How Do Amyloid Pathology and Aberrant Neuronal Activity Disrupt Plasticity and Memory in Alzheimer’s Disease?

Abstract

Alzheimer’s disease (AD) is increasingly understood as a disorder of network-state and plasticity-capacity, in which amyloid-β and tau pathologies disrupt the activity-dependent mechanisms that build and stabilize memory engrams. Here, I review how amyloid-β–driven neuronal hyperactivity contributes to plasticity and memory deficits in AD. I also discuss how various cellular pathologies reinforce one another, leading to a cellular environment that is impermissive to plasticity. I relate these cellular and circuit-level disturbances to failures in memory encoding, consolidation, and recall, emphasizing the role of interference arising from coexisting hyper- and hypoactive neuronal populations. Finally, I discuss the relevance and limitations of amyloid mouse models in understanding the cognitive decline in AD.

Introduction

Alzheimer’s disease (AD), the most common form of dementia, is associated with the loss of well-consolidated memories and an inability to form new memories. In the healthy brain, a long-term memory is encoded in a distributed network across the brain, referred to as an engram (Josselyn and Tonegawa 2020). The engram is generated through activity-dependent changes in synaptic connections between neurons. Altered connectivity reshapes the brain’s high-dimensional space of possible neural firing patterns, such that they generate activity patterns that the brain can reliably return to when triggered by partial cues, allowing for memory recall (Rolls and Treves 1998). The ability to reinstate precise activity patterns necessary for memory recall would be impaired when network connectivity is disrupted by synapse and neuronal loss, as well as by aberrant activity patterns found in AD (Palop et al. 2006; Busche and Konnerth 2016). Consistently, synapse density correlates strongly with memory scores, and epileptiform activity further accelerates cognitive decline in patients with AD (Terry et al. 1991; Vossel et al. 2016) (Fig. 1).

Figure 1.

Amyloid-β pathology contributes to cognitive decline by promoting synapse loss and aberrant activity. Amyloid-β (amyloid) oligomers, which are enriched around the plaques, lead to excitatory synapse loss. The dotted lines indicate a lost dendritic spine and an axonal bouton closer to the amyloid. Amyloid-β elicits aberrant activity in neurons (red triangles with connected lines represent aberrantly active neurons; black triangles with connected lines represent neurons with activity levels within the nonpathological range). Synapse loss and aberrant activity accelerate cognitive decline.

Despite neurodegeneration and cognitive decline, evidence suggests that some memories may not be entirely lost in AD but instead become inaccessible. For instance, some patients exhibit mental clarity and the ability to recognize loved ones for a brief period before their death, a process referred to as terminal lucidity (Nahm et al. 2012). Similarly, in mouse models of AD with impaired memory recall, optogenetic activation of memory engrams successfully restored recall despite synapse loss (Roy et al. 2016). While the mechanisms of terminal lucidity remain to be elucidated, neural population encoding a memory is redundant, and therefore, the same functional outcome (memory recall) can be achieved by different combinations of neurons (neuronal degeneracy; Noppeney et al. 2004). Therefore, under extraordinary circumstances, the neural representation of a memory can reemerge, even after partial loss of the network, due to degeneracy, leading to similar downstream activations.

While loss of network integrity due to synapse loss (Fig. 2) and neurodegeneration may contribute to memory decline, the inability of patients with AD to form memories suggests that they are unable to modify their circuit to store and represent new information, presumably due to impairments in synaptic plasticity mechanisms (Fig. 2). Synaptic plasticity mechanisms shape the formation and maintenance of memory-associated neuronal networks. The seminal hypothesis proposed by Donald Hebb posits that the efficacy of a synaptic connection between 2 neurons is enhanced when the presynaptic axon is close enough to the postsynaptic neuron and repeatedly contributes to its activation (Hebb 1949). Subsequent in vitro studies showed that high-frequency activity (Bliss and Lomo 1973) or presynaptic activity preceding postsynaptic activity within a specific temporal window (Markram et al. 1997) leads to long-lasting potentiation (LTP) of synapses. In contrast, low-frequency activity, or postsynaptic activity preceding presynaptic activity within the same temporal window, induces long-term depression (LTD) of synapses (Dunwiddie and Lynch 1978). This frequency-dependent modulation arises from variations in postsynaptic calcium influx mediated by N-methyl-D-aspartate (NMDA) receptors, although NMDA-independent forms of plasticity also exist (Malenka and Bear 2004). Additionally, engagement of metabotropic glutamate receptors via glutamate binding, typically during elevated or prolonged synaptic activity, can facilitate potentiation or depression in a receptor subtype- and circuit-specific manner (Niswender and Conn 2010). These calcium transients, modulated by the frequency and temporal correlation of pre- and postsynaptic activity, engage distinct intracellular cascades. For acute adaptations, they alter the phosphorylation status of synaptic proteins, such as α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptors at postsynaptic sites, enabling rapid trafficking and adjustments in synaptic efficacy (Citri and Malenka 2008). For long-lasting modifications, they regulate transcriptional processes to sustain structural changes (Citri and Malenka 2008).

Figure 2.

