Longevity Bottlenecks: Part I — Dementia

But there is nothing in biology yet found that indicates the inevitability of death. This suggests to me that it is not at all inevitable, and that it is only a matter of time before the biologists discover what it is that is causing us the trouble and that that terrible universal disease or temporariness of the human’s body will be cured.

Richard P. Feynman

Human life expectancy has seen a significant surge since the beginning of the 20th century. However, the so-called «maximum lifespan» is remarkably stable. Despite significant advancements in medicine, no one has surpassed the record set by Jeanne Calment, who passed away in 1997 at the age of 122 (there are some doubts, though, about the authenticity of the record). Over the past 27 years, life expectancy in France has risen by five years, yet we have not witnessed a single individual reaching the 120-year milestone, let alone surpassing Calment’s record.

But why can’t we extend maximum lifespan? I have a hypothesis.

«Longevity bottlenecks» hypothesis

Imagine we have a hypothetical population of some animals that only die from two diseases (A in red and B in blue). Mortality rates for both diseases grow exponentially but at different rates. Look at these curves and try to remember the difference: 

Now let’s imagine that scientists discovered how to completely cure disease A (red curve). How would it affect the survival? Let’s have a look: 

Here in black we have the survival curve (a simple guide how to read survival curves) of the original population dying from both causes. We can see that by the age of 2.5 there’s virtually no one left.

Now let’s have a look at the red curve which depicts the survival of the population after complete elimination of one of the causes of death. We can see that the median lifespan has increased but still almost no one survives more than 2.5 years. This is called «rectangularisation». Now let’s add one more curve:

(you can see all the graphs and formulas here)

Can you guess what the green curve represents? The answer is: it’s the survival curve for a scenario where, instead of curing disease A, scientists cured disease B. The difference is striking! Not only do we see a greater increase in median lifespan, but more importantly, we see an even larger increase in maximum longevity. Why is this the case? If we look at the graph showing these causes of death, we can see that although the death rate for disease A dominates early on, disease B surpasses it later. And its dominance later is substantial due to the exponential nature of death rates. We can confidently say that at older ages, overall mortality is mostly determined by cause B. This explains why the complete elimination of cause A has very little effect on maximum lifespan.

I call these faster-growing causes of death, which are lower at younger ages, «longevity bottlenecks». The «longevity bottlenecks» hypothesis suggests that current medicine does not target these bottlenecks, which is why we observe an increase in average lifespan without any significant progress in maximum longevity.

Let’s look at the graph from Chetty et al:

This looks familiar, doesn’t it? These curves are very similar to an earlier example where the elimination of a slow-growing cause of death led to the rectangularization of the survival curve. Given that wealthier individuals have access to better healthcare, we can hypothesize that even the most advanced medicine today struggles to address certain fast-growing «longevity bottlenecks». But which age-related diseases could be such bottlenecks? Let’s have a look at the graph below: 

n the graph we can see log-transformed mortality rates for men and women and (also log-transformed) incidence rates of dementia and Alzheimer’s disease. And it is quite obvious from the graph that the dementia incidence grows very quickly (while being virtually absent at ages <60) hinting that it could serve as a «longevity bottleneck».

In contrast to many other studies which grouped together all people older than 90 years into a single group Corrada et al. discovered that dementia incidence continues to grow exponentially after the age of 90 years and the incidence doubling time was striking: just 5.5 years reaching 40.7% per year in the 100+-year age group!

What does it mean in practice? Let’s imagine a hypothetical scenario where humanity cured all diseases except dementia. What would be the probability of having dementia by the age of 110 years? For simplicity let’s assume that a person has already reached 90 years and that the incidence plateaus after 100 years (likely a lower bound). Then we plug numbers from the paper above and get the following number:

P(dementia by 110|age=90)=1−e−cumulative hazard≈1−e−(0.125⋅5+0.212⋅5+0.407⋅11)≈1−e−6.16≈99.8%

This is a back-of-the-envelope hazard integration using today’s observed incidence rates; it’s not a precise forecast because competing mortality, selection effects, and changing biology under radical life extension would shift these hazards.

It turns out that even if no other diseases intervene, the vast majority of people will develop dementia—and within a few years, die or lose much of their personality. In practice, that means radical life extension is impossible unless we can prevent or cure dementia. But what exactly is dementia, and can it ultimately be defeated?

Dementia is an umbrella term encompassing various neurodegenerative diseases that lead to cognitive decline and, eventually, death. Clinicians often divide it into familiar “subtypes” — Alzheimer’s disease, vascular dementia, dementia with Lewy bodies, frontotemporal dementia — as if these were cleanly separated boxes. Neuropathology tells a messier story: in late life, dementia is frequently the sum of multiple pathologies accumulating in the same brain, each nudging cognition closer to failure until a threshold is crossed. [ref]

One way to see this is to look not at clinical labels but at what autopsies actually report. In a 30-year retrospective survey of 524 neuropathological dementia diagnoses from Lund (1974–2004), Alzheimer’s disease accounted for 42.0%, vascular dementia for 23.7%, and mixed Alzheimer + vascular pathology for 21.6% of cases [ref]. The authors highlight the crucial implication behind those percentages: cerebrovascular pathology corresponded with the dementia syndrome, “entirely or partly,” in almost half of the demented patients [ref].

This matters because “vascular dementia” can look small in some epidemiological graphs, yet vascular injury still plays a large causal role. Vascular pathology often acts as an accelerant: it lowers the amount of neurodegenerative damage required before symptoms become clinically obvious. Remove the vascular “second hit,” and many cases are pushed to later ages — but the underlying neurodegenerative processes can continue to progress in the background.

