Skip to main content
Menu

Clinical Trials

NeuroSearch’s Decline

If you looked at the timelines of a clinical trial, you’ll notice that there’s often a surprisingly long gap between when the trial actually ends and when the results of it are ready to announce. If you’ve ever been involved in working up all that data, you’ll know why, but it’s usually not obvious to people outside of medical research why it should take so long. (I know how they’d handle the scene in a movie, were any film to ever take on such a subject – it would look like the Oscars, with someone saying “And the winner is. . .” within the first few seconds after the last patient was worked up).
The Danish company NeuroSearch unfortunately provided everyone with a lesson in why you want to go over your trial data carefully. In February of 2010, they announced positive results in a Phase III trial of a drug (pridopidine, Huntexil) for Huntington’s (a rare event, that), but two months later they had to take it back. This move cratered their stock price, and investor confidence in general, as you’d imagine. Further analysis, which I would guess involved someone sitting in front of a computer screen, tapping keys and slowly turning pale and sweaty, showed that the drug actually hadn’t reached statistical significance after all.
It came down to the varying genetic background in the patients being studied, specifically, the number of CAG repeats. That’s the mutation behind Huntington’s – once you get up to too many of those trinucleotide repeats in the middle of the gene sequence, the resulting protein starts to behave abnormally. Fewer than 36 CAGs, and you should be fine, but a good part of the severity of the disease has to do with how many repeats past that a person might have. NeuroSearch’s trial design was not predicated on such genetic differences, at least not for modeling the primary endpoints. If you took those into account, they reached statistical significance, but if you didn’t, you missed.
That’s unfortunate, but could (in theory) be worse – after all, their efficacy did seem to track with a clinically relevant measure of disease severity. But you’ll have noticed that I’m wording all these sentences in the past tense. The company has announced that they’re closing. It’s all been downhill since that first grim announcement. In early 2011, the FDA rejected their New Drug Application, saying that the company needed to provide more data. By September of that year, they were laying off most of their employees to try to get the resources together for another Phase III trial. In 2012, the company began shopping Huntexil around, as it became clear that they were not going to be able to develop it themselves, and last September, Teva purchased the program.
This is a rough one, because for a few weeks there in 2010, NeuroSearch looked like they had made it. If you want to see the fulcrum, the place about which whole companies pivot, go to clinical trial design. It’s hard to overstate just how important it is.

7 comments on “NeuroSearch’s Decline”

  1. Keitje says:

    The key question is, did the folks at the top cash in their stock options on time. If so, the company was highly successful and this is a working model for neuro-based biotech in general.

  2. gwern says:

    > In February of 2010, they announced positive results in a Phase III trial of a drug (pridopidine, Huntexil) for Huntington’s (a rare event, that), but two months later they had to take it back. This move cratered their stock price, and investor confidence in general, as you’d imagine. Further analysis, which I would guess involved someone sitting in front of a computer screen, tapping keys and slowly turning pale and sweaty, showed that the drug actually hadn’t reached statistical significance after all.
    >
    > It came down to the varying genetic background in the patients being studied, specifically, the number of CAG repeats. That’s the mutation behind Huntington’s – once you get up to too many of those trinucleotide repeats in the middle of the gene sequence, the resulting protein starts to behave abnormally. Fewer than 36 CAGs, and you should be fine, but a good part of the severity of the disease has to do with how many repeats past that a person might have. NeuroSearch’s trial design was not predicated on such genetic differences, at least not for modeling the primary endpoints. If you took those into account, they reached statistical significance, but if you didn’t, you missed.
    I’m not following here… Your description makes it sound like they didn’t take into account the genetic differences, failed to reach significance, correctly re-analyzed with the genetic differences in mind, and reached significance. But the events seem to go the other way. Yet that seems nuts: if the genetics are so important, why would you re-analyze while *not* taking them into account and treat that re-analysis as more correct than the original?!

