Microglial checkpoint collapse in Alzheimer’s disease: a tri-axial framework for biomarker-informed neuroimmune therapy

Abstract

Background

Anti-amyloid antibodies have validated amyloid-β (Aβ) as a disease-relevant target in Alzheimer’s disease (AD), but their modest clinical effect, efficacy largely restricted to early disease, and amyloid-related imaging abnormalities (ARIA) indicate that Aβ removal alone does not resolve the glial, lipid, and inflammatory programmes that sustain neurodegeneration. Microglia sit at the centre of this therapeutic gap. Single-nucleus and spatial profiling has resolved several AD-associated microglial states, yet state labels remain descriptive and do not explain why adaptive engagement becomes maladaptive.

Main body

We frame AD-relevant microglial dysfunction as checkpoint collapse: progressive failure of regulatory nodes that coordinate lipid sensing, lysosomal competence, neuronal restraint, and inflammatory threshold control. The central nodes are TREM2-mediated lipid and apolipoprotein sensing, progranulin-associated lysosomal regulation, CX3CR1-dependent neuron–microglia restraint, and CD33/Siglec-3 inhibitory tone. When these controls destabilise, downstream pathology can be organised around three coupled effector axes: a lipid axis centred on APOE-biased cholesterol trafficking, ACSL1/DGAT2-driven lipid-droplet accumulation, and impaired lysosomal flux; an iron/ferroptosis axis involving labile iron, phospholipid peroxidation, and insufficient GPX4/FSP1 defences; and an inflammation/complement axis linking NLRP3 activation, type-I interferon signalling, and C1q/C3-dependent synaptic engulfment to tau pathology and synapse loss. White-matter injury, astrocyte–microglia crosstalk, and cGAS–STING-linked senescence are integrated as cross-axis amplifiers.

Conclusions

This framework is proposed as a hypothesis-generating scaffold for biomarker-informed translational studies, rather than as a validated clinical stratification system. It may help organise stage-aware therapeutic hypotheses, including regulatory-node preservation in early disease, lipid-handling restoration and ferroptosis control at intermediate stages, and complement- or senescence-directed modulation in later disease. Current glial, iron, inflammatory, and imaging biomarkers remain insufficiently specific to assign individual patients reliably to discrete pathological axes in clinical practice.

Background

The diagnostic and therapeutic landscape of Alzheimer’s disease (AD) has changed markedly in the past three years. The 2024 revision of the NIA-AA biological framework defines AD primarily by biomarker-confirmed amyloid-β (Aβ) and tau pathology, while also recognising neuroinflammation, vascular contribution, and α-synuclein co-pathology as biologically informative dimensions [1, 2]. In parallel, anti-amyloid monoclonal antibodies have reached clinical maturity. Lecanemab in CLARITY-AD [3] and donanemab in TRAILBLAZER-ALZ 2 [4] have confirmed that fibrillar Aβ is a tractable and disease-relevant target. Their clinical effects, however, remain modest in absolute terms and are concentrated in early symptomatic disease. Treatment is also accompanied by amyloid-related imaging abnormalities (ARIA), a safety signal whose incidence and severity increase with cumulative drug exposure and APOE ε4 dose [5, 6]. ARIA should not be viewed as an incidental adverse event. Its association with microglial Fc-receptor engagement and APOE-dependent perivascular handling of cleared amyloid implicates the same immune and vascular circuitry that shapes broader AD pathology [7, 8]. Thus, the limitation of Aβ-directed therapy is not that amyloid targeting is biologically misplaced, but that fibrillar Aβ removal alone may leave unresolved the glial, lipid, and inflammatory programmes that continue to sustain neurodegeneration once disease is established.

Microglia occupy the biological space exposed by this limitation. In the steady state, they regulate synaptic remodelling, debris clearance, lipid turnover, and immune surveillance through homeostatic transcriptional programmes shaped by transforming growth factor-β (TGF-β), MEF2C, and related regulatory modules [9–11]. In AD, microglia are chronically exposed to plaques, dystrophic neurites, oxidised lipids, damaged myelin, and inflammatory cues. The resulting cellular states are heterogeneous, regionally patterned, and temporally staged, which explains why the binary M1/M2 vocabulary is no longer adequate for AD microglial biology [12–14]. Single-cell, single-nucleus, and spatial profiling studies have resolved several AD-associated microglial states. These include disease-associated microglia (DAM) and the related microglial neurodegenerative phenotype (MGnD) [15, 16], lipid-droplet–accumulating microglia (LDAM) [17], interferon-responsive microglia, dystrophic states, and white-matter-associated microglia [18–22]. These states distribute differently across Aβ-rich, tau-rich, and white-matter compartments. The 2022 Paolicelli consensus emphasised that such labels describe dynamic states rather than fixed cell types, and that each state should be interpreted in relation to the eliciting stimulus and surrounding tissue context [14]. State annotations are therefore useful but incomplete: they describe what microglia look like at a given time point, but they do not explain why some responses remain adaptive whereas others become self-reinforcing and maladaptive.

This review therefore focuses not only on how AD-associated microglial states are named, but also on how microglia shift from adaptive engagement to maladaptive dysfunction. We use “transition control” to describe the regulatory mechanisms that normally keep microglial responses within a recoverable, tissue-protective range. Microglial homeostasis is not maintained by a single switch. It depends on multiple regulatory nodes that together determine how microglia sense lipid-rich cargo, sustain lysosomal degradation, restrain inflammatory activation, and communicate with neurons. TREM2-mediated sensing of lipids and apolipoproteins supports phagocytic competence and the transition into a sustainable disease-associated response [23, 24]. Progranulin maintains lysosomal acidification and degradative capacity during chronic phagocytic load [25]. CX3CR1–fractalkine signalling provides neuron-derived restraint of microglial reactivity and synaptic engulfment [26]. CD33/Siglec-3 contributes inhibitory tone that limits innate immune activation thresholds [27, 28]. Recent work further suggests that APOE ε4 can engage a TGF-β-linked microglial checkpoint programme that constrains disease-associated responses and modifies tau-driven neurodegeneration [29]. Together, these observations support the view that AD-relevant microglial dysfunction reflects progressive failure of regulatory control, not simply excessive activation.

The contribution of this review is conceptual integration rather than the description of a new pathway. We use “checkpoint collapse” to describe the progressive failure of regulatory nodes that normally maintain adaptive microglial responses under chronic pathological load. The framework is intended to be testable at the level of pharmacodynamics, with each downstream axis linked to candidate biomarker and pharmacodynamic readouts that may support stage-aware therapeutic hypotheses. Single-node disruption can shift microglial thresholds, but symptomatic AD is unlikely to reflect failure of one pathway alone. More plausibly, TREM2-coupled lipid sensing, progranulin-supported lysosomal competence, CX3CR1-mediated neuronal restraint, and CD33-dependent inhibitory tone become compromised in overlapping sequence. Once this regulatory layer weakens, microglia lose the capacity to complete phagocytosis, recycle lipid-rich cargo, restrain synaptic engulfment, and return to an adaptive set-point.

Downstream pathology can then be organised along three interacting effector axes: lipid, iron/ferroptosis, and inflammation/complement. These axes are separated here because each has a distinct evidence base and a different translational readout. The lipid axis centres on APOE-biased cholesterol trafficking, impaired lysosomal flux under chronic phagocytic load, and the emergence of LDAM, particularly in APOE ε4/ε4 brains where ACSL1-positive lipid-laden microglia accumulate near plaques and exert non-cell-autonomous neurotoxicity through tau-associated secretory programmes [17, 30, 31]. The iron/ferroptosis axis reflects expansion of the labile iron pool under phagocytic and ferritinophagic stress, enrichment of polyunsaturated-phospholipid membranes, and insufficient GPX4/FSP1 antioxidant defences. Human-tissue evidence links AD lipid rafts to iron-associated lipid peroxidation and loss of ferroptosis suppressors [32], while microglia-restricted ferroptotic stress can drive neuronal injury in a non-cell-autonomous manner [33, 34]. The inflammation/complement axis includes chronic NLRP3 activation, inflammasome-driven tau hyperphosphorylation [35], ASC-speck cross-seeding of Aβ [36], and C1q/C3-dependent synaptic engulfment that mediates pre-plaque synapse loss in AD models [37, 38].

These axes are not parallel narratives. They are connected by biochemical bridges that explain why lipid droplets, iron retention, complement activation, white-matter injury, and astrocyte reactivity often co-occur in AD. Lipid-handling failure promotes lysosomal congestion and disrupts ferritinophagic flux, increasing the labile iron pool. Iron-driven membrane peroxidation generates oxidised lipid species that lower the threshold for inflammasome priming. Inflammation and complement activation then feed back onto lipid handling through APOE upregulation, lipid-receptor remodelling, and astrocyte–microglia crosstalk. This review therefore reorganises AD-relevant microglial biology around a staged regulatory logic: from checkpoint collapse, through lipid, iron/ferroptosis, and inflammation/complement axes, to biomarker-informed therapeutic translation. It also integrates white-matter pathology, astrocyte–microglia crosstalk, and microglial senescence as cross-axis amplifiers, including the cGAS–STING–type I interferon pathway recently linked to APOE ε4 and TREM2-R47H risk biology [39].

To support this narrative synthesis, we searched PubMed and Web of Science through April 2026 using combinations of the terms “Alzheimer’s disease”, “microglia”, “TREM2”, “APOE”, “progranulin”, “CX3CR1”, “CD33”, “lipid droplet”, “ferroptosis”, “NLRP3”, “complement”, “cGAS–STING”, “senescence”, “white matter”, “astrocyte”, “biomarker”, and “clinical trial”. We prioritised human post-mortem studies, single-cell and spatial transcriptomic analyses, genetic studies, mechanistic preclinical work, biomarker studies, and clinical-intervention reports. Older landmark studies were retained when they defined core mechanisms, nomenclature, or translational context. This review is therefore a critical mechanistic synthesis rather than a systematic review. The search was intended to support a narrative mechanistic review and was not conducted according to systematic-review or meta-analysis methodology. Because the evidence base spans human AD tissue, animal models, in vitro systems, AD clinical trials, non-AD clinical studies, and clinical-trial reports, we indicate the evidentiary context where it is important for interpretation.å.

