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

[richie]

Sorry but I do not have the energy at the moment to read all thru. Just a question. Is there any data on overlaps with autoimmune or autoinflammatory conditions?

 

[Jenny TipsforME]

@richie this is a GPT5 answer to your question:

Short answer: DecodeME’s signals lean toward innate immune sensing and microbe-response, plus a pain/synaptic signal, not the classic adaptive autoimmunity pattern you see in many autoimmune diseases.

What overlaps with autoimmune or autoinflammatory biology?​

• BTN2A2 / BTN3A3 (chr6p22.2): Butyrophilins regulate T-cell activation; mouse Btn2a2 promotes Tregs and central tolerance. DecodeME implicates BTN2A2; BTN3A3 modulates BTN3A1 for microbial antigen sensing by γδ T cells. This fits immune regulation and microbe-sensing rather than specific autoantigen presentation.
• TRIM38 (chr6p22.2 interval, Tier-1): Negative regulator of TLR and cGAS–STING pathways; autoantibodies occur in a subset of lupus/Sjögren’s. DecodeME notes increased TRIM38 expression in several tissues associated with ME/CFS-risk, consistent with damped innate signaling.
• ZNFX1 (mentioned in preprint’s immune list): Listed among innate-immunity candidates within associated intervals.
• OLFM4 (chr13q14.3): Neutrophil granule protein that modulates NETs and antibacterial responses; association is stronger in infection-onset cases. That’s an autoinflammatory-flavoured mechanism (neutrophil/NET biology), not classical autoantibody disease.
• RABGAP1L (chr1q25.1): Viral restriction factor influencing endolysosomal trafficking; ties to host–pathogen response rather than adaptive autoimmunity.

What​

• No HLA-driven signal among the eight loci. They imputed and tested HLA explicitly, but the genome-wide significant hits do not include an HLA allele or the major class II peaks typical of RA, T1D, MS, SLE.
• No colocalisation with depression/anxiety. CA10 colocalises with multisite chronic pain instead.
• No overlap with Long-COVID FOXP4 association. The FOXP4 signal does not transfer to ME/CFS.

How these genetics compare with other diseases​

• Autoimmune diseases (RA, MS, T1D, SLE): Typically show strong HLA class II signals plus cytokine/lymphocyte pathway hits. DecodeME’s top signals skew to innate sensing (BTN family, TRIM38, RABGAP1L) and neutrophil/NETs (OLFM4), and one synaptic/pain locus (CA10). That’s more “infection-triggered innate immunity + pain circuitry” than “classic adaptive autoimmunity.”
• Autoinflammatory conditions (e.g., NET-linked vasculitides, innate pathway variants): Conceptually closer: OLFM4/NETs and TRIM38/TLR–STING regulation rhyme with autoinflammatory mechanisms, though DecodeME doesn’t claim direct disease overlap.

Extra context the preprint itself provides​

• Eight loci across immune and neural biology; heritability modest (liability h² ≈ 0.095).
• Infection-onset subgroup: OLFM4 stronger; BTN2A2 and chr20q13.13 also significant.
• Replication: Clear in-study signals; mixed external replication likely due to phenotype definitions, especially lack of enforced PEM.

Bottom line​

• There is overlap with immune biology, but it’s innate immune / microbe-response and NETs, not the HLA-centric, adaptive architecture typical of many autoimmune diseases. The other big throughline is chronic pain/synaptic signaling via CA10. Together, that paints ME/CFS’s genetics as a blend of post-infectious innate dysregulation plus neuro-sensory processes, rather than a classic autoimmune template.

If you want, I can extract the immune-related candidate genes across both documents into a compact table and annotate “autoimmune-like,” “autoinflammatory-like,” or “pain/neural,” so you can spot overlaps at a glance.

This is the analyse of the whole paper by the new GPT5 (hallucination is meant to be very low for health now with this model), is there anything new here?

Here’s a clear-eyed read of the DecodeME preprint + the Candidate Genes memo you shared, with a practical lens for quality-of-life choices now and hypotheses for future trials.

