Nov. 16, 2026

Huntington’s Disease: Can Gene Therapy Slow It Down?

Huntington’s Disease: Can Gene Therapy Slow It Down?
Huntington’s Disease: Can Gene Therapy Slow It Down?
Normal Curves
Huntington’s Disease: Can Gene Therapy Slow It Down?

Could a new gene therapy slow the progression of Huntington’s disease? Early results suggested an astonishing 80% reduction in disease progression. But there’s a catch: the evidence comes from just a handful of treated patients, compared with patients from an entirely different study. How can we know whether we're comparing like with like? We dig into the science behind the treatment, the statistical challenges of finding a fair comparison group, and the heated controversy between the drug company and the FDA. Along the way, we take a statistical detour into propensity score matching—and wrestle with a difficult question: How much evidence should we demand when patients are running out of time?

Statistical topics

  • External control groups
  • Randomized controlled trials
  • Sham surgery controls
  • Propensity scores
  • Restricted full propensity score matching
  • Inverse probability of treatment weighting (IPTW)
  • Confounding
  • Selection bias


Methodologic Morals

  • “Seventeen patients is still 17 patients, no matter how you cut it and no matter how many controls you have.”
  • “With external controls, the question is always, are we seeing the treatment or are we seeing the difference between the studies?”

References

Kristin and Regina’s online courses:

Demystifying Data: A Modern Approach to Statistical Understanding

Clinical Trials: Design, Strategy, and Analysis

Medical Statistics Certificate Program

Writing in the Sciences

Epidemiology and Clinical Research Graduate Certificate Program


Program that Kristin teaches in::

Epidemiology and Clinical Research Graduate Certificate Program


Kristin’s manuscript writing course:

https://instats.org/seminar/from-data-to-discussion-writing-a-scient

Find us on:

Kristin - LinkedIn & Twitter/X

Regina - LinkedIn & ReginaNuzzo.com


  • (00:00) - Cold Open: FDA Demands a Bigger Trial
  • (08:17) - Huntington's Genetics: Dominant Inheritance and CAG Repeats
  • (18:05) - AMT-130: Gene Therapy Delivered by Brain Surgery
  • (20:31) - Early Trials, Sham Surgery, and Safety
  • (26:55) - External Controls: A Statistical Detour
  • (29:07) - Two-Year Headline: 80% Reduction
  • (36:42) - Propensity Scores: Barbie vs. Oppenheimer
  • (49:14) - FDA Reversal and the Ethics Fight
  • (56:03) - June 2026 Reversal and Four-Year Results
  • (01:00:36) - Rating the Claim: ...

00:00 - Cold Open: FDA Demands a Bigger Trial

08:17 - Huntington's Genetics: Dominant Inheritance and CAG Repeats

18:05 - AMT-130: Gene Therapy Delivered by Brain Surgery

20:31 - Early Trials, Sham Surgery, and Safety

26:55 - External Controls: A Statistical Detour

29:07 - Two-Year Headline: 80% Reduction

36:42 - Propensity Scores: Barbie vs. Oppenheimer

49:14 - FDA Reversal and the Ethics Fight

56:03 - June 2026 Reversal and Four-Year Results

01:00:36 - Rating the Claim: The Smooch Scale

01:02:16 - Methodological Morals

01:03:04 - Fold Me Once: Data Colada on Gino

[Kristin] (0:00 - 0:11)
The FDA looked at these three year results and said, nope, the data are not strong enough, and you need to do a larger, randomized, double-blind, sham surgery-controlled study. Wow.


[Regina] (0:12 - 0:13)
Whoopsie. Okay, that's quite a reversal.


[Kristin] (0:19 - 0:28)
Welcome to Normal Curves. This is a podcast for anyone who wants to learn about scientific studies and the statistics behind them. I'm Kristin Sainani.


I'm a professor at Stanford University.


[Regina] (0:28 - 0:34)
And I'm Regina Nuzzo. I'm a professor at Gallaudet University and part-time lecturer at Stanford.


[Kristin] (0:35 - 0:40)
We are not medical doctors. We are PhDs, so nothing in this podcast should be construed as medical advice.


[Regina] (0:40 - 0:45)
Also, this podcast is separate from our day jobs at Stanford and Gallaudet University.


[Kristin] (0:45 - 0:54)
Today, we're shifting gears a bit. We've been doing a lot of statistical sleuthing lately, but today we're going to talk about an experimental therapy for Huntington's disease.


[Regina] (0:55 - 1:01)
Big shift. I feel my gears grinding. A little bit.


But this will be interesting.


[Kristin] (1:01 - 1:37)
Yes, this is one of two episodes we're planning to do on experimental therapies for rare diseases. Both therapies have been in the news lately, and these cases bring up some interesting ethical and statistical questions. Regina, before we get to that, though, a quick update on the Francesca Gino case.


We released our second episode on that case on November 2nd, and coincidentally, that same day, a federal judge in Massachusetts dismissed all eight of her remaining claims against Harvard, because remember, she was suing Harvard, but Harvard's defamation claims against her are still pending. Which is just like, wow.


[Regina] (1:37 - 1:42)
That seems like a clear win for Harvard, not a happy day for Gino and Lessig.


[Kristin] (1:43 - 2:09)
Our timing was impeccable. I don't know how we managed to get that episode out right on the right day. There's another update, though, Regina, because yesterday, November 10th, Data Colada published another bombshell blog post, this time identifying suspicious data alterations in a fifth Gino paper from a 2011 study, and they called their blog post Fold Me Once, because the outcome dependent variable of this study was t-shirt folding.


[Regina] (2:11 - 2:21)
They are very clever, aren't they? Kristin, how about we do a little mini-summary of that blog post at the end of this episode?


[Kristin] (2:21 - 2:27)
Oh, I love that idea, Regina. Everybody needs to stick around for that, because it's fascinating, but we'll just tease it here.


[Regina] (2:27 - 2:38)
Okay, so shifting my rusty gears again, today we are talking about an experimental treatment for Huntington's disease, and what is that treatment?


[Kristin] (2:39 - 2:58)
The drug is called AMT-130. It's made by a company called UniQure, and it's a gene therapy for Huntington's disease. Regina, we're back to a straightforward scientific claim today.


The claim we're going to evaluate today is the claim of the company, which is AMT-130 slows the progression of Huntington's disease.


[Regina] (2:58 - 3:06)
Oh, good. Something easy for a change? No suspicious Excel files, just a drug, and whether it works, easy peasy.


[Kristin] (3:06 - 3:54)
Well, as we'll see, it's not quite that simple. I knew it was too good to be true. The complications are what makes it more interesting, Regina.


We're going to look today at the limited data we have addressing this claim, and AMT-130 has been controversial, and the company and the FDA have actually had a years-long back and forth. It's a pretty dramatic scientific feud, and it also highlights a lot of ethical questions, because at one point, an FDA official called AMT-130 a, quote, failed product, and they accused the company of, quote, performing a distorted or manipulated comparison. And then a former FDA official actually took issue with the current FDA, calling the FDA's decision, quote, truly evil.


[Regina] (3:54 - 4:15)
Wow. That is a great teaser, Kristin. And we also want to thank David Rind for the idea for this episode.


He is a patron of our podcast and also the chief medical officer of the Institute for Clinical and Economic Review, and we love listener ideas.


[Kristin] (4:15 - 4:27)
Yes. Please keep sending them. There are actually a bunch of listener ideas that we have on our list.


We haven't gotten to them yet because they are evergreen, but we will get to them in the future. We have not forgotten about them.


[Regina] (4:28 - 5:17)
Okay. So, Kristin, with rare and deadly diseases like Huntington's, right, when they're looking to approve a treatment for it, I can imagine there's a real dilemma that we are talking about here today, right? Because statisticians like us, of course, we want to make sure we have really good evidence that the treatment works before you give it to people, right?


But if it were me personally, if I had a bad disease that was just going to progress and debilitate me or kill me, or if it was my loved one, I don't know. I am not sure I would really care quite so much about having perfect statistical evidence. I'd probably want to try anything and it's hypocritical.


I know, but I am human.


[Kristin] (5:17 - 5:36)
Yeah. It's very human, Regina. And actually, I've been on the other side of the equation.