Synaptic plasticity modifies networks, and amyloid-β pathology disrupts it. A hypothetical network of 6 neurons (ovals A–F). Neurons B to D are highly connected (number of lines indicates synaptic strength) to neuron A, whereas neuron E is weakly connected to A and highly connected to F, a neuron in a downstream region (long-distance connection indicated by longer lines). A to D neurons form an ensemble (red fill). Right: Long-term potentiation (LTP) of A and E synapses would recruit E and F to the A to D ensemble but, upon global downscaling to preserve overall connection strengths for neuron A, would lead to circuit reorganization. Amyloid-β (amyloid)–mediated synapse loss would weaken this reorganization. Left: Amyloid-induced hyperactivity in neurons A and C could lead to hyperconnectivity (dark red), such that neurons B and D become hypoactive and less connected to neuron A. Hyperactivity disrupts plasticity and circuit reorganization.

The synaptic plasticity mechanisms identified in vitro have been validated in vivo, where artificial induction of plasticity occludes subsequent natural plasticity, confirming the role of similar pathways (Whitlock et al. 2006). Nonetheless, the millisecond-scale temporal precision required for these processes diverges from behavioral paradigms, in which associations between stimuli, such as conditioned and unconditioned cues or actions and rewards, occur over seconds or longer (Magee and Grienberger 2020). For instance, synaptic activity representing an experience may precede the release of dopamine, which signals the valence of the experience. To reconcile such temporal delays, synapses exhibiting correlated activity are proposed to acquire a transient eligibility trace, such as altered calcium levels, rendering them amenable to modification upon the arrival of neuromodulatory signals during salient or novel events (Magee and Grienberger 2020). Depending on the neuromodulatory receptor and the prevailing brain state, such signals can induce potentiation or depression of the synapses with an eligibility trace. However, these processes lack a specification for the magnitude of synaptic weight changes needed to represent memory. Such a specification could be provided by an instructive signal (Magee and Grienberger 2020). One classic example is the error signal related to motor discrepancies provided by climbing fibers to Purkinje neurons in the cerebellum that leads to LTD of parallel fibers that were activated 50 to 250 ms earlier (Ito 1972). More recently, behavioral timescale synaptic plasticity has been identified as a mechanism associated with place field emergence in hippocampal CA1 neurons (Bittner et al. 2017). Here, CA3 afferent activity precedes a plateau potential, a long-lasting depolarization evoked by entorhinal inputs and modulated by neuromodulators such as acetylcholine and norepinephrine. This enhances synaptic strength such that place fields emerge at locations traversed 1 s prior to the plateau’s onset. This plasticity mechanism integrates correlative firing, an eligibility trace, and an instructive plateau potential that occurs in behaviorally pertinent timescales (Magee and Grienberger 2020).

Whether synapses undergo potentiation or depression depends on the history of activity of these synapses. Synapses that have been depressed or potentiated are less likely to undergo further depression or potentiation, respectively (Abraham 2008). These functional changes, in turn, affect the structure of synapses, where functional enhancement is associated with an increase in the size of existing synapses or the formation of new synapses, and functional weakening is associated with shrinkage or loss of existing synapses (Holler et al. 2021). Together, the actions of neuromodulators, glutamate-mediated signaling, and intracellular calcium dynamics determine whether excitatory connectivity between neurons is strengthened or lost, thereby shaping which neurons will be part of a network.

While coordinated neuronal activity strengthens connectivity between neurons, enhanced connectivity, in turn, would further amplify their coordination. This positive feedback loop can potentially escalate into runaway excitation. Conversely, reduced neuronal coordination and decreased connectivity, through a negative feedback loop, can potentially prevent the network from reestablishing its previous configuration. To balance firing rates while allowing for network reorganization, neurons undergo compensatory synaptic adjustments, either locally through the spread of signaling molecules or globally via multiplicative scaling of synaptic strength (Turrigiano 2012). Such adjustments are not confined solely to excitatory synapses; inhibitory synapses also exhibit plasticity, acting either synergistically or in opposition to excitatory changes, thereby preserving overall network stability (Ravasenga et al. 2022).

Many of the synaptic plasticity mechanisms are disrupted by the 2 major pathological hallmarks of AD, the accumulation and spread of amyloid-β peptides and hyperphosphorylated tau, which generate amyloid plaques and neurofibrillary tangles, respectively. These hallmarks disrupt the plasticity mechanisms that modify the network organization and maintain firing rates, thereby contributing to cognitive decline (Spires-Jones and Hyman 2014). By themselves, amyloid and tau pathologies are neither required nor sufficient to drive cognitive decline in AD. For example, up to 44% of cognitively normal older adults exhibit significant cortical amyloid deposition, as determined by positron emission tomography scans, without any measurable cognitive impairment (Jansen et al. 2015). However, cognitively normal individuals with abnormal amyloid levels have a higher risk of future cognitive decline (Mormino and Papp 2018). Similarly, primary age-related tauopathy, characterized by neurofibrillary tangles without amyloid plaques, is frequently observed in medial temporal structures of cognitively normal individuals (Crary et al. 2014). However, when amyloid and tau pathologies are present together, they increase the likelihood of the development of dementia (Teylan et al. 2020).