From microglial state heterogeneity to checkpoint collapse

Binary M1/M2 activation models capture broad features of inflammatory behaviour but do not map onto the cellular states resolved in human AD tissue [13, 14]. Single-nucleus and spatial profiling have identified partially overlapping microglial states that coexist within the same brain region and often within the same anatomical neighbourhood: disease-associated microglia (DAM) and the closely related neurodegeneration-associated phenotype MGnD [15, 40]; lipid-droplet–accumulating microglia (LDAM), enriched near plaques in APOE ε4/ε4 brains [17, 30]; interferon-responsive microglia [20, 21]; dystrophic and senescence-associated microglia in advanced disease [41, 42]; and white-matter-associated microglia surrounding myelin breakdown [19, 20]. The 2022 nomenclature consensus is explicit that these labels denote dynamic states defined by the eliciting stimulus and the surrounding tissue rather than fixed cellular subtypes [14, 22]. Human AD microglia are not phenocopies of the murine DAM programme; cross-species comparison shows both TREM2-dependent and TREM2-independent transcriptional responses, and a distinct human-enriched AD-associated subpopulation is incompletely captured by current mouse models [42–44]. An apparently inflammatory signature in AD therefore often reflects failed adaptation to chronic phagocytic and lipid-handling load rather than a primary pro-inflammatory identity [9, 45].

State annotations describe what microglia look like at a moment in time but do not specify what determines whether a given state remains adaptive or drifts into self-reinforcing dysfunction. Microglial homeostasis is enforced by a TGF-β–dependent transcriptional programme anchored on MEF2C, MAFB, and related modules; this signature constrains microglial reactivity in the steady state and is progressively eroded under chronic disease conditions [9–11]. APOE ε4 itself induces a TGF-β–coupled microglial checkpoint that limits the AD response, and disabling this checkpoint accelerates tau-driven neurodegeneration in tauopathy models — direct genetic evidence that the homeostatic programme is an actively maintained, AD-relevant control system rather than a passive reference state [29]. The discrete receptor and signalling systems that operate within or alongside this programme to set thresholds for phagocytic engagement, lysosomal flux, neuron–glia restraint, and inflammatory activation are the regulatory nodes whose collective operation determines whether disease-associated states remain reversible or progress to self-reinforcing dysfunction.

Three regulatory nodes carry most of the disease-relevant signalling load in human AD microglia, with a fourth receptor system contributing additional inhibitory tone. TREM2 binds anionic lipids and apolipoproteins, including APOE and CLU, and signals through DAP12 to couple ligand engagement to phagocytic competence, mTOR-dependent metabolic rewiring, and survival under sustained phagocytic demand [23, 24, 46, 47]; the TREM2–APOE pathway determines whether microglia can sustain an engaged, MGnD-like response without exhausting metabolic and lysosomal capacity, and links node function directly to AD genetic risk [40]. Progranulin maintains lysosomal acidification, cathepsin maturation, and degradative throughput; insufficiency under chronic phagocytic load impairs handling of lipid-rich and aggregate cargo and produces lysosomal storage-like phenotypes in microglia [25]. Progranulin dysfunction is therefore considered here as one component of a broader lysosomal regulatory network, rather than as the dominant determinant of microglial lysosomal failure in AD. This network also includes TFEB/TFE3-dependent lysosomal biogenesis, mTOR signalling, autophagy-related pathways, lipid metabolic state, and TREM2-associated metabolic adaptation. Neuron-derived fractalkine signals through CX3CR1 to restrain microglial reactivity and synaptic engulfment, and loss of this restraint expands aberrant pruning [26]. CD33/Siglec-3 sets activation thresholds for innate immune signalling within the inflammation axis, and AD-risk-associated CD33 variants alter receptor expression and amyloid handling [27, 28]. Each of these systems operates on a distinct biochemical layer — receptor-coupled sensing, lysosomal degradation, neuronal restraint, and inhibitory threshold control — and disabling any one of them shifts microglia toward a constrained operating range without by itself reproducing the multi-axis dysfunction characteristic of symptomatic AD.

Checkpoint collapse denotes the joint, progressive failure of multiple regulatory nodes during AD rather than the loss of any single node. Loss of TREM2 sensing alone shifts activation thresholds, and loss of progranulin alone alters lysosomal handling, but joint compromise of the regulatory layer — TREM2-coupled sensing, progranulin-supported lysosomal handling, and CX3CR1-mediated restraint, together with weakening of CD33-set inhibitory tone within the inflammation axis — deprives microglia of both the capacity to contain ongoing plaque, debris, and lipid load and the capacity to recover an adaptive set-point; this combined impairment may create a broader and more self-reinforcing failure state than isolated single-node disruption, although direct evidence for synergistic multi-node failure in human AD remains limited. The defect is biochemically self-reinforcing: failed phagocytic completion expands oxidised lipid and labile iron, which in turn further compromises lysosomal flux and licenses inflammasome priming. The concrete bridges connecting these effects across the lipid, iron–ferroptosis, and inflammation–complement axes are developed in the chapter that follows. Collapse is also temporally staged: early disease is dominated by reversible or partial node dysfunction and remains therapeutically tractable, whereas established disease is dominated by multi-node failure compounded by senescence-associated amplification, narrowing the window for restoration. This framing accounts for why the same molecular pathway can appear protective in one context and pathogenic in another, and for why broad anti-inflammatory strategies have repeatedly underperformed in AD prevention and treatment trials [48–50]: indiscriminate suppression cannot rescue a system whose primary defect is failure of regulation rather than excess of activation.

Tri-axial framework and links between axes

The cross-axis relationships discussed below should be read as a synthesis of human-tissue associations, experimental evidence, and testable mechanistic hypotheses, rather than as a fully validated causal sequence in human AD. Once regulatory-node restraint is destabilised, AD pathology can be organised along three interacting effector axes that share substrates, signalling routes, and cellular outputs: lipid, iron/ferroptosis, and inflammation/complement. The lipid axis is anchored in APOE-biased cholesterol and phospholipid handling, defective lysosomal flux under chronic phagocytic load, and consolidation of the LDAM programme, which combines impaired degradative capacity with neurotoxic secretory output [17, 30, 31, 51]. The iron/ferroptosis axis emerges when ferritinophagy-stressed microglia accumulate redox-active iron and become vulnerable to peroxidation of polyunsaturated phospholipids as GPX4- and FSP1-centred defences become insufficient [52–55]. The inflammation/complement axis is driven by chronic NLRP3 priming and activation, inflammasome-coupled tau-kinase regulation, and C1q- and C3-dependent synaptic pruning, which reactivates a developmental clearance programme in mature circuits [35, 37, 38, 56]. Treating these axes as separate literatures obscures their most important feature: dysfunction propagates between them through shared substrates and feedback loops, rather than through parallel and independent pathways.

Several cross-axis bridges can be organised directionally. Lipid-axis failure promotes iron-axis stress through lysosomal congestion and impaired ferritinophagic flux. Under chronic phagocytic load, defective NCOA4-dependent ferritin turnover expands the microglial labile iron pool, creating a substrate for Fenton-driven peroxidation when antioxidant defences are inadequate [53, 54]. Iron-axis activity then lowers the threshold for inflammation. Peroxidised PUFA-OOH species and oxidised cholesterol metabolites generated under ferroptotic stress engage inflammasome priming and HMGB1-mediated DAMP signalling, thereby facilitating NLRP3 activation and complement output [53, 55, 57]. Inflammation-axis output can then re-enter the lipid axis through astrocyte–microglia crosstalk. Microglia-derived IL-1α, TNF-α, and C1q induce a reactive astrocyte programme that secretes long-chain saturated fatty acids capable of injuring neurons and oligodendrocytes, while also worsening the lipid-handling burden imposed on microglia [58, 59]. These bridges are supported by mechanistic evidence rather than by co-occurrence alone, and they generate testable predictions about how intervention at one axis may shift the operating state of the others.

Two amplifier layers cut across all three axes rather than linking a single pair of pathways. Senescent microglia accumulate lipofuscin and intracellular iron, restrict ferritinophagic turnover, and release a senescence-associated secretory phenotype enriched in IL-6, IL-1β, and complement components [53, 60, 61]. The result is a lower adaptive ceiling across lipid handling, iron control, and inflammatory restraint. Senescence is therefore treated here as a cross-axis amplifier rather than as a fourth parallel axis, and is discussed in detail in Section "Microglial senescence as an amplifier of tri-axial failure". The cGAS–STING–type I interferon pathway is one of the best-characterised upstream links between AD risk biology and this senescence amplifier, with genetic evidence connecting APOE ε4 and TREM2-R47H to cGAS-associated microglial senescence and tauopathy progression [39].

Each axis can be linked to candidate biomarker families, but these markers should be interpreted as probabilistic anchors rather than axis-specific clinical classifiers. APOE genotype and plasma or CSF lipidomics may help identify lipid-axis vulnerability. QSM-MRI, CSF ferritin, and lipid-peroxidation readouts may provide context for iron/ferroptosis biology. CSF C1q, C3, inflammasome-linked measures, and microglial PET ligands may support assessment of inflammatory or complement engagement. sTREM2, progranulin-related measures, GFAP, and YKL-40 may reflect regulatory-node status or broader glial activation. However, none of these biomarkers, including GFAP, YKL-40, ferritin, or TSPO imaging, is sufficiently specific to assign individual patients to a single pathological axis in current clinical practice. Their main value is in multimarker interpretation, cohort enrichment, and pharmacodynamic monitoring when combined with Aβ, tau, neurodegeneration, vascular, and disease-stage measures. The integrated logic of the framework — regulatory-node failure, three interacting effector axes, cross-axis amplification, and candidate biomarker and pharmacodynamic readouts — is summarised in Fig. 1.

Fig. 1.