Having had a serious disease, I had breast cancer, and I have to say that statistics and numbers look totally different when you are on the patient side or when you're on the side of the family for the patient.


[Regina] (5:37 - 5:47)
Right. But of course, then again, you can't just approve anything that might work because what if it does not work or what if it makes things worse? Right.


[Kristin] (5:47 - 6:09)
That's why we have a rigorous clinical trial process before approval. But, Regina, that process can take years, right? The gold standard for establishing a treatment's efficacy is through large, blinded, placebo-controlled, randomized phase 3 trials.


And with a rare disease, it might take a decade to get enough patients for the ideal phase 3 trial.


[Regina] (6:10 - 6:52)
Right. And it's not like you can just skip straight ahead to a big phase 3 trial either, right? So even if it's just, quote, experimental treatment, you still can't just go ahead and give it to hundreds of people without having some idea that it's at least safe and that you've figured out kind of the right dose.


Right. And so that's why we have all these smaller studies first to look at safety and dosing. Those are the phase 1 and phase 2 trials, which come before phase 3 trials.


But when we do that, when we have those earlier studies, there we are. We're adding more time to the whole process. And we're talking about patients who are running out of time.


[Kristin] (6:52 - 7:49)
Exactly. It is a really difficult balance. I mean, patients want something.


So they may want access to a treatment, even if there is no good evidence that it works. And it's hard to say no to patients who are looking for hope. So the standards of evidence can be different in cases like this, right?


The FDA can exercise some flexibility. Like maybe you can get an accelerated approval based on phase 1 and phase 2 data while you're still running and haven't completed the bigger phase 3 trial and don't have those results yet. Or the FDA may accept a non-placebo, non-blinded control group.


Or they may accept something called external controls, which we're going to talk about today. And Regina, this back and forth between the FDA and the company, we're covering this because it turns out that statistics is actually at the heart of the controversy. So we have some exciting statistical topics today.


We're going to talk about external controls, propensity scores, and propensity score weighting and matching.


[Regina] (7:50 - 8:16)
That sounds nerdy, super fun. Propensity scores were a favorite back when we talked about the Barbie movie. So I can't wait to get to the statistical goodies.


But I think maybe first we should talk about Huntington's disease itself. And I actually have a friend whose father died of Huntington's. And it was just devastating.


It's a terrible disease.


[Kristin] (8:17 - 8:41)
Oh, wow. I'm sorry to hear that. Yes, because it is a devastating disease.


And one of the reasons it's so devastating is the genetics. You know, Regina, I think most people learned about this in high school biology that this is an autosomal dominant genetic disease. And Regina, I know it's been a while since you've had biology.


So I'm going to give you a little biology quiz here, see what you remember from high school biology. What does autosomal dominant mean?


[Regina] (8:41 - 9:11)
Yeah, I haven't had biology since 1987, by the way. At least they had discovered DNA by that point. But it has been a few years.


But I do know dominant means you only need to get one mutated copy of the gene to get the disease. And that's different from the more typical recessive genetic diseases where you need two bad copies of the gene. And autosomal, that's the regular chromosomes, right?


Not the sex chromosomes.


[Kristin] (9:11 - 9:34)
Exactly. Very good. Yes.


So what this means is a parent with Huntington's disease has only one copy of the bad gene, but their child has a 50% chance of inheriting the disease because each parent, of course, passes on half their genes. Huntington's is unusual because symptoms typically don't appear until adulthood, often in a person's 30s or 40s.


[Regina] (9:34 - 9:55)
Yeah, that late onset is especially awful because it means that people might have already had children at that point before they even know that they have the disease. So they've potentially passed it along without even knowing it. And so my friend that I talked about, he got tested and fortunately he did not have the mutation.


[Kristin] (9:55 - 10:22)
Oh, wow. That's really good luck. That's great to hear.


And Regina, the progression of this disease is really brutal. It progressively damages the brain, affecting your movement, thinking, mood. People have increasing difficulty walking, speaking, even swallowing.


And most people die about 15 to 20 years after symptoms begin. So if someone starts symptoms in their 40s, they might die in their 50s or 60s after years of progressive disability. Yeah.


[Regina] (10:23 - 10:31)
So what does the Huntington's gene actually do that causes all of this stuff that you just described?


[Kristin] (10:32 - 11:31)
Right. It's called the Huntington gene and it codes for the Huntington protein. Huntington's disease, by the way, was named after George Huntington, who was a physician who first described the disease in 1872.


And the gene and the protein are both named after the disease after him. It turns out, Regina, we all make Huntington protein and scientists actually don't know everything it does in the body, but we do know that it helps transport materials inside neurons. And the mutation that occurs in Huntington's disease is unusual because when I think of a mutation, what comes to my mind is the typical like single letter change in the DNA.


But that is not what the mutation is here. The Huntington gene has a stretch where the letters C-A-G repeat. So C-A-G, C-A-G, C-A-G.


Most people have 26 or fewer repeats and people can have up to 35 without getting the disease. 36 to 39 is kind of a gray zone. With this many repeats, some people will get the disease and some people won't.


But almost everyone who has 40 or more repeats will develop Huntington's disease.


[Regina] (11:32 - 11:49)
Wait, so you're saying everyone has these C-A-G repeats, right? And it's not that people with Huntington's have some completely different gene. It's that they have too many repeats of something that's already there.


So why does that matter, the extra repeats?


[Kristin] (11:50 - 12:05)
Yeah, those extra repeats matter because C-A-G, that's a codon. It's an instruction to add a particular amino acid called glutamine to the protein. So if you have extra C-A-G repeats, you get an extra long stretch of glutamines in the Huntington protein.


[Regina] (12:05 - 12:19)
Okay. So if I'm understanding this right, glutamine itself, perfectly normal. It's just that this protein ends up with this crazy long stretch of it because of all those extra C-A-G repeats.


Exactly.


[Kristin] (12:19 - 12:33)
And here's what makes this a dominant gene. This mistake in the protein does not cause the protein to just disappear altogether as is often the case in the recessive disease. Rather, it's the mutated protein itself that actually does damage.


[Regina] (12:33 - 12:53)
Oh, okay. So even if you have one perfectly good copy of the gene, that doesn't rescue you, right? Because the mutated version is the thing that's actively doing the damage.


You are still making that abnormal protein. Yeah, exactly. Okay.


So what do all of those extra glutamines do?


[Kristin] (12:54 - 13:15)
So the glutamines are involved in protein folding, right? Proteins have to fold into particular shapes to do their jobs. And this extra long stretch of glutamines makes the Huntington protein prone to misfolding.


And the misshapen proteins, or sometimes it's just pieces of misfolded protein, they can stick together and form clumps inside nerve cells and disrupt the nerve cells.


[Regina] (13:15 - 13:21)
So protein misfolding, that sounds familiar. Isn't that in a lot of neurological diseases?


[Kristin] (13:21 - 14:12)
Yeah, exactly. This is a recurring theme in brain diseases. Proteins fold incorrectly.


They stick together and form clumps. And we see this in Alzheimer's and Parkinson's, for example. And Regina, a little aside here, just because I think it's cool biology, this is actually how prion diseases, like mad cow disease, work.


So prions are these little proteins that misfold. And if you eat contaminated meat containing these prions, the prions themselves, the protein, can bind to normal copies of the same protein in your body and cause those normal proteins to also fold into this abnormal shape. And then those misfolded proteins make more proteins misfold, and it becomes a chain reaction.


And there is a fascinating book that I want to recommend to everyone here. You can get it as an audio book too, and it's called The Family Who Couldn't Sleep. And that family has an inherited prion disease.


[Regina] (14:12 - 14:28)
So you used the word fascinating. I would also say terrifying. This sounds like a fascinating, terrifying, nerdy bedtime story.


I remember when you were reading this, and I was thinking, I'm not sure I'd be able to sleep. Horrible disease.


[Kristin] (14:29 - 14:36)
But fascinating story, the way the author winds all the pieces together. Yeah, it's a really interesting biology. Okay.


All right.


