Alzheimer’s
disease

01 · The clinical problemA disease and a lived loss of independence

Alzheimer’s disease is a cause of dementia; dementia describes a loss of thinking abilities that interferes with everyday life. Memory difficulties often bring a person to care, but the question is broader than forgetting names: are they losing the ability to manage familiar tasks, understand information, find their way or communicate? Symptoms need a clinical assessment rather than a diagnosis from a single complaint. NIH disease overview.

Memory-predominant disease is a useful starting point, not a complete patient definition. Our research representation needs to separate the presenting problem, the underlying biological evidence, and the person's ability to function. A change in a cognitive test is not automatically a change in independent living. Likewise, an amyloid-positive scan does not tell us which daily activities a person can perform.

What belongs in the starting record

A meaningful study begins with symptoms and their time course, information from someone who knows the person, cognition and function assessed separately, and the clinical evaluation of other explanations. Medicines, depression and other health problems can produce memory difficulties; their presence should not be overwritten by a protein biomarker. NHS symptom guidance.

The course beyond the first symptoms

As disease advances, difficulty managing complex tasks may give way to increasing confusion, communication problems and dependence for daily activities. Severe disease can involve loss of mobility and difficulty swallowing. Stages describe a pattern of needs; they are not a fixed clock for an individual. The research must follow function and care needs across that course, rather than treating the early-disease trial population as representative of every stage. NIA signs and stages.

02 · Definitions and stagesBiological change is not the same as clinical stage

Definition under debate The 2024 Alzheimer’s Association criteria define Alzheimer’s biologically and distinguish biological from clinical staging. The International Working Group argues for a clinical-biological diagnosis and cautions against labeling asymptomatic people as having the disease on biomarkers alone. This is a real disagreement about terminology and clinical use, not a missing checkbox in our data. Jack et al., 2024; IWG, 2024.

For Luminary, the practical solution is explicit metadata: record the framework used, date of assessment, assay and clinical stage. Do not silently equate “preclinical,” “early,” “mild cognitive impairment” and “mild dementia.” In particular, early symptomatic treatment trials cannot answer the prevention question in people without symptoms. Keeping these populations distinct prevents an attractive treatment result from acquiring a broader claim than the trial tested.

03 · Burden and representativenessCount the right population

The CDC records 116,022 US deaths with Alzheimer’s as the underlying cause in 2024, with a crude rate of 34.1 per 100,000 population. These are mortality measures, not the number of people living with Alzheimer’s, the incidence of a new diagnosis, or all deaths among people with dementia. The year of the underlying data matters as much as the year a page is read. CDC FastStats, 2024 mortality data.

Our disease study should never manufacture a single burden total by combining unlike definitions. A clinical diagnosis count, a biomarker-positive population estimate, a death-certificate count and a trial recruitment pool answer different questions. Geography, age distribution and access to assessment also affect who appears in a dataset. We retain those denominators before using a number to design an experiment or estimate an addressable population.

The evidence gap inside a large trial

A trial can have many participants and still describe a selected group. Entry rules, ability to attend repeated visits, MRI eligibility and availability of a study partner can shape that group. The question for each result is therefore both “was the comparison credible?” and “who does it represent?” Our treatment table preserves the studied disease stage and outcome rather than presenting a population-wide efficacy score.

A harmonized prevalence and access analysis is still open work. This edition does not claim a complete global epidemiology dataset or patient-specific prognosis.

04 · GeneticsInherited susceptibility is not one mechanism

Rare disease-causing variants and common risk variants require different interpretation. APP, PSEN1 and PSEN2 are central to inherited forms; APOE ε4 increases susceptibility but is not a guarantee that a person develops dementia. A family history alone cannot identify the variant or establish its meaning. NIA genetics fact sheet.

Human genetics Bellenguez and colleagues’ large association study identified 75 risk loci and implicated processes beyond amyloid alone, including immune and lipid-related biology. An associated locus does not automatically identify the causal gene, the direction of a useful intervention, or a drug target that can be safely changed. Bellenguez et al., 2022.

Our existing TREM2 record provides a concrete example. The rare R47H variant was associated with higher Alzheimer’s risk, with a reported odds ratio of 2.92 (95% confidence interval 2.09–4.09) in the Icelandic analysis. An odds ratio compares groups; it is not penetrance or a personal probability. Nor does it prove that every TREM2-associated microglial state is harmful. Jonsson et al., 2013 · research L85.

05 · Risk and environmentAssociation, prevention and treatment are different claims

Age, genetics, health and environmental influences contribute to Alzheimer’s risk. Their effects should not be interpreted as a single deterministic pathway or as a person's fault. NIH describes a combination of factors rather than one sufficient lifestyle explanation. NIA frequently asked questions.

