When will ME be cured?
By drawing historical comparisons to other diseases, researching AI progress, and taking into account funding, research growth acceleration, and clinical trial reform, we've? built the world's first public prediction model for when ME will be cured.
Year by year
Read the story to see how this could play out through the upcoming years, then, after scrolling down and toggling the inputs in the model, come back and see how the story changes!
How our forecast works
To find the most accurate prediction, we worked backwards from a potential cure.
-
The cure
A registered, placebo-controlled trial shows that a treatment puts at least half of all people with ME into remission: they no longer meet ME criteria, have no significant PEM, and have at least 90% of their former function. Our cure date is the day the trial result is published, including as a preprint.
Exact rules
- Patients meet the most accepted ME criteria at the time: e.g. Canadian Consensus, International Consensus or the 2015 IOM criteria
- At least 20 patients get the treatment, and they are randomly assigned against a placebo. Significantly more patients reach remission on the treatment than on placebo
- Patients who drop out count as not in remission
- If the trial enrolls only patients picked by a biomarker, the remission rate counts for that share of ME. It is a cure for ME only if it adds up to half of all people with ME
-
...needs an effective trial
ME could possibly be several different subtypes, and a treatment usually works for only one at a time. The largest subtype sets a limit on how far any single drug can reach, so two things raise the probability that a trial reaches the 50% remission bar:
- a biomarker: a validated test, such as a blood test, that shows which patients a treatment is likely to work for, so the trial can enroll only those patients
- a combination: two or three drugs that each help a different group, tested together. This is how HIV was beaten in 1996 as well as cystic fibrosis in 2019
Trials also need funding, and even once secured, the trial trial takes time: recruiting patients, administering the treatment, measuring the results, analysis, and finally publishing. Expert-level AI (AI that can do the work of top researchers) can speed up recruiting, analysis and writing, so trials get shorter as AI improves.
-
...needs an effective treatment
Researchers must first find a candidate treatment: a new or repurposed drug that has a theoretical pathway to remission. We've programmed the model to output a chance of finding one that updates with new progress:
starting rate x amount of research (compared with 2025) x AI speed-upThe starting rate is the chance per year of finding a candidate treatment, at the 2025 amount of research and before any help from AI: we predict a 0.35% chance per year while ME has no biomarker, and a 1.1% chance per year once it has one.
-
...needs mechanistic research and patient data
The research that matters most toward a cure is mechanistic: it looks for the pathways and the root cause of ME, so that a drug can target them and reverse the illness. More and more of this work starts from large patient datasets. ChronicleBio and Amatica, for example, collect samples and data so that researchers and AI can find disease pathways, and then work with drug companies to target them.
The chance of finding a candidate treatment rises with the amount of human research, which has grown 18% per year over the last 3 years, and with AI acceleration, which grows modestly as AI improves but explodes once once expert-level AI arrives.
Model inputs
AResearch growth
Mechanistic research on ME is growing faster over time: while only 3% per year from 2010 to 2019, it has accelerated to 12% per year from 2019 to 2025, and 2026 is on track for a promising 28% year-over-year growth. Over the last 3 years, from 2023 to 2026, it grew 18% per year.
We model the growth in human research as depending on:
- the number of researchers working on ME
- how much experience they have
- the funding the field receives
- how fast approvals and publishing move (the less bureacracy, the faster)
- how many large patient datasets they have access to
How fast will human research grow each year?
How much extra money will ME research get each year, beyond today's growth trend?
We estimate the world spends roughly $35-50 million a year on ME research today. Germany's Decade Against PAIS program is most of our additional $25 million guess. Other governments or private funders would increase this number.
Our reasoning
- The 2026 research speed-up is a good sign. Mechanistic papers are on track to grow about 28% this year, well above the 12% a year from 2019 to 2025.
- The field is maturing. More researchers now have years of ME work behind them, so their research should get better, not just more plentiful.
- Paper counts may not be correlated with meaningful reseach. Some of the recent jump is likely AI-assisted writing.