[Regina] (14:37 - 14:53)
Back to Huntington's. So we know what goes wrong at the genetic level, but knowing what's going wrong is one thing, being able to fix it is another. So how can researchers actually get in there into the muck and fix things?


[Kristin] (14:54 - 15:16)
Yeah. It turns out to be not as easy as you'd hope. And Regina, there are no approved treatments right now that stop or slow the disease.


There have been many attempts, some of which are in clinical trials that we don't have the results from yet. The basic idea of these treatments is to try to reduce the levels of the mutant protein. And there are a lot of ways that you might do that because remember the central dogma of molecular biology, DNA to RNA to protein.


[Regina] (15:16 - 15:28)
Yeah. I even remember that back from 1987. DNA makes RNA, RNA makes protein.


Not literally makes something, something, but I'm close, right?


[Kristin] (15:28 - 15:47)
Right, right. Yeah. The DNA is transcribed into messenger RNA and the messenger RNA is translated into protein.


And that means you can intervene at different points, right? You could try to change the DNA in Huntington's, or you could try to destroy the messenger RNA, or you could wait until after the protein is made and try to shut down the protein or clear away the protein.


[Regina] (15:47 - 15:58)
Okay. All of this sounds really promising, right? You have a lot of different options for how to kind of turn the dial down on the mutant protein, right?


So anything promising?


[Kristin] (15:58 - 16:33)
Well, there have been some promising findings and there are several treatments in clinical trials, not just AMT-130. For example, there was a drug called Tominersen that targeted messenger RNA using a technique called antisense oligonucleotides. And this is where they create a strand of DNA or RNA, that's the oligonucleotide, and that binds to the Huntington messenger RNA, the antisense, and that recruits an enzyme which cuts up the messenger RNA.


That was a drug that was injected into the spinal fluid and they found that it actually did lower levels of the mutant Huntington protein in spinal fluid.


[Regina] (16:33 - 16:38)
Oh, that sounds like very good news. That's what we wanted, right? Lower the protein that's causing the problem.


[Kristin] (16:39 - 16:55)
Well, you'd think so, but it was actually, the results were disappointing. They did a large randomized trial and it was stopped in March of 2021 because patients were not benefiting. In fact, patients who received the drug in the most frequent dosing schedule actually did worse than the placebo group.


[Regina] (16:56 - 17:10)
Wait a minute. Okay. So they actually did manage to lower the mutant protein just like they were trying to do, but the patients themselves did not get better and some of them actually got worse.


So how does that happen?


[Kristin] (17:10 - 17:26)
We don't really know. Maybe because it was in the spinal fluid and not the brain, maybe because the drug was causing inflammation, or maybe because it lowered too much of the normal Huntington protein along with the mutant protein. Remember, that normal protein has important jobs to do.


[Regina] (17:27 - 18:05)
Okay. I find this fascinating. I feel like it's a good reminder why we need clinical trials in the first place, like we were talking about earlier, because you've got this great biology and this theory and which lever you need to pull in order to fix the disease, and you develop a drug, and then it turns out patients are not actually any better off.


I feel like we see this often with phase three clinical trials. Yeah. So all right.


Now with that backdrop, I'm feeling a little less optimistic about getting rid of that mutant protein now. So enter AMT-130. What is that about?


[Kristin] (18:05 - 18:29)
So AMT-130 also tries to reduce messenger RNA, but it uses something called micro RNA, same idea as the antisense oligonucleotide, but different pathway. So you make a complementary strand of RNA that binds to the messenger RNA, and this recruits some other proteins to cut up the messenger RNA. Now, like Tominersen, AMT-130 lowers both the mutant and the normal Huntington protein.


[Regina] (18:29 - 18:38)
Okay. So you are saying we're trying to do the same thing that did not work with Tominersen? So what is different about this one?


Why do they think it might work any better?


[Kristin] (18:38 - 18:56)
So again, different mechanism of destroying the mRNA, and Tominersen was injected into the spinal fluid, which of course is easier to access. And AMT-130, as we'll see, is delivered directly into the brain. So maybe getting it to the brain is the tricky part.


We don't really know, but there's some reasons to believe that this would be more effective.


[Regina] (18:56 - 19:01)
Okay. Directly into the brain cells. How do they do that?


[Kristin] (19:01 - 19:09)
So they use a modified virus to deliver instructions for producing this micro RNA into the brain cells.


[Regina] (19:09 - 19:19)
Okie dokie. Modified virus. How are you getting the virus into the brain cells?


Are brain cells tend to be kind of tucked away and not easily accessible?


[Kristin] (19:19 - 19:32)
It requires brain surgery, actually. So surgeons drill small holes in the skull and they use these thin catheters to deliver the drug directly into the striatum, which is a part of the brain that is particularly affected by Huntington.


[Regina] (19:33 - 19:56)
Yikes. Okay. So we are talking about literally drilling into people's heads here.


Right. Not just like a pill or something. And for an experimental treatment, that is kind of a leap of faith to undergo brain surgery.


So it must be that people are hoping for a pretty big benefit to justify taking that kind of big risk, brain surgery.


[Kristin] (19:56 - 20:07)
Yeah. I mean, one good thing is it's designed as a one-time treatment. So you're giving the brain cells the instructions and they can keep making the micro RNA themselves.


And the hope is that the effect could last for years.


[Regina] (20:07 - 20:31)
So you are turning the brain cells into little drug factories. Yep. The idea.


Okay. All right. That's fun.


So one big risk, hopefully years of benefit if it doesn't work. So how do you even start testing something like this in people? So I'm assuming that we are back to that phase one and phase two trials we talked about earlier, where you're trying to figure out, is it even safe to do this?


Exactly. You got to do that first.


[Kristin] (20:31 - 20:42)
So in 2020, they started testing AMT130 in humans with a combined phase one, phase two trial. And the goal was to look at safety and tolerability at two different doses.


[Regina] (20:43 - 21:04)
Okay. So we are talking here about a very small study, totally makes sense, right? You don't want to give hundreds of people experimental brain surgery before you have some idea that it's safe.


But then the flip side, of course, with so few patients, you're not going to have much statistical power to show whether the treatment actually works. That's not the point.


[Kristin] (21:04 - 21:20)
Exactly. These studies were very small. The initial US study enrolled 26 patients in two stages.


So first, they randomly assigned six patients to receive a lower dose of AMT130 and four to actually receive sham surgery.


[Regina] (21:20 - 21:53)
Ooh, sham surgery. We have talked about sham surgery before. So you need the sham group partly because you don't want patients to know if they got the actual drug or not, right?


Because even just that knowledge can affect how they feel and what kind of symptoms they experience and report when you're talking about safety and tolerability. But especially for the safety outcomes, right? If someone gets a terrible headache afterwards, I'm thinking, how do you know whether that's because of the drug or just because they had someone drill holes into their head?


[Kristin] (21:53 - 22:28)
Good question, Regina. Yeah. And that's one reason having a comparable comparison group is so useful.


The sham surgery patients were put under general anesthesia and they had small superficial holes drilled into the skull. Of course, nothing was actually injected into their brain. So you can't completely separate out the gene therapy from the brain surgery delivery.


Still, it's a much better comparison than doing nothing to them. All right. So Regina, after those first 10 patients, an independent safety board reviewed the safety data and they decided it was safe to go on and try a higher dose.


So they then randomly assigned 10 patients to receive the higher dose and six more for sham surgery.


[Regina] (22:29 - 22:37)
Okay. So baby steps. It sounds like they're being pretty cautious here.


Try it with a few patients at low dose. Turn up the dial a tiny bit with a few more patients.


[Kristin] (22:37 - 22:40)
Yeah, exactly. And they followed everyone initially for 12 months.


[Regina] (22:41 - 22:54)
Okay. So we've got 26 patients altogether, if I counted right. Very small, but at least it's randomized and we have a good control, sham surgery control.


Are we ready to hear what they found?


[Kristin] (22:54 - 23:10)
Yeah. They began reporting safety results in 2022. The lower dose was generally well tolerated.


They did though have four serious treatment related adverse events in the higher dose group. So three cases of central nervous system inflammation and one severe headache.


[Regina] (23:11 - 23:21)
Okay. So that is four kind of adverse events in a small study. A little worrisome, right?


