rs34626694	2.1e-09	cases: 34.6% controls: 32.9% OR 1.084	CA10	CA10	MODERATE CA10 is the only gene within ~900 kb; the SNPs lie close to its start, in DNA that is active in brain, enhancer–gene predictions point only to CA10. FLAMES also picks CA10.
6: 26,239,176 A/G rs9358913	4.1e-09	cases: 27.7% controls: 26.0% OR 1.086	BTN2A1	undecided	MODERATE We could not perform our analyses at this locus is inside the HLA region that has complex LD. A rare gene analysis of ME/CFS by AstraZeneca pointed to BTN2A1 which is also close to the signal here.
15: 54,866,724 A/G rs7165327	7.6e-09	cases: 32.7% controls: 31.1% OR 1.082	UNC13C	UNC13C	WEAK The SNPs are in a gene desert but folded into one DNA domain that includes UNC13C as the only protein-coding gene. FLAMES also picks UNC13C. The top SNP (rs7165327) is probably mis-imputed; without it the strongest SNP has a p-value of 1.1 x 10⁻⁷, so this locus is probably not genome-wide significant.
1: 173,846,152 C/T rs12071663	2.6e-08	cases: 68.7% controls: 67.4% OR 1.079	DARS2 GAS5 ZBTB37 SERPINC1	undecided	UNCLEAR This is a gene dense region that points to possible immune involvement and T-cells. Just like FLAMES, we couldn’t pinpoint a causal gene.
6: 97,984,426 CA/C rs59732267	4.9e-08	cases: 47.0% controls: 45.3% OR 1.071	POU3F2	MMS22L	WEAK The SNPs are in a gene desert but folded into one DNA domain that includes POU3F2 as the only protein-coding gene. We believe FLAMES’ prediction (MMS22L) is incorrect because it is restricted to a 750 kb window.
17: 11,325,637 C/G rs1546559	8.3e-08	cases: 50.1% controls: 48.4% OR 1.071	SHISA6	SHISA6	MODERATE The whole set lies inside SHISA6 and there are few other protein-coding genes nearby. FLAMES also picks SHISA6.
11: 16,217,844 C/G rs17538814	1.1e-07	cases: 10.8% controls: 9.8% OR 1.117	SOX6	SOX6	MODERATE The whole set lies inside SOX6 and there are few other protein-coding genes nearby. FLAMES also picks SOX6.
13: 53,194,927 GT/G rs35306732	1.2e-07	cases: 30.0% controls: 28.6% OR 1.077	PCDH8	OLFM4	WEAK 3D contacts strongly pointed to PCDH8 whose neuronal function better matches ME/CFS’ brain enrichment and the phewas results (depression, insomnia) than OLFM4. PCDH8 was also the prioritized gene for this region in a recent fibromyalgia GWAS.
1: 73,126,414 C/CA rs34330896	1.2e-07	cases: 51.1% controls: 49.2% OR 1.070	NEGR1	not scored	WEAK NEGR1 is the closest protein-coding gene but it is still far off from the credible set and the 3D contacts are conflicting. ERICH3 and LRRC53 are other possibilities. All protein-coding genes lie outside FLAMES’s 750 kb window, so it gave no score.
12: 118,336,076 C/T rs11068910	1.9e-07	cases: 17.2% controls: 15.8% OR 1.093	PEBP1	PEBP1	WEAK No credible set was found. The most likely SNPs lie mainly in TAOK3 (74% of the probability), with smaller parts in SUDS3 (6%) and PEBP1 (3%). SUDS3 has the best eQTL support. FLAMES is confident about PEBP1, possibly because it fits better into the pattern of other ME/CFS genes (it’s part of synaptic pathways that show enrichment).
1: 91,028,158 C/T rs358676	1.9e-07	cases: 45.5% controls: 44.4% OR 1.069	ZNF644	ZNF644	WEAK About half of the credible set lies in ZNF644, although BARHL2 is also a possibility: it lies in the same DNA folding domain and has contacts in lab-grown neurons.