Tri-axial model of microglial checkpoint collapse in Alzheimer’s disease. Lipid dysregulation, iron-associated ferroptotic stress, and inflammation/complement signalling are shown as three interacting effector axes organised around regulatory-node failure. The central dashed circle depicts TREM2, progranulin, CX3CR1, and CD33/Siglec-3 as regulatory nodes that normally support lipid sensing, lysosomal competence, neuron–microglia restraint, and inhibitory immune tone. Red prohibition symbols indicate weakening of these regulatory brakes, leading to unrestrained lipid, iron, and inflammatory cascades. The inter-axis arrows illustrate reciprocal propagation rather than parallel pathway activation. Microglial senescence, characterised by iron/lipid inclusions and a senescence-associated secretory phenotype (SASP), is positioned as a cross-axis amplifier. Candidate biomarker and pharmacodynamic readouts are shown in the perimeter boxes: LDAM, ABCA1/ABCG1, LXR/PPAR, and plasma/CSF lipidomics for lipid-axis biology; ACSL4–PUFA, GPX4/FSP1, QSM-MRI, and CSF ferritin for iron/ferroptosis biology; and NLRP3–IL-1β/IL-18, ASC specks, C1q/C3, and inflammasome readouts for inflammation/complement biology. These readouts are intended to support cohort enrichment and pharmacodynamic monitoring rather than axis-specific clinical classification. The lower band outlines a stage-aware therapeutic logic across the AD continuum, from TREM2 agonism and inflammasome restraint in preclinical/MCI stages, to lipid/iron modulation in early AD, and complement or senescence-directed strategies in later disease. This figure is intended as a conceptual map rather than a literal pathway diagram

Lipid axis: APOE, lipid-droplet–accumulating microglia, and the path to non-cell-autonomous neurodegeneration

Microglia are among the principal lipid-processing cells of the central nervous system. In AD, this homeostatic burden is intensified by plaque-associated phospholipids, dystrophic neurites, oxidised lipid species, and accelerated myelin turnover. The lipid axis captures how this chronic cargo load becomes pathological when APOE-biased trafficking, ABCA1- and ABCG1-dependent efflux, and lysosomal–autophagic flux become rate-limiting [62–66]. APOE ε4 contributes to this failure in two linked ways: as a lipid-transport variant with reduced efflux capacity, and as an immune modifier that biases microglial responses towards chronic engagement rather than resolution [63–67]. When lipid-rich cargo exceeds degradative and efflux capacity, microglia consolidate an LDAM programme characterised by reduced phagocytic competence, elevated reactive oxygen species, and a secretory profile capable of injuring neurons non-cell-autonomously [17, 51]. This makes the lipid axis a central point of convergence between genetic risk, plaque-associated stress, white-matter burden, and downstream neuronal injury.

Recent studies identify two enzymatic entry points into the AD-relevant LDAM programme. ACSL1, a long-chain fatty-acyl-CoA synthetase, is enriched in lipid-laden microglia from APOE ε4/ε4 brains and channels free fatty acids into triglyceride and phospholipid pools rather than supporting degradative flux. ACSL1-positive microglia accumulate near plaques and produce factors that exert non-cell-autonomous neurotoxicity in iPSC-derived neurons, linking APOE ε4 risk biology to tau-relevant injury [30]. DGAT2, a diacylglycerol acyltransferase, represents a distinct but convergent entry point. It is induced by fibrillar Aβ and converts excess free fatty acids into triglyceride deposits that suppress microglial phagocytic capacity; pharmacological DGAT2 inhibition restores phagocytosis and reduces plaque burden in mouse models [68]. Together, these findings argue that LDAM is not simply a passive storage phenotype. It is a metabolically organised state produced by convergent APOE ε4-driven and Aβ-driven inputs, with ACSL1 and DGAT2 marking partially distinct opportunities for intervention [17, 30, 31, 68].

The therapeutic implication is that the lipid axis should be judged by restoration of flux rather than by reduction of droplet burden alone. LXR and PPAR signalling, ABCA1/ABCG1-dependent cholesterol efflux, TFEB-directed lysosomal biogenesis, and TREM2-supported adaptation to chronic phagocytic load are the most coherent mechanistic levers [69–72]; LXR agonism has provided proof of concept in tau- and APOE ε4-linked models, where activation of lipid efflux can reduce glial lipid accumulation and normalise related inflammatory changes [69]. A lipid-axis intervention should therefore demonstrate more than a lower LDAM count. It should show improved degradative throughput, restored apolipoprotein-mediated efflux, or recovery of lysosomal–autophagic handling. This distinction matters because lipid-axis failure also restricts ferritinophagy and weakens phagolysosomal resolution, thereby feeding the iron/ferroptosis and inflammation/complement axes. Restoration of lipid-handling flux is therefore a cross-axis intervention rather than a siloed lipid-storage strategy. The architecture of this axis, from APOE ε4-linked lipid-handling stress through LDAM formation to tau-relevant neuronal injury, is shown in Fig. 2.

Fig. 2.

Lipid-axis model of APOE ε4-linked LDAM and tau-relevant injury in Alzheimer’s disease. This figure illustrates how APOE ε4-associated lipid-handling stress may shift microglia towards lipid-droplet–accumulating microglia (LDAM) and downstream tau-relevant neuronal injury. The upper inset places LDAM within a broader spectrum of dynamic microglial states, including homeostatic, DAM, interferon-responsive, white-matter-associated, and dystrophic/senescent-like states. (a, b) APOE ε3- and APOE ε4-linked cholesterol handling through ABCA1 and ABCA1/ABCG1 transporters; APOE ε4 is depicted as reducing lipid clearance and favouring intracellular lipid retention. (c, d) Under chronic myelin-debris and Aβ-associated cargo load, impaired lysosomal–autophagic flux promotes LDAM formation, reduced clearance capacity, and oxidative stress. (e, f) The right panel links microglial lipid failure to tau-axis vulnerability, including tau-associated microtubule destabilisation, GSK3β/CDK5-associated phosphorylation pathways, neurofibrillary tangle formation, and diseased neuronal morphology. The red arrow denotes a non-cell-autonomous injury bridge from LDAM-associated metabolic and secretory stress to tau-relevant neuronal injury, not direct proof of linear tau propagation. Candidate biomarker anchors include APOE ε4 genotype, plasma/CSF lipidomic profiles, ceramides, sphingomyelins, and neutral lipids, with Aβ/tau markers used to define disease context rather than to classify lipid-axis dominance

Translational interpretation requires biomarkers that reflect the rate-limiting biology rather than adjacent disease processes. APOE genotype identifies the subgroup in which lipid-axis vulnerability is most likely to be prominent. Plasma and CSF lipidomic signatures, including selected ceramides, sphingomyelins, and neutral lipids, are tractable enrichment and staging markers and have been linked to AD risk and progression in independent cohorts [73, 74]. Recent work in human iPSC-derived microglia further shows that inflammation-induced lysosomal dysfunction is exacerbated by the APOE ε4/ε4 genotype, providing a cellular bridge between APOE status, lysosomal flux failure, and lipid-axis vulnerability [75]. By contrast, tau-axis markers such as plasma p-tau217 and tau-PET define where a patient sits along the Aβ–tau continuum, but they do not by themselves identify lipid-axis dominance [76]. Trials targeting LDAM-related biology should therefore be designed around APOE ε4 stratification and prespecified lipidomic pharmacodynamic markers; otherwise, the subgroup most likely to depend on lipid-axis pathology may be diluted within biologically heterogeneous cohorts.

Iron/ferroptosis axis: from labile iron to non-cell-autonomous neuronal injury

Iron dysregulation in AD is better understood as a compartmental and cell-state problem than as simple bulk-tissue overload. Under physiological conditions, microglia take up transferrin-bound iron, store it in ferritin, and release it into the labile pool through NCOA4-dependent ferritinophagy at rates calibrated to metabolic demand [52, 53, 77, 78]. AD perturbs this balance at several points. Chronic phagocytic load and lysosomal congestion impair ferritinophagic flux, redox-active Fe²⁺ accumulates in the labile iron pool, and this pool may be further enlarged in senescent microglia [53, 54, 78, 79]. Imaging and post-mortem studies support the clinical relevance of this process: regional brain iron rises across the AD continuum, QSM-detectable iron predicts cognitive trajectory independently of Aβ burden, and CSF ferritin is associated with faster decline at intermediate disease stages [79–83]. The disease-relevant signal is therefore unlikely to be total iron load alone. It is more likely to reflect the fraction of iron that remains labile, redox-active, and positioned to drive Fenton chemistry on lipid substrates.

The downstream consequence of an expanded labile iron pool is selective peroxidation of polyunsaturated-phospholipid–containing membranes (Fig. 3). ACSL4-dependent incorporation of polyunsaturated fatty acids into membrane phospholipids increases peroxidation susceptibility, particularly when system Xc⁻, GPX4, and FSP1-centred antioxidant defences become insufficient [55, 84–88]. This connects iron stress to inflammatory amplification. Oxidised phospholipids, including POVPC, and HMGB1 released from ferroptotic cells can prime NLRP3 in nearby microglia and support complement output, while lipid-peroxidation adducts such as MDA and 4-HNE provide measurable readouts of the same peroxidative stress [57, 89, 90]. Importantly, ferroptotic stress confined to microglia can injure neurons non-cell-autonomously. In models in which ferroptosis is selectively induced in microglia, animals develop progressive cognitive impairment and neuronal loss without primary neuronal lipid peroxidation, indicating that the iron axis can act through microglia-derived injury signals rather than through direct neuronal ferroptosis alone [33, 34]. This iron–lipid-peroxidation state may also converge with astrocyte–microglia crosstalk, because microglial IL-1α, TNF-α, and C1q are established drivers of neurotoxic reactive astrocyte conversion, which can subsequently mediate cell injury through saturated-lipid release, as developed further in Section "Astrocyte–microglia crosstalk: the bidirectional axis with context-dependent readouts" [58, 59].

Fig. 3.