Especially that inflammation thing. Big deal. Yeah.


[Kristin] (23:22 - 23:29)
The good news is all of them improved with steroids. So the independent safety board did decide that they could keep going with the high dose treatment.


[Regina] (23:29 - 23:44)
Ah, okay. Safety part, not bad enough to stop the whole trial. And what about efficacy?


So I know it was not designed for that, but often researchers will look at the data for hints of efficacy anyway. Yeah. So did they find any hints?


[Kristin] (23:44 - 23:54)
Well, they didn't give us much to work with here. So in June of 2023, the company put out a press release that did discuss efficacy, but there were not a lot of concrete numbers.


[Regina] (23:54 - 24:05)
Okay. That is frustrating. But Kristin, knowing you, I'm wondering, did you manage to figure out anything from whatever they did release to the press?


[Kristin] (24:05 - 24:19)
I went through the limited materials pretty carefully and they focused on one secondary efficacy outcome, which is called the total motor score has to do with movement. And they said that this score worsened in sham patients, but it was stable in the treated groups.


[Regina] (24:19 - 24:27)
Hmm. Okay. One secondary outcome and they're implying treatment helped, right?


Did they give any actual numbers for this?


[Kristin] (24:27 - 24:57)
Well, that's where it gets tricky. We didn't get much in the way of stats. They did provide a graph though.


And when I look at the graph, I really don't see a convincing difference. To me, it really looks like noise. It's not robust.


And there were two other outcomes that they mentioned, an overall composite score and a daily functioning score, but they give us absolutely no numbers or no graphs. They just told us that the sham patients, quote, remained stable on these other two outcomes, which to me implies that they saw no differences in the groups on these outcomes.


[Regina] (24:58 - 25:18)
Yeah. Otherwise they probably would have said, okay, so you are saying you did not see any convincing evidence that it actually helped. Yeah.


But of course, we're talking about tiny trial here, 26 patients altogether. So even if there were some benefit, we might not be able to see it with such a small sample. That's good to keep in mind.


[Kristin] (25:18 - 25:23)
Exactly. Small sample, short follow up. The study was not designed to establish efficacy.


[Regina] (25:23 - 25:34)
So we can't say that it worked, but we can't say that it didn't work either. Right. Okay.


So they got the green light on safety to keep going. What did they do next?


[Kristin] (25:34 - 25:44)
So after 12 months, they broke the blinding, but they continued to follow all of the treated patients. And some of the sham patients were then allowed to get AMT-130 if they wanted. Oh.


[Regina] (25:44 - 25:57)
So at this point, they fundamentally changed the whole compared to what part of the trial here, right? Because once the sham surgery patients started getting the real treatment, you can't use them as a control group anymore.


[Kristin] (25:57 - 26:16)
Right. And they also started a separate study in Europe and the UK where they had no control group, but they gave 13 more patients the drug, six at the lower dose and seven at the high dose. So altogether, that gave them 12 total low-dose patients and 17 high-dose patients, 29 total treated patients.


[Regina] (26:17 - 26:35)
Okay. So basically, we started with a randomized trial with a sham surgery control group, and now we're just following a bunch of treated patients over time with no comparable untreated patients in the same study. That feels like a shift.


[Kristin] (26:35 - 26:36)
It is.


[Regina] (26:36 - 26:55)
Right. So if you don't have a control group anymore, Kristin, how do you even start to figure out whether these treated patients are actually doing any better than they would have without the treatment? Because Huntington's disease progresses over time, but not necessarily at the same rate for everyone.


This is why we like to have a control group.


[Kristin] (26:55 - 27:00)
Right. And this is where the company turned to external controls.


[Regina] (27:00 - 27:09)
Ooh, external controls. This is a new topic for us, Kristin. So let's do a little statistical detour on external controls.


Sure.


[Kristin] (27:09 - 27:26)
An external control means you look at data from people who were not part of the original trial, and their data might come from another study, maybe from a patient registry or from medical records. In this case, the company used data from a completely separate observational study of Huntington's disease.


[Regina] (27:27 - 27:40)
Okay. So you're going to look at how fast those external control patients are getting worse, and then you're comparing that with how fast the treated patients, here the AMT-130 patients, are getting worse. Yeah.


But there's a big problem.


[Kristin] (27:41 - 27:50)
The reason we prefer controls from the same study, ideally randomly assigned, is to make the groups comparable. Maybe different studies are looking at different patient populations.


[Regina] (27:50 - 28:04)
Which is important, because if you happen to have healthier patients in your treatment trial, then their disease progress might happen more slowly than it does in the controls, even if the drug is doing absolutely nothing.


[Kristin] (28:04 - 28:23)
That is exactly the central problem that we are going to run up against here, whether these two groups of patients were really comparable. That became the center of the controversy and the back and forth between UniQure and the FDA. Because when UniQure compared the treated patients with these external controls, they reported some promising results.


[Regina] (28:23 - 29:07)
Okay. Promising results, but with these potentially weird controls. I feel like that is a great teaser, Kristin.


Nice job. Good time for a short break. Welcome back to Normal Curves.


Today, we are discussing AMT-130, an experimental gene therapy for Huntington's disease. We were about to get to some efficacy results.


[Kristin] (29:07 - 29:20)
Yeah. In July of 2024, the company reported two-year results. Their banner headline was that AMT-130 demonstrated an 80% reduction in disease progression compared to the external controls.


[Regina] (29:20 - 29:32)
80% reduction in disease progression. That actually sounds pretty dramatic and impressive. How are they measuring this?


What's the outcome measure here?


[Kristin] (29:32 - 30:02)
Their primary measure of efficacy was a composite score called the Composite Unified Huntington's Disease Rating Scale, which is widely used. It captures how patients are doing on movement and thinking and everyday functioning and higher is better. These patients started at an average of about 15 and the company reported a 0.2, two-tenths of a point decline in this score in the high-dose group versus a full one-point decline in the external controls.


[Regina] (30:03 - 30:21)
Okay. Treated patients still got worse on average, not unexpectedly, but they got worse a lot more slowly. That 80% comes from comparing that two-tenths of a point decline with a one-point decline.


Even I can do that math in my head.


[Kristin] (30:22 - 30:41)
Exactly, exactly. That's easy math because of the 1.0 point is even. Regina, the p-value was statistically significant.


Now, for the low-dose group, they found a 0.7 point decline, seven-tenths of a point decline versus 1.0 point in the external controls and that was not statistically significant.


[Regina] (30:41 - 30:53)
Okay. More exciting for high-dose than low-dose, but getting back to that magnitude, that eight-tenths of a point slowing in progression, is that enough to matter to patients?


[Kristin] (30:53 - 31:20)
Potentially, yes. If this was sustained over many years, it could be very meaningful, but of course, Regina, the data here are not robust by any stretch. We're not even talking about the full 29 treated patients, so when they released this analysis, only nine high-dose and 12 low-dose patients had a full two years of follow-up data.


The statistical model still includes everyone, but essentially, the estimate at two years for the high-dose group is based on nine treated patients.


[Regina] (31:21 - 31:39)
Okay. This dramatic 80% reduction, essentially, just nine patients here. That means there's a lot riding on just nine people and a lot riding on these controls, especially figuring out whether they are really a good comparison group.


[Kristin] (31:39 - 31:51)
Right. There's a lot riding on this because the company took these results, the statistically significant difference on the main efficacy measure in the high-dose group, they took it to the FDA to discuss a faster path to approval.


[Regina] (31:52 - 32:00)
Okay. Small group, potentially weird external controls, the company is excited, but what does the FDA say?


[Kristin] (32:00 - 32:21)
We don't really know because we don't have a transcript from the FDA, but in December of 2024, UniQure, the company, issued a press release saying that the FDA had agreed that these small trials compared with external controls could provide the main evidence for an accelerated approval application, meaning they wouldn't have to do another study before submission.


[Regina] (32:21 - 32:29)
The company was telling investors that the FDA was willing to consider approving the drug without doing another trial.


[Kristin] (32:29 - 33:13)
Yes, and investors were happy. The stock more than doubled during trading that day. As we'll see, Regina, an FDA official later disputed the company's account saying that the agency had never said this, that they had always wanted a bigger trial with an internal control group.