If radical life extension is the goal, the uncomfortable next step is that vascular prevention — powerful as it is — cannot be the whole solution. Even if strokes and small-vessel disease were minimized, the brain would still face the age-associated rise of proteinopathies: distinct families of misfolded proteins that accumulate, spread through neural networks, and erode cognition.

Three proteinopathies dominate the late-life dementia landscape often enough to look like a “minimal set” for any serious attempt at dementia eradication:

First, amyloid-β and tau — the defining lesions of Alzheimer’s disease. In the Lund autopsy series, Alzheimer pathology either alone or as part of the Alzheimer+vascular mixed category already represented nearly two thirds of neuropathological dementia diagnoses [ref]. Across community cohorts, Alzheimer-type lesions are even more pervasive when lower thresholds are counted.

Second, TDP-43 pathology in the form now called LATE (limbic-predominant age-related TDP-43 encephalopathy). LATE is the quiet giant of the “oldest old.” The 2019 consensus report emphasizes that LATE neuropathologic change is present in more than one in five — and in some series up to one in two — people past age 80, and that around one in four brains in community autopsy cohorts show enough LATE burden to be associated with discernible cognitive impairment [ref]. Clinically, LATE often presents as an amnestic syndrome that resembles Alzheimer’s, which means a substantial fraction of “Alzheimer-like” dementia in advanced age likely contains an unrecognized TDP-43 component [ref]. Large multi-cohort autopsy analyses reinforce the same qualitative picture: LATE-NC shows up in roughly two out of five very old brain donors overall, and it frequently coexists with Alzheimer pathology [ref]. The consensus authors go further, arguing that the overall public-health impact of LATE is on the same order of magnitude as Alzheimer neuropathologic changes — not because it always replaces Alzheimer’s disease, but because it so often adds to it [ref].

Third, α-synuclein pathology (Lewy body disease), which underlies dementia with Lewy bodies and commonly appears as a comorbidity alongside Alzheimer pathology. Even in cohorts where “pure” DLB is undercounted by referral and diagnostic tradition, Lewy pathology still surfaces as an important additional burden within clinically Alzheimer-like dementia [ref].

Put differently: the question is not whether a single dominant dementia subtype can be prevented, but whether the brain can be protected from a recurring trio of late-life proteinopathies that often co-occur and reinforce one another. The prevalence of mixed pathology means that eliminating only one mechanism is unlikely to “cure dementia” in the strong sense; it would more often delay dementia, reshape its clinical presentation, and leave a residual risk driven by whatever remains [ref].

In that light, dementia starts to look less like one enemy and more like a convergence point: different upstream insults — amyloid/tau, TDP-43/LATE, α-synuclein, plus vascular injury — funnel into the same downstream catastrophe of synaptic loss, network collapse, and eventual failure of the mind. Defeating dementia at the timescales implied by radical life extension would therefore require more than “better cardiovascular health.” It would require reliable ways to prevent, neutralize, or compensate for the dominant proteinopathies of brain aging — especially the ones, like LATE, that are both common and still largely invisible to routine diagnosis during life [ref].

Alzheimer’s disease

On 3 November 1906, an audience of psychiatrists in Tübingen, Germany, were the first to hear the story of Auguste Deter. She had been admitted to Frankfurt Psychiatric Hospital five years earlier with memory loss and other cognitive symptoms. On examination of her brain, physician Alois Alzheimer described two striking lesions: dense deposits outside neurons (what we now call amyloid plaques) and twisted filaments within neurons (neurofibrillary tangles). That pairing—plaque plus tangle—has framed Alzheimer’s disease ever since, and it set up a question that still drives the field: which lesion starts the chain reaction, and which is downstream wreckage [ref]?

The amyloid cascade hypothesis is the most influential answer. In its modern, less cartoonish form, it doesn’t claim amyloid explains everything about dementia. It makes a narrower—and testable—bet: that abnormal accumulation of amyloid-β (Aβ) is an early, upstream event that triggers a sequence of pathologies (synaptic dysfunction, inflammation, tau misfolding/spread, neurodegeneration), and that if you intervene early enough in the Aβ part of the chain, you can slow the clinical trajectory. The hypothesis crystallized in the late 20th century as the molecular identity of plaque material became clearer (for example, the 1984 work identifying Aβ sequences in plaque-related amyloid), and it was explicitly formulated in 1992 by Hardy and Higgins [ref].

Want to learn the history of Alzheimer’s disease research? Here’s the beautiful timeline from Nature!

Why did the idea feel so compelling? Because human genetics—nature’s own randomized experiment—kept turning the Aβ “dial” and moving Alzheimer’s risk with it. Mutations in APP or presenilin genes (PSEN1/PSEN2) can cause autosomal-dominant, early-onset Alzheimer’s; many of these variants shift APP processing toward aggregation-prone Aβ species or otherwise increase amyloidogenic burden, and the disease often begins decades earlier than typical late-onset AD. Even more telling, there are rare protective APP variants, such as A673T (“the Icelandic mutation”), associated with reduced amyloidogenic processing and lower AD risk—an existence proof that less amyloid biology can mean less disease biology [ref].

Mechanistically, amyloid begins with a mundane piece of cell biology: APP is a membrane protein that can be cut by different enzyme “scissors.” In the amyloidogenic route, β-secretase makes the first cut, then γ-secretase (whose catalytic core involves presenilins) makes the second, producing Aβ peptides of varying lengths. The key point isn’t that Aβ exists—brains make some Aβ normally. The key point is that certain forms, especially Aβ42, have a higher tendency to self-associate. Once Aβ concentration and residence time cross a threshold, it can undergo nucleation and seeding: small aggregates form, then recruit more peptide, then assemble into larger structures. That physical chemistry matters because it shifts the disease conversation away from “plaques are junk piles” toward the more mechanistically loaded question: which Aβ assemblies are doing the damage?