Amyloid-β accumulation begins diffusely in the neocortex (Thal phase 1), subsequently involving the hippocampus (phase 2), the striatum and other subcortical structures (phase 3), the brainstem (phase 4), and finally the cerebellum (phase 5; Thal et al. 2002). The level of amyloid-β accumulation in brain regions positively correlates with their activity levels (Bero et al. 2011). Neural activity brings together amyloid precursor protein (APP) and BACE1, an enzyme that cleaves APP to produce amyloid-β (Das et al. 2016). Neurons release amyloid-β into the interstitial fluid, which diffuses passively or is degraded by metalloproteases (Iwata et al. 2005). The cerebrospinal fluid enters the brain tissue through aquaporin channels on astrocytic end-feet, facilitated by the pulsation of arteries, resulting in the mixing of interstitial and cerebrospinal fluids. As this mixture exits the brain through perivascular spaces, it carries secreted amyloid-β, effectively clearing it from the brain tissue (Iliff et al. 2012). However, impairment in the drainage system, known as the glymphatic system, due to reduced pulsations of stiffened arteries associated with aging, can lead to the accumulation of amyloid-β along the drainage routes, including distant sites, where it seeds new plaque formation. Amyloid-β also exhibits prion-like propagation, in which preexisting aggregates promote the formation and deposition of new amyloid-β, thereby spreading pathology beyond the original inoculation site (Watts and Prusiner 2018). Amyloid-β accumulation is promoted by multiple factors, including genetic, epigenetic, lifestyle-related, infectious, metabolic, hormonal, or physiological disruptions (Silva et al. 2019).

In contrast, hyperphosphorylated tau pathology typically originates in layer II of the entorhinal cortex (Braak stages I–II; Braak et al. 2006) and can extend into hippocampal and adjacent limbic regions independently of amyloid-β deposition. Their spread to neocortical areas (Braak stages IV–VI) is facilitated by extracellular amyloid-β (Pooler et al. 2015; Lee et al. 2022; Giorgio et al. 2024; Roemer-Cassiano et al. 2025), although it can happen without amyloid-β, such as in primary age-related tauopathy. Once established in the cortex, tau exhibits prion-like seeding behavior through the secretion and uptake of soluble, hyperphosphorylated tau species, enabling self-sustained propagation (Clavaguera et al. 2009). The propagation of tau seeds is facilitated by neuronal activity (Wu et al. 2016). Therefore, amyloid-β–driven hyperactivity, which requires functional tau (Roberson et al. 2007), promotes tau spread in the brain (Roemer-Cassiano et al. 2025). For instance, hyperactivity in the default mode network, driven by elevated amyloid-β, leads to downstream hyperactivity in the medial temporal lobe, resulting in tau accumulation in that region (Giorgio et al. 2024). Pathological tau, in turn, leads to neuronal hypoactivity, which dominates amyloid-β–associated hyperactivity when both pathologies are combined (Busche et al. 2019). Thus, amyloid-facilitated, pathological tau spread could contribute to the transition of hyperactivity to hypoactivity observed with AD progression (O’Brien et al. 2010). Together, amyloid and tau pathologies disrupt neural activity, contributing to impaired synaptic plasticity and neuronal dysfunction.

Multiple factors contribute to the emergence of aberrant neural activity in AD, including excessive glutamatergic activity, decreased GABAergic drive, elevated intracellular calcium release, and altered ratios of excitatory to inhibitory synapses. These factors have been extensively reviewed (Palop and Mucke 2010a; Busche and Konnerth 2015; Stargardt et al. 2015; Busche and Konnerth 2016; Ambrad Giovannetti and Fuhrmann 2019; Targa Dias Anastacio et al. 2022; Vicente et al. 2024). Intriguingly, dendritic spine loss, which would be expected to decrease neural activity, can also contribute to neuronal hyperactivity (Siskova et al. 2014; Henderson et al. 2019). When dendritic spines are lost, neuronal surface area decreases, which increases the input resistance, causing synaptic currents to produce larger voltage changes than normal (Henderson et al. 2019). Here, I discuss how aberrant neural activity, exacerbated by amyloid-β pathology, contributes to deficits in plasticity in a vicious cycle with different cellular-level pathologies (Figs. 3, 4), ultimately interfering with the formation and stabilization of new memory traces. Finally, I consider the merits and limitations of mouse models of AD.

Figure 3.

Various pathologies that contribute to hyperactivity and amyloid-mediated plasticity deficits in a vicious cycle. Arrows indicate the relationships between the connected boxes. The listed pathologies are not exhaustive but cover many of the cellular mechanisms known to be affected in mouse models of Alzheimer’s disease.

Figure 4.

Mechanisms that weaken an excitatory synapse in amyloid-β pathology. Hyperactive neurons release excess glutamate, which is not properly cleared by astrocytes, resulting in excessive glutamate receptor activation, including extrasynaptic N-methyl-D-aspartate (NMDA) receptors and metabotropic glutamate receptors. Aberrant calcium influx activates signaling through calcineurin and calpain, among other mechanisms, leading to the endocytosis of α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptors, loss of gephyrin scaffold (inhibitory synapses), breakdown of microtubules, and loss of nuclear signaling through CREB. Furthermore, excessive calcium triggers calcium depletion in the endoplasmic reticulum, leading to protein misfolding and reactive oxygen species production in mitochondria. With reduced ATP generation, misfolded proteins, damaged microtubules, and reduced transcription of plasticity genes, synapses will not effectively respond to changes in experience and could be engulfed by microglial processes for elimination.