Ooh, disagreement. Now, we don't have a public transcript of those discussions, so we're relying on the two sides accounts, but at the time, the company was moving forward on the understanding that it could apply without another trial first. And then in September of 2025, it released its three-year results, and Regina, just one detail to keep in mind, when they reported their three-year results, they changed the external control group, so we're not necessarily even comparing the same groups across these different analyses.


[Regina] (33:14 - 33:23)
Okay, interesting. So we're following the same treated patients longer, but beneath the surface that whole compared to what control group just got changed on us. Yes.


Okay.


[Kristin] (33:24 - 33:44)
And this three-year analysis now included 12 high-dose and 12 low-dose patients who had completed a full three years of follow-up, and the high-dose results still looked impressive. They reported a 75% reduction in disease progression in the treated group compared with the external controls. Note that that is very similar to the 80% figure that they had reported at two years.


[Regina] (33:44 - 33:48)
Yeah, okay, so still sounding promising. So what about the low-dose group?


[Kristin] (33:49 - 33:56)
Well, the low-dose group is a different story. At three years, the low-dose patients had declined at almost exactly the same rate as their external controls.


[Regina] (33:56 - 34:18)
Ooh, okay, so basically no benefit in the low-dose group, at least not an apparent one that we could see, but this is disappointing in a bigger way, right, Kristin? Because if the drug really does work, you might expect at least some benefit in the lower dose, even if it's not as much of a benefit as the high dose, but we're not seeing that here.


[Kristin] (34:19 - 34:39)
Right, this makes the whole thing look less robust, right? But, Regina, some doctors did find the high-dose results impressive, because these patients had, on average, declined on that composite score by only about four-tenths of a point over three years in a disease where you expect progressive worsening, four-tenths not so huge.


[Regina] (34:39 - 34:53)
Okay, so this is promising. We're still back to that whole compared to what problem, though, right? Because maybe, Kristin, these patients would have progressed slowly anyway.


We can't know.


[Kristin] (34:53 - 35:10)
Exactly. That's the central dilemma here, and there's a particular reason to worry about that, because to receive AMT-130, patients had to have sufficient volume in two brain regions affected by Huntington's, and it turns out that patients with larger volumes in those brain regions do tend to progress more slowly.


[Regina] (35:11 - 35:37)
Oh, interesting. So, the trial was actually selecting patients for a characteristic that has been associated with a better prognosis. That is potentially a big problem when you're working with external controls who were not selected for this, right?


So, how did the company try to make sure that those external controls really were a good comparison group and comparable to the patients?


[Kristin] (35:37 - 36:24)
Yeah, so these controls came from ENROLL-HD, which is a large ongoing international observational study. It began in 2012, and this is a study where researchers have been following people from Huntington's families over time, generally with annual visits, and they collect the same movement, thinking, and functioning measures used in the AMT-130 trial, which lets the company compare changes on the same scores, at least. The company selected a subset of about 3,000 participants from ENROLL-HD, using most of the eligibility requirements from the AMT-130 trial, but importantly, there was one eligibility requirement that they could not apply, because it was not measured in ENROLL-HD, and that was the brain volume requirement.


You need an MRI for that, and ENROLL-HD didn't have that.


[Regina] (36:24 - 36:32)
Oh, that does seem very important. Yeah. So, the patient populations here have similarities, definitely not the same.


Right.


[Kristin] (36:33 - 36:42)
So, to further make the groups comparable, because they're not totally comparable, they then applied a statistical technique called restricted full propensity score matching.


[Regina] (36:42 - 36:49)
Ooh, Kristin, let's take a statistical detour here to talk about propensity scores and propensity score matching.


[Kristin] (36:49 - 36:50)
Yes, let's do it.


[Regina] (36:51 - 37:01)
We talked about propensity scores back in the shingles vaccine episode, and that was where we discussed inverse probability of treatment weighting, IPTW.


[Kristin] (37:02 - 37:13)
Right, and we decided to rename IPTW to sex statistical exception expansion, because that just sounds more exciting. Unfortunately, I don't think it's caught on yet, Regina.


[Regina] (37:13 - 37:24)
I don't think it's caught on. I'm still quite proud of that. I mean, we did have to stretch the whole acronym thing here, but I feel like it's worth it.


It's worth it. Exactly.


[Kristin] (37:24 - 37:33)
But let's, Regina, to illustrate things, revive the example that we used when we talked about IPTW. We used the Barbie and Oppenheimer movies.


[Regina] (37:33 - 38:06)
Which I love. We said the audiences were likely to be very different. This was Barbenheimer from a few summers ago.


So, maybe Barbie was all pink and teenage girls, and Oppenheimer was, I don't know, older men from engineering. And if we wanted to compare the two movies by asking people who were leaving the theaters to rate them, we would have a big problem, we said, because are the ratings different because of the movies or because of the people who went to see them?


[Kristin] (38:06 - 38:15)
Right. Ideally, we would randomly assign people to the two movies, and then we would compare their ratings. That would be comparable, but that's a much harder study to do.


[Regina] (38:15 - 38:56)
It is. Okay. So, instead, we use propensity scores, and let's just, Kristin, say 500 people show up to each movie.


That's a nice round number. So, we fit a statistical model using all 1,000 people. And let's say the outcome that we are predicting is which movie they saw.


And the predictors are the characteristics that might influence that choice of movie that they went to see. So, age, or sex, or social media use, or how much they like the color pink, or whether they like engineering, a score on a nerdiness questionnaire.


[Kristin] (38:57 - 39:22)
Right. And let's pause for a minute, Regina, because I think this can be confusing. So, we are saying we are fitting a statistical model to predict who saw Barbie versus Oppenheimer.


And that might seem strange because we already know who saw Barbie and Oppenheimer. But we're doing this so we can give everybody an estimated probability of being in the Barbie group. That probability is their propensity score, and that turns out to be a useful quantity.


[Regina] (39:23 - 39:42)
Right. So, let's say a young woman who loves social media and pink might have an 80% probability of seeing Barbie. And let's say a 60-year-old male engineer with a big toolbox in his workshop, he might have a 15% probability of seeing Barbie.


Exactly.


[Kristin] (39:42 - 40:15)
And then we can use those propensity scores to help us balance the groups. So, in the shingles episode, we use those probabilities, those propensity scores for weighting. Right.


So, imagine there was a young girl who liked pink, but happened to go to the Oppenheimer movie, maybe with her dad. She's unusual for that group. So, we want to give her a bigger weight in the Oppenheimer group, right, because she is representing the young people, which might be very sparse in that group.


So, if we upweight her, that's going to help us balance out things like age and how much you like pink, and it's going to make the comparison fairer.


[Regina] (40:15 - 40:52)
Right. We call this giving more voice, upweighting the statistical oddballs. That was the statistical exception part of the sex acronym.


But there is another way to use those propensity scores. And instead of giving people different weights, we can do some matching. So, imagine someone who saw Barbie had a 70% probability of seeing Barbie based on those characteristics that we measured.


We could then match them with someone who saw Oppenheimer, but who also had about a 70% probability of seeing Barbie.


[Kristin] (40:52 - 41:06)
Right. Even though they saw different movies, they had similar probabilities of choosing Barbie based on the characteristics we measured. So, they are more comparable than two randomly chosen moviegoers.


And that helps balance the groups and make the groups more comparable.


[Regina] (41:06 - 41:18)
Right. So, that is the idea behind matching. Now, Kristin, where did this restricted full propensity score matching come in?


That's what we are talking about here. That is a mouthful to say, by the way.


[Kristin] (41:18 - 41:22)
Right. And it doesn't make a nice acronym, RFPSM.


[Regina] (41:23 - 41:24)
Definitely not.


[Kristin] (41:24 - 42:05)
Yeah. Regina, the simplest matching that people often do is a one-to-one matching. So, you find one Oppenheimer viewer with a similar propensity score for every Barbie viewer.


That's just kind of the easiest. But, Regina, there are cases when a one-to-one matching actually isn't optimal. Right.


