Reading · Brain Trials · RSS Amplifier

A couple of weeks ago I wrote about CELIA, Biogen’s Phase 2 trial of the anti-tau ASO diranersen, which missed its primary endpoint and had its largest clinical signal at its lowest dose.

A question I got often was some version of: so did the tau hypothesis just fail?

No, it did not. In fact, it’s the opposite: that at least one dose showed evidence of clinical slowing means the tau hypothesis is well alive.

But… the primary endpoint failed right? Yes it did.

Which brings us to the topic of this post: “failed” has different interpretations, at least eight, and the differences among them are everything.

A trial can miss because the biology is wrong. It can miss because the biology is right and the drug never reached the tissue. It can miss because the drug reached the tissue but the dose was too low, or high enough to cause harm. It can miss because the drug worked but the measuring instrument couldn’t see it. It can miss because nothing went wrong at all but chance had its say.

Each points to a different next step: abandon the target, change the molecule, change the dose, change the patients, change the endpoint, or run the same study again. Putting them together into the same “failed” basket throws away useful information the trial might have produced.

Six checks on a negative trial. Checks 1–3 are Morgan's three pillars of survival (Drug Discovery Today, 2012); checks 4–6 are how the clinical trial was built. Fail any one and your negative trial may have not tested the hypothesis it intended, regardless of what it may have taught you.

In 2012 a team at Pfizer, led by Paul Morgan, went back through 44 of their own Phase 2 programs and asked why each had succeeded or failed.¹ The finding was not the failure rate. It was that in almost half of the programs (43%), it was impossible to tell whether the mechanism had been tested at all.

Not just “the drug didn’t work.” But whether the drug had even engaged the target it was supposed to.

They also found something about the programs that did work: every success had demonstrated that the drug engaged its target. None of the failures had.

From this they proposed three things a program should establish before anyone reads a negative result as meaningful, which they called the three pillars of survival:

Exposure: the drug reaches the tissue where the target lives, at adequate concentration, for long enough.
Binding: it engages the target once it gets there.
Pharmacology: engaging the target produces the biological effect it was supposed to produce.

The framework is now 14 years old and has held up. Pfizer reconfirmed it internally in 2020, and it has become standard in brain drug development, where all three pillars are harder to establish than anywhere else in medicine.

The same author later helped build a broader version at AstraZeneca, the 5Rs: the right target, right tissue, right safety, right patient, and right commercial potential, plus what they called the right culture.² Under it, success rates from candidate nomination through completion of Phase 3 rose from 4% in 2005–2010 to 19% in 2012–2016.

Each pillar has a standard measurement in neurodegeneration clinical trials, and each has a catch.

Exposure is usually cerebrospinal fluid (CSF) drug concentration, or its ratio to blood, which tells us how much of a given drug crosses into the CNS. Of course, for intrathecal or intracerebral administration, this says very little: we know that the drug gets into the CSF. The tricky thing is to determine if the drug gets into the relevant brain area. Those programs would need imaging of labeled drug, or an honest admission that distribution into specific brain tissue is unknown.

Binding is cleanest when the target is a receptor, because we can use PET occupancy studies to measure it directly. Nowadays, most neurodegeneration targets aren’t receptors, so engagement has to be inferred from the nearest biochemical consequence. Which creates a problem: in neurodegeneration, binding and pharmacology are often the same measurement. Tominersen’s CSF huntingtin protein levels, or total tau CSF levels for MAPT-targeted therapies, serve as both.

Pharmacology means the biology moved: tau, alpha-synuclein, or huntingtin falling in CSF, or plaque and tangle burden falling on PET. Clearing this, however, does not guarantee that the clinically-relevant endpoints will move. We have many examples for this. For instance, tominersen, Roche’s anti-HTT ASO, lowered both mutant huntingtin and neurofilament light in CSF, and still produced no clinical benefit in HD patients.⁵ Verubecestat, Merck’s BACE1 inhibitor, lowered Aβ42 by up to 90%, and patients got worse.⁹

There is a third possibility this pillar tends to hide: the biomarker moves, but in the direction opposite to the one the mechanism predicts. Roche’s NLRP3 inhibitor selnoflast suppressed its cytokine targets convincingly — stimulated IL-1β down about 91% against placebo — and then raised the TSPO-PET neuroinflammation signal by roughly 41% in the midbrain, the region an anti-inflammatory drug should be quieting.¹⁷ That was presented as a pharmacodynamic effect supporting further study. I’ve written separately about this pattern, which I’d call hypothesis flipping. A wrong-way biomarker is not a nuance to be explained. It means the pillar was not cleared.