Iron/ferroptosis axis and microglia-mediated neuronal injury in Alzheimer’s disease. This figure summarises the iron/ferroptosis axis as a compartmental process in which labile iron, lipid peroxidation, and glial inflammatory amplification converge. (a) Holo-transferrin carrying Fe³⁺ enters through transferrin receptor-mediated uptake; STEAP3 reduces Fe³⁺ to Fe²⁺, and DMT1 supports endosomal iron release. Iron is stored in ferritin and mobilised through NCOA4-dependent ferritinophagy, with senescent microglia contributing to expansion of the labile iron pool. (b) Labile Fe²⁺ promotes lipid oxidation through Fenton chemistry, lipoxygenase activity, and P450-linked oxidoreductase reactions. (c) Ferroptosis susceptibility increases when system Xc⁻, glutathione synthesis, and GPX4-centred antioxidant protection become insufficient, allowing PUFA-OOH accumulation, oxidised lipid generation, MDA/4-HNE formation, reactive oxygen species, and DNA damage. Oxidised lipids and HMGB1 are shown as danger signals that can support inflammasome and complement-related amplification. (d) Iron-loaded and SASP-like microglia can promote neurotoxic reactive astrocyte conversion and neuronal degeneration through IL-1α, TNF-α, C1q, and related inflammatory mediators. (e) Translational anchors include QSM-MRI for regional brain iron, CSF ferritin for iron-storage biology and decline risk, and exploratory pharmacodynamic readouts such as MDA and 4-HNE

Therapeutic strategies for the iron/ferroptosis axis fall into three mechanistic classes. Iron-chelation strategies aim to reduce the labile iron pool or sequester redox-active Fe²⁺, with selective microglial-iron approaches moving through preclinical evaluation [91]. GPX4- and FSP1-supportive strategies seek to raise the antioxidant ceiling of the same compartment; lipophilic radical-trapping antioxidants and selenocompounds have shown protection in AD-relevant models [92, 93]. ACSL4-directed approaches target the substrate side of ferroptosis by limiting incorporation of polyunsaturated species into membrane phospholipids, with pharmacological proof of concept established mainly in non-AD ferroptosis models [55, 88]. These strategies are mechanistically complementary but individually incomplete. Chelation without compartmental selectivity may restrict physiological iron supply. Antioxidant support may be insufficient if substrate vulnerability remains high. ACSL4 modulation may reduce peroxidation liability but does not by itself resolve the upstream labile-iron driver. The therapeutic objective is therefore not generic iron lowering, but correction of the iron–lipid-peroxidation compartment most closely linked to microglial dysfunction and neuronal injury.

The deferiprone TEAM-AD trial is an important cautionary example. In mild-to-moderate AD, deferiprone reduced QSM-detectable iron but did not produce cognitive benefit and was associated with increased volumetric loss in some regions [91]. These findings should not be read simply as evidence against iron biology in AD. A more cautious interpretation is that non-selective iron lowering may fail, or even become harmful, when it does not distinguish physiological iron pools from labile, lipid-peroxidation-competent compartments. Several explanations remain possible. The iron pool reduced by deferiprone may not have been the pathogenic compartment; functional iron deficiency may have occurred in cells that require iron for normal metabolism; or the treatment may have insufficiently engaged microglial labile iron at the tested exposure [91]. Human-tissue evidence showing iron-associated lipid peroxidation in AD lipid rafts, together with loss of ferroptosis suppressors in the same compartment, supports the need to define the relevant iron pool more precisely rather than to abandon the axis [32].

Translational interpretation for this axis should combine regional iron imaging with direct or proximal measures of lipid peroxidation. QSM-MRI can index regional iron burden, CSF ferritin can report iron-storage biology, and exploratory MDA or 4-HNE measurements can provide evidence of lipid-peroxidation engagement. These readouts should be interpreted alongside Aβ and tau biomarkers, because iron/ferroptosis biology is unlikely to define disease stage by itself. Future trials should prioritise compartmentally selective interventions, enrol patients with biomarker evidence of iron–lipid-peroxidation biology, and define pharmacodynamic success in terms of regional iron modulation, ferritin dynamics, and lipid-peroxidation suppression rather than bulk iron reduction alone. These translational anchors are summarised in Fig. 3.

Inflammation/complement axis: NLRP3, complement, and type-I interferon

The inflammation/complement axis in AD brings together three linked processes: receptor-coupled Aβ sensing, NLRP3 inflammasome activation, and complement-mediated synaptic remodelling. Type-I interferon signalling adds a further layer by connecting inflammatory activation to the senescence amplifier. Aβ engagement of CD36 with TLR4/TLR6 drives Src and NF-κB signalling and primes NLRP3 transcription, while oxidised lipid DAMPs generated by the iron/ferroptosis axis can raise the priming tone further [94, 95]. Receptor balance is also part of this axis. TREM2-coupled engagement promotes phagocytic containment of Aβ through DAP12 and Syk–PI3K/Akt signalling [46, 96, 97]. CD47–SIRPα–SHP1 signalling provides homeostatic restraint and is reduced under AD conditions [98, 99]. CD33–SHP1 contributes inhibitory tone, with receptor expression and function influenced by AD risk variants [27, 28].

NLRP3 activation is not merely a generic inflammatory response. It is bidirectionally coupled to tau pathology. Assembly of the NLRP3 inflammasome with ASC, caspase-1, and NEK7 generates mature IL-1β and IL-18 and promotes tau hyperphosphorylation through inflammasome-linked kinase regulation. ASC-speck release can also support extracellular Aβ cross-seeding [35, 36, 100]. The reverse direction is also documented: aggregated tau activates microglial NLRP3–ASC through endolysosomal cathepsin-B-dependent processing, creating a feed-forward loop in which inflammasome activation amplifies tau pathology and tau pathology further stimulates inflammasome activation [101]. Pharmacological inhibition of NLRP3 with MCC950, OLT1177, and related inhibitors reduces inflammasome output, Aβ burden, synaptic dysfunction, or cognitive impairment in AD models [102–107]. More recent work further supports the potential value of NLRP3 inhibition after symptom onset or with improved CNS-penetrant compounds, although clinical efficacy in AD remains unproven [108].

Complement-mediated synaptic engulfment represents a developmental clearance programme that re-emerges aberrantly in AD. C1q tags vulnerable synapses, C3 and iC3b deposition opsonises them, and CR3 on microglia mediates phagocytic engulfment. This pathway normally supports developmental synaptic pruning, but in AD models it can be reactivated before frank plaque accumulation, with synapse loss closely tracking complement engagement [37, 38, 56, 99, 109–111]. Anti-C1q antibodies, C3-targeted approaches, and CR3 modulation interrupt this loop at distinct points [111–113]. The therapeutic objective is not broad complement suppression. Complement remains essential for antimicrobial defence and tissue homeostasis; the more defensible goal is to reset pathological engulfment thresholds while preserving protective complement functions.

Type-I interferon signalling is an integral part of this axis, not a peripheral inflammatory by-product. CX3CR1 downregulation activates NF-κB and contributes to amplified inflammasome and interferon output, linking loss of neuronal restraint to a more permissive inflammatory state [26, 114]. Pathogenic tau can also activate microglial cGAS through cytosolic mitochondrial DNA leakage. This response reduces neuronal MEF2C expression and weakens a broader cognitive-resilience network in tauopathy models [115]. The same pathway is connected to AD risk biology: APOE ε4 and TREM2-R47H have been linked to increased cGAS-associated microglial senescence and accelerated tauopathy progression [39]. These findings suggest that NLRP3- or complement-only strategies may leave part of the inflammatory–senescence interface untouched. Pharmacodynamic panels should therefore include IL-1β, IL-18, ASC-related measures, C1q/C3, and, where relevant, type-I interferon signatures.

The therapeutic implication is selective node-level modulation, not broad anti-inflammatory suppression. NLRP3-directed approaches target inflammasome assembly and cytokine maturation. MCC950 and OLT1177 provide AD-model evidence for reduced inflammasome output, Aβ burden, synaptic dysfunction, or cognitive impairment [105, 106]. NT-0796 and post-symptomatic NLRP3 inhibition support the broader rationale for brain-penetrant or later-stage NLRP3 modulation [107, 108]. Complement-directed strategies act at a different step. C1q-targeted antibodies interrupt complement-mediated synapse loss [111, 112], whereas C3-pathway modulators target downstream convertase and amplification biology [113]. These approaches should not be treated as interchangeable anti-inflammatory interventions. Their value in AD will depend on disease stage, CNS target engagement, and biomarker-defined inflammatory or complement enrichment. Figure 4 summarises the convergence of receptor balance, NLRP3 amplification, complement-mediated synaptic pruning, type-I interferon signalling, and checkpoint collapse.

Fig. 4.

Inflammation/complement axis and checkpoint collapse in Alzheimer’s disease. The figure integrates receptor-coupled Aβ sensing, NLRP3 inflammasome activation, complement-mediated synaptic pruning, and regulatory-node failure. (a) In adaptive plaque engagement, TREM2 signalling supports Syk- and PI3K/Akt-linked containment and degradation of Aβ, whereas CD47–SIRPα–SHP1 and CD33–SHP1 provide inhibitory restraint. CD36-dependent Aβ sensing through TLR4/TLR6 and Src promotes NF-κB and NLRP3 priming. (b) Reactive oxygen species and Aβ-associated stimuli facilitate NLRP3 oligomerisation with ASC, caspase-1, and NEK7 to form the inflammasome complex. (c) Selective NLRP3 inhibition is shown as a therapeutic entry point; downstream inflammasome activity includes IL-1β/IL-18 maturation, GSDMD cleavage to N-GSDMD, ASC-speck release, and Aβ/tau-associated cross-seeding or propagation. (d) Complement-mediated synaptic pruning is depicted as sequential C1q tagging, C3/iC3b deposition, and CR3-dependent engulfment, counterbalanced by CD47/SIRPα “don’t-eat-me” signalling that may be weakened in AD. (e) Checkpoint-collapse mechanisms include progranulin loss, CX3CL1–CX3CR1 downregulation with NF-κB/type-I interferon-linked outputs, and ADAM10/17-mediated TREM2 shedding with altered soluble TREM2 signalling. (f) Candidate translational readouts include CSF C1q/C3, inflammasome readouts, TSPO and exploratory microglial PET ligands, together with therapeutic approaches such as C1q/C3 inhibition, NLRP3 antagonism, and checkpoint stabilisation. These PET and fluid readouts are intended for multimarker interpretation and pharmacodynamic monitoring, not as stand-alone axis-specific classifiers

White-matter pathology and myelin burden: the lipid axis at scale

White-matter pathology should not be treated as a passive by-product of cortical degeneration. Myelin turnover imposes a large and continuous lipid-processing burden on microglia. In AD, this burden is increased by myelin debris, plaque-associated lipids, dystrophic neurites, and local inflammatory stress. Under physiological conditions, TREM2- and MerTK-dependent phagocytosis helps clear cholesterol-rich myelin fragments and supports remyelination [116–120]. With ageing and AD, this clearance system becomes less efficient. Persistent myelin debris promotes lysosomal congestion, lipid-droplet accumulation, and local inflammatory activation, while impaired microglial lipid metabolism can further disrupt myelin turnover [51, 121]. White-matter biology is therefore not separate from the lipid axis. It is one of the clearest tissue settings in which the microglial adaptive ceiling is tested.

