Skip to content

Neuroinflammation and glia

TL;DR — Genetics forced immunity into the centre of AD: TREM2 R47H carries OR ~2.9 for AD (Jonsson 2013, PMID 23150908; Guerreiro 2013, PMID 23150934), and pathway analyses of the two largest GWAS implicate immunity and microglia alongside amyloid, tau and lipid processing (Kunkle 2019, PMID 30820047; Bellenguez 2022, PMID 35379992). Mechanistically, complement-tagged synapses are engulfed by microglia before overt plaque deposition in mouse models, and blocking C1q, C3 or CR3 reduces early synapse loss (Hong 2016, PMID 27033548); activated microglia in turn induce a neurotoxic reactive astrocyte state via IL-1α, TNF and C1q (Liddelow 2017, PMID 28099414). But the direction of effect in humans is stage- and context-dependent rather than uniformly harmful: higher baseline TSPO-ligand binding predicted slower decline, while a rising signal over two years predicted faster decline (Hamelin 2018, PMID 29608645), and higher CSF sTREM2 was associated with slower hippocampal atrophy and cognitive decline independent of p-tau181 (Stephenson 2026, PMID 41605308). Single-nucleus atlases show human glial responses that differ substantially from mouse models and are attenuated in TREM2 risk-variant carriers (Zhou 2020, PMID 31932797). The one large randomised test of blanket anti-inflammatory prevention failed: across ADAPT and its ~7-year follow-up, neither celecoxib (HR 1.03, 95% CI 0.72–1.50) nor naproxen (HR 0.92, 0.62–1.35) reduced AD incidence, and naproxen showed weak evidence of cognitive harm (ADAPT 2008, PMID 18474729; ADAPT-FS 2013, PMID 23562431).

Why immunity is not a side story

Three independent lines converge. First, the genetics: rare coding variants in the microglial receptor TREM2 raise risk two- to three-fold; pathway analyses of common-variant GWAS repeatedly recover immunity and microglial gene sets; and fine-mapping confirmed an HLA haplotype (HLA-DR15) as a risk factor (PMID 30820047; PMID 35379992). Second, the tissue: microglia and astrocytes are the cell types whose transcriptomes change most in AD cortex. Third, the imaging: TSPO-ligand PET signal is elevated in prodromal and demented AD versus controls and predicts outcome, though not in the direction a simple "inflammation is bad" model would predict (PMID 29608645).

The unresolved question is whether glial activation is a cause, a consequence, or — most likely on current evidence — a protective response that becomes maladaptive.

Microglial states

Finding Model / population Result
Complement-dependent early synapse pruning 5XFAD and related mouse models C1q is increased and synapse-associated before plaque deposition; inhibiting C1q, C3 or CR3 reduces phagocytic microglia and early synapse loss; C1q is required for soluble Aβ-oligomer toxicity to synapses and hippocampal LTP; adult microglia engulf synaptic material CR3-dependently when exposed to Aβ oligomers (Hong 2016, PMID 27033548)
DAM state is TREM2-dependent in mouse; human response differs Single-nucleus RNA-seq, 5XFAD mice and human AD Trem2-dependent disease-associated microglia confirmed in mouse, plus a novel Serpina3n⁺C4b⁺ reactive oligodendrocyte population; human AD microglial signature instead resembled IRF8-driven reactive microglia seen in peripheral-nerve injury; the reactive microglial phenotype was less evident in TREM2-R47H and R62H carriers than non-carriers (Zhou 2020, PMID 31932797)
Cell-type-specific early changes; sex differences 80,660 single nuclei, prefrontal cortex, 48 individuals Strongest disease-associated changes appear early and are highly cell-type specific; late-stage upregulated genes are shared across cell types and reflect a global stress response; female cells were over-represented in disease-associated subpopulations; myelination-related processes recurrently perturbed (Mathys 2019, PMID 31042697)
Regional atlas and a resilience programme 1.3 million cells, 6 regions, 283 samples, 48 individuals 76 cell types identified; region-specific vulnerable excitatory and inhibitory neuron populations with Reelin signalling implicated in vulnerability; an astrocyte programme linking choline metabolism and polyamine biosynthesis to preserved late-life cognition — i.e. a glial resilience signature, not only a damage signature (Mathys 2024, PMID 39048816)

The Zhou result is the most important caution in this literature: the mouse DAM programme, on which much of the microglial-therapeutics rationale rests, is not the programme human AD microglia run. Any TREM2-agonist strategy has to be validated against the human signature.

