THE CELLULAR ARCHITECTURE OF COLLAPSE

Abstract

Its companion paper, The Temporal Architecture of Collapse, resolved Alzheimer's disease in time: a three-phase front of failure crossing the brain over half a century, from a small nucleus in the pons in the third decade to the inhibitory scaffolding of the cortex in the eighth, joined by two mechanistically specified bridges. That paper answered the question when. The present paper answers the questions which and how many. It is a census — an enumeration, cell type by cell type and synapse class by synapse class, of exactly what is destroyed across the arc, in what order, in what number, and through which molecular chain. Where the temporal architecture is a chronology, the cellular architecture is a casualty list.

The list is the argument. Alzheimer's disease does not thin the brain uniformly, as a fog dims a city; it demolishes a specific and identifiable set of cellular structures while leaving their neighbours, sometimes microns away, untouched. The human brain holds on the order of eighty-six billion neurons and something near one hundred and fifty trillion cortical synapses, and the disease destroys a minute, exquisitely selected fraction of them in a stereotyped sequence. We document that sequence quantitatively: the noradrenergic neuron of the locus coeruleus, whose population erodes by roughly a third before a patient ever complains and by some sixty per cent by end-stage; the cholinergic projection neuron of the basal forebrain, depleted by more than three quarters; the layer-II stellate cell of the entorhinal cortex, lost by sixty per cent in the mildest clinically detectable disease and by ninety per cent at its end; the CA1 pyramidal neuron, halved; and beneath all of these the synapse itself, whose loss — not the plaque, not the tangle — remains the single strongest structural correlate of dementia ever measured.

Two principles emerge from the count. The first is selectivity with a signature: the neurons that die are not a random sample but share a phenotype — extreme and sustained metabolic demand, an extensive axonal arbor or an outsized connectivity load, high oxidative exposure, and dependence on a protective envelope (the neuromelanin of the coeruleus, the perineuronal net of the interneuron) whose loss removes the last buffer between the cell and its own activity. The second is the primacy of the synapse over the soma: across every phase, functional synaptic disconnection precedes and exceeds frank cell death, the count of lost neurons exceeds by manyfold the count of tangles that supposedly killed them, and the disease's terminal event is not the death of the parvalbumin interneuron — whose somatic loss is genuinely contested — but the digestion of the perineuronal net and the silencing of the perisomatic synapse that together collapse the gamma rhythm on which cognition runs.

We assemble these counts from the primary stereological and synaptic-quantification literature, state the confidence and the method-dependence of each, and decline to launder uncertainty into false precision: where a baseline is method-dependent we give the range, and where a loss is contested we say so. The result is the temporal architecture rendered in the currency of cells — a disease one can, at last, count. And what can be counted can be measured, staged, and defended.

I. The Question of Number

What the field has counted, and what it has not

For four decades the quantitative energy of Alzheimer's research has been spent counting two things: plaques and tangles. The neuritic plaque and the neurofibrillary tangle are countable, stainable, stageable; they are the substrate of Braak staging and of CERAD scoring and of the entire neuropathological apparatus by which the disease is confirmed at autopsy. They are also, as measures of the thing that matters, strangely disappointing. The correlation between plaque burden and the severity of dementia in life is weak; the correlation between tangle burden and dementia is better but imperfect; and the most famous single finding in the quantitative neuropathology of the disease is that neither is the best predictor of how demented a patient was. The best predictor is the loss of synapses (Terry et al., 1991; DeKosky & Scheff, 1990). The field has spent a generation counting the lesions that are easiest to see rather than the casualties that best explain the illness.

This paper proposes to count the casualties. It takes as given the temporal architecture established in the companion synthesis — three phases, two bridges, one homeostatic system failing across the brain in sequence — and asks, at each station of that arc, a set of deliberately concrete questions. Which cell, defined by which molecular markers and which anatomical address? How many of them are there in a healthy brain, and by what method do we know? What does each connect to, and across how many synapses? How many are lost, at what stage of the disease, and with what confidence? And by what mechanistic chain — which named sequence of molecules — is each population brought down? The answers, assembled, constitute the cellular and synaptic census of the disease.

The disease is selective, countable, and ordered

Three properties of the census deserve stating at the outset, because they are what make the exercise worth performing rather than a mere tabulation.

The disease is selective. It does not kill neurons at random. Of the eighty-six billion neurons in the brain (Azevedo et al., 2009), the overwhelming majority — the granule cells of the cerebellum, the neurons of the primary sensory and motor cortices, the great mass of subcortical grey — are spared until very late or entirely. The casualties are drawn from a short list of specific, nameable populations, and the membership of that list is itself a clue to mechanism.

The disease is countable. Each of those populations has been subjected, over the past thirty years, to design-based stereology — the unbiased counting methods (the optical fractionator, the Cavalieri estimator) that replaced the biased two-dimensional profile counts of the older literature and made absolute neuron numbers trustworthy for the first time. We therefore possess, for the load-bearing populations, real numbers and real loss fractions, not impressions.

The disease is ordered. The populations do not fall together. They fall in a sequence — the same rostro-limbic-cortical sequence that the temporal architecture describes — and the order is reproducible enough that the identity of the dying cell is itself a clock. To know which population is currently being lost is to know, within a decade, where on the arc a given brain has travelled.

The census that follows is organized along that order. After a note on the instruments by which the counting is done, we proceed station by station — from the first neuron to fall to the last synapse to be silenced — giving for each its identity, its standing number, its synaptic commerce, its documented toll, and the mechanistic chain that destroys it.

II. The Instruments of the Census

How a neuron is counted

A census is only as trustworthy as its method, and the history of neuron counting in Alzheimer's disease is in large part a history of correcting for method. The older literature counted profiles — neuronal cross-sections visible in a thin histological section — and converted them to densities. This is biased in ways that matter: large cells are over-counted relative to small ones, shrinkage distorts density, and a density is not a number unless one also knows the volume it inhabits. A disease that shrinks the brain can appear to raise a neuronal density even as it destroys neurons, simply because the remaining tissue has contracted around them.