A problem in many brain trials is that there is no good biomarker of target engagement. The temptation here is to skip ahead to clinical endpoints. “If we cannot measure a biomarker, let’s just go ahead and measure clinical endpoints.” That’s precisely what the 43% of Pfizer trials in Morgan’s paper did. So the answer is to build the biomarker first (slow, expensive, and likely to benefit your competitors as well), or to proceed with the clinical endpoint: you may get lucky and hit it (particularly if you expect to see movement in a few weeks, in fact a lot of symptomatic drugs were approved this way over the decades), but if you get a negative result, you will not know what failed.

So here is the sorting I actually use. It’s just two questions, not a list.

Six things have to be true. Morgan’s three pillars cover the pharmacology. Three more cover the trial itself.

Did the drug get where it’s needed? This is the hardest pillar in brain disease and the one most often assumed rather than shown. Antibodies against intracellular targets are one example. First, antibodies given intravenously reach CSF at a fraction of a percent of their blood concentration. Not only is their CNS concentration low, but they are unable to reach the intracellular space, where relevant protein aggregates occur. Cinpanemab, an antibody against alpha-synuclein, was negative across essentially every endpoint in Parkinson’s disease.⁷ Like other antibodies, it targeted alpha-synuclein outside cells. A drug that never reaches the compartment where the pathology sits has not tested whether that pathology matters. We’ll see how the prasinezumab Phase 3 trial turns out.

Did it engage the target? UniQure's AMT-130, an intracerebral AAV therapy for Huntington disease, did not show reductions in mutant huntingtin in CSF — the entire mechanistic point of the drug.⁶ UniQure argued that local delivery might not register in CSF, yet its own animal studies (minipig) showed exactly that reduction. A trial can report some inconclusive, potential benefit and still leave you unable to say whether the mechanism was engaged.

Did engagement do anything? Gantenerumab's twin Phase 3 trials enrolled 1,965 people and produced CDR-SB differences of 0.31 and 0.19, neither significant.⁸ Roche also disclosed that amyloid removal came in lower than expected, at roughly half of projection. Turbocharging gantenerumab by conjugating it to a transferrin-receptor shuttle — trontinemab — did result in dramatically more robust amyloid plaque clearance: at the higher dose, 92% of participants dropped below the amyloid positivity threshold within 28 weeks, with ARIA-E under 5%.¹⁶ Two identical Phase 3 trials are now underway, and it would be surprising if the CDR-SB difference weren't more robust than that with gantenerumab.

Did the patients have the disease? Clinical diagnosis of Alzheimer's is wrong against autopsy in 25 to 30% of cases.¹⁰ In one research cohort, about a quarter of patients with clinically diagnosed early-onset Alzheimer's had no amyloid on PET. For 20 years, anti-amyloid drugs were given to trial populations recruited on clinical grounds, meaning a substantial fraction of participants had no amyloid plaque to remove. We have, unfortunately, too many examples for this.

Could the clinical endpoint see the effect? Some neurodegenerative disorders are slowly progressive. Parkinson's disease, HD, some SCAs — you need several years to see progression on current clinical scales.⁴ In Parkinson's disease, MDS-UPDRS parts 1 and 2 barely move after 1–2 years, and part 3 moves a little more.¹¹ But the harder question is whether that movement is clinically meaningful and reflects how the patient feels and functions, rather than what the physician observes. This is a complicated question that deserves its own post.

Could the design and statistical plan answer the question? CELIA’s primary endpoint was a dose-response relationship on CDR-SB.¹⁵ But Phase 1b had already shown spinal fluid tau lowering flattening out around 60% across doses, so the pharmacology predicted no dose-response. Everybody is wondering whether they should have powered the study to detect individual differences in each dosing arm versus placebo. The answer is, of course, yes. That would have required more patients (likely double) but then Biogen might have had a “positive Phase 2” instead.