A major revision of the original hypothesis is that the most toxic species may often be soluble Aβ oligomers / protofibrils, not the mature plaque itself. These soluble assemblies can impair synaptic plasticity (the cellular basis of learning) and perturb neuronal signaling long before neurons die. Several converging models point to Aβ oligomers interfering with receptor systems involved in excitatory transmission and plasticity, promoting abnormal calcium (Ca²⁺) dynamics, oxidative stress, and spine destabilization—effects that look, in network terms, like a slow erosion of synaptic “bandwidth” before frank neuronal loss. Meanwhile, plaques may act as both a sink and a reservoir: they can reflect the burden of aggregation, but they can also maintain a local environment that sustains a toxic soluble pool and a chronic glial response.

That glial response is not just scenery. Aβ aggregates recruit and activate microglia and astrocytes, pushing innate-immune pathways that can be double-edged: helpful for attempted clearance, harmful when inflammation becomes chronic or misdirected. This is also where the “cascade” becomes a literal cascade: inflammatory signaling, synaptic pruning programs, and metabolic stress can amplify each other.

So why does amyloid accumulate in most people who get Alzheimer’s, given that they don’t carry rare APP or presenilin mutations? The late-onset story is often less about “overproduction” and more about impaired clearance: aging shifts the balance between Aβ generation and the brain’s ability to remove it. The brain’s clearance routes include enzymatic degradation, transport across the blood–brain barrier, cellular uptake by glia, and the glymphatic/paravascular system—a flow of cerebrospinal fluid and interstitial fluid along perivascular spaces that is strongly linked to sleep and vascular pulsatility. In mice, sleep states are associated with markedly increased solute clearance, including Aβ-relevant molecules, providing a mechanistic bridge between “sleep matters” and “waste removal fails” [ref].
Even more directly, an aging-brain study reported that clearance of injected amyloid-β was impaired by ~40% in old mice compared with young, accompanied by reduced vessel wall pulsatility and loss of normal perivascular AQP4 polarization—plausible plumbing-level reasons why Aβ’s residence time increases with age [ref].

Now add one more twist: what if amyloid is not merely “garbage,” but part of an evolved innate immune response—useful acutely, harmful chronically? This is the logic behind the antimicrobial/protective hypothesis. In a 2016 Science Translational Medicine paper, Aβ behaved like an antimicrobial peptide in models, and infection could seed and accelerate amyloid deposition in transgenic systems [ref].
A particularly provocative extension connects this to herpesviruses. In a 2018 study, Aβ oligomers were reported to bind herpesvirus surface glycoproteins, accelerating Aβ deposition while also entrapping viruses—suggesting plaques could form, at least in part, as a byproduct of anti-infective activity [ref].
And the “infection → defense → pathology” story is no longer just about amyloid. A 2025 Nature Neuroscience paper reports that tau becomes hyperphosphorylated in response to viral infection and can directly bind HSV-1 capsids, neutralizing infectivity in human neurons—a startling reframing in which a signature “pathological modification” may also be a form of antiviral biology that goes wrong when sustained [ref].

Epidemiology, predictably, is messier than cell biology—but there are signals that keep these hypotheses alive. A nationwide Taiwanese cohort study reported an association between HSV infection and increased dementia risk (adjusted hazard ratio about 2.56 for dementia in HSV-infected vs non-infected in that analysis) [ref]. A Nature Reviews Neurology commentary discussing these Taiwanese data highlighted an additional striking association: among HSV-infected individuals, anti-herpetic treatment was associated with a much lower dementia risk (reported HR ~0.092), with stronger reduction in those treated longer—still observational, still confoundable, but difficult to ignore [ref]. A meta-analysis in the European Journal of Epidemiology reported a pooled association between periodontal disease and dementia risk: relative risk (RR) 1.38 across high-quality studies, and RR 1.18 within cohort studies specifically [ref]. Mechanistically, Dominy and colleagues (2019) reported P. gingivalis and its proteases (gingipains) in AD brains, with gingipain levels correlating with tau/ubiquitin pathology in their dataset, and showed that oral infection in mice could lead to brain colonization alongside increased Aβ1–42 [ref].

It is important to note that infection findings don’t “replace” amyloid cascade logic; they can be read as candidate upstream triggers that repeatedly provoke an Aβ/tau-linked innate immune program—protective in bursts, damaging when chronic.

If amyloid plaques are the most photogenic feature of Alzheimer’s disease, tau is the one that tends to explain the plot. Even many amyloid-friendly researchers will concede the uncomfortable clinical fact: tau pathology correlates better with neuronal loss and cognitive impairment than amyloid burden does.

Tau is a neuron-enriched, microtubule-associated protein (encoded by MAPT) that normally stabilizes axonal transport “tracks.” In disease, tau undergoes a series of changes—hyperphosphorylation, misfolding, truncation, altered solubility, and mislocalization from axon to soma and dendrites—before assembling into insoluble filaments and ultimately neurofibrillary tangles. What made tau central to the “tau hypothesis” is not just that tangles are common at autopsy; it’s that tau burden is tightly linked in space and time to neurodegeneration and symptoms. Modern biomarker work (tau PET, CSF, plasma p-tau species) consistently supports the clinical intuition that as tau pathology expands beyond the medial temporal lobe into association cortex, cognition and function tend to fall with it [ref].

That correlation is so strong that many researchers view Alzheimer’s as a two-stage biological story: amyloid sets conditions; tau executes decline. In vivo data reinforce that staging model. In that view, Aβ (plus age/inflammation/vascular stress) helps initiate tau mislocalization and aggregation and may accelerate tau spread through vulnerable neural circuits; once tau pathology is established broadly, it can become partially self-propagating and more tightly coupled to neurodegeneration than amyloid is [ref]. PET studies increasingly show that amyloid positivity (and even the estimated duration of being amyloid-positive) helps predict the extent and topography of tau spread across the cortex, often recapitulating classical neuropathological staging patterns [ref].