A complete account of white-matter vulnerability requires attention to both sides of the myelin–microglia relationship. Microglial clearance of myelin debris represents the demand-side problem. Oligodendrocyte cholesterol production represents the supply-side problem. APOE ε4 affects this second arm directly. iPSC-derived oligodendrocytes from APOE ε4/ε4 donors show reduced cholesterol synthesis, abnormal myelination of co-cultured neurons, and impaired recovery of lipid handling under stress [122]. APOE ε4 should therefore be viewed not only as a modifier of microglial lipid handling, but also as a contributor to oligodendrocyte lipid and myelin vulnerability. In APOE ε4 carriers, impaired debris clearance and impaired myelin lipid support may operate together, increasing the likelihood that white-matter injury becomes self-sustaining.

Vascular contribution adds a further layer to this lipid-axis stress. Cerebral small-vessel disease, blood–brain barrier dysfunction, and chronic hypoperfusion expose oligodendrocytes and white-matter tracts to metabolic strain. These insults increase the lipid and debris burden that microglia must handle, while also compromising the metabolic conditions required for effective clearance [123–126]. The SPRINT-MIND experience shows that intensive blood-pressure control can reduce white-matter hyperintensity progression and lower the risk of mild cognitive impairment [127, 128]. This does not mean that vascular treatment is a direct lipid-axis therapy. It does, however, support a coupled interpretation: improving perfusion and barrier integrity may reduce the chronic tissue stress that feeds microglial lipid-handling failure, whereas restoring lipid-handling capacity may improve tolerance to vascular injury.

The translational implication is that white-matter burden should be incorporated into microglia-directed trial design rather than treated as background noise. Diffusion-tensor imaging, white-matter hyperintensity volume, myelin water imaging, and molecular markers of oligodendrocyte stress can help identify patients in whom this dimension is clinically meaningful. Baseline white-matter burden can serve both as an enrichment variable and as a stratification factor, distinguishing predominantly cortical from white-matter–burdened phenotypes. Combination strategies that pair vascular optimisation with lipid-axis support are biologically plausible, especially in APOE ε4 carriers or patients with prominent white-matter injury. At present, these should be framed as testable trial hypotheses rather than established therapeutic algorithms. Their value will depend on prespecified myelin, vascular, and lipid-handling pharmacodynamic readouts, not on global cognitive endpoints alone.

Astrocyte–microglia crosstalk: the bidirectional axis with context-dependent readouts

Microglial dysfunction in AD is never purely microglial. Astrocytes participate in both feed-forward injury and residual defence. The same dynamic-state logic applied to microglia should therefore also be applied to astrocytes. The 2021 consensus on reactive astrocyte nomenclature moved away from the A1/A2 binary and recommended context-, stimulus-, and time-resolved descriptions of reactive states [129]. Systematic analyses of reactive astrocyte biology further support this shift away from fixed binary labels [130]. In this review, when the original neurotoxic-reactive astrocyte literature uses A1-style terminology, we refer instead to the relevant process as IL-1α–, TNF-α–, and C1q-driven reactive astrocyte conversion. This wording preserves the underlying biology without treating the state as a categorical cell identity.

One harmful loop runs from microglia to astrocytes and back. Microglia-derived IL-1α, TNF-α, and C1q induce neurotoxic reactive astrocyte conversion in vivo [58]. Once converted, reactive astrocytes can injure neurons and oligodendrocytes through saturated-lipid release, with ELOVL1-dependent lipid synthesis providing a defined molecular handle for this toxicity [59, 131]. These astrocyte-derived outputs then feed back onto microglia. Saturated lipids and complement components can worsen debris-handling stress, increase oxidative and ferroptotic vulnerability, and reinforce CR3-dependent engulfment of complement-opsonised synapses. The result is a bidirectional amplification loop rather than a one-way astrocytic response. This pathogenic branch is summarised in Fig. 5a.

Fig. 5.

Astrocyte–microglia crosstalk as a cross-axis amplifier in Alzheimer’s disease. The figure depicts pathogenic and protective branches of astrocyte–microglia communication in AD using dynamic-state terminology rather than the retired A1/A2 binary. (a) Pathogenic feed-forward loops: microglial IL-1α, TNF-α, C1q, oxidised lipids, and DAMPs promote neurotoxic-reactive astrocyte conversion. Reactive astrocytes release ELOVL1-dependent long-chain saturated fatty acids (LC-SFAs) and increase C3 output, thereby amplifying lipid-mediated toxicity, oxidative vulnerability, ferroptotic risk, and impaired TREM2-dependent clearance. Complement-opsonised synapse phagocytosis is shown as a sequential process: C1q marking, C3/iC3b deposition, and CR3 (CD11b/CD18)-mediated engulfment. Anti-C1q intervention with ANX005 is indicated as one potential entry point for interrupting this loop. (b) Protective nodes and therapeutic entry points include augmentation of CX3CL1–CX3CR1 signalling to restrain microglial activation, IL-3–IL-3R signalling that synergises with TREM2-associated plaque-proximal phagocytic programmes, and TGF-β–SMAD2/3-associated homeostatic restraint. Astrocyte-linked context markers include plasma GFAP as an early amyloid-associated reactivity marker and CSF YKL-40 as a marker of later inflammatory remodelling. These markers should not be interpreted as stand-alone indicators of a specific pathological axis

Astrocyte–microglia signalling is not uniformly harmful. Several protective branches act through mechanisms distinct from the injury loop. Astrocytic IL-3 signals through microglial IL-3Rα/βc and promotes plaque-proximal microglial responses with enhanced TREM2-linked phagocytic competence; IL-3 deficiency accelerates AD-like pathology in mouse models [132, 133]. TGF-β signalling is a central regulator of steady-state microglial identity, and TGF-β-dependent transcriptional programmes help maintain the homeostatic microglial signature [10]. Recent work extends this regulatory logic to AD risk biology: APOE ε4 has been linked to disruption of a TGF-β-associated microglial checkpoint, positioning APOE ε4 not only as a lipid-handling allele but also as a modifier of regulatory restraint [29]. Neuron-derived CX3CL1 acting through microglial CX3CR1 provides an additional restraint pathway that complements astrocyte-derived regulatory signals [26]. These protective nodes and therapeutic entry points are summarised in Fig. 5b.

The bidirectional structure of this axis matters for biomarker interpretation. A rise in an astrocytic biomarker does not by itself distinguish compensatory glial activation from maladaptive crosstalk. Astrocyte-facing readouts should therefore be interpreted as context markers rather than stand-alone mechanistic verdicts. Plasma GFAP is especially sensitive to early amyloid-associated astrocytic change and is now widely used as a biomarker of preclinical AD [134]. GFAP-defined astrocyte reactivity also modifies the relationship between Aβ and tau pathology in preclinical disease, suggesting that astrocytic activation may participate in Aβ-to-tau progression rather than simply report downstream injury [135]. CSF YKL-40 may better capture later inflammatory remodelling and is increasingly interpreted alongside GFAP in multimarker panels [136]. Meta-analytic and systematic-review evidence supports the robustness of these astrocyte-linked readouts across cohorts, although their biological direction remains context dependent [137, 138]. Figure 5 integrates the harmful feed-forward loops, protective regulatory nodes, and context-dependent readouts that make astrocyte–microglia crosstalk a discrete translational layer within the broader tri-axial framework.

Microglial senescence as an amplifier of tri-axial failure

Ageing remains the dominant background risk factor for AD not only because pathology accumulates over time, but also because glial adaptive capacity declines. Senescent or senescence-prone microglia show impaired phagocytosis, altered lysosomal trafficking, mitochondrial stress, iron retention, dystrophic morphology, and a senescence-associated secretory phenotype (SASP) enriched in IL-6, IL-1β, and complement components [53, 60, 61]. These features lower the adaptive ceiling of all three axes. Lipid-handling capacity declines, favouring LDAM-like states. Ferritinophagic flux becomes less reliable, expanding the labile iron pool. SASP-associated cytokines and complement programmes further amplify NLRP3 and C1q-linked inflammatory circuits.

Human post-mortem studies support the relevance of senescence-marker-positive cells in AD brain, although these markers should not be equated uncritically with irreversible cellular senescence. Senescence-eigengene analysis identified CDKN2D/p19-positive excitatory neurons overlapping with neurofibrillary tangle pathology [139]. Direct fluorescence immunohistochemistry showed an approximately three-fold increase in p16^INK4A-positive NeuN+ neurons in AD prefrontal cortex [140]. Complementary imaging mass cytometry and single-nucleus RNA-sequencing studies demonstrated increased GLB1- and p16^INK4A-positive glial populations in AD cortical tissue [141]. Together, these data support a senescence-associated tissue environment across neuronal and glial compartments, while leaving cell-type-specific causality to be tested experimentally.

The cGAS–STING–type-I interferon pathway provides a mechanistic link between this amplifier layer, tau pathology, and AD genetic risk. Pathogenic tau can activate microglial cGAS in part through cytosolic leakage of mitochondrial DNA. The resulting type-I interferon response reduces neuronal MEF2C expression and weakens a broader cognitive-resilience gene network; genetic or pharmacological cGAS inhibition in tauopathy mice reduces microglial IFN-I output and restores MEF2C-linked resilience [115]. This pathway is also connected to risk-allele biology. APOE ε4 and TREM2-R47H have been linked to elevated cGAS-associated microglial senescence and accelerated tauopathy progression in mouse models, with reversal of the senescence phenotype after cGAS inhibition [39]. cGAS–STING should therefore be viewed as one of the best-characterised links between microglial senescence, type-I interferon signalling, and AD risk biology, rather than as a universal upstream cause of AD.

Senolytic strategies follow a different logic. They do not target a single initiating cause of AD. Instead, they aim to reduce the chronic senescent-cell burden that can make checkpoint collapse, oxidative stress, and inflammatory amplification self-sustaining. Preclinical evidence supports this rationale. Clearance of senescent glia in MAPT P301S tau-transgenic mice reduces tau pathology and improves cognitive performance [142]. Senolytic therapy also alleviates Aβ-associated oligodendrocyte progenitor cell senescence and cognitive deficits in AD models [143]. Additional murine AD paradigms support the broader feasibility of senescence-targeted intervention [144]. These studies justify senescence as a therapeutic target, but they do not establish clinical disease modification.