Aberrant Activity, Glucose Hypometabolism, and Plasticity Deficits

One of the earliest clues that aberrant neural activity contributes to synaptic and neurodegeneration in AD came from observations that excitatory neurotransmitters, such as glutamate and aspartate, cause excitotoxic damage in cultured neurons and in hypoxic brains (Greenamyre 1991). These effects include regression of dendritic arbors and reduced RNA content, which are pathological features similar to those observed in AD (Greenamyre and Young 1989). Early studies also found that the regions of the brain affected early in the disease exhibit a loss of glutamate binding sites (Dewar et al. 1991), reinforcing the idea that disruption of excitatory neurotransmission is a key feature of the disorder.

The damaging effects of excitatory amino acids are amplified when the energy status of neurons is compromised. At nontoxic concentrations, glutamate causes excitotoxicity in glucose-deprived cultured neurons. During glucose deprivation, neurons lack the energy required to maintain membrane potential. As a result, the magnesium block on NMDA receptors is more easily removed, leading to excessive calcium influx (Greenamyre 1991). In fact, glucose hypometabolism is one of the earliest known features of AD, indicating that these neurons could be particularly vulnerable to excitotoxicity (Mosconi et al. 2008).

Amyloid-β contributes to impaired glucose metabolism through several mechanisms, including alterations in glucose transporter expression, the inhibition of glycolytic enzymes, mitochondrial dysfunction, and reduced surface expression of insulin receptors (Wang et al. 2025). In turn, disrupted glucose metabolism due to insulin resistance and hyperglycemia increases amyloid-β levels by reducing the activity of the insulin-degrading enzyme, which plays a role in amyloid-β clearance (Zhao and Townsend 2009). Furthermore, decreased insulin receptor signaling releases the inhibition on glycogen synthase kinase 3 (GSK3). The α subunit of GSK3 (GSK-α) triggers gamma-secretase activity, which cleaves the amyloid precursor protein and increases amyloid-β production (Phiel et al. 2003). This establishes a cycle in which impaired glucose metabolism and increased amyloid-β production mutually reinforce one another, leading to glutamate excitotoxicity.

Insulin resistance, which is associated with hypertension and dyslipidemia, also leads to vascular damage. Impaired vascular function will reduce cerebral blood flow and waste clearance as well as neurovascular coupling, a process by which blood flow is regulated to match neural activity (Kellar and Craft 2020). These changes exacerbate glucose hypometabolism, elevate amyloid-β accumulation, and increase the risk of aberrant neural activity. In addition, amyloid deposits in the walls of cerebral arteries, a condition known as cerebral amyloid angiopathy, impair blood circulation and glucose availability in the brain (Noto et al. 2025). Aggravating these effects, insulin resistance is associated with decreased estrogen levels. Estrogen regulates the expression of glucose transporters and enzymes involved in cellular respiration (Brinton 2008). Therefore, reduced estrogen levels, such as during menopause, could impair energy metabolism and make neurons vulnerable to excitotoxicity and contribute to the heightened risk of AD in postmenopausal women (Mosconi et al. 2021).

Chronic hyperglycemia also drives nonenzymatic glycation of proteins and lipids, leading to the formation of advanced glycation end products (AGEs). AGEs bind to the receptor for AGEs (RAGE) and activate signaling pathways, including different mitogen-activated protein (MAP) kinases, that stimulate NADPH oxidase, which converts molecular oxygen into superoxide radicals in the cytoplasm (Akhter et al. 2021; Boccardi et al. 2025). The cytosolic reactive oxygen species (ROS) will disrupt mitochondrial function and exacerbate oxidative stress. Mitochondrial oxidative stress can trigger the release of danger-associated molecular patterns (DAMPs), including mitochondrial DNA and ATP, which activate pattern recognition receptors on microglia, leading to inflammation, further accelerating the vicious cycle of ROS generation and inflammation (Lin et al. 2022).

Reduced glucose availability and impaired ATP production affect neurons, astrocytes, and microglia. When neurons are active, they rely heavily on lactate produced by astrocytes as an alternative energy source (Magistretti and Allaman 2015). However, when glycolysis is suppressed in astrocytes, both lactate production and overall energy support to neurons will be compromised. As a result, neuronal energy supply will decline during hyperactivity, but at the same time, astrocytic glutamate clearance becomes impaired, due to both energy limitations and amyloid-β–induced downregulation of glutamate transporter surface expression (Scimemi et al. 2013). The situation is further worsened by amyloid-β–mediated increased glutamate release from astrocytes, which occurs through the activation of nicotinic receptors (Talantova et al. 2013). The low-energy, hyperactive state is incompatible with the demands of synaptic plasticity, which requires precisely timed neural activity and energy to maintain synaptic function.