Mechanistically, the tau hypothesis is also more than “tangles are toxic.” It proposes specific ways tau biology can kill neurons. When tau detaches from microtubules and mislocalizes, axonal transport becomes less reliable; mitochondria and synaptic cargo don’t reach where they’re needed. Mislocalized tau can disrupt synaptic function, alter neuronal excitability, and interact with stress and inflammatory pathways. As insoluble tau assemblies accumulate, they can impair proteostasis (the cell’s protein quality-control systems) and trigger a cascade of cellular stress responses. Importantly, tau pathology looks like a network disease: it tends to appear first in select regions and then progress along anatomically and functionally connected circuits, a pattern that is hard to explain by “global toxicity” alone [ref].

This is where the most conceptually powerful idea enters: tau may spread by templated misfolding, in a “prion-like” manner. In multiple experimental systems, injecting brain extracts containing aggregated tau can induce tau pathology in recipient brains, and the induced pathology can propagate beyond the injection site—consistent with the idea that misfolded tau can act as a seed that converts soluble tau into the same misfolded conformer. The classic demonstration is associated with work such as Clavaguera et al. (2009) and subsequent “tau propagation” experiments and reviews, which explicitly frame tau spread as seed-dependent and self-amplifying [ref, ref, ref]. The “prion-like” label is meant in a mechanistic sense (templated conformational replication and spread), not in the infectious sense of classical prion disease; the important point is that it offers a plausible reason why pathology can start focally yet become widespread over years.

The tau hypothesis becomes even harder to dismiss once it’s remembered that tau pathology can exist without amyloid at all. Alzheimer’s is not the only tauopathy, and tau is not merely a downstream byproduct of amyloid. The cleanest proof is genetic: dominantly inherited MAPT mutations cause frontotemporal dementia with parkinsonism (FTDP-17) and related syndromes, establishing that abnormal tau is sufficient to drive neurodegeneration and dementia in humans [ref]. Beyond MAPT mutation families, there are primary sporadic tauopathies such as progressive supranuclear palsy (PSP) and corticobasal degeneration (CBD), and there is also a very common aging-associated entity, Primary Age-Related Tauopathy (PART), defined by Alzheimer-type neurofibrillary degeneration in medial temporal lobe in the absence of significant β-amyloid pathology. PART matters here because it breaks the simplistic narrative that “tau only follows amyloid.” It shows that tau can arise as a primary process—often limited in topography and typically producing milder cognitive effects than full Alzheimer’s—while still being “real” tau pathology [ref].

So where does that leave the amyloid–tau relationship? A coherent synthesis is that tau has two faces. In aging, tau pathology can begin locally (entorhinal/medial temporal regions) even without amyloid—this is the PART/early-tau end of the spectrum. In Alzheimer’s disease, amyloid pathology appears to be associated with a qualitatively different trajectory: tau moves beyond its usual aging confines and spreads more extensively through neocortex, tracking neurodegeneration and symptoms as it goes. In other words, amyloid may not be necessary for tau to start, but it may be important for tau to go systemic (in brain-network terms) [ref].

That framing also creates room for a provocative—yet logically consistent—initiating hypothesis: what if a single rare cellular event creates the first tau seed? I propose the following hypothesis: a somatic genetic mutation arising in a single neuron (or a small clone of neurons) that alters tau itself (MAPT) or the cellular machinery that controls tau phosphorylation, splicing, clearance, or folding. Neurons accumulate somatic mutations and other genomic alterations across life, and the brain is a mosaic; this is now a mainstream concept in neuroscience, even if the disease relevance is still being mapped [ref]. If tau propagation is seed-amplified, then the bar for an initiating event can be low: it might only take one unlucky neuron producing a particularly aggregation-prone tau conformer to start a process that eventually becomes macroscopic. The strongest version of this idea would predict detectable signatures—e.g., disease-specific somatic variants enriched in tangle-bearing neurons. Early work is beginning to ask related questions at single-cell resolution; for example, a 2025 preprint explicitly tests whether NFT-positive neurons show disease-specific somatic genomic changes [ref]. At the moment, though, it’s fair to describe “single-neuron somatic MAPT mutation as a general initiator of sporadic tauopathy” as plausible but unproven: we have strong evidence for tau seeding/spread and strong evidence for brain somatic mosaicism, but not yet a definitive bridge that says this is the common trigger in humans.

The same conceptual machinery—templated misfolding, seeding, spread, and possibly rare initiating events—doesn’t stop at tau. It is relevant to α-synuclein and TDP-43 as well. For α-synuclein, a substantial experimental literature supports prion-like propagation models, and the idea is often discussed in relation to Braak-style staging of Lewy pathology, even while careful reviews emphasize that definitive proof of intercellular propagation as the driver of human disease progression remains challenging [ref]. For TDP-43, there is likewise a growing body of in vitro and in vivo evidence consistent with seeded aggregation and cell-to-cell transfer, and multiple reviews now explicitly frame TDP-43 proteinopathies through a prion-like lens [ref].

That broader relevance is not just conceptual—it is clinical. Mixed pathologies are common in late-life brains: Alzheimer’s biology frequently overlaps with Lewy body pathology (α-synuclein) and with TDP-43–related syndromes such as LATE/FTLD-TDP, which means real patients can have multiple self-propagating proteinopathies in parallel. In that setting, the most useful role of the tau hypothesis is not as a competitor trying to overthrow amyloid, but as a clarifying statement about where disability comes from: tau burden and spread are among the best biological correlates of neurodegeneration and cognitive decline, while amyloid is often best understood as an early staging marker and upstream accelerator that can shape tau’s trajectory [ref].