For this study, a risk association is a prompt to investigate. A prevention claim requires an intervention comparison in a defined at-risk population, with an outcome and follow-up long enough to answer the question. A treatment claim requires evidence in people who already have the condition being treated. These steps cannot be collapsed by putting all three under a label such as “brain health.”

A practical evidence template

For any proposed factor, record the exposure definition, how and when it was measured, what preceded what, and which confounders were addressed. Then ask whether intervention data exist. If an intervention changes a cognitive composite over two years, retain that finding without rewriting it as prevention of Alzheimer’s dementia decades later. The prevention trials in chapter 16 illustrate this distinction.

Environmental and vascular influences remain part of the larger study even when our current fibril model cannot represent them. We should preserve them as research questions with provenance, rather than inventing hidden effects to make the animation look more comprehensive. “Mapped, not executed” is an honest and useful state.

06 · Amyloid biologyWhich form of amyloid, in which place?

Laboratory mechanism Amyloid is not a single interchangeable object. Our library distinguishes soluble species, oligomers, fibrils and deposited material. Shankar and colleagues isolated soluble amyloid-β dimers from human Alzheimer’s brain and tested effects on synaptic plasticity and memory in experimental systems. The human origin of the material does not turn those assays into a clinical intervention trial. Shankar et al., 2008 · A07.

Human tissue structure Cryo-electron microscopy of brain-derived Aβ42 filaments found structures that differ from some laboratory-assembled fibrils. This matters for our benchmark: reproducing growth on mica is not proof of reproducing the forms present in a person’s brain. Structure, soluble species and growth conditions all need to travel with a measurement. Yang et al., 2022 · L47.

Why length alone is insufficient

Length tells us something about a particular visible object. It cannot, by itself, specify total amyloid mass, the number of particles, soluble concentration, accessible antibody binding sites or toxicity. Two environments with the same total fibril length could differ in their number of fibrils, ends and molecular structures. This is a limitation of the measurement, not a reason to discard it.

Our experimental question is intentionally narrower: can a model reproduce a measured trajectory under the original imaging conditions? A favorable answer would improve one piece of the larger research. An unfavorable answer, as we currently see for parts of S5, identifies a missing or poorly represented process before we make tissue-level claims.

07 · Tau and protein structureA second pathology with its own evidence

Human tissue structure Fitzpatrick and colleagues resolved Alzheimer’s tau filaments by cryo-EM. Their ordered core included residues 306–378, with paired helical and straight filaments sharing a protofilament fold but differing in packing. This is structural evidence; it does not supply a rate of formation, neuronal injury or clinical progression. Fitzpatrick et al., 2017 · A06.

A subsequent study found a common Alzheimer tau fold across additional sporadic and inherited cases. That supports a reproducible structural feature while leaving other tau states and disease contexts open. A common fold should not become an assumption that all patients share identical soluble tau concentrations or response kinetics. Falcon et al., 2018 · A08.

Experimental propagation Our research also includes tau seeding after inoculation of brain material into mice. Such experiments test the capacity of material to induce pathology under a specific intervention; they do not establish ordinary person-to-person transmission or a complete mechanism for the regional progression of human Alzheimer’s. Clavaguera et al., 2013 · L49.

08 · Production and clearanceAccumulation is a balance, not just growth

Human kinetic observation Mawuenyega and colleagues used metabolic labeling to investigate amyloid-β production and clearance in humans. Their late-onset Alzheimer’s group showed impaired clearance, without a group difference in average production in that study. This does not establish that all Alzheimer’s is a clearance-only disorder or identify a single failed clearance route. Mawuenyega et al., 2010 · L46.

The modeling consequence is important. A longer fibril can result from a local growth process, while the amount of amyloid in a tissue also depends on supply, conversion among species, removal and movement between compartments. We cannot use a movie’s elongation rate as a replacement for this mass balance. Concentration, volume and flux must have compatible units and a defined anatomical location.

Mouse physiology Xie and colleagues reported sleep-associated changes in interstitial space and clearance in mice. That is a mechanism worth studying, not a calibrated human Alzheimer’s clearance constant or proof that a sleep intervention reverses the disease. Xie et al., 2013 · L267.

For now, production, transport and clearance are explicit gaps in the fibril sub-study. Before adding them, we need to specify the compartment and molecular species, obtain time-resolved measurements and preserve the original simpler comparison. Otherwise several adjustable rates could fit the same trace without telling us which explanation is correct.