Human translation remains early. A phase 1 feasibility trial of dasatinib plus quercetin in mild AD demonstrated CNS penetration and acceptable short-term safety, with subsequent fluid-biomarker analyses suggesting target engagement [145, 146]. A pilot study in older adults at risk for AD has also been initiated [147]. Outside AD, senescence-targeted intervention has shown early phase 1 signals in diabetic macular oedema, supporting the broader feasibility of this therapeutic class [148]. These findings should be interpreted cautiously. They establish feasibility, safety, and preliminary biological engagement; they do not yet demonstrate cognitive efficacy or disease modification in AD [149, 150].

Senescence is therefore best positioned as an amplifier layer within the tri-axial model, not as a replacement for lipid, iron/ferroptosis, or inflammation/complement biology. This distinction has trial-design consequences. Senolytic strategies may be most defensible when chronicity-driven amplification is prominent, especially in later-stage or biologically senescent subgroups. cGAS–STING inhibition may be more relevant where tau-linked cGAS activation and type-I interferon output are demonstrable before resilience networks are exhausted [115]. APOE ε4 and TREM2-R47H carriers provide biologically plausible enrichment groups, but this hypothesis still requires prospective validation [39]. In clinical translation, senescence-directed approaches should therefore be paired with stage-appropriate biomarker panels rather than advanced as stand-alone solutions.

Therapeutic implications: from pathway lists to stage-aware combination design

The therapeutic value of this framework lies in discrimination, not maximal targeting. It does not imply that every patient requires simultaneous intervention across all axes. Rather, treatment should be aligned with the biology most likely to be rate-limiting at a given disease stage. In early or preclinical disease, the most defensible objective is preservation of regulatory-node function while limiting inflammasome escalation. TREM2-directed strategies fit this objective conceptually. TREM2-activating antibodies, including approaches that engage the receptor stalk region or use blood–brain barrier transport vehicles, have shown microglial activation, improved metabolic competence, and reduced amyloid pathology in preclinical models [97, 151, 152]. AL002 advanced through peer-reviewed preclinical and first-in-human evaluation and demonstrated target engagement in healthy volunteers [153]. In the subsequent peer-reviewed INVOKE-2 phase 2 trial in early AD, AL002 produced sustained target engagement and central pharmacodynamic responses, including reduced CSF sTREM2 and increased CSF osteopontin, but did not meet the primary CDR-SB endpoint or demonstrate clinical benefit over placebo [154]. TREM2 agonism therefore remains biologically rational, but it is unlikely to be sufficient as a stand-alone strategy. Earlier-stage intervention will probably require regulatory-node support combined with selective inflammatory restraint or with axis-specific treatment selected through biomarker enrichment.

As disease progresses towards chronic tissue stress and tau propagation, lipid-handling restoration and ferroptosis control become more relevant therapeutic objectives. The deferiprone experience is instructive here. A target class can fail clinically when the intervention is too broad or when enrichment does not identify the subgroup in which the targeted biology is dominant [91]. Consistent with the iron-axis analysis above, the disease-relevant compartment is more likely to involve microglial labile iron and lipid-peroxidation-competent membranes than bulk-tissue iron. Future trials should therefore verify target engagement with compartment-sensitive pharmacodynamic readouts rather than rely on global iron lowering alone. When synaptic attrition and complement engagement are prominent, complement-directed strategies become biologically plausible: C1q-dependent synapse elimination has been demonstrated in AD mouse models, and terminal-pathway targeting reduces brain complement activation, amyloid load, synapse loss, and cognitive impairment in a dementia model [56, 155]. Senescence-directed strategies are also plausible when chronic glial amplification is detectable, but the human evidence remains early. Exploratory biomarker analyses from a phase 1 senolytic trial in mild AD support biological engagement, whereas phase 1 senescence-targeted intervention in diabetic macular oedema supports broader feasibility outside AD; neither should be interpreted as established AD disease modification [146, 148]. These later-stage approaches should therefore be framed as testable translational hypotheses, not mature treatment algorithms.

Table 1 compiles clinical, near-clinical, and mechanistically grounded preclinical programmes across the three axes, together with the candidate biomarker and pharmacodynamic readouts needed for interpretation. Its purpose is not to rank therapeutic readiness, but to map each intervention to the axis it is most likely to engage. This distinction is important. A therapy that shifts a biomarker without engaging the intended axis may be biologically interesting but clinically difficult to interpret. Conversely, a modest pharmacodynamic effect can be informative if it occurs in the appropriate biomarker-enriched subgroup.

Table 1.

Therapeutic programmes across key pathological axes in Alzheimer’s disease: stage, status, and biomarker interpretation