Current and future therapies

A decade ago, “Alzheimer’s treatment” mostly meant symptom management: boosting acetylcholine with cholinesterase inhibitors or damping glutamate signaling with memantine—helpful for some patients, but not meaningfully changing the underlying biology. That has changed recently with a somewhat successful trial outcomes of two drugs targeting beta amyloid

Lecanemab and Donanemab

The headline drugs are lecanemab (Leqembi) and donanemab (Kisunla)—both monoclonal antibodies designed to help the brain remove amyloid aggregates. Lecanemab received traditional FDA approval after confirmatory evidence of clinical benefit [ref]. In the phase 3 CLARITY-AD trial (early symptomatic Alzheimer’s: mild cognitive impairment due to AD or mild dementia due to AD, all amyloid-positive), the primary outcome—the CDR–Sum of Boxes—worsened by 1.21 points over 18 months on lecanemab versus 1.66 on placebo (difference −0.45, p<0.001) [ref]. That’s the “27% slowing” figure often quoted, but the raw numbers are worth seeing: the drug doesn’t stop decline; it slightly changes the slope. Biologically, though, it does something very concrete: in an amyloid PET substudy, amyloid burden dropped by about 59 centiloids relative to placebo [ref].

Donanemab (Kisunla), approved by the FDA for Alzheimer’s disease, follows the same general idea—clear amyloid, slow clinical progression but now we have some data about how patients respond to the therapy based on the stage of their disease. In the phase 3 TRAILBLAZER-ALZ 2 trial, participants had early symptomatic AD and were required to have amyloid pathology and tau PET staging. Main takeaways:
— among participants with low-medium levels of tau, donanemab treatment significantly slowed decline by 35% on iADRS and 36% on CDR-SB which is higher than for all participants combined (22% on iADRS and 29% on CDR-SB)
— In the low-medium group: in participants with mild cognitive impairment, donanemab slowed decline by 60% on iADRS and 46% on CDR-SB, while for those with mild dementia due to AD, donanemab slowed decline by 30% on iADRS and 38% on CDR-SB, respectively [ref].

These results give us a clear picture: the earlier we start donanemab — the better. If we see a clinically meaningful (>20%) slowing of AD progression in those patients who already have tau pathology, could we do better initiating therapy even before tau tangles started to form and accumulate (see graph below)?

This is where biomarkers stop being a futuristic add-on and become the entire strategy. Both CLARITY-AD and TRAILBLAZER-ALZ 2 were fundamentally biomarker trials: amyloid had to be present (PET or CSF), and in donanemab’s case tau burden was used to stage the disease [ref]. The logic is brutally simple: amyloid accumulates for years before symptoms; by the time dementia is obvious, tau-driven neurodegeneration may already be dominating and since tau might have prion-like self-propagating properties targeting beta amyloid might help only partially — exactly what we observed in trials . So if an amyloid drug has a window where it helps most, that window is likely earlier than we historically diagnosed AD—and blood biomarkers (plasma Aβ ratios, p-tau species, neurofilament light, GFAP) are being incorporated into the ecosystem precisely to find that window at scale [ref]. The prevention-style trials now underway (for example, lecanemab in preclinical populations) are essentially a bet on this timing argument [ref].

As usual there’s a fly in the ointment: some patients in both lecanemab and donanemab trials have experienced severe side effect of the treatment — ARIA (amyloid-related imaging abnormalities—brain swelling/edema and/or microbleeds). In CLARITY-AD, ARIA-E occurred in 12.6% of participants on lecanemab (and infusion reactions in 26.4%) [ref]. Donanemab is even punchier: ARIA-E occurred in 24% of donanemab-treated participants (52 symptomatic), and the paper reports treatment-related deaths (3 in the donanemab group) [ref]. And unfortunately ARIA hits those patients who most need the treatment — carriers of APOE-ε4.

Therapies based on APOE gene polymorphisms

APOE gene plays a critical role in lipid metabolism and is highly associated with Alzheimer’s disease (AD). It exists in three common alleles: APOE-ε2, APOE-ε3, and APOE-ε4. Since humans have two sets of chromosomes, one can have any combination of two APOE alleles.

1. APOE-ε2: Rare and may have a protective effect against AD.

2. APOE-ε3: The most common allele and considered neutral in AD risk.

3. APOE-ε4: Strongly associated with increased risk for AD. People with one ε4 allele have a 2.5x higher risk, and those with two ε4 alleles have a ~9x higher risk of developing AD (2x and ~6x higher risk of dementia correspondingly) [ref]. Some researchers say that AD associated with two copies of ε4 allele represents a distinct genetic form of Alzheimer’s disease.

Moreover, subjects with the APOE-ε2 allele had 40% lower odds of developing dementia with Lewy bodies, and the onset of the disease was delayed by 4 years. In contrast, carrying one APOE-ε4 allele increased the odds of developing DLB threefold, while having two copies of ε4 was associated with nearly sixfold higher odds.