1: 69,696,474 A/G rs76237653	2.1e-07	cases: 19.1% controls: 18.2% OR 1.088	LRRC7	LRRC7	MODERATE The most likely SNPs all lie inside LRRC7 and there are few other protein-coding genes nearby. FLAMES also picks LRRC7.
12: 123,924,955 G/A rs7133378	2.4e-07	cases: 33.4% controls: 31.8% OR 1.072	CCDC92 ZNF664 DNAH10	undecided	UNCLEAR The credible set lies above CCDC92 and ZNF664 and to a lesser extent DNAH10. This is a gene dense region. Just like FLAMES, we couldn’t pinpoint the most likely candidate.
18: 53,232,948 C/T rs72923287	2.5e-07	cases: 54.3% controls: 52.6% OR 1.068	DCC	DCC	MODERATE The whole set lies inside DCC and there are few other protein-coding genes nearby. DCC was also found in a recent fibromyalgia GWAS. FLAMES also picks DCC.
6: 4,336,259 C/T rs11753380	2.9e-07	cases: 81.9% controls: 80.4% OR 1.086	CDYL	undecided	WEAK The credible set is far away from protein coding genes but 3D folding domains point to CDYL as the most likely candidate.
20: 33,363,039 G/A rs291671	5.4e-07	cases: 90.4% controls: 89.2% OR 1.113	CDK5RAP1 SNTA1	undecided	UNCLEAR This is a gene-dense region, and the credible set is very spread out. Half of the probability lies in CDK5RAP1 (including the lead SNP) and a fifth in SNTA1, so these are the most likely (but not only) candidates.
10: 21,748,880 G/A rs6482190	6.3e-07	cases: 29.6% controls: 28.0% OR 1.072	MLLT10	MLLT10	WEAK The credible set lies mostly in MLLT10, but also in DNAJC1, and SKIDA1 starts close by; both are also possible candidates. FLAMES is confident about MLLT10. The phewas found more than 200 matches, some suggesting this locus might be affected by participation bias.
2: 197,882,813 G/A rs57862683	6.6e-07	cases: 50.9% controls: 49.4% OR 1.065	PLCL1	undecided	WEAK The credible set consists of 249 SNPs but the vast majority lie in the gene PLCL1. FLAMES picks PLCL1 but wasn't confident enough to call it causal. There's also some eQTL support for PLCL1 in various tissues such as immune cells.
4: 3,240,118 C/T rs362307	8.0e-07	cases: 8.5% controls: 7.6% OR 1.122	HTT; MSANTD1; HTT-AS	HTT	WEAK No credible set could be made but most probability lies in HTT. One of the variants (a deletion that removes one amino acid) changes the HTT protein. HTT is also picked by FLAMES and the gene was the top finding in a recent fibromyalgia GWAS. eQTL data also support HTT, mostly in immune cells. There are other candidate genes nearby and HTT isn’t fully supported by the Hi-C data. The phewas found dozens of matches, suggesting this locus might be affected by participation bias.
1: 181,676,091 A/G rs3753737	8.9e-07	cases: 31.5% controls: 30.2% OR 1.070	CACNA1E	CACNA1E	MODERATE The whole set lies inside CACNA1E and there are few other protein-coding genes nearby. FLAMES also picks CACNA1E.
13: 58,456,743 C/T rs116940660	9.4e-07	cases: 98.8% controls: 98.4% OR 1.313	PCDH17	not scored	WEAK The SNPs are in a gene desert. PCDH17, 680 kb away, is the most likely gene: the few DNA domains around the set start at its promoter, and contacts in lab-grown neurons reach it, but the evidence is weak. A shared signal with birthplace in UK Biobank suggests possible population structure. No protein-coding gene lies within FLAMES’s 750 kb window, so it gave no score.
2: 57,808,420 G/A rs2683616	9.5e-07	cases: 64.4% controls: 62.9% OR 1.067	VRK2	VRK2	WEAK The main signal lies at the start of VRK2, including its promoter. FLAMES picks VRK2 but the neighboring FANCL is also possible.

