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

[SNT Gatchaman]

Can we make a table of the proteins and functions for these and maybe check further down for sister genes where they show up?

DNMT3A. DNA methyltransferase 3 alpha Epigenetic DNA methylation -gene expression control
ADCY10 adenylyl cyclase 10. Formation of cAMP
PPP2R2A. Protein Phosphatase 2 Regulatory Subunit Balpha. Cell cycle
NLGN2. Neuroligin 2. Synapse formation
LEP Leptin Weight control/appetite
SYNGAP1. Synaptic Ras GTPase Activating Protein 1. Synapses MAPkinase signalling
AHCYL2. Adenosylhomocysteinase Like 2. Brain signalling
NLGN1. Neuroligin 1 Synapse formation
DLGAP4. DLG Associated Protein 4. Synapses
HDAC1. Histone Deacetylase 1. Regulation of gene transcription
AMPD2 adenosine monophosphate deaminase 2
AHCYL1. Adenosylhomocysteinase Like 1. Anti-inflammatory cytokine production
SHARPIN. SHANK Associated RH Domain Interactor. Signalling in auto inflammation. Synapses
NME2. NME/NM23 Nucleoside Diphosphate Kinase 2. DNA transcription. Risk factor for EBV-associated lymphoma
NME1-NME2. NME1-NME2 Readthrough. DNA transcription
CACNA2D3. Linked to brain development and autism TCR signalling
NME3. NME/NM23 Nucleoside Diphosphate Kinase 3. Goes with NME1
ZC3H13. NME/NM23 Nucleoside Diphosphate Kinase 3. RNA splicing
CAMK2A. Calcium/Calmodulin Dependent Protein Kinase II Alpha. Synapses on dendrites in brain
PIK3CA. Phosphatidylinositol-4,5-Bisphosphate 3-Kinase Catalytic Subunit Alpha. Insulin responses, brain development
MAX. MYC Associated Factor X. DNA transcription
HLA-C HLA-C (MHC I) CD8 T cell receptor and NK cell receptor recognition events
ACE. Angiotensin I Converting Enzyme

We should keep a wide purview. While many of these genes are flagged as neural and relating to synaptic function, they can have important non-canonical roles. The potential link with autism development is fascinating but autism spectrum disorder has other features beyond the neurodevelopmental, eg gastrointestinal dysfunction. So while their effect on synapse formation and maintenance would be important in the primary neurodevelopmental abnormalities we observe, the very common comorbid problems with gut function might be more due to epithelial tight junctions and barrier integrity than with the gut's neural connections.

Tons of us have OI. Does anyone know what synapses would have to be messed up for OI to occur? Are any OI-type genes showing up in the Zhang results?

We should also consider that genes identified in OI might relate more directly to vascular (esp. endothelial cell) function rather than neuronal synapses. Eg NLGN1 and NLGN2 are expressed in vascular endothelial cells. There's also potentially an endocrine aspect, as NLGN2 is also expressed in pancreatic beta cells (insulin secretion), so maybe there's also a link between the suggested GIP secretion (splanchnic vasodilator) and post-prandial increased POTS/OI symptoms.

Neuroligin 1 Induces Blood Vessel Maturation by Cooperating with the α6 Integrin (2014, Journal of Biological Chemistry)

Modulation of Angiopoietin 2 release from endothelial cells and angiogenesis by the synaptic protein Neuroligin 2 (2018, Biochemical and Biophysical Research Communications)

Altered Pancreatic Islet Function and Morphology in Mice Lacking the Beta-Cell Surface Protein Neuroligin-2 (2013, PLOS ONE)

Worsening Postural Tachycardia Syndrome Is Associated With Increased Glucose-Dependent Insulinotropic Polypeptide Secretion (2022, Hypertension)

 

[wastwater]

Notch3 and cadasil was something I looked at and in the early stages a lot of people have knock out fatigue and mood disorders

 

[forestglip]

After looking at the paper again, I realized I should have done my GSEA analysis using p-values, not attention scores. I don't fully understand their method for interpreting the model, but p-values are the metric they used for choosing the top 115 that they say are the most important. It didn't really change the results much since p-value rankings and attention score rankings are pretty similar. (But an exception, for example, is that HLA-C is ranked 22 using p-values and 2077 using attention scores). Again, very similar results on the Genebass CFS data, but the run with p-values can be seen on the last link here.