Aberrant Activity, Synaptic Signaling, and Plasticity Deficits

Hyperactive neurons would release more glutamate, and with reduced glutamate clearance, the excessive glutamate would leak to extrasynaptic sites. The extrasynaptic NMDA receptors are more sensitive to glutamate levels and are more likely to become activated. Tonic extrasynaptic NMDA receptor activation promotes LTD, a phenotype observed with amyloid-β exposure (Li et al. 2011). Calcium entering through extrasynaptic NMDA receptors activates calcineurin, which, through downstream phosphatases, dephosphorylates CREB, a transcription factor that regulates the expression of genes associated with synaptic plasticity, such as BDNF and c-Fos (Hardingham et al. 2002). With CREB phosphorylation removed, high-frequency neural activity will not be signaled to the nucleus, delinking the activity–plasticity relationship in neurons. Furthermore, extrasynaptic NMDA receptor-mediated calcium influx activates the protease m-calpain, which cleaves striatal-enriched protein tyrosine phosphatase (STEP) 61 to STEP33 (Xu et al. 2009). STEP33, unlike STEP61, will not dephosphorylate p38 MAP kinase to inactivate it. Active p38 MAP kinase, through downstream signaling, leads to excitotoxicity (Xu et al. 2009). Another kinase, GSK3-β, activated by extrasynaptic NMDA receptor signaling, excessively phosphorylates tau and destabilizes microtubules (Hooper et al. 2008). An unstable cytoskeleton cannot support synaptic potentiation, as the delivery of mRNA and ribosomes for local translation, as well as organelles such as mitochondria for energy, will be affected. On the other hand, chronic extrasynaptic NMDA receptor activation, via CaMKIV, produces alternatively spliced amyloid precursor protein transcripts, resulting in increased amyloid-β levels (Bordji et al. 2010), which would contribute to a reinforcing loop of amyloid-β-extrasynaptic NMDA receptor activation.

Excessive glutamate also activates mGluR receptors, which trigger a signaling cascade, culminating in AMPA receptor endocytosis, LTD, and synapse loss (Luscher and Huber 2010). G-protein-coupled signaling also induces LTD by increasing the phosphorylation of eukaryotic elongation factor 2, a calcium-calmodulin–dependent kinase (Park et al. 2008). This slows translational elongation and reduces global protein synthesis. Despite this overall reduction, local dendritic translation of Arc/Arg3.1 is increased, which in turn promotes clathrin-mediated endocytosis of AMPA receptors, resulting in LTD (Park et al. 2008). However, the role of amyloid-β in LTD depends on whether the exposure is acute or chronic, as well as the age. Unlike the effect of acute exposure, which consistently facilitates LTD in reduced preparations and in vivo, chronic exposure in AD mouse models shows that amyloid-β inhibits mGluR-mediated LTD, particularly in older age (>8 mo; Valdivia et al. 2023). Chronic amyloid-β exposure could lead to nonspecific LTD, which would increase the threshold needed for further depression, thereby preventing learning-specific LTD (Fig. 5). Memory formation would be disrupted, particularly at later ages, when LTD becomes critical for successful learning (Lee et al. 2005).

Figure 5.

Chronic long-term depression (LTD) elicited by amyloid-β disrupts learning-associated LTD, leading to cognitive decline. Amyloid facilitates α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptor internalization through mechanisms illustrated in Figure 4, leading to nonspecific LTD of synapses and impaired long-term potentiation (LTP). Plasticity threshold for depression is increased (metaplasticity), such that already weak synapses cannot undergo learning-specific LTD.

Extrasynaptic calcium entry also disrupts inhibitory synapses. Calpain 1 proteolytically cleaves gephyrin, a scaffolding protein for GABA-A receptors. GSK3b and CDK5, activated by extrasynaptic calcium, phosphorylate gephyrin at Ser270 and disrupt gephyrin interaction with microtubules and destabilize it (Tyagarajan and Fritschy 2014). Gephyrin expression is altered in AD models (Lauterborn et al. 2021), and the removal of the gephyrin scaffold would reduce GABA-A receptor function (Limon et al. 2012), thereby decreasing tonic inhibition and enhancing excitatory neuronal hyperactivity. This, along with the altered ratio of excitatory and inhibitory synapses (Lauterborn et al. 2021; Niraula et al. 2023) and their functional imbalance (Scaduto et al. 2023), would feed into an amyloid-β–hyperactivity vicious cycle.

In addition to inhibitory synapses on excitatory neurons, inhibitory neurons themselves are affected by amyloid-β pathology (Melgosa-Ecenarro et al. 2023). The activity of parvalbumin neurons, which controls gamma oscillations, is reduced in part due to impairments in their sodium channels (Verret et al. 2012), whereas the activity of somatostatin neurons is increased in amyloid mouse models (Algamal et al. 2022). Reduced gamma oscillations, resulting from impaired parvalbumin neuronal activity, are associated with epileptic discharge and network hypersynchrony (Palop and Mucke 2016), which in turn further increases amyloid-β levels, leading to an aberrant network activity–amyloid-β vicious cycle.