- Cost per paper. Today's $35-50 million produces about 383 mechanistic papers a year, so one paper costs roughly $100,000 to $150,000. (A typical NIH grant, which is about $500K a year for 2 to 4 papers, roughly agrees with our math).
- Trial breakdown by type. We assume 70% of new funding goes to mechanistic research and 30% to trials. In the model, trial slots are rarely the bottleneck: the hard part is finding a treatment worth testing. So the trial share moves the date less than the research share.
- Money matters most if AI progress disappoints. An extra $100 million a year brings the cure about 3 months earlier along with our other predictions, and about 6 months earlier if AI can make research only 3x faster.
- Our guess is 18% a year. Single years jump around: papers fell in 2022 and 2024, then rose 39% in 2025. So we use the growth over the last 3 years, from 2023 to 2026, which is 18% a year. This smooths out single years such as 2026 but keeps the recent speed-up.
- This growth is not only money. Public funding for ME stayed flat from 2019 to 2025, but papers still grew. The growth came from Long COVID researchers joining the field, new patient datasets, and a more experienced field. New money is a separate input above, so we do not count it twice.
- Our funding guess is +$25 million a year. The paper trend does not yet show Germany's new program, which started in 2026, or new private funders. We count them as about $25 million a year of extra money for ME.
BHistorical disease comparisons
How hard is ME to cure? To estimate this, we looked at 23 other diseases and counted the research papers that were published for each one before its first cure.
Historically, every disease that was cured had a biomarker first. So until ME has a biomarker, we compare it with the 8 diseases that have none. Once ME has a biomarker, we compare it with all 15 diseases that have one, cured or not.
Is ME harder or easier to cure than the diseases it is compared with?
"Same" means as hard as the average of its group: the 8 uncured diseases with no biomarker, then all 15 with one.
Our reasoning
- We think ME sits in the middle. It is probably harder to cure than HIV, hepatitis C or peptic ulcers, which each have one known cause. It is probably easier than Alzheimer's, Parkinson's or ALS, which kill nerve cells that do not grow back. ME does not, and some people recover. ME looks most like immune diseases such as MS, rheumatoid arthritis and Crohn's: complex, with more than one cause.
- If ME is an immune disease, it is easier than our guess. The 7 immune diseases in the table are cured about 5 times faster per paper than the average of all 15. CAR-T treatments are putting lupus patients into drug-free remission now, and an anti-CD19 trial in ME reports in 2029. We keep the average of all 15, because ME autoimmunity is unproven and the rituximab trial failed. Choosing "easier" moves the cure about a year earlier.
- Where the rates come from. The 8 diseases with no biomarker have 0 cures in 108,885 papers. The 15 with one have 10 cures in 772,008 papers. Scaled to the 772 ME papers published in 2025, that gives a 0.35% chance per year of finding a candidate treatment without a biomarker, and 1.1% with one. We count zero cures as half a cure, a standard estimate. More research and AI multiply both rates.
CWhen will expert-level AI arrive?
Expert-level AI means AI that is as good as the best human experts at almost all thinking work, including biology research. Lab work, patients and trials still take time, so research does not speed up all at once (see the next section). We base our best guess on two forecasts: the team behind AI 2027, and prediction markets.
Our reasoning
- Two independent groups point to about the same date: a team that forecasts AI full time with an excellent track record and prediction markets where many people bet real stakes.
- By their own grading, real AI progress has run at about 70-90% of the pace the AI 2027 team forecast.
- AI is strongest today at coding and math. But the AI 2027 forecasts are for AI that matches top experts in every field, biology included. The hands-on part of biology (lab work, patient samples, trials) stays slow, and sections D and F model that.
- Our best guess is the middle of the two forecasts, early 2030.
DHow much can AI speed up medical research?
While the previous section gives the date when AI can think as well as a top ME researcher, this section asks how much that actually speeds of research. It is one of the biggest inputs in our model, because AI acceleration multiplies the chance of finding a candidate treatment every year.
