Up to 40% of children in the “failed” phase 3 bumetanide trial were actually responders, according to AI reanalysis of the data – Treating autism in the real world

 

In some parts of the world even the words “treating autism” can still get you into trouble and some people have to go to quite extreme lengths to get their child’s developmental trajectory back on track.

I did note that in the US big changes have been made to their Interagency Autism Coordinating Committee (IACC) that coordinates all efforts within the Department of Health and Human Services (HHS) concerning autism. Now it includes some readers of this blog. Will this make a difference?

https://iacc.hhs.gov/ 

 

Over in France, the Bumetanide researchers Ben-Ari, Lemonnier and pals published their AI driven reanalysis of the “failed” phase 3 autism trial. They found that using AI they could actually predict who did actually respond; and many did. Nonetheless this large trial of all-jumbled-together kids with an autism diagnosis showed that overall bumetanide was no better than a placebo. Sounds strange to you? This is a common theme in autism trials because they do not narrow down a specific type of autism that they are trying to treat.

Over where I am, I keep getting positive reports of success. Some people are lucky and find that much of what works for my son works for theirs. There is a lot in this blog about other types of autism.

Why autism remains untreatable?

Autism is not simple to treat. Autism has no biological definition and measurement scales are all likely not fit for purpose. What would treatment success even mean?

From the perspective of severe autism with apparent ID (the old “Classic autism”) the biggest issues are to do with the slow rate of acquiring new skills. There are very well established tools to measure the skillset of such kids, such as  ABLLS (Assessment of Basic Language and Learning Skills). There are also non-verbal IQ tests. 

For young kids with classic autism you want them to add these basic skills ASAP, so that they can move on with their lives. In our case Bumetanide was the key to unlock new skill addition.

This is not what the phase 3 bumetanide trial was trying to measure.

Indeed one of the recurring comments from parents and teachers is the child has become more “present.” How do you quantify something like that?

For most children with autism in 2026, they do not have a problem with skill acquisition, they are a bit quirky, nervous, resistant to change, stim a bit, do not make friends. It is a very different condition. These issues are very real and genuinely concern some parents, but they are very different problems.

The modern cookie-cutter, protocol-driven, approach does work for most of medicine. But it will never work on an ill-defined category like autism. It actually becomes ridiculous when you look at all the varied types of autism. Even people with cerebral palsy or Down syndrome can be given an “autism” diagnosis on top, but they are completely different biological conditions.

Where to from here?

What does Ben-Ari do now?

Start again with another phase 3 trial? Paid for by who?  Will Servier come back and fund the second attempt?

In the meantime the clock keeps ticking.

I read Ben-Ari’s initial study and made my n=1 trial in 2012. My trial met its primary endpoint (Peter satisfied) and therapy started.

Academic performance went from complete basket-case to passing his high school public IGCSE exams a decade later.

Now it is 2026 and therapy still continues. No side effects,  heart ultrasound (echocardiogram) all normal.

Crazy world.

40% “disabled” at Stanford

I was surprised to read that almost 40% of undergraduates at Stanford University are claiming disability, to get extra time in exams. It does tell you a lot about the current generation of 20 year olds.

I would give them an E on their final diploma (I passed but needed Extra time). It is perfectly reasonable for a small number of clever students to need extra time, they might have a physical disability with their hands, be deaf, or blind, or dyslexic. It is perfectly reasonable to give some people extra time, but 40%?

It really is not fair on the remaining 60%. Maybe just give everyone an extra hour, those that finish early just leave early. They could get E on their results, for “I work fast and finish Early - hire me!"

What is annoying is the trivialization of the word disability.

40% of Stanford undergrads receive disability accommodations—but it’s become a college-wide phenomenon as Gen Z try to succeed in the current climate

So many people claim a disability like autism that theme parks in the US and Europe have had to roll back their privileged access schemes.

When I visited Charlotte International airport a while back and had to stand in a very long line for the passport control, I was amazed to see a never-ending procession of people appearing in wheelchairs to skip the queue. I have never seen this in Europe, but I suppose it will eventually come.

 

Back to those 40% in the Bumetanide trial.

New Analysis of the Bumetanide Phase 3 Trials: Were Responders Hidden in a “Failed” Study?

Approximately one-quarter to one-third of participants fit validated clinical profiles in which bumetanide showed statistically significant benefit on SRS-2, despite the overall trial being negative. The abstract itself says up to 40%.

Treating autism with Bumetanide: Identification of responders using Q-Finder machine learning algorithm

Bumetanide, a specific NKCC1 co-transporter inhibitor, restores deficient GABAergic inhibition implicated in various brain disorders, including Autism Spectrum Disorders (ASD). In keeping with this mechanism, nine successful phase 2 clinical trials, conducted by seven independent teams using an identical protocol, have shown significant improvements in ASD symptoms among individuals treated with Bumetanide. Despite these promising results, two large phase 3 clinical trials (over 400 children recruited in approximately 50 centers and covering age groups 2–6 and 7–17 years) failed with no significant difference between patients treated by placebo or Bumetanide. This failure may stem from the substantial heterogeneity of ASD symptom profiles across the study population, potentially diluting the overall observed treatment effect. To address this, we reanalyzed the phase 3 data using Q-Finder, a supervised machine learning algorithm, aiming to identify subgroups of patients who responded to the treatment. This analysis was based on clinical parameters collected at the baseline of trial and used the same standard endpoints and success criteria defined in the original phase 3 protocol. It enabled the identification of responder subgroups showing a statistically significant difference between placebo and Bumetanide treatment arms. We report detailed descriptions and statistical evaluations of these subgroups. The discovered responder subgroups, representing up to 40% of participants, were cross validated between the two study populations. These findings suggest that meaningful treatment responses can be uncovered within negative phase 3 trials, highlighting the limitations of a one-size-fits-all approach for heterogeneous conditions such as ASD. Machine learning appears to be a promising tool to support this precision medicine strategy.