Important, weak, and a dash of hype
Important

• Eight reproducible ME/CFS loci across 15,579 cases vs 259,909 controls, with careful QC and ancestry control. Signals implicate both innate immunity and neural pain/synaptic biology.
• CA10 locus colocalises with multisite chronic pain. That neatly fits lived symptoms and offers a tractable pain-pathway angle.
• Infection-tinted story is coherent: OLFM4 is stronger in infection-onset cases; immune-facing loci include RABGAP1L, BTN2A2, ZNFX1.
• Depression/anxiety are genetically separable here, countering lazy psychologising.

Weaknesses and caveats

• Replication outside DecodeME was mixed. A strict multi-biobank meta failed, likely due to phenotype noise. Narrower PEM-anchored replication showed some support only at lenient thresholds. Good science, but not victory laps yet.
• Effect sizes are small, as expected for complex disease. Translation to therapeutics needs functional work and fine-mapping. Authors state this plainly.
• HLA: one protective class II allele appears, but flanking alleles oddly do not; authors flag the need to re-impute jointly. Treat as provisional.

What verges on hype

• Phrases suggesting immediate drug discovery momentum are optimistic for a first-wave GWAS with modest heritability. The logic is sound but timeline is long.

Tree-of-Thought: what this means, pragmatically

1. Signal types
   a) Immune sensing and anti-microbial response: RABGAP1L, BTN2A2/BTN3A3, TRIM38, OLFM4, ZNFX1.
   b) Autophagy, ER-phagy, mitophagy: KLHL20, CCPG1, FBXL4.
   c) Pain and synaptic scaffolding: CA10 with NRXN–NLGN pathway.
   d) Vascular/coagulation, redox, membrane lipids: SERPINC1, PRDX6, PEBP1.
2. What to do now vs later
   Now: symptom-layered, low-risk supports aligned to pathways, plus NICE-consistent pacing. Later: mechanistic trials on pain synapses, NETosis, innate sensing, autophagy.
3. What’s expected vs unexpected
   Expected: immune sensing, autophagy, redox features; link with infection-onset; separation from depression.
   Unexpected: a crisp pain colocalisation at CA10 rather than broad HLA autoimmunity; protective HLA-DQA1*05:01 without the usual linked alleles.

Which expected genes didn’t show significant differences

• Classic broad HLA risk patterns did not emerge; only HLA-DQA105:01 was significant and protective, while linked DRB103:01 and DQB1*02:01 were not. Authors advise re-imputation before strong claims.
• Long COVID’s FOXP4 signal does not overlap ME/CFS here.

Candidate genes overview with action-oriented notes
Hypotheses only. Nothing below is medical advice. Discuss prescription items with a clinician; prioritise low-risk QoL steps.