09 · BiomarkersMeasure pathology without confusing it with benefit

Human diagnostic research Our p-tau217 record includes discrimination between Alzheimer’s and other conditions, with an area under the receiver-operating curve of 0.96 in a clinical cohort. AUC is a measure of discrimination across thresholds; “96% accurate for every patient” would be a different, unsupported statement. The assay, population and reference standard matter. Palmqvist et al., 2020 · L50.

In May 2025 the FDA cleared the Lumipulse G pTau217/β-Amyloid 1-42 Plasma Ratio for use in adults aged 55 and older with signs and symptoms of disease, alongside other clinical information. Its clearance does not authorize general screening of healthy people or use as a stand-alone diagnosis. The evaluation included an indeterminate category, which should remain visible rather than be forced into a positive/negative result. FDA clearance announcement.

A future study export should preserve assay identity, units, thresholds, uncertainty and missingness. A change in a biomarker may show that a treatment reached or influenced its target; clinical benefit still needs its own evidence.

10 · Glia and immunityCell state matters more than a good-or-bad label

Mouse mechanism Hong and colleagues connected complement and microglia with early synapse loss in Alzheimer’s mouse models. Blocking components of that pathway reduced loss in those experiments. This supports a testable mechanism, not a universal human complement concentration or proof that broadly suppressing immunity will preserve cognition. Hong et al., 2016 · A04.

Keren-Shaul and colleagues described a disease-associated microglial program with stages differing in TREM2 dependence. The useful lesson is that immune cells change their programs with context; a single “activated microglia” actor hides that diversity. The original work was in a transgenic mouse setting, and a label does not establish a uniform effect in people. Keren-Shaul et al., 2017 · A05.

Our astrocyte records likewise include both an experimental induction mechanism and a later nomenclature consensus. Liddelow and colleagues implicated microglia-derived signals in inducing a particular astrocyte response. The subsequent consensus cautions against treating “A1” and “A2” as fixed, comprehensive identities. We should specify the measured program and function. Liddelow et al., 2017 · L195; Escartin et al., 2021 · L196.

In a future executable experiment, the necessary inputs include cell type, measured state, local abundance, intervention timing and a defined output such as synapse loss. Merely adding more cell shapes would not provide those relationships. Until the interaction is implemented and checked, these actors belong in the research map with the missing numerical references visible.

11 · Vessels and deliveryThe environment includes the route to the target

Human association Our APOE4 blood–brain barrier record links barrier dysfunction and pericyte injury markers with cognitive decline. It points to vascular biology as part of the research landscape rather than a passive backdrop. Association does not identify every causal step or provide a patient's permeability coefficient. Montagne et al., 2020 · V18.

Experimental vascular mechanism Nortley and colleagues studied amyloid-associated capillary constriction involving pericytes, reactive oxygen species and endothelin signaling, including human tissue observations. The findings motivate a mechanism-specific question; they do not establish how often it dominates blood flow in an unselected patient population. Nortley et al., 2019 · V15.

Delivery is another separate question. Our transferrin-receptor shuttle research tests ways to transport biologics into brain. Niewoehner and colleagues reported substantially improved target engagement in mice with a monovalent shuttle design. The reported engagement ratio is not a general multiplier on human tissue drug concentration or efficacy. Niewoehner et al., 2014 · D02.

12 · Vulnerability and progressionFrom a molecular lesion to a failing network

Human post-mortem atlas The 2024 multiregion study in our library profiled about 1.3 million nuclei from 283 samples across six brain regions and 48 individuals. It identified region-specific vulnerable populations and an astrocyte program associated with cognitive resilience. These are rich human observations, but cross-sectional tissue sampling does not follow the same cells through a person’s disease. Mathys et al., 2024 · A02.

The earlier single-nucleus work also showed that disease-associated transcriptional changes depend on cell type and pathological stage. Neither study's recovered cell fractions should be copied directly into a simulated tissue as absolute cell density; sample preparation and analysis influence what is recovered. Mathys et al., 2019 · A01.

Circuit mechanism in mice Verret and colleagues connected inhibitory interneuron dysfunction with network abnormalities and tested Nav1.1 restoration in a transgenic mouse model. This supports investigation of circuit balance. It does not make that manipulation an established human Alzheimer’s treatment. Verret et al., 2012 · L190.

Our next disease model therefore needs an explicit outcome layer. Neuron count, synapse function, network activity and daily function are different variables. We should identify which layer an experiment measures and resist assigning an assumed conversion between them. An attractive animated chain from fibril to memory loss would currently imply a calculation we do not have.