Aberrant Activity, Neuroinflammation, and Plasticity Deficits

In addition to AGE and DAMP-mediated neuroinflammation described earlier, neuronal hyperactivity directly activates microglia. Microglia extend their processes toward active neurons, and calcium levels increase when these processes target hyperactive and hypoactive neurons (Eyo et al. 2014). Hyperactive neurons secrete ATP, which activates the purine receptor on microglia, leading to the expression of proinflammatory cytokines (Inoue 2002). Hyperactive neurons also secrete more amyloid-β and glutamate, whose accumulation triggers signaling pathways, including NF-κB, which activates genes that express proinflammatory cytokines, such as TNF-α, IL-1β, and IL-6 (Sun et al. 2022). These cytokines up- or down-regulate neural activity and modify synapses in a context-dependent manner. For instance, IL-1β enhances neuronal excitability through the NMDA receptor but downregulates it by reducing sodium or calcium channel activities or by boosting inhibitory tone through GABA receptors (Nemeth and Quan 2021). Similarly, IL-1β promotes plasticity at lower concentrations, whereas it disrupts it at higher levels; however, the mechanisms underlying this effect are not fully understood (Nemeth and Quan 2021). In contrast, TNF-α increases the surface expression of surface AMPA receptors and contributes to synaptic homeostatic upscaling (Stellwagen and Malenka 2006). This could compensate for synapse loss elicited by amyloid-β. Furthermore, as observed in the dentate gyrus, a local rise in TNF-α levels activates TNF receptors on astrocytes, leading to glutamate release (Santello et al. 2011). Astrocyte-released glutamate, in turn, binds to presynaptic NMDA receptors and enhances neuronal glutamate release (Santello et al. 2011). Together, enhanced excitability of neurons would increase the threshold needed for further potentiation and disrupt plasticity.

Microglial recruitment toward hyperactive neurons also contributes to achieving activity homeostasis (Badimon et al. 2020). Synaptic hyperactivity increases the complement component C1q at synapses and marks them with the C3 fragment via the classical complement pathway, facilitating microglial engulfment (Crowley et al. 2024). Furthermore, phosphatidylserine on the inner leaflet of the plasma membrane becomes exposed on the outer leaflet on hyperactive synapses. This also acts as an “eat-me” signal for microglial engulfment (Rueda-Carrasco et al. 2023). Thus, microglial activation produces signals that increase neuronal excitability, as well as those that curtail it. This could lead to a continued increase in the levels of hyperactive and hypoactive neurons, both of which further trigger microglial activation. If further exacerbated by glucose hypometabolism, this could result in a vicious cycle of aberrant neural activity and microglial activation.

Aberrant Activity, Protein Misfolding, and Plasticity Deficits

Neuronal hyperactivity in amyloid-β pathology can trigger intracellular calcium release from the endoplasmic reticulum through a process known as calcium-induced calcium release (Bezprozvanny and Mattson 2008). This process, which requires the ryanodine receptor and NMDA-mediated calcium, can lead to excessive intracellular calcium. Endoplasmic reticulum (ER) calcium depletion can lead to misfolding of nascent proteins. This, along with protein misfolding caused by oxidative stress, genetic mutations, or aging, initiates the unfolded protein response (UPR). UPR halts protein translation to lessen ER load, promotes chaperone expression to refold misfolded proteins, and promotes protein degradation pathways (Lim et al. 2023). With age, the accumulation of misfolded proteins activates PERK, which phosphorylates eIF2-α, involved in translation initiation (Hetz and Mollereau 2014). eIF2-α phosphorylation impairs protein synthesis and, therefore, late-phase LTP dependent on protein synthesis (Costa-Mattioli et al. 2007), while at the same time increasing translation of mRNA with a 5′-untranslated region, such as BACE1, which increases amyloid-β levels (O’Connor et al. 2008). Thus, protein misfolding and amyloid accumulation are amplified through a self-perpetuating feedback loop.

Ubiquitin tags misfolded proteins to target them for degradation, and proteosomes degrade these ubiquitin-tagged proteins. However, an overload of misfolded proteins can overwhelm the proteosomal capacity. When proteosomal degradation is insufficient, autophagocytic clearance of damaged proteins and organelles via lysosomes will be initiated (Pandey et al. 2007). However, lysosomal integrity is impaired in amyloid-β pathology for various reasons, including reduced lysosomal acidification, calcium dyshomeostasis, impaired lysosomal trafficking, altered lysosomal membrane permeabilization induced by oxidative stress, and disrupted autophagosome–lysosome fusion (Nixon and Rubinsztein 2024). With UPR-mediated suppression of protein translation, impaired ER function, and protein clearance through lysosomes, synaptic plasticity, such as LTP and LTD, will be affected, as they rely on these processes. On the other hand, chronic activation of the UPR upregulates β-secretase (BACE1; O’Connor et al. 2008), which accelerates APP cleavage and amyloid-β production, which would worsen aberrant neural activity and fuel a vicious cycle of neuronal hyperactivity and protein misfolding.

Aberrant Activity, Mitochondrial, and Plasticity Deficits

Excessive cytoplasmic calcium is also sequestered by mitochondria (Rizzuto et al. 1992). Mitochondria, in addition to their well-known role in generating ATP, also regulate the neural firing rate (Ruggiero et al. 2021). The integrity of mitochondrial proteins is disrupted by excessive calcium entering mitochondria due to aberrant neuronal activity (Calvo-Rodriguez et al. 2020). In addition, soluble amyloid-β interacts with mitochondrial proteins, such as voltage-dependent anion channels, mitochondria-associated ER membranes, and cyclophilin D, a key component of the mitochondrial permeability transition pore (mPTP), and increases calcium influx into mitochondria (Du and Yan 2010). The persistent opening of mPTP causes a loss of mitochondrial membrane potential and impairs ATP production. mPTP opening will also release proapoptotic signals, contributing to neurodegeneration (Du and Yan 2010).

