In March 2025, a six-year-old girl in Shanghai became the first person to receive a gene-editing therapy aimed at the brain. Her parents had raised $860,000 to pay for its development. She died seven days after the infusion. The hospital's own ethics board found the death was definitely related to the treatment and named the cause: thrombotic microangiopathy.

None of that was disclosed. The team published its animal work in Nature eleven months later without mentioning the trial, the family, or the money. State media called the paper a first ray of hope. A neuroscientist wrote an admiring commentary. Other families in the same parent support group started making calls.

The story surfaced in July 2026, when Science and Retraction Watch published a joint investigation built on documents and recordings the parents had kept. What follows is not a summary of that reporting. It is what the primary record shows when you go read it yourself, including the parts a nephrologist notices first.

The thing that makes this story hard to hold in your head is that it is not one failure with complications. It is one failure repeated at five different scales. The consent form named the complication but never named death. The ethics committee approved the trial before the safety report existed, and never went back for it. The paper omitted the funding. The paper omitted the child. The registry sat untouched for sixteen months. Same shape every time, and the shape is a disclosure that somebody decided not to make.

Thrombotic microangiopathy is not an exotic complication of AAV gene therapy. It is a listed one. Complement gets activated, endothelium gets injured, platelets get consumed in microvascular thrombi, and the kidney, being an organ that is mostly capillary bed with ambitions, fails first. It is aHUS by another route.

The US prescribing information for onasemnogene abeparvovec, the most widely used AAV9 product in the world, describes TMA occurring generally within two weeks of infusion, characterized by thrombocytopenia, microangiopathic hemolytic anemia, and acute kidney injury. It instructs the clinician to obtain a baseline creatinine and complete blood count, monitor platelets weekly for the first month, and consult a pediatric nephrologist immediately. There is at least one published fatality in the literature, and case series across Duchenne micro-dystrophin, Danon disease, and SMA programs going back years.

So the syndrome was known, the timing was known, the organ was known, and the monitoring was specified somewhere in the field's published memory. She presented on schedule. Fever, then anuria, then thrombocytopenia, then death, inside the window the label predicts.

One high-dose monkey also had kidney injury, which the investigators could not characterize because no data were collected during the first month after infusion, the exact window when TMA appears. That is not a subtle gap. That is the observation period being switched off during the interval where the thing you are looking for happens.

Prophylaxis is worth a line of its own. She received prednisone alone. High-dose AAV programs have been steadily adding heavier immunosuppression, rapamycin and rituximab among them, precisely because steroids alone are thin cover against a complement-mediated event. She screened negative for anti-AAV neutralizing antibodies before dosing, which is reassuring for some mechanisms of TMA and irrelevant to others.

And then the entry criteria. The protocol required normal liver, heart, and immune function, and normal coagulation and platelet counts. A trial's inclusion criteria are a risk assessment wearing a disguise. This one screened carefully for exactly the organ systems that would fail.

The underlying research is not junk science, and the argument gets weaker if you pretend otherwise. Here is what was actually established before the girl was treated, without the vocabulary.

The mouse work was real. The team bred mice carrying the girl's exact mutation, gave them the treatment, and the mutation was corrected. The mice got better at the tasks they had been failing.

A mouse is not a child, so the next step was monkeys. There were two of them, one given a high dose and one a low dose. And because the monkey version of this gene differs from the human version, the treatment could not actually correct anything in them. All the monkeys could show was that the virus reached the brain.

The mice and the monkeys also received the treatment two different ways. The mice got an injection into a vein, carried by a virus that slips into the brain in rodents but would not do so in a person. The monkeys got an injection into the spinal fluid, carried by a different virus, the one the girl would eventually receive.

So one experiment showed that the edit works. A separate experiment showed that a delivery method reaches the brain. No single experiment showed both things at once, in the same animal, with the same virus. The girl was where those two half-answers were put together for the first time.

The same story with the vocabulary restored, for readers who want the specifics. A humanized knock-in model of the R1025W variant reproduces the phenotype, and a split adenine base editor delivered by dual AAV-PHP.eB corrects roughly ten to fifteen percent of alleles in bulk brain tissue with modest bystander editing at A13. That is a legitimate result.