And AMT130 is one of those cases because it has very few treated patients, but lots of potential control patients. So, we want to use more control patients if we can. In the movie example, suppose for some reason we had just 50 Barbie viewers, but 1,000 Oppenheimer viewers, maybe the day Oppenheimer came out.


If we matched one-to-one, we'd just end up with 50 people from each group, just 100 people altogether. Which is just a waste.


[Regina] (42:05 - 42:16)
Yeah. We've got those 1,000 nerds watching Oppenheimer. Why throw them all away?


And this is where we say, can't we use more than one match for each Barbie viewer?


[Kristin] (42:16 - 42:45)
Yeah, we can. We could match each Barbie viewer with four Oppenheimer viewers, or five, or six, right? As long as we can find people with similar propensity scores from the Oppenheimer group.


That gives us extra information from the control group, and that can help make the comparison more precise. Okay, so full matching takes that flexibility even further. Instead of requiring the same number of matches, like exactly four for every Barbie viewer, we could be more flexible and allow some Barbie viewers to have more controls than others.


[Regina] (42:45 - 43:06)
Right. So, Kristin, we might have some people in the treated group, those are the Barbie viewers, who have lots of good potential control matches. Those are the Oppenheimer viewers, and other Barbie viewers might not have a lot of good potential matches, for whatever reason.


They're a combination of characteristics. But with the full matching, we can totally be flexible about that. Right.


[Kristin] (43:07 - 43:24)
We can optimize. But then, Regina, we do have to account for the different group sizes, and we do that through weighting. So, if one of the control groups is really large, we don't want to count those controls extra than another control group that's really small, so we use some weights to balance out so all the count equally.


[Regina] (43:25 - 43:43)
Right. I feel like this is so clever. We're not letting someone who has a whole crowd of matches count like 10 times as much as somebody with just one match, but we're still making sure that all of those controls contribute information to the analysis anyway.


Yeah, exactly. It's pretty cool.


[Kristin] (43:43 - 43:59)
And the restricted part of restricted full propensity score matching, that restricted just means that we're putting some limits on the matching. For example, we might cap how many controls can be matched to each treated patient. Maybe we allow 10, but not 100.


[Regina] (43:59 - 44:27)
So, Kristin, as we're talking about this, I imagine that exactly where you set those limits could really affect other results downstream, right? It could affect who gets included as a match, and who doesn't, and how much they get weighted. Exactly.


All right. So, we have matched our movie audiences. We've decided how much everyone is counting in our analysis.


Now, we can compare the ratings more fairly. Yes.


[Kristin] (44:28 - 44:43)
Keeping in mind, of course, that we've only addressed the characteristics that we've measured. So, if we didn't measure something important that affects both which movie people go to and how they rate the movies, we could still have a bias comparison, right? We can't match on things we didn't measure.


[Regina] (44:43 - 44:56)
I feel like we say this a lot. Statistical approaches like this are amazing, but they're not magic. Matching can make the groups more comparable, but it's not perfect.


We still have to worry about what we might have missed.


[Kristin] (44:57 - 45:44)
Statistical wizardry, it's great, but it's not magic, and it doesn't totally compensate for a weaker study design. Yes. All right.


Let's get back now to AMT130. So, they used eight baseline characteristics to calculate the propensity of being in the treated group of receiving AMT130. They then did this restricted full propensity score matching that we just explained, and they ended up with 940 controls for the 17 high-dose patients, and they used a separate control group, 626 controls for the 12 low-dose patients.


And they show us that the measured characteristics of the groups do look pretty similar. For example, the high-dose group averaged about 46 years old with a composite score of about 15 points at baseline, and for their match controls, it was 45 years old on average with about the same average composite score.


[Regina] (45:44 - 46:03)
Okay. This is encouraging and kind of nifty because they started with thousands of potential controls, and they did all this fancy statistical matching, and they actually did end up with groups that looked pretty similar, again, at least on the things that they measured, which of course we know is not everything. Right.


[Kristin] (46:03 - 46:18)
But, Regina, we are missing a lot of statistical details about this propensity score matching. Like, I don't know, did one treated patient have 100 controls and another one have 5? Did a handful of controls end up driving the comparison?


If you change the matching rules, are you going to get the same result? Right. There's a lot we don't know.


[Regina] (46:19 - 46:44)
It would be nice to have information on that because there's a lot of judgment involved in all of this matching, like we talked about, especially with only 17 treated patients, right? Different decisions might change results a lot. And, Kristin, it's what stats nerds like you and I call researcher degrees of freedom or the garden of forking paths, which just both of those are super nerdy.


[Kristin] (46:45 - 47:08)
They are super nerdy. And that's why it would be great to have more details about exactly how they did the matching, what things were pre-specified, what were not, but we just don't have that. And just to remind everyone, even though we started with 17 high patients and 940 controls, not everybody made it out to three years.


So there were only 12 treated and 568 controls that had data out to three years when they did this three-year analysis.


[Regina] (47:08 - 47:31)
Okay. Good to remember. So let me see where we are, Kristin.


Company reported this really impressive 75% reduction in decline for high-dose patients, but we've got a tiny treatment group, we've got some questions about the missing follow-up you just talked about, a whole bunch of decisions about how they constructed these potentially weirdo external controls. Yep.


[Kristin] (47:31 - 48:09)
Yes. And I should say that one secondary outcome was significantly better in the low-dose group versus controls. But remember, the primary outcome was totally not different and they tested multiple secondary outcomes.


So this could just be a chance finding. So, Regina, I'm looking at this and saying, yeah, maybe there's something there, but it's not terribly convincing. And there is another wrinkle because remember I said that one of the eligibility criteria for the AMT-130 trial is that patients needed this sufficiently large brain volume and Enroll HD just didn't measure that.


So researchers could not apply the same volume requirement to the external controls and they also could not match on that characteristic.


[Regina] (48:10 - 48:34)
We talked about this earlier. Treated patients had to pass this extra brain volume test to qualify to be in the study. External control didn't have that hurdle.


So the problem is if that larger brain volume is associated with slower disease progression, that means you could end up with your treatment group that's just healthier to begin with and progress more slowly no matter what. Right.


[Kristin] (48:35 - 48:49)
And actually, I went down a whole rabbit hole on this brain volume requirement because the company did do some additional analyses. We don't have time to get into that in our podcast episode today, but I'll leave some of those details for the show notes.


[Regina] (48:49 - 49:14)
Okay. So figuring out where we are, propensity score matching, very clever. Makes sense why the company would use it, but we still have this big question, again, about whether those external controls are really comparable, which means that whole 75% slowing in disease progression, that was really impressive.


We can't be nearly as confident about it as we'd like.


[Kristin] (49:14 - 49:54)
Which brings us now back to the FDA. So we are now, Regina, in fall of 2025. And remember, back in December of 2024, after those two-year results, UniQure had put out this press release saying, hey, the FDA has given us the green light to go ahead with these external controls as the basis of an accelerated approval application.


Well, it turns out that the FDA was not impressed with the three-year results because in the fall of 2025, according to UniQure, the FDA looked at these three-year results and said, nope, the data are not strong enough and you need to do a larger, randomized, double-blind, sham surgery-controlled study. Wow.


[Regina] (49:55 - 50:01)
Whoopsie. Okay, that's quite a reversal from what the company at least reported that it thought it had been told.


[Kristin] (50:01 - 50:38)
Right. And again, as I said before, the FDA disputes that they had agreed to something different back in 2024. In fact, in March of 2026, in a press briefing, an unnamed senior FDA official is quoted saying, did FDA ever agree to accept this distorted comparison?


No, I've asked our people to look for any written or verbally transcribed or minutes of a meeting that documents such a promise and there is no such promise made. And they also said, quote, moreover, FDA as a general rule never makes such assurances. FDA will always say, well, we have to see the data when we get it.


[Regina] (50:38 - 50:42)
Ooh, they are totally contradicting the company here.


[Kristin] (50:42 - 50:59)
They are. And the FDA official, this unnamed official was actually pretty harsh. They also are quoted saying, instead of doing the right thing and running the correct clinical study, UniQure is performing a distorted or manipulated comparison in the mind of FDA.


[Regina] (51:00 - 51:05)
Distorted or manipulated comparison. Ah, that is strong. Strong language.