The correction was the design-based, or unbiased, stereology developed in the 1980s and 1990s: the optical fractionator, which samples a known fraction of a structure's volume with three-dimensional counting frames and counts whole cells under strict, geometry-independent rules; and the Cavalieri estimator, which measures the structure's total volume without assuming its shape. The combination yields an absolute neuron number with a calculable error and no dependence on cell size or tissue shrinkage. Essentially every load-bearing count in this paper — the entorhinal figures of Gómez-Isla, the hippocampal figures of West, the coerulean trajectory of Theofilas — rests on these methods, which is why those figures can be trusted where the density estimates of the 1970s cannot.

Why baselines are ranges, not points

Even unbiased stereology yields a range for a baseline, not a point, and the honest census reports the range. The clearest example is the very first population on our list. The locus coeruleus is reported to contain about twenty thousand neurons per side when one counts only the tyrosine-hydroxylase-positive, neuromelanin-bearing noradrenergic cells (Manaye et al., 1995), but nearer fifty thousand per side when one counts every medium-to-large neuron in the nucleus, pigmented and unpigmented alike (Theofilas et al., 2017). Neither number is wrong; they count different things. The companion paper's "roughly fifty thousand cells" sits at the upper, inclusive end of this range. Throughout this census we give such ranges explicitly and name the inclusion criterion, because a baseline stated as a single false-precise figure is a small dishonesty that compounds into a large one when loss fractions are applied to it.

How a synapse is counted

The synapse is harder to count than the soma, and the difficulty is worth respecting, because the synapse is the unit that matters most. Three methods underwrite the figures in this paper. Electron microscopy counts synapses directly, by their ultrastructure — the gold standard, laborious, and the basis of the biopsy counts of DeKosky and Scheff and the stereological synapse counts of the Scheff series. Immunohistochemical densitometry quantifies a presynaptic protein, classically synaptophysin, as a proxy for synaptic density — the basis of the Terry correlations. Array tomography, the most recent, reconstructs synapses in three dimensions from serial ultrathin sections labelled for pre- and post-synaptic markers, and has confirmed that synapse loss concentrates in the immediate vicinity of pathology. The convergence of three independent methods on the same conclusion — that synapse loss is early, regional, and the best structural correlate of cognition — is what gives that conclusion its unusual strength.

The ledger

The census that follows is summarized in the ledger below; the sections that follow it defend each line in turn, giving for each population its molecular identity, its synaptic commerce, and the mechanistic chain that destroys it. The ledger is deliberately conservative — baselines are given as method-dependent ranges, and where a loss is contested the entry says so.

Population	Standing number	Documented loss	Earliest stage detected	Primary source
Locus coeruleus (noradrenergic)	~20,000–50,000 / side	~30% by MCI; ~60% in AD	Braak 0–I (pretangle, 3rd decade)	German 1992; Kelly 2017; Theofilas 2017
Nucleus basalis (Ch4, cholinergic)	not cleanly established	>75%	early symptomatic	Whitehouse 1982
Entorhinal cortex, layer II (stellate)	~650,000	~60% (very mild) → ~90% (severe)	CDR 0.5 (very mild)	Gómez-Isla 1996
Hippocampal CA1 (pyramidal)	~14 million	~48–68%	symptomatic (spared preclinically)	West 1994 / 2004
Association cortex (superior temporal sulcus)	—	>50%; exceeds tangle count manyfold	with disease duration	Gómez-Isla 1997
Parvalbumin interneuron (hippocampus)	—	~60% immunoreactivity (contested; largely functional)	late	Brady & Mufson 1997
Perineuronal net (aggrecan–brevican coat)	enwraps ~60–80% of PV cells	extensive, plaque-proportional; precedes PV loss	Phase III	Crapser 2020
Neocortical synapses (the unit)	~150 trillion	~18% (MCI) → ~55% (CA1, mild AD); best correlate of dementia	MCI	Terry 1991; Scheff 2006 / 2007

With the instruments stated, the census proper begins.

III. The First to Fall — The Coerulean Neuron

Identity

The first neuron to show Alzheimer-type pathology is among the most distinctive in the brain. The noradrenergic neuron of the locus coeruleus is defined by its catecholaminergic enzymatic machinery — tyrosine hydroxylase and dopamine-β-hydroxylase, the synthetic enzymes for noradrenaline, and the noradrenaline transporter that recovers it — and by a feature visible to the naked eye: the dense blue-black pigment neuromelanin, the polymerized by-product of catecholamine metabolism, which accumulates in these cells across the lifespan and gives the nucleus its name (locus coeruleus, the blue place). Many of these neurons co-express the neuropeptide galanin. The cell is autonomously pacemaking, firing tonically at roughly one to three hertz throughout waking life and modulating that rate with arousal, and it is this ceaseless activity, sustained over decades, that sets the metabolic terms of its destruction.

The standing population

Counted as tyrosine-hydroxylase-positive, neuromelanin-bearing cells, each locus coeruleus holds on the order of twenty thousand neurons in young adulthood (Manaye et al., 1995); counted inclusively, nearer fifty thousand (Theofilas et al., 2017). It is, by either count, a tiny nucleus — a few tens of thousands of cells out of eighty-six billion, less than one neuron in a million — and yet from it long, thin, sparsely myelinated axons ascend, branching into varicose fibres that reach virtually the entire forebrain.

Synaptic commerce

The coerulean neuron's mode of connection is as distinctive as its chemistry. It signals very largely by volume transmission: its varicose boutons release noradrenaline not into discrete one-to-one synaptic clefts but diffusely into the parenchyma, bathing wide territories in noradrenergic tone. A single neuron thereby influences enormous numbers of downstream cells, including the microglia, which express the β2-adrenergic receptor and listen to coerulean tone as a standing brake on their own activation. The relevant "synapse," in other words, is partly a chemical atmosphere rather than a wire — which is exactly why the loss of these few thousand cells deregulates territories vastly larger than their number would suggest.