While AI can make the non-physical steps of research almost instant, picking areas of researching and desiging studies, other steps must happen in real-time: cells grow at a set pace, and patients must be monitored for a set amount of time when taking a drug.
One research cycle
| Step | Today | Expert-level AI |
|---|---|---|
| Read the research, pick an idea | months | hours |
| Design the study | months | days |
| Funding and ethics approval | 6-18 months | faster, but people still decide |
| Lab work and patient samples | months to years | faster with robot labs, but cells grow at the same speed |
| Analyze the data | months | hours |
| Publish | 6-12 months | weeks |
| Test in patients | years | still years: patients take the drug in real time |
The rate at which research accelerates depends on how much of the physical work AI can replace and still run a believable trial:
- Up to 3x faster: AI is much more skilled at reading literature, picking ideas, designing studies, analyzing data and writing papers, but lab work, approvals and patient studies still run at today's speed. Thinking is about two thirds of a research cycle, so a study that takes 3 years today would take about 1.
- Up to 10x faster: AI also runs robotic labs where robots test stored blood and tissue from biobanks for thousands of markers at once. They screen thousands of drugs on patient cells, or on healthy cells exposed to patient blood, day and night. AI also picks better ideas that fail less often. Research that takes 10 years today takes about 1.
- Up to 100x faster: AI simulates cells and the body well enough to test most ideas "in silico". Only the best ideas go to the lab, and AI is so effective that the drugs it picks to test in patients usually work the first time. Research that takes 10 years today takes about a month.
Current situation
Note: the AI-designed drugs tested so far were designed before expert-level AI existed, so results may look better yet. 15 more efficacy results for AI-designed drugs are due by the end of 2028.
How much faster can AI make medical research, at most?
Our reasoning
- AI is already accelerating ME research. The AI Futures team estimated that AI made coding about 2x faster by August 2026. Medical research has more physical steps, so we start at 1.25x today and let it grow smoothly to 3x by the time expert-level AI arrives.
- ME does not yet have a lab model. No mouse or cell model shows post-exertional malaise, so robot labs can only work with samples and cells from patients. This limits how far lab work can speed up.
- Our guess is AI speeds research up to 10x. 100x would need AI to predict ME biology well enough to skip most lab work. With no lab model, and with trials that run in real time, we think that is unlikely before a cure arrives.
- AI matters so much because it multiplies. In 2035, with a biomarker and AI at 10x, the model expects about one candidate treatment every 2 to 3 years. With no AI, it expects one about every 24 years.
EWhen will a biomarker for ME be validated?
A biomarker is a lab test that separates ME from healthy controls and other diseases, and can be used in clinical trials to pre-screen patients. 29 candidate biomarkers have been proposed since 2005, but none have been independently validated.
To validate, an independent group must repeat the result, blinded, in 200 or more people, and tell ME apart from other illnesses correctly most of the time (a statistical score, called an AUC, of at least 0.80).
The leading candidates today:
- EpiSwitch: a blood test that correctly identified 92% of patients and 98% of healthy people in its first study. Not yet tested against other illnesses, and the same DNA pattern also appears in Long COVID, PTSD, rheumatoid arthritis and MS
- Nanoneedle (Stanford): measures how white blood cells react to stress. In 2019, it separated all 20 patients from 20 healthy people. A UK-funded study is working to repeat it.
- Raman test (Oxford): a laser reads single blood cells. In 98 people, it told ME apart from healthy people and other illnesses with 91% accuracy. It still needs a larger study.
- Amatica: an RNA test that patients can order, which also builds a large research database. Not yet published, but preliminary results show a promising AUC.
Why a cure needs a biomarker
Of the 23 diseases above, every one that got a cure had a biomarker first. The 8 with no biomarker still have no cure after 108,885 research papers. There are four reasons why having a biomarker would speed this up:
- The right patients. ME probably consists of several subtypes. Without a biomarker, a trial mixes patients a drug can help with patients it cannot, and what could be a cure for one group comes out looking like a weak result. Cystic fibrosis is a perfect example: the first drug that fixes the root cause helps only 4% of patients, but it only succeeded because its trials enrolled exactly that group of patients.