[Kristin] (51:06 - 51:31)
Very strong. And the FDA also pointed out a contradiction in UniQure's positions. So, UniQure was trying to explain away the fact that they didn't see much in the sham, that 12 months sham surgery control trial.


But they were saying, well, it wasn't long enough. It was nearly impossible to show a benefit because there wasn't time for the controls to decline. But then in their other data, they reported a benefit at one year against external controls.


[Regina] (51:32 - 51:42)
So, they're saying one year is too short to see an effect when you're comparing patients in an actual randomized trial, but long enough if you're using these external controls. Yeah. Interesting.


[Kristin] (51:43 - 52:04)
Yeah. The FDA picked up on that contradiction and the official is saying, if both of those statements are true, why do we see a benefit at one year in the external control data but not in the randomized data? Zing.


Zing, yeah. So, when UniQure confirmed the FDA's position on March 2nd of 2026, its shares plummeted more than 40%.


[Regina] (52:04 - 52:06)
But this was not the end of the drama?


[Kristin] (52:06 - 52:24)
No, Regina. And this is where those other ethical questions that we alluded to earlier really come up. So, we as statisticians, I think we look at data like this and think, okay, well, maybe there's a signal here, but there's an awful lot of uncertainty.


I personally would not invest in the company based on these results.


[Regina] (52:25 - 52:36)
Right. But then also, what we were saying at the beginning, if I actually had Huntington's disease or my loved ones did, I'm not sure I would be thinking about this in that same way.


[Kristin] (52:36 - 52:49)
Right. I mean, I completely sympathize here with the patient's perspective. It's one thing to look at data from on high, from afar, and pontificate about what the numbers show, what the evidence shows, but it's totally different to be on the other side.


[Regina] (52:50 - 52:51)
So, patient advocate groups, would they say?


[Kristin] (52:52 - 53:19)
So, not surprisingly, the patient advocate groups pushed back pretty hard because from their point of view, here is a treatment that potentially, maybe, has some chance of slowing progression. And now patients are being told that they have to wait for a complicated randomized trial that could take years. The company and the patients also complained about the ethics of doing sham surgery, right?


Is it ethical to subject a participant to anesthesia and cuts in their scalp without actually giving them any treatment?


[Regina] (53:19 - 53:35)
Right. We talked earlier why this is good statistically, but I can see how this would be a hard sell to patients, right? You're asking patients with this already devastating progressive disease to get brain surgery, and at the end of the day, maybe not even get the treatment.


That's tough.


[Kristin] (53:35 - 53:54)
Yeah. And so, the company and the patients were not happy, and there was actually a New York Times article in which Janet Woodcock, who is a former senior FDA regulator, she said this of the FDA's decision, the Huntington's refusal, I thought, was truly evil. I just feel so bad for those people.


[Regina] (53:54 - 54:03)
Truly evil. That is a pretty, wow, extraordinary thing to say in print for a former regulator.


[Kristin] (54:03 - 54:59)
Yeah, Regina, that is really a moral indictment of statisticians in some way, right? We are on the side of evil because we say, hey, no, we want evidence. And patients put together a petition to try to encourage the FDA to back down, basically.


So, we just want to read a quote by a mother whose husband had Huntington's disease and whose children were then at risk, and she wrote, as a mother, the single most terrifying thing is that I cannot protect them, meaning her children, from this fatal disease. My daughter, like many other young people, is getting ready to start the process of testing for the Huntington's disease gene. If she tests positive, this disease will have defined her entire life since she was four years old and her dad was diagnosed.


A new therapy could give her more time where Huntington's disease does not define her, not to sway her career, her aspirations, her hope, or her ability to dream.


[Regina] (55:00 - 55:15)
Wow. You know, hearing that, you can really feel and understand why a family in that situation would be willing to just go ahead and accept more ambiguity, more uncertainty than we would normally be comfortable with.


[Kristin] (55:15 - 55:22)
Yeah. I mean, the uncertainty about the treatment is real, but so is the cost of waiting, and the difficult question here is how to weigh those things.


[Regina] (55:22 - 55:35)
Right. So, we're kind of stuck with two competing risks here, right? Approve something that does not actually work or delay something that might help people who themselves don't have time to wait.


[Kristin] (55:35 - 56:03)
Right. It's the cost of false negatives versus the cost of false positives, right? But then, Regina, there was another development.


In June of 2026, the FDA's position changed again. So, according to UniQure, the FDA agreed that the existing three-year analysis could serve as the main basis for an accelerated approval application, and they also discussed allowing a confirmatory phase three trial using a standard of care control group rather than the sham surgery control group.


[Regina] (56:03 - 56:16)
Okay. So, FDA now agreed that the company could apply for accelerated approval using the evidence it already had, but they also needed to plan a larger trial to confirm whether the treatment helped.


[Kristin] (56:17 - 57:18)
Yes. This reopened the path to an application. It is, of course, not an approval.


They still need to get it approved, but it is a path forward. But then, Regina, we get another twist. Whiplash.


I've got whiplash. Yes. So, in September of 2026, the company released new four-year results, and those four-year results were a little less impressive.


So, at this point now, 12 high-dose patients had four-year observations. Their reported decline was about nine-tenths of a point on that composite score compared with a 1.6-point decline in the external controls. That translates to just a 44% reduction in the progression of disease, no longer 80% or 75%, and it was no longer statistically significant.


Now, there were two secondary outcomes, everyday functioning and one cognitive test that still looked favorable with nominally significant results, but the primary outcome, the composite score, and some other secondary outcomes were not significant. Ouch.


[Regina] (57:18 - 57:36)
Okay. So, when we take it out to four years, that original banner headline result, 75%, slower progression. Now, we're down to 44% and no longer statistically significant.


Okay. Not quite as convincing now, is it?


[Kristin] (57:36 - 58:14)
Right. It's bouncing around a lot, as you might expect with results that are not robust. Now, of course, the company had a counter-argument.


They argued that, hey, wait a minute, we might be underestimating the treatment effect because some of the controls are dropping out. And so, they noted that by four years, 53% of the controls had no observations, and their analysis suggested that controls who stopped coming back had been progressing faster. So, they said, hey, we're essentially underestimating how bad the controls are doing because we're dropping the sick controls, and that's making the treatment look less effective.


[Regina] (58:14 - 58:20)
Okay. It seems not entirely implausible, but couldn't the same thing be happening in the treated group?


[Kristin] (58:21 - 58:35)
Right. We don't have observations on everyone in the treatment group either. So, we really need to understand the missing follow-up in both groups, and this is where it would have been really nice to have those internal controls where everything was more comparable.


[Regina] (58:35 - 58:43)
Right. So, again, interpretation depends heavily on the controls, which is why I am getting a tattoo compared to what.


[Kristin] (58:43 - 58:59)
Compared to what, yes. And this is why internal controls are so much better. And Regina, investors weren't buying this though.


So, when the four-year results were released, the stock went down again, plummeted down about 40%. Ouch.


[Regina] (58:59 - 59:02)
Okay. So, where are we now? Where's the application?


[Kristin] (59:04 - 59:55)
So, the four-year results were released. It was based on the original three-year analysis. And on November 2nd of 2026, the company announced that the FDA had accepted the application for priority review, and they would make a decision in May of 2027.


Okay. So, they accepted it after the four-year results were public. Right.


So, they're willing to look at the application. Of course, that doesn't guarantee that they're going to approve anything. By November, the company also reported that they were on track to do a larger randomized trial, about 200 patients, randomized two to one to gene therapy or standard of care, not the sham surgery.


And they changed their primary outcome. They said they would make the primary outcome now everyday functioning at three years, rather than the composite score. They chose a different primary outcome.


I'm guessing that outcome was looking better in their existing data. Right.


[Regina] (59:56 - 1:00:03)
So, okay. If I'm hearing all of this correctly, the drug may get approved before we have those results.


[Kristin] (1:00:03 - 1:00:31)
It's possible. Yeah. Regina, that trial won't be done certainly by May of 2027.


And so, the FDA is going to have to weigh a whole bunch of trade-offs, right? Remember, there were more serious adverse events in the high-dose group than either the sham surgery or low-dose group. So, the treatment may have some risks, right?