Increased mitochondrial calcium activates enzymes in the tricarboxylic acid (TCA) cycle and oxidative phosphorylation (Brookes et al. 2004). However, excessive calcium impairs the electron transport chain, a process that is worsened by a disrupted mitochondrial membrane potential. The combination of higher activation of the TCA cycle and oxidative phosphorylation, along with the impaired electron transport chain, leads to leakage of electrons, resulting in increased reactive oxygen species (ROS; Brookes et al. 2004). Amyloid-β and ROS also disrupt the normal fission and fusion dynamics of mitochondria, tilting the balance toward more fission (Wang et al. 2014). Fragmented mitochondria have reduced ATP-producing capacity and impaired calcium-buffering capacity. ROS can damage lipids, proteins, and DNA, further worsening protein-folding machinery and gene expression (Perez Ortiz and Swerdlow 2019). Thus, decreased bioenergetics and increased oxidative stress would impair neuronal activity levels, particularly high-frequency firing (therefore, a higher ATP requirement) in parvalbumin neurons (Kann 2016), and synaptic plasticity, leading to a vicious cycle of aberrant activity–mitochondrial disruption and contributing to the buildup of more amyloid-β.

How Is Memory Disrupted in AD Due to Aberrant Neural Activity?

Encoding Defects

Aberrant neuronal activity, coupled with a reduced energy state driven by glucose hypometabolism and mitochondrial deficits, abnormal calcium signaling, protein misfolding, and neuroinflammation, collectively impedes the implementation of synaptic plasticity, as discussed earlier. Aberrant neural activity further exacerbates memory impairment by disrupting the precise temporal patterns essential for initiating plasticity pathways. Beyond disrupting the plasticity of synapses with elevated activity, the plasticity threshold of nearby synapses could also be affected by the local spread of elevated calcium. Furthermore, the spread of calcium would disrupt the “eligibility trace” in nearby synapses. Thus, plasticity deficits can propagate beyond the affected subnetwork and prevent the encoding of related memories. Additionally, amyloid-β affects cholinergic neurons in the basal forebrain (Whitehouse et al. 1982), which would weaken the instructive signal needed for synaptic weight optimization. Without appropriate synaptic plasticity, neural circuits fail to reorganize effectively, thereby impairing the encoding and storage of new memories.

Synaptic plasticity requires energy, but the brain operates under a fixed energy budget, raising the question of how it meets energy demands when neurons become hyperactive. In certain brain regions, such as the visual cortex and hippocampus of AD mouse models, hyperactive and hypoactive neurons coexist (Busche and Konnerth 2016). Hypoactivity, through excitatory synapse loss, could compensate for hyperactivity (Palop and Mucke 2010b), allowing the brain to maintain its fixed energy budget despite elevated neuronal activity. However, this coexistence can result in hyperactive neurons dominating responses to experience. In the visual cortex of AD mouse models, hyperactive neurons participate in multiple ensembles and become less selective (Fig. 6), leading to interference from memory-irrelevant experiences that disrupt memory encoding (Niraula et al. 2023). A similar loss of coding specificity is observed in the hippocampus, where ensemble imaging in 3xTg-AD mice found that neurons were hyperactive and had less precise spatial representations (Lin et al. 2022). Interference in memory encoding can also occur when an imbalance in excitation–inhibition in default mode network circuits impairs its deactivation (Anticevic et al. 2012; Giorgio et al. 2024), which is necessary for reducing interference from self-referential thoughts during the encoding of external experience.

Figure 6.

A model for how interference impairs plasticity. Left: Hypothetical ensembles of neurons (oval) responding to two stimuli (stimulus 1, red ensemble; stimulus 2, green ensemble) or none (empty). Repeated experience with stimulus 1 weakens connectivity between red ensemble neurons, which are not coactivated during stimulus 2 or other stimuli. In the absence of continued coactivity during all times other than the stimulus 1 experience, their connectivity would not be reestablished. Thus, the circuit would selectively reorganize for red ensembles, reflecting repeated experience with stimulus 1. Right: In amyloid-β pathology, hyperactive neurons participate in both ensembles (yellow), thereby reducing the selectivity for a stimulus. This would disrupt the weakening of connectivity elicited by repeated experience with stimulus 1, because the same neurons would be coactive during stimulus 2 or other stimuli. Thus, reduced selectivity due to hyperactivity would disrupt experience-selective circuit reorganization.

Consolidation Defects

Sleep is critical for memory consolidation. Amyloid-β–induced neuronal hyperactivity in the locus coeruleus (Kelly et al. 2021) would increase norepinephrine levels, thereby increasing arousal and reducing sleep quality. Moreover, amyloid-β disrupts orexin neurons in the lateral hypothalamus, promoting daytime sleepiness and nighttime wakefulness, which would increase neuroinflammation (Wang and Holtzman 2020) and amyloid levels (Carvalho et al. 2018). Sleep dysregulation is worsened by amyloid-β’s disruption of the suprachiasmatic nucleus, the brain’s master clock in the hypothalamus, which disrupts circadian rhythms (Wang and Holtzman 2020). Furthermore, cholinergic neuron activation facilitates rapid eye movement sleep, and reduced cholinergic tone in amyloid-β pathology leads to fragmented sleep patterns (Wang and Holtzman 2020). Meanwhile, degeneration of prefrontal cortex neurons impairs slow-wave sleep (Parhizkar and Holtzman 2025). Slow-wave sleep is essential for the replay of hippocampal activity involved in memory consolidation in the cortex (Lee and Wilson 2002). Disrupted sleep would also increase amyloid-β levels due to reduced clearance through the glymphatic system (Wang and Holtzman 2020), thereby perpetuating the cycle of sleep disruption and amyloid-β accumulation, leading to cognitive decline.