The toll

The coerulean census is the clearest demonstration in the disease that volume loss, functional loss, and cell death are three different curves. Stereology across Braak stages shows that the nucleus loses volume early and almost linearly — on the order of eight per cent per Braak stage, with roughly a quarter of its volume gone before clinical onset — while its actual neuron number is comparatively preserved through the early-to-middle stages and only declines significantly from Braak stage III onward, accelerating thereafter (Theofilas et al., 2017). Counted across the clinical continuum, the population falls by roughly thirty per cent already at the transition into mild cognitive impairment and by a further quarter into established disease (Kelly et al., 2017); disease-specific stereology places the loss in established Alzheimer's at around sixty per cent, concentrated in the rostral and middle thirds of the nucleus with relative caudal sparing (German et al., 1992). The reconciliation of these figures — volume shrinking from Braak 0, number holding until Braak III, then collapsing — is itself a finding: the cell is sick, atrophic, and pathology-bearing for decades before it dies, which is precisely the "clinically silent ignition" the temporal architecture predicts.

The mechanistic chain

Why this cell first? Because it lives perpetually at the ceiling of its metabolic capacity, with no reserve to spend on the slow accumulation of unrepaired damage. The destructive chain is the one detailed for Phase I in the companion work: hyperactivation of PARP-1 in response to accumulating DNA damage depletes NAD⁺, starving the sirtuins and crippling mitochondrial biogenesis; the obstruction of mitochondrial protein import as Aβ binds the import channel TOM40 (Devi et al., 2006) leaves damaged mitochondria unrepairable; and the PINK1/Parkin mitophagy that should clear them is itself defeated, both because PINK1 import depends on the channel that is now blocked and because the autophagy–lysosomal apparatus downstream is failing. The terminal morphology is the flower-like, autolysosome-swollen neuron Nixon and colleagues named PANTHOS, and the plaque that forms where such a neuron ruptures is its gravestone, not its cause (Lee et al., 2022). The coerulean neuron, in short, is killed by the failure of its own quality control under a metabolic load no other cell sustains for so long.

The cholinergic twin

The coeruleus does not fall alone. Its companion in early vulnerability is the cholinergic projection neuron of the basal forebrain — the large neuron of the nucleus basalis of Meynert (Ch4) whose ascending acetylcholine supplies the cortex and hippocampus. It shares the coerulean phenotype in its essentials — a neuromodulatory rather than point-to-point mode of action, and an early, severe vulnerability — and adds a vast, energetically expensive axonal arbor of its own. Its toll is the most severe of any single projection population in the disease — a profound and selective degeneration exceeding seventy-five per cent of the neurons (Whitehouse et al., 1982), the finding that launched the cholinergic hypothesis and the cholinesterase-inhibitor therapeutics that remain in use. Two ascending modulatory systems, both extravagant, both diffuse, both first to fall: the pattern is already visible in the disease's opening move.

IV. The Cortical Bridgehead — The Entorhinal Stellate Cell

Identity and the gateway it guards

When the disease crosses from brainstem to cortex — the transition the companion paper assigns to the first bridge, the coerulean projection carrying both the withdrawal of the noradrenergic brake and a stream of templated tau — it makes landfall at a specific and famous cell: the stellate (and pyramidal) projection neuron of layer II of the entorhinal cortex. This is the origin of the perforant path, the great fibre bundle that carries cortical information into the hippocampus, and it is therefore the literal gateway of memory: essentially everything the neocortex tells the hippocampus passes through these cells. Many of them are reelin-expressing. Their position at the head of the hippocampal circuit makes them the single most consequential small population in the disease.

The standing population and its toll

Layer II of the entorhinal cortex holds roughly six hundred and fifty thousand neurons in the healthy aged brain (Gómez-Isla et al., 1996). The figure that made this cell famous is its loss fraction in the mildest clinically detectable disease. In individuals at a Clinical Dementia Rating of 0.5 — the threshold of detectable impairment — layer-II entorhinal neurons are already reduced by about sixty per cent; in severe disease the loss reaches roughly ninety per cent (Gómez-Isla et al., 1996). Sixty per cent of the gateway is gone before the disease is unambiguously diagnosable. No statistic in the cellular census more sharply contradicts the intuition that early Alzheimer's is a mild disease: at the level of this one critical population, "very mild" dementia is already a near-decimation.

The synaptic consequence — deafferentation of the dentate

The death of the entorhinal stellate cell is experienced by the hippocampus as deafferentation: the perforant-path synapses onto the granule cells of the dentate gyrus, which terminate in the outer molecular layer, are withdrawn as their presynaptic partners die. The Scheff series quantified this directly. In the dentate outer molecular layer — the entorhinal afferent zone — synapse numbers are significantly reduced in early Alzheimer's relative to both unimpaired and mild-cognitive-impairment brains, and even at the mild-cognitive-impairment stage the great majority of individuals already sit below the unimpaired group mean (Scheff et al., 2006). The first synaptic casualty of the cortical disease, in other words, is the very connection that lets experience enter memory.

The chain, and the count that indicts the tangle

The mechanistic chain delivering pathology to this cell is the tau-seeding arm of the first bridge: pretangle tau released from coerulean and entorhinal terminals, internalized through LRP1 and heparan-sulfate proteoglycans, templating the misfolding of the host's own tau and propagating trans-synaptically along the perforant path. But the entorhinal census also supplies the disease's most important quantitative argument against the sufficiency of tau as the proximate cause of death. In high-order association cortex, more than half the neurons are lost in Alzheimer's, and the amount of neuronal loss exceeds, by manyfold, the number of neurofibrillary tangles ever accumulated (Gómez-Isla et al., 1997). If neurons died only by becoming tangles, the two counts would track; instead the dead outnumber the tangled many times over. The tangle is part of the story, but most neurons that die in this cortex die without one — by the upstream metabolic and synaptic failures the census is tracing, with the tangle as one visible marker of a process that kills by other means.

V. The Limbic Resident — The Hippocampal Pyramidal Neuron

Identity and selectivity within the hippocampus

The hippocampus is not lost as a whole; it is lost subfield by subfield, and the selectivity is itself informative. The principal casualty is the pyramidal neuron of the CA1 field — a glutamatergic projection neuron receiving the Schaffer-collateral input from CA3 and supplying the chief hippocampal output to the subiculum and entorhinal cortex. CA1 is the hippocampal subfield most specifically vulnerable to Alzheimer's disease, and its vulnerability is disease-specific rather than age-related: in normal ageing CA1 is largely spared, so that CA1 loss is a signature of the disease and not of the years (West et al., 1994).