And so, if there's not a lot of good evidence that it works and it has some risks, are we giving people an expensive brain surgery for nothing, right? This is a really, really tough one.


[Regina] (1:00:31 - 1:00:36)
We might need to do a follow-up episode on this. We probably are going to have to follow up on this.


[Kristin] (1:00:36 - 1:00:36)
Yes.


[Regina] (1:00:36 - 1:00:48)
All right, Kristin, I feel like we're up to present day times. So, are we ready to wrap up this episode and rate the strength of evidence for the claim? Yes.


And the claim was?


[Kristin] (1:00:48 - 1:00:54)
The claim is the claim of the company that AMT-130 slows the progression of Huntington's disease.


[Regina] (1:00:54 - 1:01:19)
Okay. And this is rated using our smooch rating scale. Very, very scientific.


We need to get around to trademarking this at some point, Kristin, I'm sure. We're registering it or whatever one does, patenting it. One to five smooches, one means little to no evidence for the claim.


Five smooches means lots of evidence for the claim. So, Kristin, you go first. Consider just it.


What are you doing?


[Kristin] (1:01:19 - 1:01:35)
I am not going to diss it completely, but I'm going to go two smooches. So, I think there is maybe a little hint of some evidence here, but it's not very robust. It's not very strong.


And I may even be giving two smooches just because I feel for the patients. What about you, Regina?


[Regina] (1:01:36 - 1:01:55)
That's interesting. I am going with two smooches as well, much for the same reason. I feel like don't have enough evidence to say it does work, but we also don't have enough evidence to say it does not work.


So, the two smooches is my optimism in there.


[Kristin] (1:01:55 - 1:02:07)
At least it didn't seem to go in the other direction. So, it doesn't seem that the treated patients did worse on the outcomes as with that earlier drug, Tominersen. So, maybe it's at least not worse.


[Regina] (1:02:08 - 1:02:16)
Right, right. Okay. So, methodological morals, there are lots of interesting things in here.


What do you have ready for us?


[Kristin] (1:02:16 - 1:02:34)
Yeah, I'm going to focus on the fact that no matter what, at the end of the day, there were a total of 17 high-dose patients. And no matter what statistical wizardry you do, you can't really ever totally make up for that. So, mine is 17 patients is still 17 patients, no matter how you cut it and no matter how many controls you have.


[Regina] (1:02:35 - 1:02:55)
Oh, I like that one, Kristin. Very nice. Very nice.


What about you, Regina? I'm going to focus on the compared to what, because that's my tattoo. So, how about this?


With external controls, the question is always, are we seeing the treatment or are we seeing the difference between the studies? Yeah, good point.


[Kristin] (1:02:55 - 1:03:04)
Regina, I just want to remind you that we teased at the beginning, the latest blog post from Data Colada, and it's pretty interesting. So, I want to make sure we don't forget to add that in in the end here.


[Regina] (1:03:04 - 1:03:37)
Okay, let's have at it. You said this was called Fold Me Once. What a great name for a blog post.


So, here's a question. Data Colada published this blog post suspiciously immediately after the results from the court case were announced. You think they were waiting, sitting on this, waiting to publish it until the lawsuit was settled to avoid, oh, I don't know, being sued again for $25 million?


I wondered that too.


[Kristin] (1:03:37 - 1:05:01)
I mean, I'm not sure, but yeah, the timing is awfully coincidental, so I do wonder if they held off publishing this until they knew they were no longer in legal jeopardy. Yeah. Okay, Kristin, give us a summary then of the blog post.


So, this was about a 2011 study that had three coauthors, and one of those coauthors sent Data Colada two different versions of the data set for this study. An earlier data file that he had received from a research assistant in March of 2010, good that he held onto these, and a later data set that he had received directly from Francesca Gino in April of 2010. And the April data set from Gino is the one that was used as the basis for the published paper.


The study had about 160 participants, and the outcome was the number of t-shirts they folded. And when Data Colada compared the two data sets, they found that 29 participants had different values for this primary outcome variable, the number of t-shirts folded, in the Gino data set compared with the earlier data set. And these changes did not appear to be random because all but one was changed in a way that favored the study's hypothesis.


Also, the differences between the values in the two data sets were large and they were round. So, for example, Regina, for a whole bunch, participants had exactly 30 or 20 or 10 more t-shirts folded in the Gino spreadsheet compared with the earlier spreadsheet.


[Regina] (1:05:01 - 1:05:11)
I feel like this is just a greatest hit compilation of all of the things that one should not do when manipulating or making up data.


[Kristin] (1:05:13 - 1:05:26)
I want to go back to Lessig's podcast where they said, surely a Harvard professor would be too smart to fake data in such a dumb way. I think we might have disproven them with this one.


[Regina] (1:05:26 - 1:05:36)
Okay. So, you said for a whole bunch of these changed data points, they had exactly 30 more folded t-shirts or 20 or 10.


[Kristin] (1:05:36 - 1:05:37)
Yes. A lot of 30s.


[Regina] (1:05:37 - 1:05:51)
Right. And all but one were now in favor of the study hypothesis. So, this is highly unlikely to be explained by, oops, someone accidentally omitting some values.


[Kristin] (1:05:51 - 1:06:03)
Or entered the wrong values and had to correct them later. No, this is really suspicious. And Regina, without these 29 changes, totally not statistically significant.


The p-value went from 0.01 to a p-value of 0.65. Oh, come on.


[Regina] (1:06:04 - 1:06:05)
This sounds very familiar.


[Kristin] (1:06:05 - 1:06:25)
I am waiting though for Lawrence Lessig to do a 10th episode of his podcast. I really want to hear the explanation, the convoluted explanation they're going to come up with to explain how this happened accidentally from the Army of Research Assistants. The story would be wild.


And I almost really want to hear that, Regina.


[Regina] (1:06:26 - 1:06:32)
I mean, it's just like the imagination cannot- The creativity involved in law.


[Kristin] (1:06:32 - 1:06:33)
Wow.


[Regina] (1:06:35 - 1:06:39)
Okay. So, data colloidal came down hard on this, I'm guessing.


[Kristin] (1:06:39 - 1:06:53)
I think they're no longer afraid of being sued because they say, we can't imagine a plausibly benign explanation for this pattern. These data were altered to achieve the hypothesized result.


[Regina] (1:06:53 - 1:06:56)
Oh, wow. They went there. They just said it.


They did not mince words.


[Kristin] (1:06:57 - 1:07:07)
Yeah. They say that this is, quote, about as cut and dried as it gets, and I tend to agree.


[Regina] (1:07:07 - 1:07:12)
What happens now? Are people just saying, oh, wow, that's interesting? Or are they going to follow up on it?


[Kristin] (1:07:12 - 1:07:31)
Well, I think there has been this debate with people defending her. It's going to become harder and harder to defend her, I think. And Data Colada is now calling for a broader audit of Gino's research.


They end their blog by saying, if there are five fraudulent papers, then there are more than five.


[Regina] (1:07:31 - 1:07:39)
Wow. That is interesting. There may be more to come in the story.


Andrew Gelman, statistician, also with the blog.


[Kristin] (1:07:40 - 1:07:41)
Oh, I haven't seen that one, Regina.


[Regina] (1:07:41 - 1:07:57)
Yeah. He said, I have a simpler way of dealing with this. Just treat any results from Francesca Gino as being suspect.


Don't trust them. So he's bringing his prior probability in there. Yes.


[Kristin] (1:07:58 - 1:08:07)
I think in all this, though, our take on the data that we presented in the last two episodes, let's say that I think it's been validated by these updates.


[Regina] (1:08:08 - 1:08:16)
It has. It's sad. It's depressing, but at least it's consistent.


This is a consistent story here.


[Kristin] (1:08:16 - 1:08:17)
Yeah, absolutely.


[Regina] (1:08:17 - 1:08:32)
All right, Kristin, this entire thing has been very, very interesting. I'm so glad that David Rind brought this to our attention. And you did a lot of statistical sleuthing on this, and I think you paint a very clear and nuanced picture of the evidence.


[Kristin] (1:08:32 - 1:08:34)
Thanks, Regina, and thanks, everyone, for listening.