Are organoids useful for Alzheimer’s research?

Life update: to celebrate the Public Data for Health program I work on at the OpenAI Foundation officially launching—with $125M+ in issued grants— I am now permanently based in San Francisco! This post, as with all of owlposting.com, remains my own opinions and not my employer's.

This essay is the last of three covering organoids. The full set is:

1. Why haven’t organoids solved all of drug discovery?
2. Where have organoids actually been useful?

1. Introduction
2. How exactly does Alzheimer’s work?
3. The case for (and against) organoids

Introduction

Alzheimer’s is a horrendous disease. You, me, and everyone else can agree on how bad it is for patients, but we should spare a moment of sympathy for Alzheimer’s drug developers too. What a job they’ve got! Imagine being tasked with curing a condition whose roots seemingly appear decades before it actually becomes symptomatic and whose pathologic tissue you cannot meaningfully access during a patient’s life. What hope could they possibly have? If we want to make any progress here, one north star really ought to be making it easier for these people to do their job.

Many people have suggested the use of organoids here. My knee-jerk reaction to such a suggestion is this: if I were asked to think of the worst possible disease to model with organoids, Alzheimer’s would lie at the top of that list. It really seems like a terrible fit. The disease seems to operate at large tissue scales and is partially a result of chronological age. Also, the primary way we measure its progression is cognitive decline! Organoids are fetal, small, and (probably) cannot think. Throw them out the door! Bring in the rats.

But you and I should be careful with instinctive pessimism. Being anti-something in translational medicine is typically only a useful sentiment if there is something else we are pro-. Sadly, there is not a lot left on the table for Alzheimer’s. The two other options for characterizing the disease are studying human brains with natural Alzheimer’s or studying animal brains—the aforementioned rat—with induced Alzheimer’s.

The former is incredibly slow, expensive, and maybe even outright intractable since the true pathogenesis of the condition emerges decades before its first symptoms.

The latter basically doesn’t work at all. As one review paper put it:

Our conclusion is that currently no animal models exist which are predictive for the efficacy of interventions for Alzheimer’s disease.

Given the situation we’re in, how confident are you that organoids don’t model Alzheimer’s well? Are you willing to stake millions of patients’ lives on it? Don’t you think we really should be absolutely positive, without a shadow of a doubt, that this model system has nothing to offer before we give up on it? Beggars really shouldn’t be choosers.

I personally think it is worth giving organoids a fair shake. How much will it help us cure the disease?

Before we get to that, we first need to understand what Alzheimer’s actually is.

How exactly does Alzheimer’s work?

This disease is complicated. If you already understand the pathophysiology of it, you can skip this section, but you may not fully understand the following organoid section. I promise to be as quick as I can here!

The dominant answer for the last three decades for ‘where does Alzheimer’s come from?’ been the amyloid cascade hypothesis, first laid out in an influential 1992 paper, which says roughly the following:

Amyloid precursor protein (APP) gets cleaved first by β-secretase and then by γ-secretase, releasing a peptide fragment called Aβ. Then Aβ accumulates. Then accumulated Aβ triggers tau—also a protein—to detach from microtubules and tangle up inside neurons. Tangled neurons die, and dead neurons make you forget things. Certain phosphorylated tau fragments can also be measured in the bloodstream, which has helped researchers develop blood tests that detect Alzheimer’s-related changes, including before symptoms appear. Some of these tests match cerebrospinal-fluid tests in their ability to identify amyloid pathology! But detecting the pathology doesn’t tell us exactly when someone will develop symptoms or how quickly they’ll decline. More on that later!

So: you need to remember APP, β-secretase, γ-secretase, Aβ, and tau, roughly in that order. Again, this is simplified, but it covers the broad strokes of the disease.

The genetic evidence behind this hypothesis is quite convincing, as genetic evidence often is. Every known mutation that causes early-onset familial Alzheimer’s sits in APP itself or in the proteins that partially make up γ-secretase. People with Down syndrome carry three copies of APP and develop Alzheimer’s pathology at extraordinary rates by their forties. There is also an Icelandic variant, A673T, that sits right next to the β-secretase cleavage site and reduces Aβ production by about forty percent and reduces its carriers’ risk of Alzheimer’s in old age fourfold.