The standing population and its toll

The healthy CA1 field holds on the order of fourteen million pyramidal neurons (West et al., 1994). In Alzheimer's disease that population is reduced by roughly half to two-thirds — a final CA1 count near 4.4 million against some 14 million in normal ageing in the original Lancet series (a deficit of about sixty-eight per cent), and a loss of about forty-eight per cent against matched controls in the same group's later, refined stereology, with the subiculum (about twenty-four per cent) and hilus (about fourteen per cent) far less affected (West et al., 1994; West et al., 2004). The defensible statement is that CA1 loses on the order of half its pyramidal neurons in established disease, more than any other hippocampal subfield. Critically, the same group found no significant CA1 loss in preclinical disease — brains bearing Alzheimer pathology without cognitive decline — which places CA1 death on the symptomatic, not the silent, side of the arc (West et al., 2004): the neuron dies as the patient declines, not decades before.

Synaptic commerce and its loss

Before the CA1 pyramidal soma dies, its synapses are stripped, and they are stripped earliest at the input the disease attacks first. Stereological synapse counts in the CA1 stratum radiatum — the Schaffer-collateral field — show that synapse number falls by roughly fifty-five per cent in mild Alzheimer's disease relative to unimpaired controls, with the mild-cognitive-impairment stage already down by some eighteen per cent (Scheff et al., 2007). The ordering matters: synapse number in these counts correlates with cognitive performance but shows no relationship to Braak stage or APOE genotype (Scheff et al., 2006; Scheff et al., 2007), a recurring signature of the census — the synaptic casualty tracks the symptom, not the stainable lesion.

The chain

The chain that kills the CA1 pyramidal neuron is the quality-control collapse of Phase I now playing out in a second cell population, compounded by the immune attack of Phase II. The endosomal–lysosomal trafficking apparatus jams — the retromer hub fails, intraneuronal Aβ accumulates in late endosomes and multivesicular bodies — even as the disinhibited, post-homeostatic microglia of the surrounding tissue begin to tag and strip the cell's synapses by the complement and C4d pathways detailed below. The pyramidal neuron is thus caught between an intrinsic failure of its own housekeeping and an extrinsic withdrawal of synaptic support, and it is the conjunction, not either alone, that brings it down.

VI. The Custodian Turned — The Microglion

A different kind of census entry

The microglion requires a different accounting, because it is not, in the main, destroyed: it is converted. The pathology of the microglion is a change of state, not a change of number, and so its census entry quantifies a transformation rather than a loss. But it belongs at the centre of the cellular architecture for a reason the temporal companion makes plain: the microglion is the agent by which most of the disease's synaptic casualties are actually inflicted. If the neurons are the victims of the census, the microglion is, in large part, its executioner.

Identity — the homeostatic signature

The healthy microglion is defined by a transcriptional signature maintained by TGF-β/SMAD signalling: P2RY12, TMEM119, CX3CR1, SALL1, HEXB, and a larger companion ensemble (Butovsky et al., 2014). This signature is a behavioural programme, not a label — it specifies a cell that surveys, supports, and prunes with restraint. P2RY12 couples the cell to extracellular ATP and ADP gradients; CX3CR1, ligated by neuronal fractalkine, carries a tonic anti-inflammatory signal. The loss of this signature is the Phase II lesion, and it is the loss of a state, measurable as the down-regulation of these markers, that the census records here rather than a fall in cell count.

The post-homeostatic trajectories

Having exited homeostasis, the microglion enters one of a family of post-homeostatic states, each defined by its own markers. The disease-associated microglia (DAM) of Keren-Shaul and Amit pass from a TREM2-independent first stage into a TREM2-dependent second stage marked by APOE, CST7, the cathepsins CTSB and CTSD, LPL, SPP1, ITGAX, CLEC7A, and TYROBP (Keren-Shaul et al., 2017). The lipid-droplet-accumulating microglia (LDAM) of Marschallinger up-regulate PLIN2, DGAT2, and ACSL1 as their cytoplasm gridlocks with lipid, becoming at once metabolically senescent and primed for inflammasome assembly (Marschallinger et al., 2020). The dystrophic microglia of Streit, driven by iron and ferritin accumulation, show cytoplasmic beading, process fragmentation, and spheroid formation as they fragment toward terminal senescence in the company of pretangle tau (Streit et al., 2009). The receptor TREM2 sits at the collision point of these trajectories, coupling lipid and apolipoprotein sensing to phagocytic and metabolic competence (Ulland et al., 2017) — which is why it recurs at every level of the disease.

Synaptic commerce — the executioner's arms

The microglion's relevant commerce, for this census, is its capacity to remove synapses, and the effector arms are quantified and causal. Post-homeostatic microglia deposit complement C1q on vulnerable synapses, activating the classical cascade through C4 to the C3 and C4d "eat-me" opsonins that license CR3-mediated engulfment; genetic removal of C1q or C3 is protective, and complement-mediated synapse loss is demonstrable in Alzheimer models before and independently of plaque deposition (Stevens et al., 2007; Hong et al., 2016). A parallel, cell-autonomous arm requires no engulfment at all: C4d, a high-affinity ligand for the postsynaptic LilrB2 receptor, drives synaptic withdrawal directly (Werneburg et al., 2025). These are the same arms that, deployed with restraint in the homeostatic state, maintain the very synapses they now dismantle. Phase II is therefore not the acquisition of a harmful cell type but the loss of the governor on a cell the brain has always possessed — and the cost of that loss is paid, synapse by synapse, by the neurons of the preceding census entries.

VII. The Iron-Bearer — The Oligodendrocyte

Identity and its dangerous cargo

The oligodendrocyte enters the census not for its number but for its chemistry. It is the myelinating glial cell of the central nervous system, and it carries the highest iron concentration of any brain cell type — iron it requires for the lipid synthesis of myelin, and which makes it both metabolically distinctive and, in death, dangerous to its neighbours. The human neocortex contains on the order of one hundred and fifty to one hundred and eighty thousand kilometres of myelinated fibre (Pakkenberg et al., 2003), an immense surface of iron-laden membrane threading the very territory the disease attacks.