But I thought it might also be useful to just see the top ranked clusters using the Zhang HEAL2 list of genes. Basically another way to do it from their method of taking their top ranked 115 genes and seeing what pathways these specific genes are enriched in, without considering ranking. Instead I uploaded all of the genes with their -log10 p values to STRING so that it looks at all of them and weighs them by the metric that is apparently useful for interpreting the most important genes.

Here is the link to the enrichment analysis with many different gene set collections. (The page is pretty slow to load and interact with.) I merged items by similarity greater than 0.5, and filtered by FDR < .001 and enrichment score > 1.0. ("Top of input" for direction is what's interesting since it relates to to genes with low p-values.) Filters can be changed near the bottom of the page.

Here are the top 10 STRING local clusters from its protein-protein interaction network, which I think should be most comparable to the enrichment analysis they did that found synaptic function and proteasomes.

Local Network Cluster (STRING)


There are some other interesting gene sets in other collections, though. For example, in DISEASES, there are the following four that match my filters. This includes COVID and two diseases that may be related to sex differences,. If I filter with a less conservative FDR, then autistic disorder is the highest enriched disease with FDR=.0078.

Disease-gene Associations (DISEASES)


For the "Tissue Expression (TISSUES)" collection, the two gene sets are "autonomic nervous system" and "brain ventricle". For "Human Phenotype (Monarch)", the top ranked phenotype is "myeloid leukemia".

Tissue Expression (TISSUES)


Human Phenotype (Monarch)

 

[tralfamadorian97]

Just finished a quick read of this paper. Overall, this seems promising. I look forward to seeing if this paper’s findings match those of analyses performed by other groups on different data. And the discussion so far in this thread has been insightful. I wish I had the bioinformatics background to contribute.

I do have a bit of background in deep learning research. Some thoughts from that perspective:

1. In the deep learning literature, when you present a complex model, you are typically asked by reviewers to perform an ablation study. This involves systematically removing different components of your system and reporting how this removal affects the system’s performance. Ablation studies reveal which components of the system are most important. I would be interested in seeing an ablation study for HEAL2. I appreciate that the authors compare HEAL1 to HEAL2, but I would like to see finer granularity. This could involve ablating the various input-preprocessing steps, as well as numerous components of the neural architecture. I expect this would shed useful light on how the model is making its predictions.
2. Much of the practical value of this paper comes from the list of “ME/CFS risk genes” that it generates. These risk genes are a consequence of attention scores. The use of attention scores for model explanation is reasonable and fairly standard, though not without some controversy (See here and here). It would be interesting to see whether other model explanation techniques (for example, integrated gradients) produce concordant lists of risk genes. If so, this would increase confidence in the results.

EDIT:
I also just realized that the authors are using a non-standard form of attention. In equation 4, the Sigmoid would be replaced by Softmax in standard attention. I am curious as to what motivated them to make this change.

The authors also say "The source code of HEAL2 has been uploaded in the Supplementary Materials.", but I don't see any source code link on medrxiv. Presumably they plan to add this later.

I think it could be clarifying to explore the code a bit. I'll try to do this once the code is available.

 

[Utsikt]

Much of the practical value of this paper comes from the list of “ME/CFS risk genes” that it generates. These risk genes are a consequence of attention scores.

Can you give an ELI5 of attention scores?

 

[tralfamadorian97]

Can you give an ELI5 of attention scores?

I can certainly try. Let me know if this helps:

• The input to HEAL2 is a bunch of statistics describing a person’s genome.
• HEAL2 operates on this input gene by gene.
  • That is: it performs one set of neural computations on gene A related input to produce a gene A output, and another set of neural computations on gene B related input to produce a gene B output, etc.
    • (Aside: This is slightly complicated by the graph structure: if gene A is known to interact with gene B according to the STRING database, the gene A neural computation is allowed to influence the gene B neural computation and vice versa)
• To predict a person’s ME/CFS status, the gene-specific outputs need to be aggregated.
• Attention is a mechanism of aggregation.
• Attention scores are numbers used as aggregation weights. So all else being equal, if the attention score for gene A is larger than the attention score for gene B, we should expect the gene A output to have more influence on the final prediction than the gene B output.
• Attention scores are dynamic. This means that the attention scores of person X for genes A and B will differ from the attention scores of person Y for genes A and B
• The authors identify ME/CFS risk genes by finding genes for which the attention scores of patients tend to be larger than the attention scores of controls.

 

[Hoopoe]

ChatGPT made me a nice summary of the highlighted genes in the article:

Here’s what current research reveals about diseases most closely associated with your gene list: ACE, NAMPT, IL12A, PSMB4, PSMB5, PSMB7, PSMD7, DNMT3A, NME1, NRAS, SYNGAP1, NLGN1, DLG2, GRM1, DLGAP1‑4.