13 · Why treatment is difficultA plausible mechanism can still fail in a trial

Negative human trial In the A4 trial, solanezumab did not slow cognitive or functional decline over 4.5 years in cognitively unimpaired people with elevated amyloid. Its target was monomeric amyloid. This result should be retained alongside positive antibody trials, with the different target, population and design intact. Sperling et al., 2023.

Failure can arise at several different steps: the intervention may not alter the intended target sufficiently, the target may not control the chosen clinical outcome, the timing may be unsuitable, or harm may offset benefit. These are hypotheses to distinguish with measurements, not explanations we can select after a disappointing result without evidence.

Five questions for every proposed therapy

1. What molecular species or cellular process is actually changed?
2. Does enough intervention reach the relevant compartment?
3. What confirms target engagement, and what remains a surrogate?
4. Does a controlled comparison show better clinical function?
5. Are benefit, burden and harm acceptable in the intended population?

Our research workflow should keep unsuccessful programs and discordant results discoverable. A literature collection containing only supportive experiments can make an immature mechanism look settled. The same rule applies internally: the S5 model’s disagreement is a result to preserve, not a curve to hide or tune away without recording the change.

14 · Treatment evidenceSlower decline is valuable—and different from recovery

Randomized human trials Lecanemab and donanemab provide evidence of slower decline in selected people with early symptomatic Alzheimer’s. The comparisons below retain the actual scales and populations. They are not a head-to-head ranking: different measures, enrollment rules and analysis populations prevent that inference.

Sources: van Dyck et al., 2023 · A09; Sims et al., 2023 · A10. CDR-SB is the Clinical Dementia Rating–Sum of Boxes; iADRS is the integrated Alzheimer’s Disease Rating Scale. The donanemab row is the low/medium tau population, not the combined population.

In both rows the treated group still declined on average. The difference is the amount of decline during the trial, not recovery of the lost ability. Relative slowing percentages can be useful within a trial but are less interpretable without the absolute changes, uncertainty and time horizon. These group averages cannot tell an individual exactly how much time or independence treatment will preserve.

For the broader donanemab population, the FDA reports a 2.92-point iADRS difference at week 76. Keeping this separate from the low/medium tau result matters: mixing subgroup and combined-population estimates can create a misleading efficacy claim. FDA donanemab approval summary.

15 · Safety, symptoms and careA treatment result includes burden and harm

Anti-amyloid treatment requires attention to amyloid-related imaging abnormalities, or ARIA: imaging changes that can include swelling and bleeding. FDA strengthened early MRI monitoring for lecanemab in August 2025, adding imaging before the third infusion to the monitoring schedule. Safety information can change after the pivotal trial; the current prescribing information, rather than this study page, governs clinical use. FDA safety communication.

Clinical implementation is not simply choosing a compound. Eligibility assessment, confirmation of pathology, safety review, monitoring access and a response plan for new symptoms are parts of delivering the intervention. Our research comparisons should include these burdens rather than interpreting a statistically significant cognitive result as sufficient by itself.

Symptomatic treatment and supportive care remain distinct from amyloid-targeting treatment. NIA describes cholinesterase inhibitors and memantine among medicines used to manage symptoms, while emphasizing that current treatments do not cure Alzheimer’s. A study about disease mechanisms should still make room for daily functioning, care partner needs and meaningful goals. NIA treatment overview.

Measured harm belongs beside measured benefit

In the CLARITY AD randomized core, ARIA with edema or effusion occurred in 113 of 898 lecanemab recipients (12.6%) and 15 of 897 placebo recipients (1.7%). These are trial-population rates, not an individualized risk estimate; the later open-label extension is a different analysis population. Retaining both denominators and the period prevents safety data from being mixed across study phases. Updated CLARITY AD safety analysis, 2024.

Outcomes worth keeping alongside a biomarker

For future clinical synthesis, we should extract daily function, adverse events, discontinuation, monitoring burden and the experience of the person and care partner whenever reported. Missing outcome data should stay missing. They should not be inferred from plaque reduction, a favorable mechanistic assay or a sponsor’s description of convenience.

Late-stage care raises additional questions about swallowing, comfort and the burden of interventions. These needs remain important even when a person is outside the population studied for disease-modifying treatment. NIA late-stage care guidance. A molecular research program should be able to explain how its chosen endpoint could eventually matter at the bedside, while being honest about the distance still to travel.

16 · Prevention, resilience and reversalBe precise about what improved

Human prevention-oriented trial FINGER randomized 1,260 older adults at risk of cognitive decline to a multidomain intervention or control. The between-group difference in annual change in the neuropsychological test battery score was 0.022 (95% CI 0.002–0.042) over two years. This supports the studied cognitive outcome; it is not proof of reversing established Alzheimer’s or eliminating future dementia. Ngandu et al., 2015.