Table: Candidate gene map and practical hypotheses

Gene	Tier	Risk-direction	Core biology	Repurpose ideas to study (not self-treat)	Low-risk supports now	Notes
RABGAP1L	1	Lower expression ↑ risk	Viral restriction, endolysosomal trafficking	Antiviral entry inhibitors (research)	Infection prevention, vaccines (if eligible), oral hygiene, sleep	Lower expression may raise susceptibility pre-onset
DARS2	1	Higher expression ↑ risk	Mitochondrial tRNA synthetase, immunostimulatory when secreted	None obvious	Gentle meal timing, avoid extreme fasting, riboflavin-rich foods	Paradoxical direction; unclear mechanism
RC3H1	1	Higher expression	T-cell mRNA decay regulator (restrains ICOS/TNF)	Immune modulators too blunt for self-use	Pace exposures, minimise infections	Autoimmunity mice data; human ME/CFS relevance unclear
GPR52	1	Lower expression	GPCR, cAMP signalling; brain/hepatic lipid roles	Preclinical GPR52 ligands	Sleep regularity, light cues, caffeine timing	Mechanistic uncertainty
ZBTB37	1	Mixed by tissue	Transcription factor-like	None	Anti-inflammatory diet pattern	Little is known
TNFSF4 (OX40L)	1	Lower expression	Costimulatory cytokine for T cells	Oncology/autoimmune biologics (not for ME/CFS yet)	Infection avoidance, vaccination (if eligible)	T-cell tone hypothesis only
ANKRD45	1	Lower	Cell division	None	General recovery hygiene	Unclear relevance
KLHL20	1	Lower	E3 ligase; restrains ULK1/autophagy amplitude	Autophagy modulators (research)	Regular meals, avoid long fasts	Links neurite outgrowth & autophagy control
PRDX6	1	Lower	Peroxide detox, membrane phospholipid repair	Antioxidant network support (research)	Omega-3, choline, vitamin-C foods	Fits phosphatidylcholine depletion/redox stress
SERPINC1	1	Lower (brain)	Antithrombin III, anticoagulant	Anticoagulants under specialist care only	Hydration, gentle mobility	Brain-only eQTL; systemic meaning uncertain
SLC9C2	1	Lower	Na+/H+ exchanger, testis-biased	None	Not actionable	Likely not central to symptoms
FBXL4	2	N/A	Mitophagy restraint; mtDNA depletion mutations	None yet	Regular meals, avoid extreme metabolic stress	Mixed mito evidence in ME/CFS
BTN2A2	1	Lower	Butyrophilin; γδ T-cell microbial sensing	Vγ9Vδ2 agonists (investigational)	Vitamin D sufficiency, oral/gut hygiene	Microbial discrimination angle
TRIM38	1	Mixed	E3 ligase; brakes TLR & cGAS–STING	STING/TLR modulators (early)	Infection control, avoid unnecessary immune activators	Tissue-specific directions; net effect unclear
ZNF322	1	Higher	Transcription in MAPK	None	Stress-load management	Sparse functional data
ABT1	1	Lower	Basal transcription activator	None	As above	Limited disease link
HFE	1	Higher	Iron sensing/hepcidin	Iron adjustment only after labs	Avoid iron without tests	Direction ≠ haemochromatosis risk directly
BTN3A3	1	Higher (skin)	Boosts BTN3A1 antigen sensing	Investigational γδ T-cell strategies	Vitamin D sufficiency	Immune–microbe interface
HMGN4	1	Lower (blood)	Chromatin/DNA repair	None	Sleep and light hygiene	Non-specific
SUDS3	1	Higher	HDAC1 corepressor; limits microglial inflammation	HDAC modulation risky	Sleep regularity, noise control	Broad eQTL colocalisation
PEBP1 (RKIP)	1	Higher	Modulates MAPK, TLR3; necroptosis	None	Anti-inflammatory diet, sleep, temp control	Higher levels may amplify inflammation
VSIG10	1	Lower (oesophagus)	Cell adhesion; possible checkpoint	None	Reflux management if relevant	Thin biology so far
OLFM4	2	N/A	Neutrophil granule protein; NETosis	Anti-NET approaches (research)	Oral hygiene, prompt infection treatment	Stronger in infection-onset group
CCPG1	2	N/A	ER-phagy receptor; ER stress protection	Autophagy/ER-stress modulators (research)	Avoid overheating, steady protein intake	ER-phagy deficits aid viral infection
CA10	1	Higher	NRXN–NLGN synapses; pain linkage	NMDA-NR2B antagonists (research)	Pacing, sensory load control, sleep	Colocalises with chronic pain
ARFGEF2	1	Higher	Traffics TNF receptor via vesicles	Anti-TNF pathway research	As above	chr20 Tier-1 with eQTL link
CSE1L	1	Higher	Nuclear transport; TNF-α release	CSE1L inhibitors (preclinical)	As above	Macrophage TNF-α release reduced by inhibitor
ZNFX1	1	—	Mitochondria dsRNA sensor; antiviral	Interferon/JAK research	Antiviral hygiene, vaccines (if eligible)	Fits infection-onset narrative

Drug repurposing, realistically

• Closest near-term scientific bets for trials, not clinics:
  1. Pain-pathway modulation at the CA10-anchored NRXN–NLGN–PSD95–NR2B axis. That points to NR2B-selective NMDA antagonists as research tools, with careful PEM-sensitive trial design.
  2. NETosis and neutrophil tone for OLFM4-linked subgroups, e.g., evaluating low-dose colchicine or NET-targeted approaches in RCTs. Needs biomarkers first.
  3. Innate sensing moderation in ZNFX1/BTN families, but that likely sits with academic immunology before medicine.