The toll and the chain — fuel for the second bridge

The oligodendrocyte's role in the cellular architecture is to supply the inorganic arm of the second bridge, the Proteolytic Turn. When oligodendrocytes die by ferroptosis — the iron-dependent, lipid-peroxidation-driven death modality — they do not die cleanly: their membranes dismantle and release redox-active ferrous iron (Fe²⁺, liberated from ferritin, from heme, from iron–sulfur clusters) into the surrounding parenchyma. The perineuronal net of the nearby parvalbumin interneuron is an avid iron sink, its sulfate and carboxylate groups binding iron tightly, and an iron-loaded net is a substrate for Fenton chemistry: Fe²⁺ and hydrogen peroxide generate the hydroxyl radical, a species so reactive that its diffusion radius is on the order of a nanometre, so that it damages whatever holds the iron that made it. The radical fragments the matrix, the fragmentation is self-amplifying, and the oligodendrocyte's death is thereby converted into the chemical demolition of the next, and final, casualty. The oligodendrocyte is not the disease's victim so much as its accelerant.

VIII. The Final Substrate — The Parvalbumin Interneuron and Its Net

Identity — the metronome of the cortex

Every thread of the architecture converges on one cell and the structure that protects it. The parvalbumin-positive (PV) fast-spiking interneuron is the metronome of cortical computation: a GABAergic interneuron that provides the perisomatic inhibition pacing pyramidal-cell firing and that generates the gamma-frequency rhythms (conventionally 30–80 Hz) on which working memory and attention depend. It is built for speed — capable of sustained firing above two hundred hertz (Hu et al., 2014) — and it constitutes roughly forty per cent of all cortical GABAergic interneurons (Rudy et al., 2011). It is also, by virtue of that firing rate, among the most metabolically demanding and oxidatively exposed neurons in the cortex, and it bears, uniquely, calcium-permeable AMPA receptors lacking the GluA2 subunit, which leave it exposed to calcium overload. These features make it the cortical analogue of the coerulean neuron: extravagant, fast, exposed — and dependent for survival on a protective envelope.

The perineuronal net — a census of a structure

That envelope is the perineuronal net, and it deserves its own enumeration, because it, rather than the cell, is the proximate casualty of Phase III. The net is a lattice of extracellular matrix condensed around the soma and proximal dendrites: a backbone of hyaluronan (synthesized by hyaluronan synthases and tethered at the membrane), decorated with the lectican chondroitin-sulfate proteoglycans — aggrecan (the obligatory, highest-density component), brevican, neurocan, and versican — cross-linked by tenascin-R and stabilized by the hyaluronan-and-proteoglycan link proteins HAPLN1 and HAPLN4 (Fawcett et al., 2019). Its chondroitin-sulfate chains carry a position-specific "sulfation code" (4-O, 6-O, and the 4,6-disulfated CS-E) that tunes its function. The majority of cortical PV interneurons are enwrapped by such a net — commonly cited at sixty to eighty per cent, and higher in some regions (Celio, 1986; Härtig et al., 1992) — and the net is at once a structural scaffold, an ion-exchange buffer for the cation fluxes of fast firing, and an antioxidant shield. Strip it, and the cell's own activity begins to poison it.

Synaptic commerce — the arithmetic of pacing

The PV interneuron's power lies in a startling divergence. A single hippocampal PV basket cell innervates on the order of fifteen hundred to twenty-five hundred pyramidal cells, placing a handful of GABAergic boutons on the perisomatic region of each; conversely, a single pyramidal soma receives on the order of sixty perisomatic terminals, of which roughly sixty per cent are of PV origin (Freund & Buzsáki, 1996; Freund & Katona, 2007). This is the anatomical basis of the metronome: a small population of fast cells, each broadcasting inhibition to thousands of principal neurons, can synchronize enormous ensembles into the gamma rhythm. It is also the reason the disease's terminal lesion is so cognitively catastrophic out of proportion to the number of cells involved — to silence one PV basket cell is to release thousands of pyramidal neurons from the timing signal that made their collective computation coherent.

The toll — and an honest accounting of what dies

Here the census must be most careful, because the literature is genuinely divided, and the division is itself the key to the whole architecture. The net is lost: in Alzheimer's disease and its models, perineuronal nets are extensively degraded in proportion to plaque burden, microglia engulf the damaged matrix, aggrecan appears within human plaques, and the impairment of net integrity precedes any loss of the PV cells themselves — and microglial depletion prevents the net loss even as plaques persist (Crapser et al., 2020). The cell, by contrast, may or may not die in large numbers. The strongest human numerical estimate reports an approximately sixty per cent reduction of PV-immunoreactive interneurons in the dentate/CA4 and CA1–CA2 subfields, with CA3, subiculum, and presubiculum spared (Brady & Mufson, 1997); but this is a count of immunoreactive cells, and a substantial and growing literature holds that much of the apparent PV deficit is a loss of parvalbumin expression and of net integrity — a functional silencing — rather than frank cell death, with PV somata relatively preserved in several models (La Barbera et al., 2024). The honest census therefore records a strong claim and a weak one: the digestion of the net is established; the death of the cell is contested, and is in significant part a functional disconnection. This distinction is not a hedge — it is the therapeutic crux, because a cell that is silenced may be reawakened, while a cell that is dead cannot.

The functional readout — gamma, and the proof of concept

What the net's loss costs, measurably, is the gamma rhythm. Causal optogenetic work established that driving PV interneurons is sufficient to generate gamma and that silencing them abolishes it (Sohal et al., 2009; Cardin et al., 2009); Alzheimer models show gamma deficits emerging early, before major plaque accumulation, in the modulation of hippocampal firing (Iaccarino et al., 2016). And the same line of work supplies the census's most striking proof that this final lesion is, for a window, functional and reversible: restoring forty-hertz activity — by optogenetic drive of PV interneurons or by non-invasive sensory flicker — reduces amyloid markedly (Aβ40 down some fifty-three per cent and Aβ42 by some forty-five per cent after a single hour of optogenetic gamma in one model, with comparable reductions from sensory entrainment) and recruits microglia to the cleared material (Iaccarino et al., 2016). That a rhythm generated by the disease's final-target cell can, when restored, reach back up the chain and clear the pathology of earlier phases is the strongest possible evidence that the parvalbumin–net system is the load-bearing node of the entire architecture.