Program / Modality	Target & Mechanism (Axis)	Population / Stage	Evidence maturity / Status	Biomarkers (Entry / PD)	Reference
Clinical & near-clinical interventions
AL002 / AL002c (anti-hTREM2 mAb)	TREM2 agonist→enhances microglial proliferation, phagocytosis, and metabolic activation; reduces neuroinflammation and neuritic dystrophy (regulatory-node axis)	Preclinical: 5XFAD (hTREM2 CV/R47H); phase 1: healthy volunteers; phase 2: early AD (INVOKE-2)	Peer-reviewed preclinical/phase 1; peer-reviewed AD phase 2: phase 1 in healthy volunteers showed safety/tolerability and target engagement; INVOKE-2 showed sustained target engagement and CNS pharmacodynamic responses but did not meet the primary CDR-SB endpoint or demonstrate clinical benefit over placebo	PD readouts: CSF sTREM2 reduction and CSF osteopontin increase; exploratory AD biomarkers/amyloid PET did not support downstream disease modification.	[151, 153, 154]
Dapansutrile (OLT1177), oral NLRP3 inhibitor**	Selective NLRP3 inhibition→↓caspase-1 activation;↓IL-1β/IL-18 (inflammasome axis)	APP/PS1 (9 mo, 3-mo dosing); no AD clinical efficacy data	Preclinical AD: rescued LTP;↓cortical Aβ, IL-1β/IL-6/TNF-α; normalised plasma metabolomics; behaviour improved; no AD clinical efficacy data	Entry: pro-inflammatory cytokines. PD: cytokine reduction;↓CD68 + microglial activation; systemic metabolomic normalisation	[106]
MCC950 (tool NLRP3 inhibitor)*	Potent NLRP3 blocker preventing ASC oligomerisation (inflammasome axis)	APP/PS1; preclinical only	Preclinical/tool compound:↑Aβ clearance; enhanced non-inflammatory phagocytosis; cognitive rescue; no AD clinical efficacy data	Entry: NLRP3–IL-1β activation. PD:↓IL-1β/IL-18;↑Aβ phagocytosis	[105]
Dasatinib + Quercetin (D + Q), senolytic cocktail**	Clears senescent glia via BCL-2-family/SCAP pathways; reduces SASP-driven neuroinflammation (senescence axis)	Phase I in mild AD; pilot in at-risk older adults	AD clinical phase 1: CNS penetration confirmed; 12-week intermittent dosing feasible/safe; exploratory biological signals only; no drug-related SAEs	Entry: SASP panel; TNF-α; WMH burden. PD: SASP reduction (fractalkine, MMP-7, IL-6, TNF-α), CSF GFAP/YKL-40; PBMC stress-response signature	[145–147]
ABT263 (navitoclax), BCL-2 inhibitor senolytic	Targets BCL-2/BCL-XL to induce apoptosis of senescent cells (senescence axis)	3xTg-AD (Aβ + tau); long vs. short course	Preclinical AD senolytic: long-course ↓ AT8 tau, cortical Aβ plaques, and microgliosis; short late-course minimal effect	Entry: p16 + burden; glial activation. PD: AT8, Aβ load, Iba1	[144]
Latozinemab (AL001), anti-sortilin mAb**	Blocks sortilin-mediated progranulin turnover→raises circulating/CNS PGRN, stabilising lysosomal competence and restraining complement/NLRP3 tone (regulatory-node axis)	Phase 1/2 in FTD-GRN; no AD trials	Non-AD clinical: sustained 2–3-fold ↑ plasma/CSF PGRN in FTD-GRN carriers; target engagement confirmed; AD clinical efficacy not established; relevance to AD remains inferential through the PGRN–lysosome–complement axis	Entry: low-PGRN or high-complement subsets. PD: plasma/CSF PGRN; lysosomal/complement panels	[156]
LXR/PPAR agonists (e.g. GW3965, bexarotene, pioglitazone)*	Activate LXRα/β or PPARγ→↑ABCA1/ABCG1 cholesterol efflux, support myelin lipid handling and APOE metabolism (lipid axis)	APP/PS1 and related models; small AD trials of bexarotene/pioglitazone	Preclinical AD + AD clinical: preclinical studies showed↓amyloid burden and improved cognition in mouse models; bexarotene AD trial failed to show consistent Aβ clearance; pioglitazone TOMMORROW study terminated for futility, indicating that translation requires stage selection and APOE-aware enrichment	Entry: APOE genotype; plasma/CSF lipidomics. PD: ABCA1/ABCG1 expression; lipidomic shifts	[69, 157, 158]
UBX1325 (foselutoclax), BCL-xL inhibitor senolytic**	Eliminates senescent vascular/endothelial cells→restores barrier integrity;↓SASP (senescence axis)	Preclinical diabetic models; Phase I (advanced DME); Phase II (nAMD)	Non-AD clinical: preclinical diabetic models showed↓vascular leakage and preserved retinal function; phase I in DME was safe/tolerable; phase II in nAMD did not meet the primary non-inferiority endpoint, although ~ 50% required no supplemental anti-VEGF at 24 weeks; DME phase IIb ongoing	Entry: senescent endothelial burden. PD:↓vascular leakage; reduced injection burden; retinal structure/function stabilisation in subsets	[148, 159]
ANX005 (anti-C1q mAb)**	Neutralises C1q→blocks classical complement cascade and CR3-mediated synaptic pruning (complement axis)	Clinical: Phase I (GBS); preclinical: CNS synapse-loss models	Non-AD clinical + preclinical CNS models: phase I in GBS showed safety and dose-dependent C1q suppression in serum/CSF; preclinical CNS models showed reduced microglial engulfment, synapse preservation, and cognitive improvement	Entry: complement markers (C1q/C3). PD: serum/CSF C1q suppression; synaptic markers (PSD95, VGLUT2); microglial phagocytosis; behavioural benefit	[111, 112, 160]
Compstatin-family C3 inhibitors (pegcetacoplan, AMY-101)**	Peptide C3/C3b binders→block convertase/amplification; limit maladaptive microglial pruning (complement axis)	Phase II, intravitreal (GA-AMD); Phase III, SC (PNH); no AD trials	Non-AD clinical: GA-AMD trials showed 20–29% slower lesion growth at 12 months with increased exudative AMD risk; PNH trials showed superiority to eculizumab for haemoglobin and transfusion independence; AD relevance remains inferential	Entry: C3/C3b fragments; CFH/C3 genetics. PD: serum/CSF C3 suppression; GA lesion growth; Hb/transfusion metrics; PD in AD: synaptic integrity, microglial states	[113, 161–163]
Deferiprone (oral Fe²⁺ chelator)	Brain-permeable chelation→lowers labile Fe²⁺ (iron/ferroptosis axis)	Phase II RCT: amyloid-confirmed MCI/early AD (12 mo; n = 81)	AD clinical phase 2: target engagement positive with↓hippocampal QSM, but clinical outcome negative with accelerated cognitive decline on NTB, especially executive function; neutropenia occurred in 7.5%, suggesting global iron chelation may be detrimental in AD	Entry: amyloid-positive early AD; baseline QSM iron. PD: hippocampal QSM change; NTB; volumetrics; neutrophils	[91]
NT-0796 (brain-penetrant NLRP3 inhibitor; prodrug→NDT-19795)*	Selective NLRP3 assembly inhibitor; selective vs. NLRC4 (inflammasome axis)	Preclinical PBMC/whole-blood; mouse brain; Phase 1 (HV) done; Phase 1b/2a in PD; no AD data	Preclinical + non-AD clinical phase 1: sub-nanomolar IL-1β potency in PBMC; whole-blood IC₅₀ ~6.8 nM; good permeability; mouse brain/blood ratio ~ 0.79 after i.v. dosing; phase 1 announcement reported favourable PK/PD with CNS penetration; AD efficacy unproven	Entry: inflammation-high phenotype; plasma/CSF IL-1β/IL-18; ex vivo ASC specks. PD:↓IL-1β (± IL-18); inflammasome panels	[107]
DFV890 (small-molecule NLRP3 inhibitor)*	Direct NLRP3 inhibitor; stabilises inactive conformation (inflammasome axis)	Non-AD indications (e.g., COVID-19); no AD data	Non-AD clinical: early RCT in COVID-19 was negative for the primary endpoint versus standard of care, with signals of earlier viral clearance and improved clinical status; generally well tolerated; AD-specific CNS efficacy not established	Entry: indication-specific inflammation. PD: class-typical↓IL-1β/IL-18; AD-specific PD not established	[164]
Preclinical concepts (mechanistic proof)
Radical-trapping ferroptosis inhibitors (Fer-1, Lip-1; phenothiazine analogues)	Scavenge lipid-peroxyl radicals; interrupt PUFA-PL peroxidation; complement GPX4/FSP1–CoQ10 systems (ferroptosis axis)	Cell/animal models; no AD clinical trials	Preclinical/tool compounds: consistent protection from erastin/RSL3-induced ferroptosis; phenothiazines sometimes ≥ Fer-1/Lip-1; efficacy context-dependent; no AD clinical efficacy data	Entry: ACSL4↑, PUFA-PL enrichment. PD:↓LPO (C11-BODIPY, MDA/4-HNE),↓ROS, restored ΔΨm; GPX4/xCT levels	[165, 166]
ACSL4–LPCAT3–15-LOX enzymatic module	ACSL4/LPCAT3 load PUFAs into PLs; 15-LOX catalyses peroxidation→ferroptosis (ferroptosis axis)	Neuronal/dopaminergic models; brain slices	Preclinical mechanistic module: AA + Fe synergise to induce ferroptosis; ACSL4 or 15-LOX inhibition blocks lipid peroxidation and cell death, supporting actionable nodes for ferroptosis-axis testing	Entry: iron burden; PUFA composition. PD: MDA/4-HNE; ACSL4/LOX dependence	[88, 165]
Tau-lipid peroxidation crosstalk: emerging concepts	Pathologic tau drives LPO and ferroptosis; TRx0237 or Fer-1 rescue (tau–ferroptosis crosstalk)	Neuronal and oncology models	Preclinical/emerging concept: supportive evidence suggests pathologic tau intensifies lipid peroxidation and ferroptosis-like stress; AD-specific validation remains limited and should not yet be overinterpreted	Entry: tau aggregation state. PD: LPO, mitochondrial morphology, pharmacologic rescue	[32, 92]
Enabling biomarkers & tools (trial design / companion)
Plasma p-tau217 (± p-tau217/Aβ42)	Aβ-driven tauopathy marker (biomarker axis)	Cognitively unimpaired, MCI, early AD	Biomarker/enabling; human clinical validation: AUC ≈ 0.95–0.96 for Aβ-PET/tau-PET; predicts decline and MCI→AD conversion	Entry: plasma p-tau217/ratio. PD: longitudinal p-tau217, cognition	[76, 167–170]
Trial enrichment (AADvac1, post-hoc p-tau217+)	Biomarker-enriched subgroup (enrichment approach)	Mild–moderate AD, Phase II	Biomarker/enrichment strategy: p-tau217 + subgroup showed↓NfL (~ 56%) and↓GFAP (~ 73%) versus placebo; no cognitive efficacy	Entry: baseline p-tau217. PD: NfL, GFAP	[168]
Anti-p-tau217 antibody (mAb2A7)	Passive immunotherapy (tau axis)	Human AD tissues + PS19	Human tissue + preclinical: human AD tissue signal correlated with atrophy/decline; in PS19 mice, treatment reduced tau pathology/atrophy and improved cognition	Entry: CSF/plasma p-tau217. PD: tau pathology, atrophy, cognition	[171]
ASC specks (inflammasome biomarker)	Aggregated ASC indicates NLRP3 activity	Inflammatory cohorts; CNS injury models	Biomarker/enabling: elevated in serum; correlates with IL-18/IL-6; responsive to IL-1 blockade; AD validation pending	Entry: serum/CSF ASC. PD: IL-1β/IL-18; caspase-1	[172–174]
TREM2-PET (⁶⁴Cu-NODAGA-ATV:4D9/14 D3)	Antibody tracer to TREM2 on DAM microglia; ATV enables BBB transcytosis	5xFAD; TfR mice; AppSAA; human AD sections	Biomarker/enabling imaging tool: specific cortical/hippocampal uptake; blockable signal; microglial specificity confirmed; autoradiography validates human AD tissue binding; clinical AD imaging validation pending.	Entry: TREM2 expression/occupancy. PD: tracer uptake; co-registration with Aβ- and tau-PET; companion for TREM2 agonists	[175]
INK-ATTAC (p16 + cell-clearance tool)	Inducible Casp8 ablation of p16 + cells	3xTg-AD	Preclinical genetic tool: p16 + cell clearance reduced AT8 tau and microglial activation; astrocyte effects were cohort-dependent.	Entry/PD: AT8, Aβ load, Iba1	[144]
B1R/B2R-TRIOZAN™ nanoparticle + anti-C1q (ANX005)	Bradykinin-receptor–targeted nanoparticles transiently open BBB;↑hippocampal/entorhinal anti-C1q delivery	Tg-SwDI; IV and intranasal	Preclinical delivery platform: IV delivery produced ~ 4–12× brain exposure versus free antibody; intranasal delivery produced ~ 3–5× exposure; liver exposure was similar; brain bioavailability improved	Entry: regional distribution. PD: PK (organ/serum ratios)	[176]

The table groups programmes that differ in maturity (authorised/late-phase clinical candidates vs. mechanistic tool compounds vs. enabling biomarkers) because the review’s intent is axis-mapping rather than ranking therapeutic readiness. The “Evidence maturity / Status” column provides a short label for each entry, such as AD clinical, non-AD clinical, preclinical, tool compound, or biomarker/enabling evidence. Each entry illustrates where along the tri-axial framework a given approach is hypothesised to operate and which candidate enrichment or pharmacodynamic readouts may help interpret it. These biomarkers are not intended as validated axis-specific clinical classifiers; clinical readiness should be inferred from the Evidence maturity / Status column, not from the table layout

Abbreviations: AA arachidonic acid, AD Alzheimer’s disease, Aβ amyloid-β, ARIA amyloid-related imaging abnormalities, ATV antibody transport vehicle, BBB blood–brain barrier, CR3 complement receptor 3, DAM disease-associated microglia, DME diabetic macular edema, FTD-GRN frontotemporal dementia with GRN mutations, GA-AMD geographic atrophy secondary to age-related macular degeneration, GBS Guillain–Barré syndrome, GSDMD gasdermin D, Hb haemoglobin, HV healthy volunteers, LPO lipid peroxidation, LTP long-term potentiation, nAMD neovascular age-related macular degeneration, NfL neurofilament light, NI non-inferiority, NTB neuropsychological test battery, PD pharmacodynamic, PET positron emission tomography, PGRN progranulin, PNH paroxysmal nocturnal haemoglobinuria, PUFA-PL polyunsaturated-fatty-acid phospholipid, RA rheumatoid arthritis, RCT randomised controlled trial, RTA radical-trapping antioxidant, SAE serious adverse event, SASP senescence-associated secretory phenotype, SC subcutaneous, SOC standard of care, WMH white-matter hyperintensity, QSM quantitative susceptibility mapping

** indicates active human or near-clinical evaluation, * indicates preclinical-only or tool compound

Anti-amyloid immunotherapy provides a useful test case for this framework. Lecanemab and related antibodies have produced the first clinically meaningful disease-modifying signals in AD trials, but their effects are not purely passive amyloid removal. Available evidence suggests that Fc-mediated microglial engagement and plaque clearance contribute to therapeutic activity, with TREM2-coupled mechanisms participating in this response [7]. ARIA, the dominant safety concern, also fits a multi-axis interpretation. It is not a generic inflammatory complication. Instead, ARIA appears to involve perivascular immune engagement, complement activation, vascular remodelling, and APOE ε4-sensitive lipid and cerebrovascular biology [6, 8, 177]. APOE ε4 carriers have the highest ARIA risk, consistent with lipid-axis and vascular vulnerability. Complement deposition and microglia-mediated vascular remodelling connect ARIA to the inflammation/complement axis. ARIA-H and microbleeds intersect with iron-related vascular injury, although they should not be treated as direct evidence of ferroptosis. This interpretation supports risk stratification by APOE genotype and baseline vascular imaging burden. It also supports future studies testing whether vascular support or selective complement moderation can improve the safety window of anti-amyloid therapy.