A gene that makes more Aβ gives you the disease; a gene that makes less protects you from it. So amyloid, or some layer between amyloid and the tangles, is causal.

But there may also be pathologies downstream of how your body handles tau.

If you’re even vaguely familiar with Alzheimer’s, you may have heard of APOE4 before, which is the largest known genetic risk factor for the disease. It is not the only one; when you widen out to the full catalog of genetic risk factors—TREM2, CD33, CR1, ABCA7, PLCG2, INPP5D, SPI1—the vast majority of them are predominantly or exclusively expressed in microglia, the brain’s resident immune cells. In other words, sporadic Alzheimer’s is not only a story about amyloid overproduction, but also a story about tau-clearing machinery.

Okay. APP, β-secretase, γ-secretase, Aβ, tau, and the microglial cleanup machinery. This is it. This is (mostly) the full universe of druggable space within this horrible disease. What have people tried to do here? There are roughly seven distinct therapeutic categories, each one poking at a different layer of it.

I’ll list them out in a footnote here, and unfortunately, it is quite long. You should feel free to skip it if you don’t care; it’s not super important for what comes next. But I think it is neat to skim through, if only to understand the toil that awaits drug developers in this space.

But perhaps you’re still a little concerned. You may be thinking the same thing I was: isn’t this particular amyloid cascade hypothesis you’ve outlined just a hypothesis? Also, didn’t that One Guy make it all up? Isn’t it all fraud? As with most things in life, there is an Astral Codex Ten (ACX) essay about this question:

In Defense Of The Amyloid HypothesisRead morea year ago · 280 likes · 188 comments · Scott Alexander

To paraphrase: there indeed was fraud, actually two distinct cases of it. One of them was about which type of Aβ is most worth poking at, and the other was about a particular Alzheimer’s drug. But neither really disproved the amyloid cascade hypothesis, so the cases aren’t really worth discussing in detail.

So is the hypothesis true? For sure?

Let’s say this: the cascade certainly exists. All of the elements I walked through are conclusively part of the disease in some capacity.

What is a little up in the air is whether being aware of the cascade has actually helped us at all. Consider that a third of cognitively normal people over 70 are amyloid-positive on PET scans, and even more bizarrely, tau can exist seemingly without any amyloid present. Finally, recall that the only three successful drug approvals for Alzheimer’s were for drugs that can almost entirely clear amyloid from a living human brain—the PET scans go from lit up to essentially clean— and their value to early-stage Alzheimer’s patients has been so unclear that one has been dropped from the market (aducanumab) and the other two drugs’ (lecanemab + donanemab) have the phrase ‘has no effect in the treatment of Alzheimer's disease’ on their Wikipedia pages. Yes, they don’t literally do nothing, but they come close enough!

That said, the ACX essay offers a few reasons for why, despite the failures thus far, the amyloid hypothesis is still likely correct:

1. Early attempts had suboptimal epitopes which didn’t successfully engage their targets.
2. Later attempts did engage their target, but have mostly been tested ∼15 years after the disease started, and so after downstream tau pathologies had already started.
3. These later attempts demonstrated a ∼30% slowdown in clinical progression, which proves at least some causal role for amyloid. Some claim this isn’t clinically meaningful but they use misleading arguments (like saying it’s a 0.5-point benefit on an 18-point scale, when even a perfect drug which halted all clinical progression could only have achieved a 1.5-point benefit vs. placebo, since the patients start near a perfect score and the placebo group only worsens by about 1.5 points).
4. In all mature antibodies so far, they have been attended with not-great side effects: brain swelling and bleeding, for reasons related to their difficulty crossing the BBB [blood–brain barrier] into brain tissue where they’re actually needed. A new generation of antibodies will cross the BBB, improving efficacy and safety.