Astrocytes

Activated microglia secreting IL-1α, TNF and C1q are necessary and sufficient to induce "A1" reactive astrocytes, which lose the capacity to support neuronal survival, outgrowth, synaptogenesis and phagocytosis and actively kill neurons and oligodendrocytes; blocking A1 formation prevents death of axotomized CNS neurons in vivo, and A1 astrocytes are abundant in human AD, Huntington's, Parkinson's, ALS and MS tissue (Liddelow 2017, PMID 28099414). The binary A1/A2 framing has since been criticised as over-simple, and the human atlas data — which found an astrocyte programme associated with resilience (Mathys 2024, PMID 39048816) — show that astrocytic reactivity is not monolithically harmful.

Fluid markers make astrocytic involvement measurable early. In 384 cognitively unimpaired ALFA+ participants, path modelling found plasma GFAP mediated the relationship between CSF Aβ42/40 and amyloid-PET (i.e. between soluble and insoluble Aβ), while CSF YKL-40 partly explained the path from amyloid-PET to p-tau181 and from p-tau181 to neurofilament light (Pelkmans 2024, PMID 37690071). Reactive astrogliosis is therefore positioned inside the pathological cascade at the preclinical stage, not merely downstream of it. Plasma GFAP's diagnostic performance is covered on fluid biomarkers.

The direction-of-effect problem

The clearest human data come from TSPO-ligand PET, and they are two-signed.

In 52 patients (33 prodromal, 19 demented) and 17 controls followed for two years with ¹⁸F-DPA-714 PET (Hamelin 2018, PMID 29608645):

  • Baseline binding was higher in patients than controls.
  • Patients classified as slow decliners had higher initial binding than fast decliners; initial binding correlated negatively with CDR-SB increase, MMSE loss and hippocampal atrophy progression.
  • On repeat scanning, binding rose faster in patients than controls (mean 13.2% per year vs 4.2%; 15.8% prodromal, 8.3% demented), and the increase correlated positively with all three clinical decline measures.
  • High initial binding predicted a small subsequent rise and favourable course; low initial binding predicted the opposite, independent of baseline severity.

The authors interpret this as two microglial activation profiles with different dynamics that differ between patients rather than between disease stages. A complementary study using ¹¹C-PK11195 in 26 patients found that temporo-parietal tau and anterior temporal neuroinflammation together predicted the rate of cognitive decline over three years, and that both PET components displaced MRI atrophy and demographics from the optimal model (Malpetti 2020, PMID 32380523).

The sTREM2 data point the same way as Hamelin's baseline signal: higher CSF sTREM2 was associated with slower hippocampal atrophy and slower cognitive decline independently of p-tau181, with the cognitive benefit slightly larger at higher p-tau181 (Stephenson 2026, PMID 41605308). If sTREM2 indexes TREM2-dependent microglial activation, then more of that particular activation is protective — which is the opposite of the therapeutic instinct to suppress inflammation. The earlier ADNI analysis that established the direction found the same association in 385 people followed a mean of 4 years: in A+T+ AD, higher baseline CSF sTREM2 attenuated subsequent memory and cognitive decline, and a higher sTREM2/p-tau181 ratio predicted slower conversion from cognitively normal to symptomatic stages or from MCI to AD dementia (Ewers 2019, PMID 31462511).