Recall Defects

Excitatory synapse loss elicited by hyperactivity will reduce the sources of excitability of a neuron in the network. Consequently, the neural activity pattern required for reliable memory recall will not be reinstated. Thus, synapse loss without neurodegeneration can weaken engrams, leading to recall deficits. Synaptic weakening often triggers compensatory strengthening of unrelated synaptic connections in an attempt to maintain overall network activity. However, such compensation would inadvertently promote the strengthening of neural connections unrelated to specific memory traces, thereby further disrupting the fidelity of memory storage. Upon cue exposure, these aberrant circuits get activated and interfere with the recall of original memory (Poll et al. 2020). Thus, aberrant neural activity disrupts the encoding and recall of memories, primarily due to interference.

The Relevance of Mouse Models of AD

Amyloid deposition occurs decades before the emergence of cognitive decline in humans, and the presence of amyloid deposits is not sufficient to elicit cognitive deficits. In contrast, mouse models of amyloid pathology develop cognitive deficits even without tau pathology. Consequently, mouse models of amyloid enable mechanistic investigations into how amyloid disrupts cognition. On the other hand, it raises the question of why humans are less vulnerable to amyloid-mediated cognitive decline than mice. Various factors may account for the species difference, ranging from purely technical reasons to complex biological underpinnings. Most animal studies report average deficits across a set of animals and do not distinguish them by the extent of susceptibility. The rate of change in cellular exposure to amyloid could also influence susceptibility. Exogenous application of amyloid peptides to cultured neurons leads to acute, pronounced synaptic disruption, whereas in transgenic mouse models, amyloid accumulation occurs more gradually, with local concentrations likely remaining lower. These models display more progressive and subtle synaptic changes. In humans, in whom there is no overexpression of amyloid precursor protein, as in many mouse models, the buildup of amyloid is even slower and less concentrated, suggesting that synaptic toxicity could depend on the rate of amyloid accumulation.

The differential susceptibility to amyloid in humans is linked to differences in brain architecture, with individuals with larger brains or higher synapse density showing greater resilience against cognitive decline (Katzman et al. 1988; Boros et al. 2017; Walker and Herskowitz 2021; Walker et al. 2024). Additional evidence demonstrating the protective role of lifestyle factors, such as education and social engagement, led to a more active form of resilience, dubbed cognitive reserve, which confers protection against amyloid pathology despite similar brain reserve (Stern 2012). Mice, with much smaller brains and synapse sizes, housed in laboratory conditions, will have limited reserves to protect cognition, particularly when assessed using ethologically irrelevant behavioral tasks, in the face of amyloid pathology.

The memory deficits in AD mouse models have been reversed multiple times by reversing any one of the multitude of deficits, such as mitochondrial, lysosomal, metabolic, astrocytic, and microglial disruptions, described earlier. How could reversing one of the many deficits be sufficient to reverse memory impairment? One possibility is that the multitude of molecular deficits observed in the AD model converge on a few key network-state and plasticity-capacity variables. The state of a network is determined largely by the gain (neural excitability) and timing of activation of individual neurons. These two variables constrain noise levels (precision), sparseness, coactivity (functional connectivity), and oscillation frequencies. Similarly, whether a network is amenable to plasticity (plasticity-capacity) depends on whether individual synapses sense changes in activity patterns, generate appropriate signals, and modulate synaptic proteins, thereby potentiating or depressing synapses. Memories would fail when network-state and plasticity-capacity variables are modified, pushing them outside the permissive zone for encoding, consolidation, or retrieval (Fig. 7). The permissive zone for mice is likely to be narrower than that for humans, as evidenced by improved plasticity and memory in mice after engraftment of human astrocytes or neurons (Han et al. 2013). The larger permissive zone for humans, due to fundamental differences in size and life span, as well as learning demands, requires a larger-scale disruption of network-state and plasticity-capacity axes before cognitive decline reaches clinical grade. In contrast, a smaller nudge in one of these variables is sufficient for mice to lose or restore their memories. Therefore, many interventions developed using mouse models do not translate well into successful therapeutics. However, this does not diminish the importance of amyloid mouse models; in fact, the reversibility of cognitive decline in these models with interventions allows us to distinguish the primary drivers from secondary effects. Furthermore, mouse models enable us to dissect the mechanisms by which these diverse molecular pathologies converge onto low-dimensional network variables. Mapping this relationship would allow us to choose therapeutics that target different low-dimensional network-state and plasticity-capacity variables, such that they provide a synergistic effect. Furthermore, understanding the molecules to network linkage will facilitate the development of more complex computational and organoid models to study human memory disruption under cognitive reserve failure.

Figure 7.

A model for the differential effect of amyloid-β pathology on cognition in humans and mice. Conceptual state space showing how network excitability (y-axis) and plasticity-capacity (x-axis) jointly determine a memory-permissive zone. The boxed region denotes ranges where neuronal activity and synaptic dynamics support circuit reorganization for memory, while the rest of the space indicates impermissive states marked by hyperactivity or hypoactivity and by excessively stable or excessively labile synapses. A wider permissive zone for humans and a narrower one for mice reflect differences in redundancy, and functional buffering.