The US POINTER trial enrolled 2,111 at-risk older adults and compared structured with self-guided multidomain lifestyle interventions. It reported greater improvement in global cognition with the structured program. Because both arms received an intervention and the endpoint was cognition, this should not be restated as a quantified reduction in Alzheimer’s incidence. Baker et al., 2025.

“Resilience” can describe preserved cognition despite pathological burden in an observational study. It is not automatically a manipulable trait. The astrocyte program in our multiregion research provides a candidate association to investigate, not a validated intervention for creating resilience. Similarly, a transient improvement on a test and sustained recovery of lost function require different evidence.

17 · Development pipelineThree different bets, with different tests

This is a selected development map, not an exhaustive or continuously updated pipeline. Trial records can change; the links are provided for protocol and status review. A registered study is evidence of a research program, not evidence that the treatment works.

Early clinical mechanism The initial MAPT-targeting antisense trial demonstrated that tau lowering in the central nervous system was possible. That phase 1b result established a basis for further evaluation, not confirmed clinical efficacy. Mummery et al., 2023.

Sponsor report Biogen’s May and July 2026 CELIA announcements reported biomarker and clinical signals but also stated that the primary dose-response endpoint was not met. This edition records that failure explicitly and does not treat secondary signals as a successful confirmatory trial or an approved treatment. May topline announcement; July conference announcement.

The larger research also includes glial mechanisms, vascular biology and delivery strategies. Those are candidate directions, not a list of clinically validated therapies. A future pipeline comparison should track primary endpoints, analysis plans, negative results, exposure and adverse events—not rank programs by publication count or animation plausibility.

18 · Supporting sub-studyA fibril, measured in time

AFM-S5-01 · Supporting AD-AMYLOID-01

The original S5 movie from Watanabe-Nakayama and colleagues is the starting observation for our runnable comparison. It shows Aβ1–42 fibril growth on mica under liquid using high-speed atomic force microscopy. The parent study asks about Alzheimer’s; this child experiment asks whether our simplified growth model matches this laboratory observation. Source paper, figures and movies · PMC4889376.

Endpoint bounds and pause-compatible windows are visible in the child experiment. They describe sensitivity to the measurement assumptions; they are not population confidence intervals or proof of complete molecular stops. The trace still needs independent review. The original observation stays fixed when model settings change.

Open the supporting fibril experiment →

The experiment’s exported packet retains this parent study, the source, settings, measurements and limitations. It supports model checking and the design of the next experiment. It does not predict cognition, ARIA, survival or a medicine dose.

19 · The next research stepsFill the evidence before extending the scene

These are planned research steps, not running experiments or completed findings. The priority is a chain of understandable comparisons, each with a measurement, a control and a reason it belongs in the larger study.

The research resources already connected

Documented baseline values are a useful start when their population and conditions remain visible. An unknown should be explicit, and a scenario assumption should have its own label. We will not invent a value to make a family look complete or treat a plausible range as patient calibration.

20 · Sources and methodA traceable synthesis, with its limits intact

This edition starts from the Alzheimer’s material already held in Luminary’s family and environment research. It selects records relevant to clinical biology, tissue context and the linked experiment, then adds primary trial publications and official clinical or regulatory sources. Inline links identify the evidence for each scientific claim. Selected current development statements were checked on 23 October 2026; this is not an assurance that every program or publication has been captured.

The existing records include abstract-level screening as well as more detailed extractions. Reusing a record is not a claim that its complete paper, supplementary materials and raw data have all been independently reanalyzed. The bibliography below preserves the original record identifiers and their extraction status. Unresolved citations and unchecked numerical claims were not promoted into model inputs.

Clinical trial values are shown with their endpoint, time horizon, population and reported uncertainty. Mechanistic evidence retains its species and experimental context. The S5 result is identified as Luminary’s measurement/model comparison, separate from the source authors’ published fitted parameters. Sponsor announcements are labeled as such; the failed CELIA primary endpoint is retained.

Existing research used in this synthesis

Terms used throughout

Amyloid-β / Aβ

A peptide studied in soluble and assembled forms; the form and compartment matter.

Fibril

An elongated protein assembly. Its measured length is not itself a measure of cognition.

Tau

A protein with Alzheimer-associated pathological forms; not simulated by the fibril experiment.

Target engagement

Evidence that an intervention interacts with or affects its intended biological target.

Confidence interval

A statistical uncertainty interval under the analysis assumptions; different from the S5 endpoint sensitivity bounds.

Mapped interaction

A research relationship shown for context. It is not executable unless a model actually calculates it.