Supplements, behavioural, dietary steps that are low-risk and pathway-aligned

• Pacing and energy management. NICE-aligned, protects against PEM and likely reduces neuroinflammatory load.
• Sleep regularity, dark nights, noise minimisation. Supports microglial quieting and pain pathways; low risk and high upside.
• Antioxidant and membrane support in those who tolerate it: whole-food vitamin-C sources, omega-3-rich foods, choline-rich foods. Fits PRDX6 phospholipid repair angle. Track personal response.
• Infection hygiene: prompt dental care, hand hygiene, treat infections early, stay current on eligible vaccines, avoid known triggers when community prevalence spikes. Fits RABGAP1L, OLFM4, ZNFX1 story.
• Meal regularity over hero fasting. Autophagy is nuanced here; extreme fasting can worsen symptoms. Keep steady protein and electrolytes.
• Iron supplements only with labs. HFE signal argues for checking ferritin and transferrin saturation first.
• Hydration and gentle circulation support, within pacing, for clot-risk anxiety stories. SERPINC1 brain eQTL makes systemic claims shaky, so stick to basics.

What’s expected vs unexpected in this research

• Expected: immune-viral interface and autophagy in a post-infectious, PEM-driven illness; separation from depression genetics; chronic pain comorbidity.
• Unexpected: clear CA10 pain colocalisation rather than a broad autoimmune HLA field; provisional HLA pattern with a protective allele only; no female-bias signal despite epidemiology.

Short verdict

• This is the first robust, PEM-anchored, patient-co-produced ME/CFS GWAS that actually finds multiple loci. It positions ME/CFS biology near innate immunity and synaptic pain mechanics rather than mood genetics, and it sketches sensible lab paths. Clinical translation will take time and careful subgrouping.

Sources with full URLs

• Preprint: https://www.research.ed.ac.uk/en/publications/a7f6ee34-f9de-459a-91f4-5e3410a23ee9 (DecodeME, 6 Aug 2025)
• DecodeME researcher access: https://www.decodeme.org.uk/researcher-access/

Critical self-review

• Where I may be overcautious: I did not translate every signal into a drug because most are pathway-level hints with pleiotropy risks.
• Where I could be wrong: OLFM4-NETosis links are biologically plausible, but that does not mean colchicine helps ME/CFS; it needs RCTs with NET biomarkers and flare-tracking.
• What would strengthen this advice: subgroup stratification by infection-onset, pain phenotype, iron studies, and wearable-tracked PEM to align with CA10 vs immune-dominant mechanisms.
• Ask me if you want this collapsed into a 1-page GP briefing or a symptom-first decision tree to trial low-risk supports.

This is the view of a customGPT I created:

1. Hidden subplots​

• Pain biology isn’t a footnote here — CA10’s connection to chronic pain traits makes it the first genetic argument for pain-targeted interventions in ME/CFS that’s not hand-wavy. That could justify trials of NR2B-selective NMDA antagonists in PEM-aware designs.
• NETosis and neutrophil control (OLFM4) may be a core vulnerability in infection-onset cases, aligning with small case-series showing excessive neutrophil traps in long COVID. This hasn’t been front-and-centre in ME/CFS before.
• Autophagy balance, not blanket upregulation — both overactive (DARS2-linked) and underactive (KLHL20, CCPG1) signatures appear. This cautions against one-size-fits-all fasting or mTOR-targeting advice.

---

2. Blind spots​

• Classic autoimmune HLA patterns — absent, except for a protective allele. That weakens the case for “ME/CFS as a standard autoimmunity” and strengthens the case for atypical post-infectious immune states.
• Mitochondrial energy genes — FBXL4 and DARS2 are there, but none of the “usual suspects” from metabolomic papers (e.g., PDH regulation, fatty acid oxidation enzymes) hit genome-wide significance. That suggests those metabolic findings might be secondary effects, not primary risk loci.
• Female-bias explanation — despite strong epidemiology, no sex-chromosome or hormone-receptor genes pop up at genome-wide significance. That implies hormonal modulation might be an amplifier, not a cause.

---

3. How it could reframe priorities​

If funders were bold, they’d:

• Build mechanism-first trials for CA10 (pain), OLFM4 (NETosis), PRDX6 (oxidative/membrane repair).
• Start pre-screened subgroup trials rather than “all comers” ME/CFS — the genetics gives starting points.
• Integrate immune-viral sensing assays into baseline patient phenotyping — BTN family, ZNFX1 are not random.