 

[ME/CFS Science Blog]

Here’s more info about our analysis and what it is based on.

Finemapping
In each locus, there isn’t just one DNA variant (SNP) that is different but hundreds that are highly correlated with each other. We use a tool called SuSiE that uses those correlations to estimate how many signals there are and which SNPs are most likely to be the causal ones. Except for the VRK2 locus on chromosome 2, all regions pointed to one signal. Our analysis focused on a credible set of SNPs that has a 95% probability of containing the causal one. For signals where this wasn’t possible, we used an 80% credible set (4 loci) or the 10 most likely SNPs (6 loci). This allowed us to zoom in on where the action lies.

Technical specifications: to estimate these correlations, the linkage disequilibrium (LD), we used the UK Biobank data by Weissbrod et al. 2020. We filtered DecodeME SNPs to those that were also available in the UK Biobank and made sure they were aligned (we flipped the z-sign when the alleles were swapped). We dropped palindromic SNPs (A/T, C/G) with a minor allele frequency above 0.4, because their strand could not be checked accurately. We applied the kriging_rss test for potential mismatch with the LD panel and dropped SNPs when the difference between the actual and predicted z-scores was both statistically surprising (|z_std_diff| > 3) and large (|z − predicted z| > 1). Each region was defined as the lead SNP ± 500 kb, and SuSiE could find up to 10 signals per region.

Chromatin openness and histone marks
DNA is wrapped around proteins called histones like a sewing thread around a wooden spool. Together they are called chromatin. Regions where chromatin is open, where the DNA is accessible, have more gene activity. We used databases to check for regions of chromatin openness in our credible set region. We also checked histone marks that indicate more activity is happening such as the active-enhancer mark H3K27ac. This gave us an impression of in which cell types and tissues the signal is active. Unfortunately, it did not provide a lot of insights as the data were often noisy and ambiguous.

Technical specifications
We used the SCREEN API to query ENCODE datasets for candidate cis-regulatory elements (cCREs) in our credible set. We looked at the cell types and tissues where these were most active, using open DNA (DNase-seq, 1,272 samples) and the active-enhancer mark H3K27ac. In addition, we used the Li et al. 2023 single-cell brain atlas (42 brain regions, 107 cell types) to look for peaks of chromatin openness in neural tissue.

3D folding
DNA is folded, which means that stretches that appear apart on a string can actually be close to each other. These 3D-contacts can be measured using a technique called Hi-C (high-throughput chromatin conformation capture). It freezes DNA contacts with formaldehyde and then uses several tricks that measure the regions that were physically close to each other. This data was often useful for signals that were in a gene desert, where no protein coding genes were nearby. The 3D-contacts sometimes helped to pinpoint a particular gene as being more likely as was the case for POU3F2, UNC13C, CDYL and PCDH17.

Technical specifications
We queried the 228 ENCODE Hi-C maps of tissues, primary cells and cell lines in which ENCODE called folding domains with the Arrowhead method, and checked which genes start in the domain that holds our credible set. We also used the loops that ENCODE calculated: contacts between our credible set and gene starts. For the 5 brain Hi-C datasets in ENCODE (2 x dorsolateral prefrontal cortex, 2 x posterior cingulate gyrus, caudate nucleus), we downloaded the data in 25 kb bins over the credible set ± 2 Mb. We plotted how often the credible set touches each gene start compared with what is expected from the distance (observed / expected). We also used two other brain datasets that measured contacts with active gene starts (PLAC-seq). These were measured in neurons, oligodendrocytes and microglia from adult human cortex (Nott et al. 2019) or in lab-grown excitatory neurons (Cui et al. 2025).

 

[ME/CFS Science Blog]

eQTL
We looked at multiple datasets that measured both DNA (the genetic code) and RNA (gene expression) of participants. This allows us to look up the DNA variants that are associated with ME/CFS to see which genes they influence. Suppose for example that people with the ME/CFS risk variants at chromosome 11 have more RNA for the gene SOX6: that would give us a clue that this is the causal gene for that DecodeME signal. Researchers use the term eQTL (expression quantitative trait locus) for this: a piece of DNA that explains some of the variance in the expression of genes. It’s what the DecodeME preprint focused on.

Unfortunately, eQTL often do not match with the findings from genetic genome-wide association studies and this is also the case for DecodeME. We rarely found a clear match here that helped us better understand a locus. In some cases, however, the eQTL data indicate if a gene is likely to be increased or decreased by the DNA versions that were more common in people with ME/CFS.

Technical specifications
We queried the eQTL Catalogue release 8 and downloaded full summary statistics, RNA-seq or microarray, for the region in question for all GTEx v10 tissue datasets supplemented with all brain and immune cell types from other studies. Together with eQTLGen (blood data of up to 31,684 people), this gave about 240–265 datasets per region. For each dataset we used every gene that was tested at one or more credible-set variants, in the credible set ± 500 kb. For every dataset - gene pair, we calculated the probability that the GWAS and the eQTL share one causal variant (PP.H4) or have different ones (PP.H3) using coloc. This resulted in approximately 2,000–4,000 tests per region. We did not correct for multiple testing, so there is a real risk of false positives, and the results should be interpreted conservatively. eQTL data cover about 1 Mb around each gene, so for genes at the edge the data can miss part of the signal. We therefore focused on tests whose data covered at least half of the credible set probability.