- The right diagnosis. ME is diagnosed by symptoms only, so some trial patients have a different illness and dilute the result.
- A target. A biomarker that holds up usually points to the biology behind the disease. That tells researchers what pathways a drug must hit.
- Faster answers. A lab measure can show whether a drug changes biology long before symptoms change. Failed drugs can be dropped sooner.
Our reasoning
- At the past rate, about biomarker 1.5 candidates a year and none validated, the chance of a validated biomarker by the end of 2028 would be about 5%.
- But things have changed. Studies are far larger now: DecodeME has 15,579 patients, where most past studies had 20 to 60. EpiSwitch reported strong early accuracy, and large new patient datasets, such as those of Amatica and ChronicleBio, give more material to find and test candidates.
- But validation takes time. A proper study takes 1 to 2 years, so only candidates already in progress can report by 2028. EpiSwitch has not yet been tested against other illnesses, and the same DNA pattern also appears in 4 other diseases.
- So we put the chance at 25% by the end of 2028, 5x the past rate, and 55% by the end of 2032, as more large studies report. That puts the median in early 2032.
FClinical trials
Even with the right drug, a trial takes time. Each trial has to:
- get ethics and regulatory approval
- find and enroll enough patients who meet the entry criteria
- administer the treatment, often for several months
- follow the patients long enough to measure remission
- analyze the results and publish them, usually after peer review
Today, a completed ME drug trial takes a median of 2.0 years from its first patient until the end of patient monitoring, and its results become public about 0.8 years later as a preprint or first report, so 2.8 years in total. Everything before the first patient, such as funding, ethics approval and opening trial sites, only adds to the total duration.
Before the trial starts
A drug that is already approved for another illness can be used in an ME trial once the money and the ethics approval are in place, which takes about 6 months. A brand new compound cannot, as it must meet:
- Safety tests in animals and cells, to find the dose that harms the liver, heart and other organs. Animals take the drug for set periods, often 28 or 90 days
- Manufacturing quality to a certified standard
- Permission to test in humans: the company files with the FDA, which has 30 days to object
- Phase 1: a first small trial in healthy volunteers, for safety and dose, about a year
Published ranges put candidate selection to an approved application at 18 to 36 months, and phase 1 adds about a year, so we use 3.5 years as our estimate for each new drug.
Expert-level AI writes the paperwork and can replace some animal tests, but the dosing periods and phase 1 stay, so we cut it to 2 years. We assume half of candidate treatments are new compounds and half are repurposed drugs.
What AI speeds up
| Trial step | Today | Expert-level AI |
|---|---|---|
| Find and enroll patients | often twice as long as planned | weeks: AI finds eligible patients in health records and patient registries |
| Treatment and follow-up | months | no change: patients respond at their own pace |
| Analyze and publish | about a year | weeks, as a preprint |
So trials fall from 2.8 years today to about 1.5 years with expert-level AI.
Trial reform
Trials are slow and expensive partly because of rules and paperwork. Opening a trial site takes more than 8 months on average, and most trials take about twice as long as planned to enroll their patients.
Proponents of clinical trial reform want to make trials faster and cheaper everywhere, and some changes have already started. This matters for ME because cheaper trials mean more drugs get tested, including repurposed drugs that companies will not risk today, and every failed trial teaches researchers something about the illness. Here are three different trial reform scenarios we see as possible:
| Drug trials a year | Trial length today | Trial length with expert-level AI | |
|---|---|---|---|
| No reform | about 1.5 | 2.8 years | 1.5 years |
| Some reformOur guess | 2x as many | 2.4 years | 1.3 years |
| Major reform | 4x as many | 2 years | 1.1 years |
- No reform: AI races ahead, but human bureaucracy refuses to budge: long consent forms, months to open each trial site, and every data point checked by hand.
