From synapses to circuits and back: Bridging levels of understanding in animal models of Alzheimer’s disease

Abstract

Alzheimer’s disease (AD) is a neurodegenerative disorder characterized by behavioural changes that include memory loss and cognitive decline and is associated with the appearance of amyloid-β plaques and neurofibrillary tangles throughout the brain. Although aspects of the disease percolate across multiple levels of neuronal organization, from the cellular to the behavioural, it is increasingly clear that circuits are a critical junction between the cellular pathology and the behavioural phenotypes that bookend these levels of analyses. In this review, we discuss critical aspects of neural circuit research, beginning with synapses and progressing to network activity and how they influence our understanding of disease processed in AD.

1 |. ALZHEIMER’S DISEASE: LEVELS OF ANALYSIS

Alzheimer’s disease (AD) is a neurodegenerative disorder that causes an array of cognitive and behavioural symptoms, including memory loss, impaired language function, visuospatial disorientation and social deficits (Greene et al., 1996; McKhann et al., 1984; Mega et al., 1996; Savva et al., 2009). Described by Alois Alzheimer at the start of the 20th century, the diagnosis of AD has relied largely on patient history, with progressive impairment in memory, language and executive function serving as the clinical hallmarks of the disease (McKhann et al., 1984, 2011). The search of the neurobiological origins of this cognitive decline has focused on the cellular and molecular basis of AD (Braak & Braak, 1991; Thal et al., 2002). Two histological hallmarks of pathology have emerged: amyloid plaques, extracellular aggregates of amyloid-β (Aβ), and neurofibrillary tangles, intracellular inclusions of aberrantly phosphorylated tau protein (Braak et al., 2011; Hardy & Higgins, 1992; Kosik et al., 1986; Mandelkow & Mandelkow, 1998; Selkoe, 1991).

Aβ is a misfolded protein formed from the metabolism of amyloid precursor protein (APP), a transmembrane protein concentrated in neuronal synapses (Masters et al., 1985). Aβ is generated due to the cleavage of APP by β-secretase and γ-secretase (Selkoe, 1994). In familial autosomal dominant forms of AD, the generation of Aβ is increased (Goate et al., 1991; Sherrington et al., 1995; Wolfe et al., 1999). In sporadic or late-onset AD, it is believed that Aβ clearance is decreased and that the propensity of Aβ, especially the Aβ₄₂ species, to aggregate into oligomers, fibrils and plaques, is increased (Genin et al., 2011; Morris et al., 2010; Sanan et al., 1994; Selkoe, 2002). Aβ pathology occurs in both cortical regions as well as the medial temporal lobe and hippocampal formation, spreading to other regions in an anterograde manner (Thal et al., 2002; Thal, Rüb, et al., 2000).

Tau is a microtubule-associated protein in axons that regulate microtubule stability (Weingarten et al., 1975) and intracellular trafficking (Ram et al., 2008; Vershinin et al., 2007). In AD, tau loses its ability to bind to microtubules and undergoes an array of post-translational modifications, including hyperphosphorylation and glycosylation (Grundke-Iqbal, Iqbal, Tung, et al., 1986; Wang et al., 1996). These post-translational modifications, as well as the increased cytosolic concentration of tau present due to its diminished ability to bind to microtubules, allow for tau molecules to interact with one another and generate paired helical fragments, neurofibrillary tangles and other aggregates (Congdon & Sigurdsson, 2018; Grundke-Iqbal, Iqbal, Quinlan, et al., 1986; Kidd, 1963; Meraz-Ríos et al., 2010). Tau pathology, first appears in the transentorhinal region (Braak Stages I and II), expands to involve the hippocampus proper and other limbic areas (Braak Stages III and IV) and then spreads throughout the cortex (Braak Stages V and VI) (Braak et al., 2011; Braak & Braak, 1991). The degree of tau pathology has been found to correlate strongly with the cognitive impairment (Bancher et al., 1993; Brier et al., 2016; Giannakopoulos et al., 2003; Thal, Holzer, et al., 2000), possibly even more strongly than Aβ pathology, though this remains an area of active research. The spatio-temporal spread of tau pathology has four distinct trajectories, with each accounting for approximately 20–30% of cases (Vogel et al., 2021) and each of which may point to specific subnetworks within cortical and subcortical circuits that are vulnerable.