One more thing to add here is endotyping: separating patients by the biology driving their disease. Alzheimer’s may look like one disease, but maybe it is actually a constellation of many different ones, each of which presents differently. Perhaps tau without amyloid is a genuinely separate subtype of Alzheimer’s, and perhaps the APOE4 homozygotes really are carrying a distinct genetic disease. And so maybe every Alzheimer’s trial to date has actually been transformative, or would be transformative, for a small subgroup of patients, and the trial simply averaged them into mush.

Maybe! And happily, the author of that essay has a testable prediction on the subject:

I would bet on the following: A therapy whose sole intended mechanism involves amyloid production or clearance, in a randomized, double-blind, placebo-controlled trial, will, in the next 12 years, achieve a slowdown of cognitive decline of at least 75%, with a p-value below 0.001, in its preregistered primary cognitive endpoint (or an average of all such endpoints if more than one exists). I’d eventually expect better than 75% efficacy, but getting stuff to work takes time, and I wanted to make a prediction which can be tested in a reasonable timeframe.

Why am I talking about this? After all, this essay is not about the amyloid hypothesis, but the raw value of organoids to curing the disease. Spoiler: it turns out, the two are quite connected to one another!

The case for (and against) organoids

When you boil Alzheimer’s down to the above components, this whole line of research doesn’t actually seem too bad. All the disease really amounts to is a tale of neurons and microglia, with much of the story going on inside each cell and in the small spaces between cells. Shouldn’t organoids do well at modeling this? Especially because, as a 2014 paper established, the three-dimensional nature of organoids is especially useful for ‘containing’ the Aβ they produce instead of letting it diffuse outwards.

The recent literature seems to confirm our beliefs. A 2025 study combined neurons, astrocytes, and microglia into a single blob, then exposed the organoids to extracts from Alzheimer’s brains. The cultures developed several disease-associated changes, including amyloid and tau aggregates, inflammation, and synapse loss. Bonus: lecanemab—an approved Alzheimer’s drug—reduced the amyloid burden and increased microglial uptake of Aβ. And yes, lecanemab doesn’t actually help patients cognitively much, if at all, but it does reduce amyloid burden!

How about familial Alzheimer’s caused by mutations in APP? Organoids seemingly can recreate this too. A 2026 study cultured familial Alzheimer’s organoids with those sorts of mutations, leading to tau accumulation and synapse loss. And again, compounds that enhanced autophagy improved several of these features. Which is all to say: genetic endotyping is somewhat possible to do via organoids.

There are many papers along these lines, all saying something roughly similar. I’m as surprised as you are! I wasn’t expecting this to be as clean as it is, but here we are. The literature suggests that we should be able to start asking whether an intervention helps an organoid function, as a way to model human responses. And, even more importantly, whether an organoid with a different genetic background responds differently.

Sure, organoids are fetal. This is probably an issue, since a brain with Alzheimer’s is not fetal. But does it matter that much? I suppose it does if you really deeply care about where Alzheimer’s comes-comes from in a platonic ideal sense. But if ‘exposing the organoid to material from an Alzheimer’s brain’ or ‘inducing familial Alzheimer’s mutations in the organoid’ can capture the broad strokes of the condition, aren’t we fine? Especially given that the organoid reacts in an expected way upon being introduced to an amyloid-reducing drug?

Something must be off here. I’ve been neglecting to finish this essay for a very long time because I couldn’t quite figure out what comes next after this sentence. I’ve tried a lot of different ways of coming at it, finding some clever reason why this all isn’t true, and nothing feels quite right. In the end, I think this means something.

I don’t think there is a strong argument againstAlzheimer’s-centric organoids.

Yes, the model certainly does not represent the totality of the disease. But nothing does. Whenever I try to list out the flaws of organoids—fetalness, lack of a peripheral nervous system, and so on—I end up with a list of limitations that every preclinical model in the Alzheimer’s field shares to some degree. And yet, organoids still seem to capture some real behaviors of the disease.