---

Disease Associations by Gene

IL12A

• Genetic polymorphisms in IL12A (e.g. rs583911, rs568408) are significantly associated with primary biliary cholangitis (PBC) and multiple sclerosis (MS) (PubMed).
• Additional links to childhood acute lymphoblastic leukemia (ALL) via immune-pathway modulation (AACR Journals).

NAMPT

• Elevated expression or genetic variants correlate with rheumatoid arthritis (RA), where NAMPT acts as a pro-inflammatory mediator in joints (PMC).
• Polymorphisms such as rs61330082 and rs9034 are linked with increased risk of dilated cardiomyopathy (DCM), hypertension, and type 2 diabetes with hypertension (PubMed).
• Elevated NAMPT is associated with atherosclerosis, pulmonary arterial hypertension, metabolic syndrome, and neurodegeneration (e.g. Alzheimer’s, Parkinson’s) (PMC).

DNMT3A

• Although specific associations with DNMT3A are not fully covered, dysregulation of DNA methyltransferases (like DNMT1 and DNMT3A) is implicated in systemic lupus erythematosus (SLE), systemic sclerosis, Type 1 diabetes, and autoimmune disorders via global hypomethylation in T cells (Wikipedia).

Proteasome Subunits (PSMB4 / PSMB5 / PSMB7 / PSMD7)

• The proteasome system plays a role in neurodegenerative diseases (e.g. Alzheimer’s, Parkinson’s, ALS) through impaired protein clearance.
• Abnormal proteasome subunit expression is also reported in autoimmune diseases (RA, SLE) and various cancers (e.g. hepatocellular carcinoma, glioblastoma).

]SYNGAP1, NLGN1, DLG2, GRM1, DLGAP1‑4

• These genes encode synaptic scaffolding proteins and have strong associations with neurodevelopmental disorders:
  • Autism spectrum disorder (ASD), intellectual disability, and epilepsy are frequently linked to mutations in SYNGAP1, NLGN1, DLG2, and DLGAP family genes.
  • GRM1 (a metabotropic glutamate receptor) has been associated with spinocerebellar ataxia and neurobehavioral syndromes in rare familial cases.

NRAS

• A well-established oncogene, with activating mutations driving cancers such as melanoma, acute leukemia, and thyroid carcinoma.

NME1

• Involved in metastasis suppression; reduced NME1 expression correlates with worsened outcomes in melanoma, breast cancer, and other epithelial tumors.

ACE

• Polymorphisms in ACE have been associated with Alzheimer’s disease, diabetic nephropathy, and migraine susceptibility, reflecting its roles in vascular, metabolic, and neurological pathways.

---

Consolidated Associations Table​

Gene(s)	Notable Disease Associations
IL12A	Primary biliary cholangitis (PBC), Multiple sclerosis (MS), Childhood ALL
NAMPT	Rheumatoid arthritis, Dilated cardiomyopathy, Hypertension in T2DM, Atherosclerosis, Pulmonary hypertension, Neurodegeneration
DNMT3A	Autoimmune disorders (SLE, sclerosis, T1D) via epigenetic dysregulation
PSMB4 / PSMB5 / PSMB7 / PSMD7	Neurodegeneration (AD, PD, ALS), Autoimmune diseases, Cancer
SYNGAP1, NLGN1, DLG2, GRM1, DLGAP1‑DLGAP4	Autism, intellectual disability, epilepsy, neurodevelopmental syndromes, cerebellar ataxia
NRAS	Melanoma, acute leukemias, thyroid cancers
NME1	Metastatic progression in melanoma, breast cancer, epithelial tumors
ACE	Alzheimer’s disease, diabetic nephropathy, migraine

---

Integrated Interpretation​

• Immune-mediated diseases: IL12A and NAMPT are involved in inflammation and autoimmune pathways, linking them to RA, PBC, MS, and metabolic–cardiovascular conditions.
• Neurodevelopmental/synaptic disorders: The NLGN, DLGAP, SYNGAP1, and GRM1 group strongly maps onto autism, epilepsy, ID, and synaptic dysfunction syndromes.
• Oncogenesis and epigenetics: NRAS, NME1, proteasome subunits, and DNMT3A converge on cancer biology and abnormalities in cell‑cycle regulation or protein turnover.
• Epigenetic dysregulation: DNMT3A dysfunction is structurally aligned with autoimmune and inflammatory conditions via methylation disruption.