Phewas
We also looked at genetic datasets of other diseases and traits to look for studies that found a similar signal as DecodeME. This is called a phenome-wide association study (phewas). For some there were none, while for others there were more than 200. These other studies help us understand what the DNA changes are doing. For example, if many immune diseases found a similar signal and no psychiatric diseases or brain traits, then it’s more likely to be an immune rather than a neurological locus.

For loci with many matching traits, we have suspicions that participation bias (in DecodeME or the UK Biobank) might have affected the results. This was the case for the regions that include MLLT10, HTT and PLCL1. At the regions of PCDH17 and SNTA1, the signal was shared with where people were born or live, which suggests population structure (small differences in ancestry across the UK).

Technical specifications
We used multiple data sources. Open Targets offers credible sets, so we checked if these overlapped with the credible set for DecodeME in this region. The GWAS Catalog lists associations with p < 10⁻⁵. We therefore looked up every DecodeME SNP in the region below this threshold to see whether it was listed for other traits. We used the lead variant and several other representative SNPs to check if the OpenGWAS API had other GWAS with a p < 10⁻⁴ for these SNPs. This approach queries tens of thousands of datasets and studies (OpenGWAS includes more than 50,000 studies). No correction for multiplicity was applied so the results should be interpreted cautiously. For each study found, we tried to download its summary statistics for the region (from the GWAS Catalog or OpenGWAS). For many traits and studies these weren’t available (often related to 23andME data that can’t be shared). For those that could be retrieved we performed coloc analysis to calculate PP.H4. We focused on results that had more than 1,000 variants in the region and that covered at least half of the credible set probability.

FLAMES
We used the FLAMES algorithm to estimate the most likely causal gene in each locus. FLAMES is a recent machine-learning tool that automatically combines many of the analyses described above using a wide range of datasets. It also uses PoPS, which looks at tens of thousands of gene categories (pathways, protein interactions, expression patterns) and ranks a gene as more plausible if it shares categories with the genes that the whole ME/CFS GWAS points to.

Technical specifications
We reached the same judgment as FLAMES in most regions although there were some exceptions. FLAMES, for example, only looks at genes within a 750 kb distance. At chromosome 6 for example, it doesn’t consider the gene POU3F2, even though it is connected through 3D contacts to the credible set. Instead, FLAMES made a confident prediction for MMS22L, because that was the only one it could ‘see’ even though it’s not particularly likely to be the causal one. We consider this a limitation of the algorithm. We also reached a different conclusion at chromosome 13 where FLAMES predicts OLFM4, but where we think PCDH8 is more likely due to Hi-C contacts and its brain enrichment. For most of the regions however, we agreed with FLAMES conclusions.

AlphaGenome
We also used AlphaGenome, a deep-learning AI system by Google DeepMind designed to decode gene regulation. We entered the SNPs in the credible set and AlphaGenome predicted for each one what the consequences would be in terms of gene expression, splicing, chromatin openness, histone marks and transcription-factor binding. This didn’t lead to much insight unfortunately because there are still many SNPs in the credible sets and we don’t know which one carries the true effect (and in which cell type or tissue). Most of the effects found in DecodeME are also small or unremarkable, so they do not really stand out. In some cases, however, AlphaGenome helped to formulate a concrete hypothesis such as a variant near ARFGEF2 on chromosome 20 (rs10547611) that is predicted to weaken the binding of MAF-family transcription factors.

Despite all these analyses the strongest pieces of information were location (how close the gene is to the credible set) and competition (how many other genes are in the region).

 

[ME/CFS Science Blog]

Over the past couple of weeks, I’ve assembled a list of genes that are likely associated with ME/CFS. As DecodeME only found 6 hits above the significance threshold, signals just below this threshold (p < 10⁻⁶) were also included, resulting in 23 loci.

An option would be to increase the number of genes and set the threshold at 5 * 10^-6 which would result in 46 loci (23 extra ones). But it's a lot of work and the noise will likely increase as I take on weaker signals. So I thought it would be best to start with this threshold first and try to delve deep into what is known about these genes.