However! One thing I must square with is the fact that, despite organoids being used for a bit over a decade, there still are no particularly useful Alzheimer’s drugs. It is a good model, yes, but it still seems to have led to nothing.

So why do I think the organoids are still useful here? Well, I don’t, not entirely. But what I do think is that we’ll know whether they will be useful pretty soon.

In the prior essay, I went through the cases where organoids were unarguably useful, and the cleanest of them was cystic fibrosis (CF). There, you assemble an organoid using cells from the patient—thereby infusing the organoid with the patient’s genetic background—and dose it with a CF drug. If the organoid responds correctly (swelling up), we can reasonably conclude that the drug will work for the patient, because we’re actually doing the thing with the organoid; the full chain of causality of what proteins are broken in cystic fibrosis and how drugs fix them is contained within it.

In Alzheimer’s, the organoid readout is amyloid and tau, the drugs act on amyloid and tau, and amyloid and tau are not sufficient to predict the clinical outcome—which is whether the patient can still think. The organoid is not at fault! It cannot model cognition, but it does the job we’ve tasked it with quite well, which is to capture a very specific slice of biology the research community has identified as worthy. Really, the failure of organoids is our fault, in that we have not yet identified a phenomenon—beyond amyloid and tau—worth deeply tracking.

Will we ever?

I don’t know. But there is an inflection point on the way!

If you believe the strong version of the amyloid hypothesis, the one expressed in the ACX essay, amyloid clearance is actually an excellent marker of clinical success. The earlier trials haven’t borne that out, but that’s (probably) because patients were dosed too late. Luckily for us, two trials are currently testing that. There’s TRAILBLAZER-ALZ 3, using donanemab, which enrolls adults aged 55–80 without mild cognitive impairment or dementia, selected using tau biomarkers. Its readout is expected in November 2027. And there is AHEAD 3–45, using lecanemab, where participants are cognitively unimpaired adults aged 55–80 with evidence of brain amyloid via scans. Its readout is expected in December 2028.

In other words, we are a couple of years away from knowing how much clinical meaning to attach to one of the main readouts of organoids. From our vantage point today, these model systems have a hazy view of the truth, and it is unclear whether we can trust what they say. But soon, we will arrive to one of two worlds.

In the first world, it turns out clearing amyloid sufficiently early substantially preserves cognition. In this timeline, we could expect a extraordinary renaissance of interest in Alzheimer’s organoids, since we know they can model the amyloid-clearing process decently well. All sorts of wacky things could be tried out upon these model systems, to see how we can reduce neuronal toxicity of existing drugs, or turn on multiple tau/amyloid-cleaning mechanisms, or something else entirely. In many ways, the organoid could become incredibly useful, perhaps even instrumental, to Alzheimer’s research. All without anyone improving the organoid!

In the second world, it turns out that clearing amyloid, even early, still has an unclear payoff for patients. What role will organoids play here? It’s unclear. New knobs of control will need to be discovered for controlling Alzheimer’s progression, and though some may indeed be found within organoids, they will lie in the same sort of no-mans land they are in today, where it is iffy whether one can trust the amyloid-based readout of it. Will better readouts be found? There are good candidates—such as synapse loss, though this cannot be measured noninvasively—but nothing ready for prime-time. In this world, organoids will continue to remain part of the awful toolkit that the knackered Alzheimer’s drug developer wields.

This whole saga may serve to teach us a lesson about organoids at large. No longer should you be asking, “can organoids recapitulate this particular human biology phenotype?”, at least not at first. Rather, you should be deeply curious about whether the phenotype being recapitulated is one that anybody ought to care about. For Alzheimer's, we’ll hopefully learn the answer to that before the decade is up. Stay tuned!