Off to one side sits a 7th question that isn’t about pharmacology or design at all: are the data real? Occasionally the problem is unblinding, misconduct, or images that don’t survive scrutiny. I’ve written about one such cases before.

It’s the one category where the answer is not a better trial.

Only three things left.

The idea is wrong. The target isn’t causal, or modifying it doesn’t change the course of the disease. Cardiology has the cleanest example. Epidemiology linked higher HDL cholesterol to fewer cardiac events, so drugs were built to raise it.¹⁴ Dalcetrapib raised HDL by about 30% and did not reduce events at all. Anacetrapib, far more potent, roughly doubled HDL and did cut major vascular events by 9%, but the benefit was due to lowering non-HDL cholesterol, not elevating HDL, and it was never approved. Potent engagement, repeated, adequately powered, nothing at the other end. That is what a refuted hypothesis looks like.

Is there any way to know in advance whether a target is worth the attempt? Partly. Drug mechanisms with human genetic support are roughly 2.6 times more likely to reach approval.³ Tau has that support: mutations in MAPT cause inherited dementia. The tau hypothesis therefore enters this discussion with a real prior in its favour, which is one more reason not to write it off on the strength of a trial that never tested it.

The effect is real but too small to detect here. Common, and hard to distinguish from the above without a bigger or longer trial. It partially overlaps with the question of whether the clinical instrument used was the right one.

Chance. Aducanumab’s twin Phase 3 trials ran to the same protocol in the same populations.¹³ In EMERGE, the high dose slowed functional decline by 40%. In ENGAGE, the high dose produced a CDR-SB difference of +0.03 — nothing, in the wrong direction. Two identical experiments, two opposite answers. The power of chance.

Huntington’s disease has just supplied the model, and it’s worth pausing on because it’s so rare.

Roche’s tominersen, an anti-HTT ASO given by lumbar puncture, was the first drug to lower mutant huntingtin in humans. Its Phase 3 was halted in 2021, with patients on the more frequent schedule doing worse than placebo. Roche then analyzed the wreckage and made a specific claim: lower exposure in younger patients with less disease burden might help, where higher exposure in more advanced patients had harmed. It built a trial to test exactly that, in adults aged 25 to 50 with early disease, at a lower dose given less often: 100 mg three times a year, against 120 mg every eight weeks in the halted study.

Top-line results came in this month. The drug was well tolerated. It lowered mutant huntingtin, and it lowered neurofilament light. And Roche discontinued the program.⁵

Every one of the four tests was passed. The explanation was named in advance, it predicted something specific, the next trial was built to test it, and there was a stopping rule the company honored.

It also leaves huntingtin lowering in an uncomfortable place. Tominersen is the only program in this space that has clearly shown both target engagement and downstream pharmacology, and with a corrected dose in a better-chosen population it produced no clinical benefit. Whether the drug ever reached the striatum in adequate amounts is still not established, so the first pillar remains open. But nobody should claim the hypothesis has been rescued by another route until a competing program demonstrates the engagement this one did.

When a trial misses, “failed” is the beginning of a finding, not the finding. Ask first whether the experiment happened: did the drug get there, engage the target, do something, in the right patients, measured by an instrument that could see it, in a design that could answer the question. Only if all of that holds does the result tell you anything about the biology.

Then ask who named the explanation, and when.

That changes how a press release reads. A company saying it didn’t meet the primary endpoint but saw encouraging trends has told you nothing, on purpose, because vagueness preserves every option. A company saying it achieved 40% of target exposure and is advancing at triple the dose has told you which pillar failed, made a prediction, and given you something to hold it to.

One of those is a vague statement. The other is a scientific hypothesis. Only the second is worth anything, to a patient deciding whether to enroll, an investor deciding whether to stay, or a field deciding whether twenty years of negative trials meant what everyone said they meant.

Two lists I'd like to build in the comments: your experience on negative trials that genuinely taught the field something, and negative trials that taught us nothing because the experiment never really happened. Please, let me know in the comments:

Anatomy of a Failed Trial

Comments