Despite all these analyses the strongest pieces of information were location (how close the gene is to the credible set) and competition (how many other genes are in the region).

Happy to change the overview if new evidence is found or someone spots errors. The certainty of evidence claims for a particular gene (UNCERTAIN, WEAK, MODERATE) are just intended as a quick guide for the locus for people unable to read all the evidence and discussion.

Here’s more info about our analysis and what it is based on.

I've used AI with most of these analyses. It was particularly helpful in coding and retrieving the locus data from databases (rather than the full data) using tools such as Parquet and DuckDB. I do have a rule that I let AI propose the code in R notebooks that I can understand (tidyverse), adjust and run myself. This way, I can see and exactly what the code is doing, test intermediary results, change the analysis or plot by adjusting some of the code myself, etc. AI, however, is getting better so fast that I'm not sure that this will remain the best approach in the near future.

 

[Jonathan Edwards]

Oh goodness, I no longer have the excuse that there are too many threads on DecodeME genes. I will have to do my homework now. Great stuff M.

 

[Axel]

Well done – that’s brilliant. It gives us hope in the face of uncertainties in the outside world, such as the temporary closure of access to the UK Biobank: there are other ways to carry on with the analysis and not waste any time.

 

[EndME]

Great work @ME/CFS Science Blog ! I am full of admiration on how you have been pursuing the genetics of ME/CFS ever since the preprint came out. Incredible how you have been powering ahead!

 

[DHagen]

Amazing work and very helpful indeed, @ME/CFS Science Blog - thank you.

 

[mariovitali]

@ME/CFS Science Blog Thank you for this work. Alphagenome2 is now available and from what I understand this version might be more helpful : https://www.alphagenomedocs.com/getting_started.html

 

[hotblack]

@ME/CFS Science Blog Thank you for this work. Alphagenome2 is now available and from what I understand this version might be more helpful : https://www.alphagenomedocs.com/getting_started.html

Agree I’m very interested in what we/someone can do with AG2 and the DecodeME and later SequenceME data. There’s some great possibilities. More discussion on that tool here https://s4me.info/threads/advancing...ome-2026-avsec-et-al.48555/page-3#post-737725

Back to this thread, amazing work @ME/CFS Science Blog and it’s great to see the different techniques brought together like this.

 

[mariovitali]

@hotblack I am already working on it via code supplied by Claude. If anyone is interested I can share the notebook

 

[hotblack]

@hotblack I am already working on it via code supplied by Claude. If anyone is interested I can share the notebook

Very interested in seeing what you come up with! Less sure I can be much help at the moment.

My big hope here is AG2 allows both greater cell specificity and the ability to narrow which SNP may be interesting by looking at the effect scores. As well as having better insight into regulatory elements a long way from a gene (quite a few of the hits seem like they could fall into this category). Do you think it would be possible to do a wide search this way?

It sounds like AG2 still misses a lot so it wouldn’t be conclusive but apparently when it finds something significant it tends to be shown to be when followed up, so it could help us get to at least some of what is interesting and important sooner. What’s going on in more closed sections of the DNA may still be hidden to us but its a start!

My question had been how and who has the expertise to do this. AI tools are definitely helpful and maybe they’d be helpful enough for a few of us to do this? I also wondered if we could get a professional research team interested as it seems like a great project for someone. What do @mariovitali and @ME/CFS Science Blog think? Maybe @tralfamadorian97 and @forestglip too?

 

[mariovitali]

I do not have any expertise Genome wise. This is why it is important to have a reality check from AI generated code

 

[Verity]

For those of us who aren’t scientifically minded—is the next thing to do with these genes finding commonalities between them? What exactly are y’all looking for?

 

[ChronicallyOverIt]

Amazing work here as always!!

I guess what are the next steps?

We try to contact researchers that have been working tangentially to these genes?

We try to bring this up in grass roots local organizations hoping they can telegraph this to bigger charities?

We contact bigger charities asking they look at this genetic evidence before starting a trial?

I’d really like to think of ways to bring attention to these moderate genes. These are all just spit balling ideas.

 

[Amw66]

Amazing work . Thank you .

 

[V.R.T.]

Great stuff. Really appreciate the two posts explaining the analyses as well, as I get quite muddled on that stuff.