If one regards the study of AD across a continuous hierarchy of nervous system organization (molecular and cellular, synaptic and neuronal, circuits and behaviour), then the behaviours that typify patient presentation in AD and the cellular and molecular pathology associated with AD lie at two ends of this hierarchy (Figure 1a). Studies that start with behaviour have their origins in function, diving down to trace disease aetiology back to cellular and molecular pathology. By contrast, studies that begin with molecular and cellular pathology adopt a bottom-up strategy, ascending across levels of analyses to behaviour. Irrespective of the path taken, all approaches pass through the level of neural circuits, a critical junction between the molecular and behavioural approaches. We broadly define the circuit as the anatomical and physiological properties of individual neurons and the functional networks they form. As a number of methods exist for monitoring the structure and function of the circuit over time, an approach for investigation that originates at the circuit would be especially powerful in tracking changes in the architecture of the brain from early to mid to late stages of disease across each of these levels (Figure 1b). Thus, instead of seeing the circuit as a waypoint between two levels of analysis, we propose that beginning at the level of the circuit and working outward towards both behaviour and cellular and molecular pathology could provide critical insight into the aetiology of AD and shed light on the gaps in our understanding. For instance, why, in certain patients, does Aβ plaque burden fail to correlate strongly with cognitive impairment (Crystal et al., 1988; Katzman et al., 1988; Price & Morris, 1999; Tomlinson et al., 1968). Alternatively, why are certain emotional and cognitive domains diminished, whereas others are preserved in AD (Concepcion et al., 2015; Greene et al., 1996; McKhann et al., 2011; Mega et al., 1996; Merriam et al., 2018; Savva et al., 2009; Serby et al., 1991; Teri et al., 1999; Warner et al., 1986), even when patients appear to share similar hallmarks of disease pathology?

FIGURE 1.

(a) Schematic of space defined by the levels at which AD is studied, the brain areas within animal models where these experiments were performed and the models of pathology that were used to recapitulate disease processes. The cellular and molecular hallmarks of AD pathology and the behavioural profile of patients with AD are well known (green areas), whereas the impact of that pathology on synapses, circuits and population activity is the focus of this review (yellow areas). (b) An added dimension of study in AD is the changes in the brain across different levels of analysis, different brain areas and different models over time. (c, d) Experiments choose specific aspects of the space of inquiry to study, thereby carving out planes within the volume. For example, studies that look at the impact of different pathology models within a single brain area across molecular, synaptic and cellular levels (c) would be orthogonal to studies that examine the impact a single pathology model has across different brain areas on synaptic and neuronal properties

Studying AD from the perspective of neural circuits is more than just an alternative epistemological strategy. Consider, for example, the primary pharmacological interventions for AD, acetylcholinesterase inhibitors and N-methyl-d-aspartate (NMDA) receptor antagonists (Massoud & Léger, 2011). These medications can alleviate some of the behavioural symptoms of dementia in some patients but do not undo the cellular and molecular damage of AD or substantially alter the course of the disease (Howard et al., 2012; Kaduszkiewicz et al., 2005; Tariot et al., 2004). Conversely, experimental ‘disease-modifying therapies’, such as monoclonal antibodies against Aβ, which demonstrably reduce Aβ plaque burden (Farlow et al., 2012; Rinne et al., 2010; Siemers et al., 2010), have nonetheless had limited success in ameliorating cognitive function (Cummings et al., 2014; Doody et al., 2014; Salloway et al., 2014) (aducanumab, an Aβ-directed monoclonal antibody-targeted therapy, being a recent controversial exception [Musiek & Bennett, 2021; Sevigny et al., 2016]). Finally, alterations in neural circuits, such as changes in dendritic morphology and disruptions in synaptic plasticity, are well documented in the context of Aβ and tau pathology (Klyubin et al., 2005; Palop et al., 2007; Šišková et al., 2014; Wei et al., 2009; Yoshiyama et al., 2007). Drugs targeting Aβ or tau that may alleviate plaque burden or remove tangles may not fix the damage by that pathology done to the synapses and networks of neurons that underlie memory, cognition and behaviour. By contrast, recent approaches that target circuits appear to not only improve cognition but also decrease cellular pathology (Adaikkan et al., 2019; Iaccarino et al., 2016; Martorell et al., 2019).

In this review, we will explore the literature on the effects of Aβ and tau pathology through the lens of the neural circuit, starting with individual synapses and single neurons and ending with populations of cells. To understand how cellular pathology alters the circuit and, in turn, cognition and behaviour, we will review frameworks at each level of analysis that may provide critical insights into the neurobiology of AD. In this regard, it is useful to think of specific experiments as slices through this space of analysis whose dimensions are defined by the brain area being studied, the model being employed and the level of analysis with which the system is being interrogated. For example, a study that focuses on different models (APP/PS1 vs. 3xTg) within a single brain area, such as the cornu Ammonis subfield 1 (CA1) region of hippocampus (Figure 1c, top), would constitute a slice in this space orthogonal to a study that looks at a specific pathological model (APP/PS1) across two different brain areas (neocortex vs. hippocampus, Figure 1c, bottom). The goal of this review is to discuss how current experiments fill in the volume of understanding of AD across regions and animal models within the levels of analysis of the neural circuit.