Pathway-to-patient map (compressed view)

4. Pathway-to-patient map (compressed view)​

Pathway	Genes	Potential trial class	Low-risk now
Pain-synapse	CA10	NR2B NMDA antagonists, PSD95 disruptors	PEM-aware pacing, sensory load control
NETosis / neutrophil tone	OLFM4	Low-dose colchicine, DNase, PAD4 inhibitors	Oral hygiene, prompt infection care
Autophagy / ER-phagy	KLHL20, CCPG1	ULK1 modulators, ER stress chaperones	Regular meals, avoid extreme fasting
Oxidative / membrane repair	PRDX6, PEBP1	Lipid replacement therapy, antioxidants	Omega-3, choline, vitamin-C foods
Viral sensing	RABGAP1L, BTN2A2, ZNFX1	TLR modulators, γδ T-cell agonists	Infection avoidance, vaccination as eligible

 

[hotblack]

I’d skip the AI stuff for speculative answers like this. Probably safe to say we don’t know yet and read the great blog post

X marks the spot where ME/CFS biology can be discovered - DecodeME

06 August 2025 Scientists, people with ME/CFS, and their charities came together to create DecodeME, the world’s biggest ME/CFS study – and its results are striking. 18,000 people with ME/CFS gave their DNA, enabling DecodeME to reveal eightgenetic signals for the illness. These signals indicate...

www.decodeme.org.uk

Or listen to it if that’s easier for you

DecodeME-Initial-DNA-Results - Shared with pCloud

Keep all your files safe, access them on any device you own and share with just the right people. Create a free pCloud account!

u.pcloud.link

 

[forestglip]

So the study found significant locations in the DNA, not necessarily significant specific genes. The trouble is figuring out which gene associated with a given location is the troublemaker in ME/CFS.

For some of these locations, there are many possible genes. For example, for the chr1q25.1 location, they identied 11 possible genes that might be causing issues. That means 10 of these (maybe all 11) are probably false leads.

So maybe until scientists do more testing on these specific locations to narrow it down, I think it might make sense to focus literature research on the locations with only one associated gene to not waste time on potential dead ends. Unless there's good reason to believe a specific gene out of a group of many is the interesting one.

And those genes would be (links to dedicated threads):

Tier 1 Gene (The mutations are known to actually affect expression of the gene)
Chr17: CA10

Tier 2 Genes (No tier 1 genes, so just the closest gene to the mutation)
Chr6q: FBXL4
Chr13: OLFM4
Chr15: CCPG1

 

[JonBiz]

• Meal regularity over hero fasting. Autophagy is nuanced here; extreme fasting can worsen symptoms. Keep steady protein and electrolytes.

I've done three rounds now of the Fasting Mimicking Diet, in the hopes of improving things This would explain my poor results.

 

[EndME]

[Since this is relevant to both this thread and the media thread, I'm posting it in both threads.]

I’ve been trying to work out the strongest arguments against the criticism that participants were not diagnosed by a medical professional.

Here’s a recap of the criticism:


It’s true that DecodeME participants were not recruited through health professionals, but speculation that they did not have a formal diagnosis of ME/CFS is...purely speculative. I don’t think it is a strong response to say that cases were all diagnosed by a health professional, because we cannot verify that. We only know that they answered “Yes” to this question:

When people are objecting to self-report (ah, the irony), then countering with another self-report won't work.

I think the strongest response is something like:

• Good-quality GWAS studies need huge numbers of participants in order to have high enough statistical power to detect robust associations that would be missed in smaller studies.
• It would not have been possible to recruit 26K participants with ME/CFS indirectly through health professionals.
• By recruiting directly, DecodeME could reach the number of participants they needed, but then they needed to make sure that they were reaching people who really did have ME/CFS as currently defined. So they took extra steps that other studies in ME/CFS have not taken.
• DecodeME participants didn’t just report they had ME/CFS, they also reported that they had been diagnosed with it by a health professional, that they had post-exertional malaise and fulfilled either CCC or IOM criteria or both.
• Over 5000 potential participants (5281) were excluded from DecodeME because they did not meet these criteria:

A point should probably be added in there about whether GWAS studies in other diseases that have found robust associations have recruited directly or through health professionals. I don't have the knowledge of those studies to make the point, but others on here probably do. If studies with similar statistical power to DecodeME have managed to recruit through health professionals, then it would be worth explaining why this wasn't possible for ME/CFS in the UK.

