You may recall our clinical trials and experimental studies based on my concept of Neuroarcheology, which posited that pathological events in utero result in misplaced and poorly connected neurons that generate immature neural activity—which is ultimately the true direct cause of autism. These must therefore be blocked. Using bumetanide—an agent that blocks this hyperactivity and restores weakened inhibition—we demonstrated its efficacy first in animals and then in humans. We successfully conducted two Phase 2 clinical trials, and these results were validated by five other teams that exactly replicated our approach (doses, conditions, etc.). As a result, according to a meta-analysis (Xiao et al.), more than 1,050 children worldwide are currently being treated or have been treated with this drug. We then conducted a final Phase 3 trial involving more than 420 children divided into two age groups—2–6 years and 7–18 years—across more than 40 centers in Europe, Brazil, and Australia. The result was negative—no significant differences were found between children who received the treatment and those who received the placebo, as was the case in all Phase 3 trials conducted worldwide (Fuentes et al). This was expected because autism spectrum disorders are, by their very nature, far too heterogeneous; this precludes a single product from being suitable for everyone.
Consequently, we reanalyzed all Phase 3 data using a machine learning approach to break down the population of children included in the trial into subpopulations of responders based on certain clinical parameters. It is important to remember that assessment scales are averages of multiple values: the DSM-5 (5 values), the CARS (15 values), and especially the SRS, which is completed by parents—65 values. It is clear that a scale averaging so many values may only be partially valid; however, a treatment that alleviates, say, agitation and social difficulties—but not other attributes of autism—is still worth considering to reduce the challenges faced by the child and their parents. This study was published in Translational Psychiatry. We were able to identify, within the failed Phase 3 trial, 30 to 40 percent of the children who responded to the treatment. In other words, Phase 3 failed because the inclusion criteria were too broad, not because of the treatment itself. A new trial including only children who meet these inclusion criteria should be successful. In fact, pharmaceutical companies are now seeking to identify subpopulations of children with autism (or other syndromes) in order to conduct large-scale trials that will yield positive results.
More generally, this work shows that it is absolutely necessary to “break down” ASDs into distinct subgroups in order to create less heterogeneous populations. Otherwise, there will be no innovative treatments for ASDs because there will be no successful Phase 3 trials. Including such diverse cases—from Elon Musk to a teenager who does not speak—within the ASD category is absurd, and we are beginning to realize this. The problem is that to make this change, we must reevaluate our understanding of what autism is and, above all, establish criteria for classifying it into different forms—which is not easy, as it requires new measurements across large populations of people with autism. The use of techniques other than behavioral ones will prove crucial. Multidisciplinary measurements, including genetics but also—and above all—biological measurements such as EEG, imaging, and blood tests measuring stress markers and immune factors, the importance of which has been demonstrated by recent studies. Perhaps our findings with bumetanide also point to subsets of children with specific characteristics. It will likely also be necessary to combine clinical criteria—such as agitation and lack of sociability—to refine the diagnosis. In summary, this is valuable research that should lead to a better understanding of this multifaceted syndrome that originates in utero. The addition of data similar to what we are characterizing in our Pelargos project focused on maternity care will also be useful.
In any case, it would be absurd to abandon an approach that is effective and inexpensive for the health insurance system (which is precisely one of the reasons why pharmaceutical companies are not interested). Our AI-based approach is clearly promising. We will try to secure funding to carry out this approach and conduct a validation trial. It must be said that pharmaceutical companies prefer to fund “modern” approaches using new methods and techniques (genetics, immunological identification, etc.), despite their repeated attempts having yielded no results.
A victory without risk is a triumph without glory….
Source : Rabiei et al Treating autism with Bumetanide: Identification of Responders Using Q-Finder Machine Learning Algorithm, Translational Psychiatry 2025
Photo Maxim Berg– Unsplash