---

 

[Hoopoe]

Here's one hypothesis by ChatGPT on how all these genetic risk factors might converge to cause a disease. It didn't know that these genes were related to ME/CFS and I wanted to see if it could come up with something that seems useful.

A congenital or early-onset disorder where:

• Synaptic and neurodevelopmental genes (SYNGAP1, NLGN1, etc.) create vulnerability to impaired neural circuit formation → leads to cognitive, behavioral, and sensory symptoms.
• Immune dysregulation (via IL12A, NAMPT, ACE) amplifies neuroinflammation, possibly through glial activation and peripheral cytokine spillover into the CNS.
• Epigenetic dysregulation (DNMT3A, NME1) reduces cellular adaptability, impairs synaptic plasticity, or enhances immune reactivity.
• Proteostasis failure (PSMBs, PSMD7) leads to accumulation of damaged or misfolded proteins, compounding neural and immune stress.

Potential Clinical Features:​

• Early-onset cognitive delay or ASD-like behavior
• Episodic neuroinflammation (similar to PANS/PANDAS or autoimmune encephalitis)
• Risk of seizures or neurodegeneration
• Immune comorbidities: autoimmunity, atopy, chronic infections
• Possible increased cancer risk (NRAS/DNMT3A involvement)

The clinical features don't seem very convincing to me. The idea that all these risk factors could be acting together on the brain to cause problems seems interesting and plausible. It is a model of multiple hits on the same system. Here I asked only about genetic risk factors, but it's easy to see how infection can easily fit in as non-genetic risk factor.

 

[Trish]

ChatGPT made me a nice summary of the highlighted genes in the article:

As soon as I see someone has used AI I stop reading. I don't trust it to know what the important points are and not to add inaccurate stuff.

 

[chillier]

I don't know if it's been discussed much in this thread what the impact of the loss of function of these genes, especially the HOMER-SHANK-DLGAP complex would have on the function of the synapse. I think there's also a plausible link to the proteasome genes that also come up in this paper directly to the SHANK complex.

This review: The DLGAP family: neuronal expression, function and role in brain disorders goes into the function of this complex. It seems to be a part of the structural integrity of post synaptic density by physically linking the glutamate receptors AMPAR, NMDAR, mGluR to the cytoskeleton. The complex is involved in synaptic scaling, which I understand is a homeostatic mechanism that takes places over hours and days to increase or decrease the sensitivity of the synapse in response to increased or decreased stimulation.

excess receptor stimulation leads to a breakdown and degradation of the SHANK complex by the proteasome, leaving the untethered glutamate receptors to be internalised via endocytosis reducing the sensitivity of the synapse. The complex is rebuilt when there is low receptor stimulation.

​

 

[chillier]

So it seems like loss of function of the proteins in this complex would if anything lead to an excessive downscaling or loss of integrity of the synapse. It's a bit complicated though because I think the loss of some Homer proteins would actually have the opposite effect, and the same can be said of SYNGAP1, the loss of which would lead to excess MAPK signalling and an increase in membrane receptors.

Because of the way this paper uses protein complexes as an input to their model, i think it's difficult to interpret each of the protein members within the SHANK complex individually and it might be you have to interpret the complex as a whole, which muddies things further.

I think you could perhaps argue that loss of core proteasome subunits would mean the dissasembled (and ubiquinated) SHANK complex members would be unable to be degraded in good time and therefore accumulate in the post synapse, kind of acting as a dominant negative preventing the complex from reforming - therefore acting to increase downscaling/ reduce firing.

 

[hotblack]

The complex is involved in synaptic scaling, which I understand is a homeostatic mechanism that takes places over hours and days to increase or decrease the sensitivity of the synapse in response to increased or decreased stimulation.

Really interesting. The questions I immediately have are

Did any of these or related show up with any significance in DecodeME? (I haven’t yet got the data, still on my todo list, but if anyone has and can share any details for DLG1, DLG4, DYNLL1, DYNLL2, SHANK2, HOMER1?)

And is there any way of measuring this, what sort of studies would we need to understand if there is something here related to ME/CFS?

 

[DMissa]

As soon as I see someone has used AI I stop reading. I don't trust it to know what the important points are and not to add inaccurate stuff.

It has its uses (mostly to suggest starting points) but yes I wouldn't accept what it says without verifying specifics. I put the decodeME genes in there as an exercise to test its sanity and the very first gene summary had two completely fabricated factual inaccuracies.