One couldn't recruit via health professionals or confirmed diagnosis because there are no trained clinicians dedicated to ME/CFS and if you'd only have a few of them then you'd struggle with sufficient sample sizes and confounding factors will be introduced by how those clinicians handle things, where they are based, where they get their referrals from etc.

So isn't the argument simply: If you want large genetic studies of ME/CFS with rectruitment based on diagnosis, you have to have clinicans decidated to ME/CFS spread across the country? Instead of focusing on a possible shortcoming should the response not be: Yes, now is the time to give us a GWAS where the diagnosis is made by dedicated clinicans that know what they are doing! We'd all happily sign up for that, please get the wheels moving now.

 

[V.R.T.]

What verges on hype

Phrases suggesting immediate drug discovery momentum are optimistic for a first-wave GWAS with modest heritability. The logic is sound but timeline is long.

Well the AI is really raining on my parade today! Lets hope its confidently incorrect on that front (not that I expect a drug tomorrow, nice as that would be...)

It's weird though, I searched the preprint for drug, treatment and theraputic and only really found this phase claiming that the DecodeME results

they improve the likelihood of finding effective drugs for ME/CFS

Which is a modest and rational claim that doesn't seem anything like hype to me.

 

[Yann04]

The more I think about it the more I think getting full sequencing data to clear things up (and hopefully replication) would be very valuable.

I see decodeME like the funding bodies building the foundations for the house. Now they need to go and start building the main structure.

 

[Eleanor]

Well the AI is really raining on my parade today! Lets hope its confidently incorrect on that front (not that I expect a drug tomorrow, nice as that would be...)

It's weird though, I searched the preprint for drug, treatment and theraputic and only really found this phase claiming that the DecodeME results


Which is a modest and rational claim that doesn't seem anything like hype to me.

The LLM isn't actually detecting any real hype, it's just generating some text about 'hype' because similar text strings are often found in the sources that it's been trained to mimic.

 

[Jonathan Edwards]

Yes, now is the time to give us a GWAS where the diagnosis is made by dedicated clinicans that know what they are doing!

One could argue that since patients appear to have a much more rational and informed approach to their illness than the few physicians and psychiatrists who dabble in it, recruiting through patients has to be a better bet.

When I looked at the original proposal I had some reservations about self-referral and thought that there would be risks of bias unless a truly population-based trawl was done. But I was clear in my mind that recruiting from clinics would be the worst of all worlds.

 

[V.R.T.]

The LLM isn't actually detecting any real hype, it's just generating some text about 'hype' because similar text strings are often found in the sources that it's been trained to mimic.

Yeah LLMs are very good at confidently spouting bollocks.

 

[Jonathan Edwards]

NETosis / neutrophil tone

OLFM4

Low-dose colchicine,

Now who was it recently recommending colchicine for something?

This is pretty dumb. There isn't any NETosis. Policeman carrying handcuffs are not criminals.

 

[Jonathan Edwards]

Yeah LLMs are very good at confidently spouting bollocks.

Remarkably like humans in fact.

 

[EndME]

One could argue that since patients appear to have a much more rational and informed approach to their illness than the few physicians and psychiatrists who dabble in it, recruiting through patients has to be a better bet.

When I looked at the original proposal I had some reservations about self-referral and thought that there would be risks of bias unless a truly population-based trawl was done. But I was clear in my mind that recruiting from clinics would be the worst of all worlds.

I think Germany may be able to do better. They have a few clinics and biobanks that seem to be somewhat ok at handing out diagnosis (but I think the data might inevitably still include a large referal bias) and if they'd run a GWAS you could split the data into the cohorts of clinicans and patients that have reported a diagnosis and see how things turn out and compare that to DecodeME.

 

[Evergreen]

One couldn't recruit via health professionals or confirmed diagnosis because there are no trained clinicians dedicated to ME/CFS and if you'd only have a few of them then you'd struggle with sufficient sample sizes and confounding factors will be introduced by how those clinicians handle things, where they are based, where they get their referrals from etc.