How can we apply this knowledge? The basic idea is relatively straightforward: we «just» need to neutralize the detrimental effects of the ε4 allele and, ideally, introduce a protective copy of the ε2 allele. Of course, this is easier said than done: a 2024 expert working group laid out a roadmap for therapeutically targeting APOE-ε4 (including lowering ApoE4 protein, altering its structure/lipidation, and shifting isoform balance), but also highlighted a theme that keeps recurring in dementia: biology is tissue- and context-dependent, so “fixing APOE-ε4” is not one intervention, but a menu of interventions with different risks [ref]. The boldest clinical attempt is APOE-ε2 gene therapy for APOE-ε4 homozygotes: NCT03634007 is testing intrathecal delivery of AAVrh.10hAPOE2 (LX1001) to express the protective APOE-ε2 isoform in the CNS [ref]. On the preclinical side, even anti-apoE antibodies have shown the ability to reduce amyloid deposition and—importantly—block aspects of tau seeding/spread in mouse models, suggesting apoE is not just an “amyloid helper” but a broader modulator of proteopathy. [ref]

I propose a different strategy: let’s edit APOE gene and turn ε4 variant into neutral and ε3 variant into protective. Sounds like a fairy tale but it might actually be possible! A group of researchers led by Michael Greicius discovered a rare genetic variant, R251G, which is co-inherited with the high-risk APOE-ε4 variant. This means that every person with the R251G mutation also carries a copy of APOE-ε4. However, unlike most people with APOE ε4, these individuals do not have an increased risk of developing Alzheimer’s. The single amino acid change caused by R251G neutralizes the risk typically associated with APOE-ε4. If we can introduce this mutation into the APOE gene carrying the ε4 allele, we may be able to protect a large number of people from Alzheimer’s disease and other forms of dementia. Moreover, a similar genetic variant named V236E was discovered for the «neutral» ε3 allele. The V236E mutation decreased Alzheimer’s risk by about 60%, offering a similar level of protection as the protective APOE-ε2 variant. Researchers have already demonstrated that in vivo (in mice) neuronal gene editing is feasible and has minimal off-target effects. But to make gene editing therapy feasible in humans our goal is to achieve virtually zero off-target effects to ensure optimal safety. So, we need a better tool than CRISPR/Cas9. And we already have some promising candidates, e. g. Cas-CLOVER that utilises two guide RNAs.

Other drugs targeting beta amyloid

Beyond lecanemab and donanemab, the anti-amyloid pipeline is trying to improve the trade-offs: clear plaque more completely, earlier, with less ARIA, and ideally in easier formats. There are next-generation antibodies (including candidates like remternetug) being developed to push efficacy or speed [ref]. Some groups are using “brain shuttle” designs to get more antibody across the blood–brain barrier (for example Roche’s trontinemab program) [ref].

And there’s renewed interest in small molecules aimed at toxic Aβ oligomers e.g. ALZ-801 [ref]. This is a prodrug of homotaurine, a modified amino acid previously developed under the names tramiprosate. ALZ-801 is converted to homotaurine in vivo, but is more easily absorbed and lasts longer in the blood than tramiprosate. Tramiprosate was reported to inhibit Aβ42 aggregation into toxic oligomers [ref, ref]. ALZ-801 was shown to have similar properties [ref]. Both ALZ-801 and tramiprosate are metabolized to 3-sulfopranpanoic acid (3-SPA), which is normally found in brain and also inhibits Aβ42 aggregation [ref]. A more recent study found that homotaurine activates GABA receptors, and suggests an alternative mechanism of action for ALZ-801 [ref]. After tramiprosate failed in Phase 3, its maker, NeuroChem, marketed it as a nutritional supplement. Years later, a subgroup analysis of the trial data indicated a potential positive effect in participants who carried two copies of APOE-ε4 [ref, ref]. Alzheon licensed ALZ-801 from NeuroChem and is developing it for Alzheimer’s disease. Unsurprisingly, the drug failed to achieve primary endpoint in 3rd phase clinical trial dubbed APOLLOE4, however there’s a glimmer of hope — treatment with valiltramiprosate resulted in a 52% benefit on the ADAS-Cog13 scale for patients with mild cognitive impairment [ref, ref]. This is in alignment with the idea of early treatment however this particular result barely reached statistical significance and if we correct for multiple comparisons the effect disappears.

Lithium — a century-old mood stabilizer — entered the Alzheimer’s conversation through a mix of molecular biology and epidemiology: in the late 1990s, researchers showed that lithium can reduce tau phosphorylation by inhibiting GSK-3, a key tau-related enzyme implicated in Alzheimer pathology [ref]. Around the same time, clinicians began noticing population signals consistent with neuroprotection: longer, continuous lithium exposure in older adults with bipolar disorder was associated with a lower subsequent risk of dementia, and Danish registry studies even linked trace lithium in drinking water to dementia incidence (with the usual caveats about confounding) [ref]. When lithium finally reached Alzheimer’s trials, results were mixed: a short 10-week study at conventional serum targets found no meaningful biomarker or cognitive benefit, while longer, lower-dose approaches looked more promising — a “microdose” 15-month trial reported stabilization on MMSE, and a 2-year randomized trial in amnestic MCI found placebo patients declined as lithium-treated patients stayed relatively stable, with a shift in CSF Aβ1-42 consistent with disease engagement [ref, ref]. Still, a 2025 meta-analysis of six RCTs concluded that evidence for consistent cognitive/functional benefit remains insufficient, despite an overall acceptable safety profile in studied regimens [ref]. The story took a sharp turn in August 2025, when a Nature paper reframed lithium not just as a drug but as a physiological brain factor: lithium was markedly reduced already at the MCI stage, and the authors proposed a mechanism in which amyloid pathology sequesters lithium, creating a local deficiency that helps drive downstream damage; in mouse models, restoring lithium with lithium orotate (described as “amyloid-evading”) reduced plaques and tangles, restored synapses, and improved memory at doses far below standard psychiatric lithium — positioning lithium orotate as a candidate worth testing in rigorous human trials [ref, ref].