  3. Bear says:

    “I know how they’d handle the scene in a movie, were any film to ever take on such a subject…”
    Oh, Bog and Ghu… Think Outbreak. I saw that movie with several medical and tech geeks. You can imagine the alternating snarky/outraged/hysterical-laughter commentary. The only thing worse is gunnies watching a blockbuster shoot-em-up (being a gunnie, I’d know).
    Back when I was on an Air Farce Special Maintenance Team, we’d spend a week analyzing/collating/and publishing the data from a 2-3 day tech eval of established comm systems. And that was established, known systems; mostly some computer number-crunching and fill-in-blank forms. Not analyzing new experimental data (or rarely) with a lot of yet-to-be-discovered environmental variables. When we did have to deal with something fairly new, we could spend months getting the data into a form that would convince the powers that be just to release a tech manual update.
    There are days that I’m amazed that new drugs ever make it all the way to the market.

  4. Sok Puppette says:

    I’m just an ignorant layman in this field, but I’m kind of bothered by what this says about presumably common statistical methods, largely unrelated to the case in point.
    First off, what’s with this “reached statistical significance” binary? Are we talking about a difference between p=0.93 and p=0.95, or are we talking about the difference between p=0.99 and p=0.50? It seems as though the thresholds most of these fields use are simultaneously too high (“The data don’t mean anything at all until you’re two sigmas out”) and too low (“We’re going to build long, complicated inference chains from two sigma data sets, by the method of collapsing 0.95 to 1.0”).
    But what really worries me is that the only reason I can actually see for it to take very long to analyze a data set isn’t a very legitimate one. Maybe somebody can explain the giant issue that I’m missing, but…
    It seems to me that the only thing that would take that kind of time would be if you were letting the data inform your analytical methods. And that’s bad. Maybe you can get away with it in some sort of advanced bayesian analysis, but I thought the orthodox view of experimental design (the one that talks about “statistical significance”) was that you were supposed to plan exactly how you would use the data before you analyzed them. If you do enough different analyses, you’re eventually going to find something “significant” in any data set.
    So, what’s really taking the time? Please tell me people aren’t routinely taking all that time trying to find something “significant” in noise…

  5. Esteban says:

    @4: For Phase 3 registration trials, a statistical analysis plan (SAP) is agreed upon with the FDA, so you are correct that there is no wiggle room in terms of the analysis that will be used for FDA approval. That’s not to say a company can’t do some post hoc analyses and present silver linings to the public. This often happens when a compound fails to meet the SAP endpoint, but this is just spin and most people realize that. As for taking so long to release results, there are legitimate and illegitimate reasons. Multi-center, multi-national phase 3 trials are legitimately complicated when it comes to producing the final analysis. Illegitimate reasons would be to hide bad news. This happened a few years back when Merck took an inordinately long time to release the results of the Vytorin/ENHANCE trial, a compound already on the market and bringing in a lot of money. They just paid a big settlement over this to investors a few weeks back.

  6. Sok Puppette says:

    So NeuroSearch’s problem would have been that their SAP was not to take the genetic differences into account, but they made their prelimimary announcements based on an analysis that did take them into account? And both the FDA and the market are using a hard p-value cutoff to judge everything?
    Am I getting that right?

  7. kastellet says:

    I think you got it all wrong. Huntexil is the only drug ever to show effects in HD, and it looks pretty promising. The fact that Teva has bought the drug means they are serious about it. So far HD drug trials have been nothing but failures. The only drug approved is Tetrabenazine (for chorea), and it is a lousy drug causing depression and sedation. And does little to help HD patients. Works more like a sedative. Huntexil has no side effects and also seem to be disease modifying. This is not the end of the story.
    The problem for NS seem to have been a novel drug with a novel moa, and a disease with no efficacy guidelines, no previous evidence based methods and an aggressive neurodegenerative disorder. And the SAP, doses etc had to be predetermined. Some guesswork I suppose. No wonder there are no new CNS drugs out there.

Comments are closed.