So isn't the argument simply: If you want large genetic studies of ME/CFS with rectruitment based on diagnosis, you have to have clinicans decidated to ME/CFS spread across the country? Instead of focusing on a possible shortcoming should the response not be: Yes, now is the time to give us a GWAS where the diagnosis is made by dedicated clinicans that know what they are doing! We'd all happily sign up for that, please get the wheels moving now.

I don't think that would be enough. Even when there were 50 specialist centres around the UK (I don't know the situation now), they saw just 8000 patients a year (according to Collin & Crawley 2017 who cite Collin et al. 2012 for this), and not all were diagnosed with ME/CFS. Some sites weren't run by physicians. Not all patients would want to participate in a study. For that matter, there might have been considerable resistance from some physicians to the study. It would have taken forever to get to 26,000. Not feasible.

I think the argument is that the only way to do a study with adequate statistical power was to do exactly what DecodeME did.

Now that they've done it, we might see some multi-country replication attempts using patients who have already been diagnosed.

 

[Hoopoe]

My impression of ChatGPT is that it's good at being optimistic, seeing connections, and overinterpreting things. Likely because it's imitating hypothesis papers that are in this style. It's a bit like a Cort Johnson with a PhD in biology.

 

[V.R.T.]

My impression of ChatGPT is that it's good at being optimistic, seeing connections, and overinterpreting things.

Again, a lot like humans!

 

[Evergreen]

One could argue that since patients appear to have a much more rational and informed approach to their illness than the few physicians and psychiatrists who dabble in it, recruiting through patients has to be a better bet.

When I looked at the original proposal I had some reservations about self-referral and thought that there would be risks of bias unless a truly population-based trawl was done. But I was clear in my mind that recruiting from clinics would be the worst of all worlds.

Agree. It would be a heterogeneous fatigue mish-mash with no associations, or bogus associations due to overly broad criteria.

What do you think now? Do you think the proof is in the pudding?

 

[EndME]

I don't think that would be enough. Even when there were 50 specialist centres around the UK (I don't know the situation now), they saw just 8000 patients a year (according to Collin & Crawley 2017 who cite Collin et al. 2012 for this), and not all were diagnosed with ME/CFS. Some sites weren't run by physicians. Not all patients would want to participate in a study. For that matter, there might have been considerable resistance from some physicians to the study. It would have taken forever to get to 26,000. Not feasible.

I think the argument is that the only way to do a study with adequate statistical power was to do exactly what DecodeME did.

Now that they've done it, we might see some multi-country replication attempts using patients who have already been diagnosed.

Yes, the only way to get any meaningful data was probably doing something like what DecodeME did. So there's no point in coming off defensive.

But if they want to complain about that data being unreliable due to patients not seeing a clinican as part of the study I think there is now a pretty simple solution to that or as Carson/Wessely/Garner one said "Just do it better yourself": Set up a small well run GWAS study where people are invited from the DecodeME cohort to see a well trained board of clinicans (similar to what happened in the intramural study) and only look at the significant genes. You'd only need a couple of hundred patients and it shouldn't take long. Wouldn't that work? Practically, the problem will probably be: You won't find enough clinicans that aren't biased in some way or even want do this.

 

[Jenny TipsforME]

My impression of ChatGPT is that it's good at being optimistic, seeing connections, and overinterpreting things. Likely because it's imitating hypothesis papers that are in this style. It's a bit like a Cort Johnson with a PhD in biology.

Interpretation does need to be taken with a slight pinch of salt, but it’s also important to understand that “AI” isn’t a static entity. The model out this week has very low error rates on health bench marks, on the graph it’s essentially the difference between last week (white) on ChatGPT vs this week on ChatGPT (pinks). The middle graph is about health information accuracy/reliability.

AI on its own has lower diagnostic errors than human doctors now (interestingly lower than human doctor plus AI which should really be the best combination).

https://cdn.prod.website-files.com/63f416b32254e8679cd8af88/6894e49ca3a1be9d17beb575_reliability.png

Context is important though and in my prompt I had asked it what was hyperbole, so that it wasn’t unrealistically optimistic. There can definitely be a tendency to guess what the user wants and be sycophantic and here it might have over corrected.