Not every “promising” therapy is a drug. A striking outsider is 40 Hz light-and-sound stimulation, often framed as “gamma entrainment” or GENUS, grew out of a simple observation from systems neuroscience: the brain’s fast “gamma” rhythms—important for attention and memory—are often disrupted in Alzheimer’s, so what if you nudge the network back into that tempo? In 2016, an MIT-led team provided the first headline-grabbing proof-of-concept in Alzheimer’s mouse models: driving circuits at 40 Hz (and later using a noninvasive 40-Hz light flicker) reduced amyloid-β and produced striking microglial changes, suggesting that a brain rhythm can recruit an immune-and-neuronal cleanup response [ref]. Human evidence arrived later and remains early but increasingly concrete: in a 2022 PLOS ONE feasibility/pilot report (including a small randomized, sham-controlled 3-month at-home study in mild probable AD), daily 40-Hz light + sound was well tolerated, produced EEG entrainment, and was associated with less ventricular expansion/hippocampal atrophy, stronger default-mode connectivity, better face-name delayed recall, and improved day–night activity rhythmicity versus controls (exploratory outcomes, small sample) [ref]. A larger 6-month double-blind sham-controlled trial from Cognito Therapeutics (OVERTURE; 76 participants with mild-to-moderate AD) likewise found good adherence and no unexpected serious device-related events; the primary clinical endpoint did not separate from sham, but some secondary measures (e.g., MMSE, ADCS-ADL, MRI whole-brain volume) suggested nominal slowing of decline—enough to justify bigger, properly powered trials (and notably, no ARIA on MRI, a contrast to amyloid-antibody safety concerns) [ref]. The longest human follow-up reported so far (an Alzheimer’s & Dementia open-label extension published in 2025) tracked five mild AD participants using daily 40-Hz stimulation for ~2 years; the late-onset subgroup showed better-than-matched expected trajectories on several cognitive measures, and two participants had decreases in plasma p-tau217—intriguing, but still very small-N [ref]. Mechanistically, the most illuminating recent work is a 2024 Nature study arguing that multisensory 40-Hz stimulation can amplify the brain’s waste-clearance “plumbing” (the glymphatic system): increasing arterial pulsatility/vasomotion, improving astrocytic AQP4 polarization, and expanding meningeal lymphatic drainage—ultimately boosting amyloid clearance in AD mouse models, with VIP interneurons implicated as key drivers [ref]. As with any hot idea, it comes with important footnotes: parameters matter, not every preclinical study replicates the same effects, and the field is still waiting for large, definitive clinical outcomes [ref].

Targeting microglia

Microglia used to be framed as the brain’s cleanup crew, but Alzheimer’s genetics and single-cell biology re-cast them as frontline decision-makers: depending on their state, they can either contain pathology or accelerate it. One striking microglial state is lipid-droplet-accumulating microglia (LDAM), described in aging mouse and human brains—cells bloated with lipid droplets that are worse at phagocytosis, crank out more reactive oxygen species, and secrete more inflammatory signals [ref]. What makes this more than a descriptive label is that we now have a plausible biochemical throttle: in a 2025 Immunity study, amyloid-β drove microglia to convert free fatty acids into triglycerides and stockpile lipid droplets via the enzyme DGAT2, and DGAT2 (along with lipid droplets) was elevated near plaques in both AD mouse models and human AD brain tissue [ref]. Even better (from a therapeutic point of view), the same work reported that pharmacologically targeting DGAT2 or promoting its degradation restored microglial Aβ uptake and reduced plaque burden and dystrophic neuronal damage in 5xFAD mice—essentially arguing that “foam-cell microglia” isn’t just a symptom, but a druggable failure mode [ref].

Sargramostim (Leukine) is a rare Alzheimer’s candidate that comes from immunology-first thinking: researchers puzzled over reports that people with rheumatoid arthritis seemed somehow “protected” from Alzheimer’s, and when anti-inflammatory drugs failed to help in trials, they flipped the hypothesis—maybe it isn’t less immunity that matters, but the right kind of immune activation [ref]. In their AD mouse work, the same group found that short-term GM-CSF treatment boosted activated microglia, cut amyloid load by ~50%, increased synaptic area, and restored spatial memory—enough to justify a human test of sargramostim, a recombinant GM-CSF already used clinically to stimulate bone-marrow recovery [ref]. In a small randomized, double-blind, placebo-controlled phase II study (NCT01409915), participants received 15 injections over 3 weeks; the headline result was a rise in MMSE at end of treatment versus baseline and versus placebo, with the between-group effect still present at the 45-day follow-up, alongside “expected” shifts in innate-immune cell markers [ref]. Exploratory blood biomarkers also moved in a direction the authors considered encouraging—plasma Aβ40 up ~10%, total tau down ~24%, and UCH-L1 down ~42% compared to placebo change [ref]. Safety mattered here as much as signal: the study reported no drug-related serious adverse events and no ARIA, although cognitive readouts beyond MMSE were not uniformly positive, underscoring that this was a “promising-but-early” result rather than a definitive win [ref]. Mechanistically, the working idea is that GM-CSF re-tunes innate immunity—expanding and activating myeloid cells and (directly or indirectly) shifting microglia toward better debris handling and repair, which could improve amyloid clearance and reduce downstream neuronal stress [ref, ref]. A longer, higher-stakes test is now underway: a 6-month phase II trial (NCT04902703) in mild-to-moderate AD is evaluating repeated dosing (five times per week) with clinical and biomarker outcomes, currently listed with an end date of November 2026 [ref, ref].

Drugs targeting Tau, TDP-43 and α-synuclein

If amyloid is the match, tau is often portrayed as the wildfire—and the tau-directed treatment landscape reflects that urgency. The most credible “big swing” in tau therapeutics is antisense oligonucleotides that reduce tau production (e.g., BIIB080 / IONIS-MAPTRx), which have shown biomarker effects in early studies and are being pushed forward in trials. [ref] Vaccination strategies exist too: the tau vaccine AADvac1 has reported safety and immunogenicity and is being explored for signals of clinical impact, though the field remains split on whether any of the current tau immunotherapies are targeting the right tau species in the right compartment at the right stage [ref]. A sobering reality is that multiple anti-tau antibodies have produced disappointing results in AD; tau may be the right target but a hard one—biologically diverse, intracellular for much of its life cycle, and entangled with neuron-to-neuron spread mechanisms.