The first large test of TREM2 agonism did not convert that association into a treatment effect. INVOKE-2 randomised 381 participants with early AD 1:1:1:1 to AL002 15, 40 or 60 mg/kg or placebo intravenously every 4 weeks for 48–96 weeks. Target engagement was demonstrated (CSF sTREM2 fell; CSF osteopontin rose). The primary endpoint was missed at every dose: week-96 CDR-SB LS mean difference versus placebo −0.31 (95% CI −1.61 to 0.98) at 15 mg/kg, +0.13 (−1.18 to 1.43) at 40 mg/kg, and −0.17 (−1.49 to 1.15) at 60 mg/kg (all P>0.05). The most frequent treatment-emergent adverse events were MRI changes resembling ARIA (Mummery 2026, PMID 41787076; NCT04592874, completed, n=356 actual). A TREM2 agonist can engage microglia and still produce ARIA-like imaging without slowing clinical decline. Whether those MRI changes reflect microglial clearance of vascular amyloid is a hypothesis the trial does not test; the report describes them only as events resembling ARIA. That is a constraint on OQ-20, not a resolution: the trial tested one antibody, at symptomatic stages, against a CDR-SB primary.

The inflammasome is a second, directionally opposite, immune target. Active caspase-1 is strongly increased in human MCI and AD brain; Nlrp3−/− or Casp1−/− APP/PS1 mice were largely protected from spatial-memory loss, with reduced IL-1β, enhanced Aβ clearance and an M2-skewed microglia phenotype (Heneka 2013, PMID 23254930). That mouse result licenses NLRP3 inhibition as a candidate; it does not license broad anti-inflammatory prevention, which ADAPT already tested.

Caveats that constrain all of this. TSPO ligands bind a target expressed by more than one cell type and are affected by the TSPO rs6971 genotype; the cohorts are small (n=52 and n=26); "microglial activation" is not one thing; and none of these designs can establish causality.

The one large randomised test

ADAPT randomised 2,528 people aged ≥70 with a family history of AD to celecoxib 200 mg twice daily, naproxen sodium 220 mg twice daily, or placebo (1:1:1.5), stopping treatment early for cardiovascular safety concerns.

Outcome Result
Cognitive function during treatment (n=2,117 with follow-up assessment) Lower global summary score over time for naproxen vs placebo (−0.05 SD, P=0.02); lower 3MS scores for both celecoxib (−0.33 points, P=0.04) and naproxen (−0.36 points, P=0.02); differences attenuated when restricted to people without dementia (ADAPT 2008, PMID 18474729)
AD incidence over ~7 years (median <1.5 y of treatment) 161 events total: 48 (6.6%) celecoxib, 43 (6.0%) naproxen, 70 (6.5%) placebo. HR celecoxib vs placebo 1.03 (95% CI 0.72–1.50, P=0.86); naproxen vs placebo 0.92 (0.62–1.35, P=0.66) (ADAPT-FS 2013, PMID 23562431)
Mortality 349 deaths: 110 (15.2%) celecoxib, 96 (13.4%) naproxen, 143 (13.2%) placebo; celecoxib HR 1.15 (0.90–1.48, P=0.27)

The observational literature that motivated ADAPT reported reduced AD risk in NSAID users; the trial did not reproduce it, and produced weak evidence of harm. A later, more specific anti-inflammatory candidate failed on the same logic: minocycline 200 or 400 mg daily for 24 months in 544 people with mild AD (mean sMMSE 26.4) did not change sMMSE decline versus placebo (combined minocycline 4.1 vs 4.3 points; difference 0.1, 95% CI −1.1 to 1.2, P=0.90) or BADLS (treatment effect −0.53, −2.4 to 1.3, P=0.57); 400 mg was poorly tolerated (28.8% completed vs 61.9% on 200 mg and 63.7% on placebo), mainly because of gastrointestinal symptoms, dermatologic adverse effects and dizziness (Howard 2020, PMID 31738372). The lesson generalises beyond NSAIDs: a mechanism that is real in tissue does not license a drug class in a population, and the exposure window in ADAPT (median under 18 months of treatment in people already aged ≥70) may simply have been the wrong one. This is the same inference problem discussed on risk reduction and prevention.