1. APP
   1. APP does a lot of useful things besides giving us Aβ, so lowering it comes with some obvious concerns. But people are trying: Alnylam’s RNAi drug mivelsiran, for example, has reached human trials.
2. β-secretase
   1. This is where we get our first hints of something promising. Lots of companies tried to reduce Aβ production via BACE1 inhibitors—Merck built verubecestat, J&J built atabecestat, and so on. In trials, they generally reduced Aβ by a fair bit, but often worsened cognition.
   2. Why? One proposed explanation is that β-secretase is involved in a lot of other useful things—like synaptic function—so inhibiting it can disrupt the brain in other ways.
3. Drugging γ-secretase
   1. The story splits here into two: inhibition and modulation.
   2. Eli Lilly tried the former with semagacestat, but halted the trials in 2010 after running into serious toxicity, including more infections and skin cancers. Again, γ-secretase is involved in an awful lot of things.
   3. How about the modulation story? This gets into the weeds a bit, but Aβ comes in a whole family of peptides with different lengths. Some of them are less prone to aggregating than others, and you can drug γ-secretase in such a way that more of the less-aggregate-y ones are created. Some earlier programs failed on efficacy or safety, but newer ones are still in play.
4. Preventing Aβ aggregation
   1. This one is simple: stop the peptide from clumping together. It’s not a bad idea, but the problem here is that stopping a protein-protein interaction with a small molecule is not particularly easy. Of course, people have certainly tried, and their efforts have been rewarded with several Phase 3 failures.
   2. That said, it is unclear whether particular drugs failed because of unfavorable pharmacology, because this is simply a bad target, or because only a particular patient subgroup benefits.
5. Clearing Aβ aggregates
   1. If you can’t prevent the aggregation in the first place, perhaps you could just clear it out after the fact. This is the one that (eventually) worked. An early attempt was a vaccine—train the patient’s own immune system to attack Aβ—whose trial was halted in 2002 after about 6% of vaccinated patients developed meningoencephalitis, dangerous inflammation of the brain and its surrounding membranes.
   2. Much of the field turned to simply infusing the antibody directly, though it’s a bit of trial and error to figure out which form of Aβ is worth aiming at—remember, there are many of them, some non-pathogenic! Two approved antibodies ended up targeting different forms: lecanemab binds Aβ aggregates, including soluble protofibrils, and donanemab binds a chemically modified form of Aβ found in deposited plaque. Both dramatically reduce the amyloid visible on a PET scan, and both slow cognitive decline by something like a quarter to a third relative to placebo over roughly 18 months in early-stage patients, depending on the measure and patient group. And both can cause brain swelling and microbleeds.
   3. This is the current state of the art.
6. Boosting clearance machinery
   1. Your body has a lot of gunk it naturally needs to deal with, and Aβ is something it already knows how to clear. So let’s just pump up that cleaning activity! But clearing gunk is a pretty hard thing to regulate, because the same cells and enzymes also do a lot of other useful things.
   2. That said, one drug in this category was Alector and AbbVie’s TREM2 antibody, AL002, which attempted to push microglia towards phagocytosis. Unfortunately, it failed its Phase 2 readout. Perhaps the clearance story here is more complicated than it seems. Consider that there is a drug called sacubitril for heart failure that could theoretically reduce the capacity for Aβ to be cleared, by tamping down one of the enzymes responsible for its clearance. Yet, one study found no extra cognitive decline with sacubitril/valsartan compared with valsartan alone over roughly two years. Weird!
7. APOE
   1. Of course, right? The strongest common genetic risk factor for late-onset Alzheimer’s; surely there is something here. Indeed there is. There are antisense approaches to lower APOE4, small “structure correctors” that aim to push the APOE4 protein into an APOE3-like conformation (which is apparently safer), and an AAV gene therapy that delivers the protective APOE2 allele into the CNS. Nothing approved, but the gene-therapy approach has shown some encouraging early biomarker changes.
   2. This all said: the human gene-therapy data come from a trial that specifically enrolled patients who are APOE4/4 homozygotes, which is 2% of the population and 15% percent of Alzheimer’s patients. Since at least one paper has argued that APOE4 homozygosity defines a distinct genetic form of the disease, it perhaps makes sense that genetic therapies are uniquely useful here.