For α-synuclein (relevant to Parkinson’s disease dementia and dementia with Lewy bodies), disease modification has been even more elusive. Antibody approaches continue; Roche has reported signals consistent with a potentially meaningful effect of prasinezumab on motor progression in early Parkinson’s across multiple analyses, while acknowledging uncertainty and the need for confirmation in ongoing/next studies [ref]. If α-synuclein trials eventually succeed, they’ll likely reshape “dementia treatment” in the same way anti-amyloid drugs did: by forcing clinicians to subtype patients biologically, not just clinically.

The situation is even more dire for TDP-43: there are currently no drug candidates at all. So, if you—dear reader—are interested in conducting research in the field of longevity or radical life extension, please consider working on therapies that target TDP-43. This is an exceptionally important target, as it is involved not only in the pathogenesis of LATE but also plays a pivotal role in frontotemporal dementia and amyotrophic lateral sclerosis (ALS, also known as motor neuron disease)—the condition that claimed the life of Stephen Hawking.

While it appears that the vast majority of effective therapies remain a matter for the future, the natural question arises: can we do anything right now? My answer is yes. I have compiled a list of potential interventions—ranked by the strength of available evidence—that may help prevent or delay the onset of dementia.

Intervention to prevent or slow down dementia

Here are interventions with credible causal evidence of lowering incident dementia risk (and when available, Alzheimer’s disease specifically). I’ve grouped them by strength of causal inference and noted typical effect sizes.

Randomized trials (strongest causal evidence)

• Lower blood pressure more intensively (treat hypertension).
  • A large cluster-randomized trial in rural China (≈34k people, 4 years) found ~15% fewer new dementia cases with intensive BP reduction vs. usual care [ref].
  • An individual-patient meta-analysis of 5 double-blind RCTs (n=28,008) reported antihypertensive therapy lowered incident dementia risk (OR 0.87, 95% CI 0.75–0.99) with ~10/4 mmHg BP reduction—Class I evidence [ref].

• Cognitive “speed-of-processing” training (ACTIVE trial).
  In a 10-year randomized trial of 2,802 initially healthy older adults, the speed-of-processing arm showed a ~29% lower incidence of dementia vs. no-contact control (dose–response with more sessions) [ref].

Quasi-experimental / natural-experiment evidence (high causal credibility)

• Shingles (herpes zoster) vaccination.
  • A regression-discontinuity “natural experiment” in Wales (birth-date-based eligibility) found ~20% fewer dementia diagnoses over 7 years among those offered vaccine—strongly suggestive of causality [ref]. These findings are corroborated by a recent Australian study which had a similar design.
  • A Nature Medicine study leveraging the rapid switch from live Zostavax to recombinant Shingrix found the recombinant vaccine is associated with lower dementia risk than the live vaccine, supporting a vaccine-specific effect beyond confounding [ref].
  • An npj Vaccines analysis (US EHR, propensity-matched) also linked AS01-adjuvanted shingles and RSV vaccines to reduced 18-month dementia risk, hypothesizing the adjuvant itself may contribute [ref]

• Post-hoc analysis of data from Wales and Australia revealed that not only shingles vaccine prevents different subtypes of dementia but it also prevents (or delays) mild cognitive impairment and slows the disease course among those already living with dementia [ref].

Randomized evidence for slowing cognitive decline (likely upstream of dementia)

• Treating hearing loss (hearing aids + audiologic rehab).
  The ACHIEVE RCT showed that, in older adults at higher risk, a best-practice hearing intervention halved the rate of cognitive decline over 3 years vs. health-education control. (The trial was powered for cognition, not incident dementia; dementia outcomes are still maturing) [ref].

• Cataract surgery.
  In a long-running cohort with robust adjustment (including marginal structural models), cataract extraction was associated with ~30% lower incident dementia risk, including Alzheimer’s disease dementia specifically. While not randomized, the design goes well beyond simple correlation [ref].

Other signals worth watching (not yet causal)

• Seasonal influenza & pneumococcal vaccines have repeatedly shown lower Alzheimer’s/dementia incidence in large observational datasets (e.g., ~40% lower AD risk after flu vaccination), but residual confounding can’t be ruled out [ref].

• Metabolic drugs (SGLT2 inhibitors / GLP-1 RAs): multiple observational studies and meta-analyses suggest lower dementia risk vs. comparators in type 2 diabetes; RCT confirmation is still lacking [ref].

Conclusion

In this article, I have argued that dementia may be the ultimate limiting factor of maximum human longevity. Because there is still no cure, we should dedicate substantial resources and effort to developing treatments for the underlying proteinopathies—such as tau, TDP-43, and alpha-synuclein.

A major obstacle is that many current models of Alzheimer’s disease and related dementias translate poorly to humans. If we want faster, more reliable progress, we need better translational models. Fortunately, some animals naturally develop conditions closely resembling Alzheimer’s disease, such as bearded capuchin monkeys and even domestic cats.

Although some interventions appear to be effective at reducing dementia risk, this effectiveness is, in many ways, deceptive. Most current interventions merely delay the onset of the disease. Given that dementia incidence roughly doubles every 5.5 years with age, such measures often amount to postponing the inevitable by only a few years. While those additional years can be critically important for some individuals, they are insufficient for achieving radical life extension. To move beyond incremental gains, we must focus on true disease-modifying therapies—or, ideally, near-perfect prevention.

Author: Alexander Fedintsev

© Copyright 2025 RLE Group. All rights reserved.