The primate efficacy data rest on one high-dose animal, one low-dose animal, and two untreated controls. The guide RNA was designed against human CHD3 and carries four mismatches at the orthologous macaque site, so the on-target edit could not be assayed at all. What the authors detected in monkey brain was bystander conversion at neighboring positions, plus measurable editing at a predicted off-target locus with two mismatches. The paper is careful about this. The reader in a hurry is not.

The delivery switch matters as much as the sample size. Mice received AAV-PHP.eB intravenously, a capsid engineered to cross the rodent blood-brain barrier and one that does not translate to primates because of receptor differences. Monkeys received AAV9 intrathecally. The efficacy evidence and the delivery evidence live in different experiments with different vectors by different routes.

The paper reports transduction and editor reconstitution in primates. It reports nothing about primate liver or kidney toxicity. Reviewers asked for controls on the key imaging figure and got them, though an image-integrity specialist noted the controls appear to have been processed at a different time with different microscope settings. The paper closes with the observation that bridging the gap between preclinical research and clinical translation remains a significant challenge. That sentence was written by people who knew what had happened in the gap.

Jesse Gelsinger died in 1999 in a gene therapy trial at Penn. The principal investigator was barred from human research for five years, in part for failing to disclose adverse reactions from primate toxicology studies on the consent form. The university paid more than half a million dollars. The field stopped and examined itself for the better part of a decade.

That investigator is James Wilson. He is quoted in this reporting saying the primate liver data alone should have triggered lower-dose studies to find a maximum tolerated dose. He is the person best positioned in the world to recognize the pattern, and he is watching it recur, and the total institutional response was a $3,600 fine and a conversation.

The regulatory mechanism is worth naming plainly, because it is not unique to China. This ran as an investigator-initiated trial at a major hospital, a track that in China's dual-system framework does not require national drug regulator review. The oversight was local. The local committee approved the study before the safety report existed. Every country has some version of this door, and the argument for keeping it open is a good one. It is how bespoke therapies for children with ultra-rare mutations get built at all. The argument is only good if someone checks what walks through.

What nobody should take from this is that the institutions failed to notice. They noticed. The hospital's ethics board did a causality assessment two days after she died. The district health commission investigated and levied a fine. The responsible physician received verbal counseling. The university reviewed the parents' complaint in April 2026 and sent a written response declining to act, explaining that the payments were research service fees and that the Nature paper bore no relation to the experiments the family funded. Every body with authority here opened a file, processed it, and closed it. The machinery ran. It just did not produce disclosure, because nothing in it was pointed that way.

Which is why the correction came from outside. Retraction Watch is not a regulator or an oversight body, whatever its name suggests. It is a journalism nonprofit, and this was a reported story funded by an investigative journalism fund, built on recordings the parents kept because the science was over their heads. The audit function that worked was two reporters and a grieving family with a phone.

Her syndrome was not fatal. Most people with a CHD3 variant have a normal life expectancy. She was on the mild end. She was in kindergarten, repeating a year, and most of the other parents did not know anything was different. Her father, a software engineer, was afraid of a future where the gap widened and she could not live alone. That is a reasonable fear and it is not a terminal diagnosis.

Which is the question underneath all of this, and it is the same one Gelsinger raised. What risk is proportionate when the condition is not going to kill the patient? Greely's framing is the honest one. It is easy to be blinded by hope, whether it is hope for your child, hope for your research, or hope for your company. All three were present in the same room, and all three pointed the same direction, and nobody in that room was assigned the job of pointing the other way.

The parents paid $860,000, including $130,000 wired directly to a researcher's personal account. They bought iPhones and an iPad and a case of Maotai. They blamed themselves afterward for trusting too completely. They were not the ones who were supposed to be checking.

The failure was not that good science was recklessly translated. The gap between the animal work and the child was papered over rather than crossed, and nobody whose job it was to see the gap was looking at it. That is what stepwise, gated research is for. Not to slow anyone down, but to make each unproven step visible before somebody stands on it.

I spend a fair amount of time telling patients that the research behind a therapy is sound and that the people who ran it were careful. Then a story like this one arrives. A family paid for the therapy, a child died of a documented complication of the delivery route, and none of it appeared in the paper. When somebody tells me they are not sure they trust medical research, I usually have an answer. On this one, they are right and I do not.