The chain

The chain that digests the net is the three-armed Proteolytic Turn, converging on the aggrecan–brevican coat: the lipid-driven proteolytic switch (NLRP3 → IL-1β → MMP-9 and the ADAMTS proteases), the iron-catalysed Fenton chemistry supplied by the dying oligodendrocyte, and the complement priming (C1q → C4 → C4d → CR3) that marks the matrix for phagocytic stripping. Three chemistries, one structure — and when that structure is gone, the most exposed neuron in the cortex is left to face its own metabolism unbuffered.

IX. The Unit of Cognition — The Synapse Itself

The true currency of the disease

Beneath every cellular casualty in this census lies a smaller and more numerous one: the synapse. The human neocortex contains on the order of one hundred and fifty trillion synapses (Pakkenberg et al., 2003); an average cortical neuron carries some thousands of them — the textbook figure of roughly seven thousand per neuron (Drachman, 2005), with human pyramidal cells running several-fold higher. It is at this scale, not the scale of the soma, that the disease is best measured, for the synapse is the unit of cognition and, as it turns out, the unit of decline.

The single strongest correlate

The cornerstone finding of the entire quantitative neuropathology of Alzheimer's disease is that synapse loss, not plaque or tangle burden, is the strongest structural correlate of dementia severity. In the canonical series, neocortical synaptic density correlated powerfully with every cognitive measure tested, and a model combining synaptic density with regional plaque counts reached a correlation with the Dementia Rating Scale of about 0.96 — within which plaque density contributed only about a quarter of the explanatory power, the rest carried by synapse loss (Terry et al., 1991). Independently, electron-microscopic synapse counts from frontal-cortex biopsies of living patients fell with cognitive severity and tracked the Mini-Mental State Examination, with a telling compensatory signature: as synaptic density fell, the mean size of the remaining synaptic contacts rose, the cortex enlarging its surviving synapses in a doomed attempt to hold total contact area constant (DeKosky & Scheff, 1990). The brain, in early disease, is visibly trying to compensate for a loss it cannot outrun.

The order of synaptic loss

The synaptic census, like the cellular one, is ordered — and the synaptic order is nested inside the cellular order, each cell's input failing as its presynaptic partner up the chain succumbs. The earliest synaptic casualty is the perforant-path input to the dentate gyrus, withdrawn as the entorhinal stellate cell dies (Scheff et al., 2006); next the Schaffer-collateral input to CA1, falling by roughly half in mild disease as the hippocampal circuit fails (Scheff et al., 2007); and finally the perisomatic PV input to cortical and hippocampal pyramidal cells, silenced as the perineuronal net is digested and the gamma rhythm collapses (Crapser et al., 2020; Iaccarino et al., 2016). The disease, read at the synapse, is a sequence of disconnections marching through the memory circuit in the order the temporal architecture predicts — input gateway, internal relay, inhibitory timing — until the network that was a brain is a set of disconnected parts.

Synapse class	Afferent pathway	Documented change	Stage of first loss	Source
Perforant-path → dentate	entorhinal layer II → dentate outer molecular layer	most MCI brains below normal mean; reduced in early AD	MCI	Scheff 2006
Schaffer collateral → CA1	CA3 → CA1 stratum radiatum	~18% (MCI) → ~55% (mild AD)	MCI	Scheff 2007
Perisomatic inhibition → pyramidal	PV interneuron → pyramidal soma	net digestion; gamma collapse; functional silencing	Phase III	Crapser 2020; Iaccarino 2016

X. The Shape of the Demolition

Selectivity has a signature

Set the casualty list side by side and a pattern declares itself. The neurons the disease destroys first and worst — the coerulean noradrenergic cell, the cholinergic basal-forebrain cell, the entorhinal stellate cell, the CA1 pyramidal cell, the parvalbumin interneuron — are not a random draw from the brain's eighty-six billion. They share a phenotype with four recurring features. Each carries an extreme and sustained metabolic demand: the tonic pacemaking of the coeruleus, the two-hundred-hertz firing of the PV cell, the relentless throughput of the entorhinal gateway. Each supports an extensive axonal arbor or an outsized connectivity load: the brain-wide projection of the modulatory neurons, the thousands of targets of a single basket cell. Each is under high oxidative exposure, whether from catecholamine chemistry, iron, or calcium-permeable receptors. And each depends on a protective envelope — the neuromelanin of the coeruleus, the perineuronal net of the interneuron — whose loss removes the last buffer between the cell and the consequences of its own activity. Selective vulnerability is not a mystery in this census; it is a phenotype, and the disease is the systematic failure of exactly those cells that live closest to their own metabolic and oxidative edge.

The primacy of the synapse over the soma

The second pattern is the consistent precedence of synaptic over somatic loss. At every station, the synapse goes first and goes further: the dentate and CA1 synapse counts fall in mild cognitive impairment, before the somata are significantly depleted; the count of dead neurons in association cortex exceeds the count of tangles manyfold; the perineuronal net is digested before the PV cell dies, if it dies at all; and the single best correlate of dementia is synaptic density, not cell number and certainly not plaque load. The disease is, in its mechanism and in its arithmetic, a disease of disconnection punctuated by death, not a disease of death that incidentally severs connections. This is why its terminal, cognition-defining event is the silencing of the perisomatic synapse and the collapse of the gamma rhythm — a functional catastrophe that can occur, and for a window be reversed, with the cells still alive.

The convergence, counted

Both patterns point to the same address. The cellular architecture, assembled from independent stereological literatures that never set out to agree, converges on the perisomatic zone of the parvalbumin interneuron exactly as the temporal architecture and the homeostatic–matrix–synaptic synthesis do — the coerulean and cholinergic projections that fail first are the very systems that regulate it; the entorhinal and CA1 circuits that fall next are the ones that feed it; the microglion and oligodendrocyte that turn in Phase II are the agents that digest its net; and the gamma rhythm it generates is the function whose loss is dementia. One address, reached from every direction, counted in every currency.