Complement assigns different work to astrocytes and microglia

In tau and combined amyloid–tau mouse models, deleting C1q reduced both astrocytic and microglial synapse engulfment and rescued synapse density. The two glial populations were not redundant: astrocytic lysosomes contained relatively more excitatory-synapse material, microglial lysosomes more inhibitory material, and loss of Trem2 shifted inhibitory-synapse engulfment toward astrocytes around plaques (Dejanovic 2022, PMID 37118504). Nptx2 supplies a candidate brake: it binds C1q, and experimental Nptx2 restoration restrained complement activity and microglial synapse loss in aged tau mice; human CSF evidence came from frontotemporal dementia rather than AD, so translation remains indirect (Zhou 2023, PMID 36989373).

Astrocytosis also has an in-vivo temporal marker. In BioFINDER-2, plasma GFAP was higher in every amyloid-positive group, predicted amyloid-PET positivity with AUC 0.76 versus 0.69 for CSF GFAP, and remained associated with amyloid after controlling for tau; the tau association disappeared after amyloid adjustment (Pereira 2021, PMID 34259835). This supports amyloid-linked early astrocyte activation, but an AUC of 0.76 is not diagnostic performance and GFAP is not AD-specific.

What a coherent model has to accommodate

  1. Complement-mediated synapse loss precedes plaques in mouse models (PMID 27033548), so the immune contribution is early, not a late reaction to established pathology.
  2. Human microglial responses are not the mouse DAM programme (PMID 31932797), so target validation must be human-tissue based.
  3. More TREM2-dependent activation is associated with better outcomes (PMID 41605308; PMID 29608645 baseline signal), while accelerating activation is associated with worse outcomes (PMID 29608645 longitudinal signal). A drug must therefore specify which activation it modulates and in which direction.
  4. Astrocytes carry both a damage programme (PMID 28099414) and a resilience programme (PMID 39048816).
  5. Broad COX inhibition in older adults does not prevent AD and may impair cognition slightly (PMID 18474729; PMID 23562431); tetracycline-class microglial modulation at the mild-AD stage does not slow decline (PMID 31738372); TREM2 agonism engages target and can produce ARIA-like MRI without clinical benefit (PMID 41787076).

Open questions

  • Is TREM2 agonism beneficial in humans, given that INVOKE-2 engaged target without slowing CDR-SB, produced ARIA-like MRI, and was tested only after symptoms (Mummery 2026, PMID 41787076; Zhou 2020, PMID 31932797; Stephenson 2026, PMID 41605308)? Would an earlier, genotype-stratified, or combination design look different?
  • Can the two TSPO dynamic profiles be identified prospectively and used to stratify anti-inflammatory or pro-phagocytic trials (Hamelin 2018, PMID 29608645)?
  • Does complement inhibition preserve synapses in humans, and at what cost to host defence (Hong 2016, PMID 27033548)?
  • Why are female cells over-represented in disease-associated subpopulations, and does this explain any part of the sex difference in prevalence (Mathys 2019, PMID 31042697; see epidemiology)?
  • Can the astrocytic choline/polyamine resilience programme be induced pharmacologically or nutritionally (Mathys 2024, PMID 39048816)?
  • Would NSAID exposure begun decades earlier, or a more selective immunomodulator, produce a different answer than ADAPT (ADAPT-FS 2013, PMID 23562431)?
  • Do plasma GFAP and CSF YKL-40 add prognostic information beyond p-tau217 in preclinical disease, or are they redundant with it (Pelkmans 2024, PMID 37690071)?