The 2026 reanalysis published in Translational Psychiatry revisited the large Phase 3 bumetanide trials that previously failed to meet their primary endpoint.

The original Phase 3 trials included more than 400 children (ages 2–17) and found no significant overall difference between bumetanide and placebo on the primary outcome measure (CARS2).

This new study asked a different question:

Instead of “Did bumetanide work for everyone?”, could it have worked for specific subgroups that were diluted in the overall average?

To explore this, the authors used a supervised machine-learning algorithm (Q-Finder) to identify baseline clinical profiles associated with treatment response.

What They Found

The original overall result remains negative

Across the entire population:

• No significant benefit on the primary endpoint (CARS2).
• No meaningful average effect.

So the trial still officially failed.

Subgroups showing benefit were identified

When the data were stratified by symptom profiles at baseline, several subgroups showed:

• Statistically significant improvement on the SRS-2 (Social Responsiveness Scale)
• Treatment effects of roughly 12–17 points in validated groups
• Coverage of about 25–36% of participants in the largest responder profiles

Importantly, these findings were cross-validated between the younger and older trial cohorts.

A Consistent Feature of Responders

Across validated subgroups, one feature repeatedly appeared:

Mildly abnormal “adaptation to environmental changes” on CARS2

This domain reflects:

• Difficulty with transitions
• Rigidity around routines
• Stress with change

Responders were typically:

• Clearly autistic (often moderate–severe social symptoms)
• With repetitive behaviours
• But not globally or profoundly impaired across all domains

Interestingly, IQ did not emerge as a defining predictor of response.

Primary Endpoint vs Secondary Endpoint

A key nuance:

• No validated responder subgroups were found using the primary endpoint (CARS2).
• Validated subgroups were found using the secondary endpoint (SRS-2).

From a regulatory standpoint, this matters: trials are judged on their primary endpoint.

From a scientific standpoint, it suggests:

SRS-2 may have been more sensitive to the type of change bumetanide produces.

What This Means

This reanalysis does not prove bumetanide works broadly in autism.

It does suggest:

• Autism is highly heterogeneous.
• A one-size-fits-all trial design may dilute effects.
• A biologically or symptom-stratified approach may be necessary.
• Around one-quarter to one-third of participants may represent a responder subtype.

However, these findings are post hoc and exploratory.

To confirm them, a new trial would need to:

• Prospectively enroll only the identified responder phenotype.
• Use appropriate primary endpoints.
• Replicate the treatment effect.

Why This Matters for Autism Research

The study reflects a broader shift toward precision medicine:

• Rather than asking “Does this drug work for autism?”
• The better question may be:

“Which subtype of autism does it work for?”

Machine learning may help identify these subgroups, but prospective validation is essential.

The original Phase 3 trial remains negative at the population level.

This reanalysis suggests that meaningful responses may have been present in specific clinical subgroups — particularly children with:

• Mild adaptation abnormalities
• Repetitive behaviours
• Significant social impairment

Whether this represents a reproducible biological subtype remains to be tested in future trials.

Conclusion

In Rett syndrome a very expensive new drug called Trofinitide was approved, even though reports suggest it is only really effective in about 20% of these girls. I was really surprised.  It costs $300,000 to $900,00 a year depending on the girl’s weight.

It looks very odd that the large bumetanide failed, even though 25-40% were actually responders. By the way, my son’s bumetanide therapy has cost about $80 a year, for the last 13 years.

It does not fill you with great confidence.

I recently saw an article saying that “paracetamol/ acetaminophen does not, after all, increase the incidence of autism.” Well theoretically it should be harmful, by depleting glutathione, which is why it should be taken with NAC. We also know that NAC taken during pregnancy can significantly reduce the risk of miscarriage and this has been studied in a clinical trial.

N-acetyl cysteine for treatment of recurrent unexplained pregnancy loss

A controlled clinical trial studied N-acetylcysteine (NAC) in 168 pregnant women with a history of recurrent unexplained miscarriage. Women received either folic acid alone or folic acid plus NAC at 600 mg per day. In the NAC group, 52% of pregnancies continued beyond 20 weeks, compared with 27% in the control group. The take-home baby rate was 47% in the NAC group, compared with 21% in the control group. This represents more than a doubling of the live birth rate. NAC works by restoring glutathione, the cell’s main antioxidant, protecting placental and fetal tissue from oxidative stress. Oxidative stress is known to impair placental function and contribute to pregnancy loss. NAC was well tolerated, with no significant safety concerns reported. These results suggest that correcting oxidative stress can directly improve pregnancy outcomes in a defined high-risk group. This study illustrates how targeting a specific biological mechanism can dramatically change developmental outcomes.

If a professionally-managed autism trial cannot detect the 25-40% who responded to some extent, do you believe a study that effectively says nobody gets autism from pre-natal acetaminophen. Not even 1%? All you likely need to do is pair it with NAC to make the risk 0%.

For decades doctors refused to believe regressive autism existed. Once people started videoing their toddlers, it became impossible to doubt that some actually had developed speech and then lost it. Parents were not imagining it. It was just an inconvenient truth, and still is.