XI. The Arithmetic of Resilience

What the resilient brain keeps

The census makes resilience legible as a set of preserved counts. There exist individuals who carry amyloid and tau burdens fully in the Alzheimer range and yet die cognitively intact, and what distinguishes their brains is not less pathology but preserved cellular and synaptic architecture: homeostatic microglia still bearing the Butovsky signature, perineuronal nets still intact around their parvalbumin interneurons, and perisomatic synapses still in place (de Vries et al., 2024). Resilience, in the currency of this paper, is the maintenance of the counts that matter — microglial state, net integrity, synapse number — in the face of a plaque-and-tangle burden that, on the old census, should have guaranteed dementia. It is the clearest confirmation that the field counted the wrong things: the resilient brain proves that one can have the lesions without the disease, provided the cells and synapses survive.

Resilience is joint, not single

And the preservation is joint. Because the homeostatic system fails as a unit — one upstream signal (TGF-β/SMAD), one molecular pivot (TREM2), one anatomical address (the PV perisomatic zone) — no single preserved layer suffices. Resilience requires homeostatic microglia and intact matrix and competent synapses together, which is why it behaves like the survival of a structure rather than the sparing of a part, and why the transition from mild cognitive impairment to dementia behaves like the crossing of a threshold rather than the descent of a slope: it is the point at which the feed-forward loop among the three layers becomes self-sustaining. Reversible by restoration before the crossing, irreversible after — and the perineuronal net around the parvalbumin interneuron is the single most compact biomarker of which side of that line a given brain is on.

XII. Falsifiable Predictions in the Currency of Number

The cellular architecture, being quantitative, generates predictions that are quantitative — about counts, ratios, orders, and rates — and several are testable with existing methods.

On the order of synaptic loss. Longitudinal or stage-stratified synaptic quantification (array tomography, synaptic PET) will find synapse loss in the fixed order the census predicts — perforant-path/dentate first, Schaffer/CA1 second, perisomatic/PV last — and the order will not vary with plaque or tangle distribution, because it is set by circuit position, not by lesion load.

On synapse before soma. Within each vulnerable population, quantified synaptic loss will reliably precede and exceed somatic loss at every stage, and the ratio of lost neurons to accumulated tangles will remain greater than one throughout — confirming death by mechanisms other than tangle formation.

On the net before the cell. In resilience and early-disease cohorts, perineuronal-net integrity around parvalbumin interneurons will predict gamma-band function and cognition better than parvalbumin cell counts do, because the load-bearing lesion is the digestion of the net and the silencing of the synapse, not the death of the soma.

On the LC dissociation. The locus coeruleus will continue to show the volume/number dissociation — early, near-linear volume loss with neuron number preserved until mid-Braak — and interventions that restore quality control will rescue volume and function before the number begins to fall, but not after.

On reversibility. Restoration of forty-hertz parvalbumin-driven activity, or of the upstream TGF-β/SMAD signal, will rescue synaptic counts and cognition in the window before the three-layer loop becomes self-sustaining, and will fail after it — a discontinuity in outcome at a definable cellular threshold, not a continuum.

Each prediction is a number or an order that a single well-designed quantification could overturn. That is the dividend of counting the right things: the casualty list becomes a set of hypotheses.

XIII. Conclusion — A Disease One Can Count

The history of Alzheimer's quantification has been a history of counting what is easy to see. Plaques and tangles are stainable, stageable, and reassuringly objective, and the field built its diagnostic apparatus around them — only to discover, again and again, that they are not the measures that best explain the illness. The thing that best explains the illness is the loss of specific neurons and, above all, of specific synapses, in a specific order, by specific chains. This census has tried to count those.

The count yields a disease of startling precision. Out of eighty-six billion neurons and a hundred and fifty trillion synapses, Alzheimer's destroys a minute, exquisitely selected set — the coerulean and cholinergic projection neurons that fail first and worst, the entorhinal gateway that is sixty per cent gone before diagnosis, the CA1 field that is halved, the parvalbumin net that is digested before its cell can die — and it destroys them in the rostro-limbic-cortical order that the temporal architecture independently derived from time. The cellular architecture and the temporal architecture are the same structure read in two currencies: the order of the casualties is the chronology of the disease. And the casualties share a face — the extravagant, fast-firing, oxidatively exposed, envelope-dependent neuron — so that selective vulnerability resolves, on counting, into a phenotype rather than a riddle.

Two consequences follow. The first is conceptual: the disease is a targeted demolition, not a diffuse decline, and its target is, at the end, not even a cell but a synapse and the matrix that protects it — which is why cognition can be lost, and for a window regained, with the neurons still alive. The second is practical: what can be counted can be measured, staged, and defended. If the load-bearing casualty is the perisomatic synapse of the parvalbumin interneuron beneath its perineuronal net, then that is what trials should measure, that is the threshold whose crossing they should aim to prevent, and that is the readout by which a restoration of the homeostatic signal — the one intervention with a claim on every phase — should be judged. The disease has always been a process one could count. The opportunity has always been to count the right things, early enough, in the one place where all the chains converge.

References

Azevedo, F. A. C., et al. (2009). Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. Journal of Comparative Neurology, 513(5), 532–541.

Brady, D. R., & Mufson, E. J. (1997). Parvalbumin-immunoreactive neurons in the hippocampal formation of Alzheimer's diseased brain. Neuroscience, 80(4), 1113–1125.

Braak, H., & Del Tredici, K. (2011). The pathological process underlying Alzheimer's disease in individuals under thirty. Acta Neuropathologica, 121(2), 171–181.

Butovsky, O., et al. (2014). Identification of a unique TGF-β-dependent molecular and functional signature in microglia. Nature Neuroscience, 17(1), 131–143.

Cardin, J. A., et al. (2009). Driving fast-spiking cells induces gamma rhythm and controls sensory responses. Nature, 459(7247), 663–667.