The broader implication is that combination therapy should be built around mechanistic complementarity, not accumulation of plausible agents. A strong combination design should specify four elements: what is being restored, what is being restrained, which biomarkers define the intended subgroup, and what pharmacodynamic evidence would count as success or futility. This standard discourages polypharmacy by accumulation. It favours combinations whose components address distinct rate-limiting steps and whose effects can be verified by stage-appropriate biomarker movement. The framework provides the axis-by-axis vocabulary in which such trial designs can be written.

Biomarker-informed translation and current specificity limits

A practical strength of this framework is that each axis can be linked to measurable biomarker families. These links, however, remain incomplete and should not be interpreted as axis-specific diagnostic assignments. The lipid axis may be informed by APOE genotype and plasma or CSF lipidomics, including ceramide, sphingomyelin, and neutral-lipid signatures [73, 74]. The iron/ferroptosis axis may be supported by QSM-MRI, CSF ferritin, and exploratory lipid-peroxidation measures, including MDA- and 4-HNE-related readouts [79–83, 89, 90]. The inflammation/complement axis may be assessed using IL-1β/IL-18, ASC-related measures, C1q/C3, and microglial PET ligands, including TSPO and emerging CSF1R- or TREM2-targeted tracers [175, 178, 179]. Regulatory-node status and glial amplification may be followed with sTREM2, progranulin-related measures, GFAP, and YKL-40 [25, 136–138, 180]. Importantly, GFAP, YKL-40, ferritin, and TSPO imaging are not sufficiently specific to distinguish among the proposed pathological axes when used alone. They should therefore be used as components of multimarker panels rather than as stand-alone clinical classifiers.

These biomarkers are not interchangeable, but they are also not fully axis-specific. Plasma p-tau217 and the p-tau217/Aβ42 ratio primarily define the Aβ–tau disease context and are increasingly useful for diagnosis, staging, and risk prediction [76, 167–170, 181]. They are not microglial pharmacodynamic markers. Conversely, increases in GFAP, YKL-40, sTREM2, ferritin, or TSPO signal may indicate glial activation, inflammatory tone, iron-storage biology, or microglial ligand binding. However, these changes do not by themselves distinguish protective adaptation from lipid-handling failure, complement-mediated pruning, ferroptotic stress, or senescence-associated amplification. The biological meaning of each marker depends on the rest of the panel and on disease stage. The aim of axis-linked biomarker use is therefore not definitive clinical classification, but reduction of biological ambiguity through convergent evidence.

Future microglia-directed trials should specify four elements before efficacy is tested at scale. First, the candidate axis or regulatory node being targeted, without assuming that a biomarker identifies a dominant axis with clinical certainty. Second, the disease stage and enrichment biomarkers used for entry. Third, at least one direct pharmacodynamic marker of target engagement. Fourth, at least one disease-context marker situating the intervention relative to Aβ, tau, neurodegeneration, vascular burden, and clinical stage. Without this structure, even biologically strong interventions may be difficult to interpret, because failed mechanism cannot be distinguished from failed implementation. This four-element specification is not a formal regulatory requirement. It is a methodological discipline for biomarker-informed translation: interventions that cannot answer these questions should undergo further pharmacodynamic clarification before definitive clinical evaluation.

Discussion

The main contribution of this review is conceptual, but its intended use is practical. By organising AD-related microglial dysfunction around checkpoint collapse and three interacting effector axes, the framework explains why lipid droplets, iron retention, complement activation, white-matter injury, astrocyte reactivity, and senescence frequently co-occur in AD. These features should not be collapsed into a single undifferentiated category of “neuroinflammation”. They reflect connected but distinct biological processes. The lipid axis can restrict lysosomal flux and ferritinophagy. The iron/ferroptosis axis can generate oxidised-lipid signals that prime inflammatory pathways. The inflammation/complement axis can feed back onto lipid handling through astrocyte–microglia crosstalk. This organisation gives the field a way to move from pathway catalogues towards testable, biomarker-informed hypotheses.

This framing also helps explain why earlier immunomodulatory strategies have often disappointed in AD. NSAIDs, IVIG, and broad anti-inflammatory approaches were not necessarily irrational, but they were biologically under-specified. Most lacked stage selection, biomarker-informed enrichment, direct CNS pharmacodynamic confirmation, and a clear strategy for preserving protective microglial functions while restraining maladaptive ones [48–50]. Newer approaches, including TREM2 agonists, selective NLRP3 inhibitors, complement-directed therapies, senolytics, brain-penetrant iron modulators, and cGAS–STING inhibitors, are mechanistically more precise. They should not, however, be assumed to succeed on target engagement alone. The INVOKE-2 and deferiprone experiences illustrate the same lesson from different directions: a plausible target and measurable biological engagement do not guarantee clinical benefit [91, 154].

The framework should therefore be used as a translational scaffold, not as a claim that AD microglial heterogeneity has been solved. The same caution applies to biomarker interpretation. Current glial, iron, inflammatory, and imaging markers can support cohort enrichment and pharmacodynamic monitoring, but they do not yet provide reliable axis-specific classification for individual patients. TREM2 is not uniformly protective. Complement is not uniformly harmful. Senolytics are not automatic upstream solutions. cGAS–STING inhibition is not a universal anti-inflammatory lever. The same pathway may be compensatory in one stage and maladaptive in another. This stage dependence is the central reason for linking each axis to biomarkers, pharmacodynamic readouts, and disease context. The value of the tri-axial framework is not that it simplifies AD into three pathways, but that it makes the assumptions behind therapeutic design explicit.

Limitations and future directions

Several limitations require explicit acknowledgement. First, the causal ordering among the three axes is unlikely to be uniform across patients. In some brains, white-matter injury, vascular dysfunction, and myelin burden may drive lipid and iron stress before cortical inflammatory amplification becomes dominant. In others, plaque-associated inflammasome activation may emerge earlier. The framework specifies connections among axes, but it does not identify the rate-limiting axis in a given patient. That determination will require longitudinal biomarker mapping.

Second, many relevant biomarkers remain probabilistic rather than cell-state or axis specific. Current microglial PET ligands, including TSPO imaging, do not yet provide clean resolution of microglial functional states. Astrocyte-facing markers such as GFAP do not distinguish protective adaptation from maladaptive crosstalk on their own. Similarly, sTREM2, YKL-40, ferritin, and lipidomic measures gain interpretive value only when read in relation to Aβ, tau, neurodegeneration, vascular burden, and disease stage. Axis-linked biomarker panels can reduce ambiguity, but they cannot eliminate it. They should not be used as validated tools for assigning individual patients to discrete pathological axes in current clinical practice.

Third, species differences remain a major constraint. Human and mouse microglia differ substantially at the transcriptomic level, including in disease-associated programmes. Human-enriched AD-associated microglial subpopulations are not fully captured by canonical mouse DAM signatures [18, 21, 43]. Mechanistic findings from rodent models should therefore be treated as hypotheses for human validation rather than as settled human biology. This point is especially important for therapeutic areas that remain mainly preclinical, including ferroptosis modulation, senolysis, complement targeting, and cGAS–STING inhibition.

These limitations define the next priorities. Longitudinal multimodal studies should jointly profile Aβ and tau biomarkers, QSM, white-matter measures, lipidomics, inflammatory panels, astrocyte markers, and emerging microglial-state imaging tools. The aim should be to identify axis dominance at the individual-patient level, not to assign a single dominant mechanism at the cohort level. Therapeutically, the most useful next step is likely to be adaptive, biomarker-enriched trial design. Such trials should include direct CNS target-engagement measures, pre-specified pharmacodynamic milestones, and explicit go/no-go criteria. At this stage, the value of the present synthesis is to clarify what should be measured, which mechanisms may be combined, and how success or futility should be interpreted.

Conclusions

A microglia-centred view of Alzheimer’s disease is most useful when it moves beyond generic neuroinflammation and addresses three stricter questions: which regulatory nodes are failing, which pathological axis is most likely rate-limiting, and which biomarkers can verify target engagement in vivo. The tri-axial lipid–iron/ferroptosis–inflammation/complement framework proposed here is intended to answer these questions in a mechanistically coherent, clinically testable, and appropriately critical way.

The main practical implication is not that any single pathway has emerged as the solution. Rather, stage-aware and biomarker-informed combination therapy can now be framed more systematically. Early disease may require restoration of regulatory-node control with selective inflammasome restraint. Intermediate disease may require improved lipid handling and ferroptosis control. Later disease may require complement moderation, senescence-directed strategies, or both. Whether these approaches succeed clinically will depend on careful enrichment, direct pharmacodynamic verification, and a clear distinction between biological plausibility and demonstrated efficacy. The value of this framework is therefore practical: it clarifies what is known, what remains uncertain, and how neuroimmune interventions may need to be deployed differently across the AD continuum.

Abbreviations

Aβ

Amyloid-β

AD

Alzheimer’s disease

APOE

Apolipoprotein E

ARIA

Amyloid-related imaging abnormalities

ASC

Apoptosis-associated speck-like protein containing a CARD

BBB

Blood–brain barrier

CNS

Central nervous system

CSF

Cerebrospinal fluid

DAM

Disease-associated microglia

FSP1

Ferroptosis suppressor protein 1

GFAP

Glial fibrillary acidic protein

GPX4

Glutathione peroxidase 4

LDAM

Lipid-droplet–accumulating microglia

MCI

Mild cognitive impairment

NLRP3

NOD-like receptor family pyrin domain containing 3

PET

Positron emission tomography

PGRN

Progranulin

QSM

Quantitative susceptibility mapping

SASP

Senescence-associated secretory phenotype

sTREM2

Soluble TREM2

TREM2

Triggering receptor expressed on myeloid cells 2

WMH

White-matter hyperintensity

YKL-40

Chitinase-3-like protein 1

Authors' contributions

JZ conceived the review, performed the literature search and synthesis, prepared the initial draft, and generated the figures and table. NY and PZ assisted with literature checking and manuscript organisation. XZ supervised the work, provided conceptual guidance, critically reviewed and revised the manuscript, and serves as the corresponding author. All authors read and approved the final manuscript.

Funding

This work was supported by the American Heart Association Career Development Award 24CDA1269588.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.