- Some reform: shorter consent forms and less paperwork, checks focused on the data that matters, published FDA rejection letters, and more early trials in faster countries such as Australia.
- Major reform: all of that, plus shared platform trials for ME that test several drugs against one placebo group, low-cost trials like the UK's RECOVERY trial, and conditional approval after strong phase 2 results, which draws in drug companies.
Reform also cuts the wait before a trial starts. Most of a repurposed drug's wait is funding, ethics and opening a site, which is exactly what reform targets, so some reform cuts it by 40% and major reform by 60%. A new compound gains less, 15% and 30%, because animal dosing periods cannot be reformed away.
How much will clinical trial rules and practice improve?
Our reasoning
- How we arrived at 2.8 years: Across 23 completed ME drug trials on ClinicalTrials.gov, the median trial ran for 2.0 years. The journal paper came a median 1.3 years after that. Preprints are more common these days, so we assume an additional 0.8 years.
- Expert-level AI cuts trials to 1.5 years. Recruiting, analysis and writing can shrink to weeks, but patients still need months on the treatment and time for follow-up.
- Clinical trial reform has already started. The FDA has cut some trial paperwork and now accepts lighter monitoring. More than two thirds of recent FDA approvals included data from trials in Australia, where an early trial can start 2 to 3 months after its protocol is final.
- But reform is slow. Reform proposals such as Clinical Trial Abundance have wide support, yet their authors say the policy change so far is small. So our guess is some reform, not major reform.
- Each extra trial from reform has a low chance of success. Most extra trials test drugs that do not work. About 43 ME drug trials have been registered, and none has put patients into remission. So we give each extra trial a 0.5% chance of finding a candidate treatment.
When a trial first shows a cure
What moves the answer most
Each bar shows how far the median moves when one factor goes from one end of its range to the other. Everything else stays at your settings.
Limitations
- The forecast is for the first trial result, not for when patients can get the treatment. Approval, price and access can add years, and differ by country.
- The biomarker candidates today tell ME apart from other illnesses. None yet predicts who will respond to a treatment, and the model assumes a validated biomarker will do both.
- Trials often leave out people with severe ME, so a cure may be shown first in milder patients.
- Some factors move together in real life but separately in the model. For example, fast AI progress could also bring more research funding.
- The model counts research papers, not their quality. Modern tools and much larger patient datasets, such as DecodeME's 15,579 patients and the databases Amatica and ChronicleBio are building, may make each paper count for more than it did in the past.
- Treatments that work in other diseases can arrive in ME from outside our research numbers. CAR-T and similar immune-reset treatments are putting lupus patients into drug-free remission now, and an anti-CD19 trial in ME is due to report in 2029. We give these a 5% chance of reaching the cure bar for all of ME and 15% for a subgroup, but they are one shared bet: if ME is not antibody-driven, they all fail together.
- The ceiling on how much of ME one drug can reach is a guess. We have no data on how large the biggest subtype is.
- The model does not take into account potential AI apocalypse.
Sources
- Research papers: OpenAlex
- Trials: ClinicalTrials.gov and ISRCTN
- AI forecasts: AI Futures Project, mid-2026, Manifold Markets
- AI-designed drugs: IntuitionLabs 2026 review
- EpiSwitch: 2025 study, 2026 study
- DecodeME: 2025 preprint
- UK Biobank blood tests: 2025 study
- Biomarker history: MEpedia
Download the model
Want to play around with our model? We've open-sourced it, so you can download our ready-made prompt and all the code, and explore it with AI. Afterwards, please give us feedback on what you find so we can keep updating our model!
Who we are
We are a group of patient forecasters with experience in trading, critical thinking, and prediction markets.
Found a mistake? Think our model is completely wrong and want to tell us? All of it is welcome: hello@lcmedata.org
LC/ME Data newsletter
Get our new patient surveys, data and forecasts on Long COVID and ME by email:
Changelog
October 19, 2026
- Our prediction is released to the public!