 

[Simon M]

So, a clean discovery cohort of 247 cases and 192 controls, and a testing cohort of 36 cases and 21 controls. 115 risk genes is a lot to come out of a relatively small sample

I didn't have the capacity to follow this thread properly when it came out, so I I'm coming back to it now. But this has been bugging me.

We know that there is a heritable component to.ME/CFS, but it's not that common. I think there's an Utah registry study that gave heritability at about 0.13. The recent EMEA survey found 13% of people reported having a 1st degree relative with the illness. DecodeME found SNP-based heritability of 0.09 (I think). And GWAS look at common variants , which are 90%-ish in non-coding regions. All of which suggests that coding Variants, likely loss of function ones, are pretty rare.

Yet here we have 115 such variants identified from a sample of only 247. Note that genetically related individuals were excluded from this analysis (I believe Stanford specifically recruited some related individuals with the illness).

I don't know much about WGS, so could somebody explain to me why such an apparently-high rate of these variants is plausible?

Thanks. And apologies if this is already been covered in this extensive thread.

 

[Kitty]

I've never been sure I really understood it.

We show that HEAL2 not only has predictive value for ME/CFS based on personal rare variants, but also links genetic risk to various ME/CFS-associated symptoms. Model interpretation of HEAL2 identifies 115 ME/CFS-risk genes that exhibit significant intolerance to loss-of-function (LoF) mutations.

My primary school-level understanding is that:

They looked for variants that the ME/CFS cases had and the controls didn't.

Then they looked at the potential effects of those variants on the clinical picture that occurs in ME/CFS. They picked out the ones that look relevant, aided by evidence from other studies.

It's possible not all those variants actually are a risk for ME/CFS, or for ME/CFS specifically. But it looks very likely some of them do have an effect, even if we can't be sure which.

So 115 genes is the long list, which still needs filtering.

(I did say primary school!)

 

[Simon M]

So 115 genes is the long list, which still needs filtering.

Thanks, that's good to know. But presumably that could substantially affect interpretation of the study if many of these repairs genes turn out not to be relevant.

Do we know how many coding variants (intolerant of loss of function) a healthy person has? I'm trying to get a feel if it's a rare thing or quite common.

 

[Kitty]

But presumably that could substantially affect interpretation of the study if many of these repairs genes turn out not to be relevant.

I imagine it'd be expected that some of them would be spurious or coincidence. When you're relying on statistics, probabilities and machine learning techniques to draw a picture because you haven't got all the data, some of it's likely to be distorted or gappy.

Do we know how many coding variants (intolerant of loss of function) a healthy person has? I'm trying to get a feel if it's a rare thing or quite common.

I've read (I don't have the source, it was a while ago) that rare variants are fairly common. I got the impression people are likely to have dozens to 100s of them.

But if they're rare we presumably don't know much about some of them. A few will be well known because they're linked to genetic disease, but it might be hard to calculate the significance of others.

This isn't from the article, but I also wonder about the designation of "rare". It's possible that if you sequenced everyone in the UK or Brazil, for instance, some of the rare variants might turn out not to be so rare. Our databases are still relatively small.

(Might also be worth reiterating that I could be talking through my hat, I've no expertise here at all!)

 

[Hoopoe]

Do we know how many coding variants (intolerant of loss of function) a healthy person has? I'm trying to get a feel if it's a rare thing or quite common.

ChatGPT says

The average person has about 100 predicted-loss-of-function variants (pLoF), but in the average healthy person, there are only 0-2 heterozygous pLoF variants in LoF intolerance genes. Homozygous pLoF variants are usually incompatible with life and extremely rare.

Most of the pLoF variants are tolerated because they occur in genes that are not dosage sensitive.

PS: loss of function tolerance is a gradient, not binary.

 

[Jonathan Edwards]

The average person has about 100 predicted-loss-of-function variants (pLoF), but in the average healthy person, there are only 0-2 heterozygous pLoF variants in LoF intolerance genes.

This seems too simplistic to be useful. I assume that some variants alter function or even lead to loss of function in a way that does not matter most of the time but may still confer significant risk for developing a disease with multifactorial causation. Complement in lupus is probably a good analogy. C4 variants that do not alter plasma C4 levels can nonetheless confer risk for lupus as I understand it.

 

[Simon M]

Does the study say how many individuals had at least one of these 115 risk genes? I'm still concerned that the number of genes is much too high, given the evidence we have on heritability.

We know that some of the inherited risk is through the genetic signals identified by DecodeME , and these are likely to be 90% non-coding, so different from these. So perhaps we can expect 10% accounted for by coding Variants. Hence my interest in the number of individuals this study identified as having risk variants here.