References

  1. Jonsson T, et al. Variant of TREM2 associated with the risk of Alzheimer's disease. N Engl J Med. 2013;368:107-16. PMID 23150908.
  2. Guerreiro R, et al. TREM2 variants in Alzheimer's disease. N Engl J Med. 2013;368:117-27. PMID 23150934.
  3. Kunkle BW, et al. Genetic meta-analysis of diagnosed Alzheimer's disease identifies new risk loci and implicates Aβ, tau, immunity and lipid processing. Nat Genet. 2019;51:414-430. PMID 30820047.
  4. Bellenguez C, et al. New insights into the genetic etiology of Alzheimer's disease and related dementias. Nat Genet. 2022;54:412-436. PMID 35379992.
  5. Hong S, et al. Complement and microglia mediate early synapse loss in Alzheimer mouse models. Science. 2016;352:712-716. PMID 27033548.
  6. Liddelow SA, et al. Neurotoxic reactive astrocytes are induced by activated microglia. Nature. 2017;541:481-487. PMID 28099414.
  7. Zhou Y, et al. Human and mouse single-nucleus transcriptomics reveal TREM2-dependent and TREM2-independent cellular responses in Alzheimer's disease. Nat Med. 2020;26:131-142. PMID 31932797.
  8. Mathys H, et al. Single-cell transcriptomic analysis of Alzheimer's disease. Nature. 2019;570:332-337. PMID 31042697.
  9. Mathys H, et al. Single-cell multiregion dissection of Alzheimer's disease. Nature. 2024;632:858-868. PMID 39048816.
  10. Hamelin L, et al. Distinct dynamic profiles of microglial activation are associated with progression of Alzheimer's disease. Brain. 2018;141:1855-1870. PMID 29608645.
  11. Malpetti M, et al. Microglial activation and tau burden predict cognitive decline in Alzheimer's disease. Brain. 2020;143:1588-1602. PMID 32380523.
  12. Stephenson HG, et al. Higher CSF sTREM2 is related to slower hippocampal atrophy and cognitive decline independently of pTau181 levels. Brain Behav Immun. 2026;134:106468. PMID 41605308.
  13. Pelkmans W, et al. Astrocyte biomarkers GFAP and YKL-40 mediate early Alzheimer's disease progression. Alzheimers Dement. 2024;20:483-493. PMID 37690071.
  14. ADAPT Research Group. Cognitive function over time in the Alzheimer's Disease Anti-inflammatory Prevention Trial (ADAPT): results of a randomized, controlled trial of naproxen and celecoxib. Arch Neurol. 2008;65:896-905. PMID 18474729.
  15. Alzheimer's Disease Anti-inflammatory Prevention Trial Research Group. Results of a follow-up study to the randomized Alzheimer's Disease Anti-inflammatory Prevention Trial (ADAPT). Alzheimers Dement. 2013;9:714-23. PMID 23562431.
  16. Ewers M, et al. Increased soluble TREM2 in cerebrospinal fluid is associated with reduced cognitive and clinical decline in Alzheimer's disease. Sci Transl Med. 2019;11. PMID 31462511.
  17. Mummery CJ, et al. The TREM2 agonistic antibody AL002 in early Alzheimer's disease: a phase 2 randomized trial. Nat Med. 2026;32:1708-1716. PMID 41787076.
  18. Heneka MT, et al. NLRP3 is activated in Alzheimer's disease and contributes to pathology in APP/PS1 mice. Nature. 2013;493:674-8. PMID 23254930.
  19. Howard R, et al. Minocycline at 2 different dosages vs placebo for patients with mild Alzheimer disease: a randomized clinical trial. JAMA Neurol. 2020;77:164-174. PMID 31738372.
  20. Dejanovic B, et al. Complement C1q-dependent excitatory and inhibitory synapse elimination by astrocytes and microglia in Alzheimer's disease mouse models. Nat Aging. 2022;2:837-850. PMID 37118504.
  21. Zhou J, et al. The neuronal pentraxin Nptx2 regulates complement activity and restrains microglia-mediated synapse loss in neurodegeneration. Sci Transl Med. 2023;15:eadf0141. PMID 36989373.
  22. Pereira JB, et al. Plasma GFAP is an early marker of amyloid-β but not tau pathology in Alzheimer's disease. Brain. 2021;144:3505-3516. PMID 34259835.