Celio, M. R. (1986). Parvalbumin in most γ-aminobutyric acid-containing neurons of the rat cerebral cortex. Science, 231(4741), 995–997.

Crapser, J. D., et al. (2020). Microglia facilitate loss of perineuronal nets in the Alzheimer's disease brain. EBioMedicine, 58, 102919.

DeKosky, S. T., & Scheff, S. W. (1990). Synapse loss in frontal cortex biopsies in Alzheimer's disease: correlation with cognitive severity. Annals of Neurology, 27(5), 457–464.

de Vries, L. E., et al. (2024). Perineuronal nets and cognitive resilience in Alzheimer's disease. Alzheimer's & Dementia, 20.

Devi, L., et al. (2006). Accumulation of amyloid precursor protein in the mitochondrial import channels of human Alzheimer's disease brain. Journal of Neuroscience, 26(35), 9057–9068.

Drachman, D. A. (2005). Do we have brain to spare? Neurology, 64(12), 2004–2005.

Fawcett, J. W., Oohashi, T., & Pizzorusso, T. (2019). The roles of perineuronal nets and the perineuronal matrix in memory and disease. Nature Reviews Neuroscience, 20(8), 451–465.

Freund, T. F., & Buzsáki, G. (1996). Interneurons of the hippocampus. Hippocampus, 6(4), 347–470.

Freund, T. F., & Katona, I. (2007). Perisomatic inhibition. Neuron, 56(1), 33–42.

German, D. C., et al. (1992). Disease-specific patterns of locus coeruleus cell loss. Annals of Neurology, 32(5), 667–676.

Gómez-Isla, T., et al. (1996). Profound loss of layer II entorhinal cortex neurons occurs in very mild Alzheimer's disease. Journal of Neuroscience, 16(14), 4491–4500.

Gómez-Isla, T., et al. (1997). Neuronal loss correlates with but exceeds neurofibrillary tangles in Alzheimer's disease. Annals of Neurology, 41(1), 17–24.

Härtig, W., Brauer, K., & Brückner, G. (1992). Wisteria floribunda agglutinin-labelled nets surround parvalbumin-containing neurons. NeuroReport, 3(10), 869–872.

Hong, S., et al. (2016). Complement and microglia mediate early synapse loss in Alzheimer mouse models. Science, 352(6286), 712–716.

Hu, H., Gan, J., & Jonas, P. (2014). Fast-spiking, parvalbumin⁺ GABAergic interneurons: from cellular design to microcircuit function. Science, 345(6196), 1255263.

Iaccarino, H. F., et al. (2016). Gamma frequency entrainment attenuates amyloid load and modifies microglia. Nature, 540(7632), 230–235.

Kelly, S. C., et al. (2017). Locus coeruleus cellular and molecular pathology during the progression of Alzheimer's disease. Acta Neuropathologica Communications, 5(1), 8.

Keren-Shaul, H., et al. (2017). A unique microglia type associated with restricting development of Alzheimer's disease. Cell, 169(7), 1276–1290.

La Barbera, L., Krashia, P., & Nobili, A. (2024). Parvalbumin interneuron dysfunction in Alzheimer's disease. Neural Regeneration Research.

Lee, J.-H., et al. (2022). Faulty autolysosome acidification in Alzheimer's disease mouse models induces autophagic build-up of Aβ in neurons, yielding senile plaques. Nature Neuroscience, 25(6), 688–701.

Manaye, K. F., et al. (1995). Locus coeruleus cell loss in the aging human brain: a non-random process. Journal of Comparative Neurology, 358(1), 79–87.

Marschallinger, J., et al. (2020). Lipid-droplet-accumulating microglia represent a dysfunctional and proinflammatory state in the aging brain. Nature Neuroscience, 23(2), 194–208.

Pakkenberg, B., et al. (2003). Aging and the human neocortex. Experimental Gerontology, 38(1–2), 95–99.

Rudy, B., et al. (2011). Three groups of interneurons account for nearly 100% of neocortical GABAergic neurons. Developmental Neurobiology, 71(1), 45–61.

Scheff, S. W., et al. (2006). Hippocampal synaptic loss in early Alzheimer's disease and mild cognitive impairment. Neurobiology of Aging, 27(10), 1372–1384.

Scheff, S. W., et al. (2007). Synaptic alterations in CA1 in mild Alzheimer disease and mild cognitive impairment. Neurology, 68(18), 1501–1508.

Sohal, V. S., et al. (2009). Parvalbumin neurons and gamma rhythms enhance cortical circuit performance. Nature, 459(7247), 698–702.

Stevens, B., et al. (2007). The classical complement cascade mediates CNS synapse elimination. Cell, 131(6), 1164–1178.

Streit, W. J., et al. (2009). Dystrophic (senescent) rather than activated microglial cells are associated with tau pathology and likely precede neurodegeneration in Alzheimer's disease. Acta Neuropathologica, 118(4), 475–485.

Terry, R. D., et al. (1991). Physical basis of cognitive alterations in Alzheimer's disease: synapse loss is the major correlate of cognitive impairment. Annals of Neurology, 30(4), 572–580.

Theofilas, P., et al. (2017). Locus coeruleus volume and cell population changes during Alzheimer's disease progression. Alzheimer's & Dementia, 13(3), 236–246.

Ulland, T. K., et al. (2017). TREM2 maintains microglial metabolic fitness in Alzheimer's disease. Cell, 170(4), 649–663.

Werneburg, S., et al. (2025). C4d, a high-affinity LilrB2 ligand, is elevated in Alzheimer's disease and mediates synapse pruning. Proceedings of the National Academy of Sciences, 122.

West, M. J., et al. (1994). Differences in the pattern of hippocampal neuronal loss in normal ageing and Alzheimer's disease. The Lancet, 344(8925), 769–772.

West, M. J., et al. (2004). Hippocampal neurons in pre-clinical Alzheimer's disease. Neurobiology of Aging, 25(9), 1205–1212.

Whitehouse, P. J., et al. (1982). Alzheimer's disease and senile dementia: loss of neurons in the basal forebrain. Science, 215(4537), 1237–1239.