2 |. SYNAPSES AND AD

Although Aβ plaques and tau tangles are the hallmarks of AD, changes in the structure and function of synapses appear to be more correlated with the cognitive decline observed in patients (Terry et al., 1991) and in animal models of the disease (Selkoe, 2002). This is perhaps expected, as synaptic connectivity forms the neural basis of learning and memory, and thus, disruptions in synaptic properties affect cognition. Over the last decade, a number of critical insights have been gained about the structure and function of synapses. When integrated with studies on synaptic damage in animal models of AD, such findings link cellular pathology with changes in neuronal activity.

On excitatory neurons in both cortex and hippocampus, dendrites are pockmarked with synapses whose structure and function provide critical insights into single-neuron computation. Excitatory synaptic connections are made on spines, small protuberances along the dendritic tree that vary in size and shape (Engert & Bonhoeffer, 1999; Hausser et al., 2000; Kasai et al., 2010; Yuste & Bonhoeffer, 2001; Zuo et al., 2005). Spine size appears to be related to synaptic strength (Arellano et al., 2007; O’Donnell et al., 2011), which, in turn, plays a critical role in memory and learning. Beginning with studies by Bliss and Lømo (1973), the use-dependent strengthening or weakening of synapses, referred to as long-term potentiation (LTP) and long-term depression (LTD), has provided a framework for connecting synaptic structure to functional plasticity (Madison et al., 1991). Bursts of high-frequency electrical stimuli in presynaptic populations resulted in persistent increases in the amplitude of excitatory post-synaptic potentials (EPSPs) between neurons. In subsequent studies, plasticity was induced by tightly coupling the timing of presynaptic and post-synaptic action potentials, so-called spike-time-dependent plasticity (STDP) (Bi & Poo, 1998; Markram et al., 1997; Sjöström & Nelson, 2002). Here, in order to induce synaptic plasticity, the post-synaptic neuron must fire precisely within a narrow interval following presynaptic spiking (Bi & Poo, 1998; Froemke & Dan, 2002). Interestingly, the rules governing how synapses are strengthened or weakened in STDP depend not only on the statistical properties of spiking (Froemke & Dan, 2002; Sjöström et al., 2001) but also on the location along the dendrite where the synapse is located (Froemke et al., 2005). LTP and STDP represent biological instantiations of Donald Hebb’s postulate on learning (Hebb, 1949). Importantly, rather than being singular, the rules for plasticity, and therefore learning, arise from a constellation of factors related to the structure of synapses and dendrites.

This diversity in synapse structure and function is critical for understanding the molecular and cellular hallmarks of AD, such as Aβ and tau pathology (Figure 2, left). Consequently, disruptions in synaptic function and the transmission of signals from one neuron to another are altered in different ways depending on the brain area and the synapse type. These effects occur both pre- and post-synaptically. For example, Aβ decreases presynaptic release probabilities (Abramov et al., 2009). On the post-synaptic side, it induces the internalization and removal of α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) and NMDA receptors (Hsieh et al., 2006; Palop et al., 2007; Snyder et al., 2005) as well as the inhibition of paired-pulse facilitation and LTP (Chapman et al., 1999; Palop et al., 2007; Wei et al., 2009). These disruptions to synaptic plasticity, specifically LTP, have been replicated when human Aβ oligomers were injected into rodent hippocampus (Klyubin et al., 2005; Shankar et al., 2008). Furthermore, in hippocampal slices from APP/PS1 mice, Aβ alters in the temporal windows required for STDP (Garad et al., 2021), suggesting that the precise plasticity rules associated with STDP may be subject to amyloid pathology. This inhibition of LTP by Aβ may be mediated by the glycogen synthase kinase-3β (GSK-3β) signalling pathway, wherein the dephosphorylation of the key synaptic scaf-folding protein post-synaptic density protein 95 (PSD-95) leads to removal of AMPA receptors, though many other mechanisms have also been proposed (Guntupalli et al., 2016; Jo et al., 2011; Kim et al., 2007; Nelson et al., 2013). Similar disruptions to synaptic function and LTP, as well as synaptic loss, have been reported in the context of tau pathology (Biundo et al., 2018; Di et al., 2016; Fá et al., 2016; Hoover et al., 2010; Lasagna-Reeves et al., 2011; Sydow et al., 2011). Though the molecular mechanisms that underlie this synaptic dysfunction are still being uncovered (Regan et al., 2017), a recent study showed that certain phosphorylated tau species may act through the protein kinase C and casein kinase substrate in neurons protein 1 (PACSIN1) to disrupt AMPA receptor trafficking (Regan et al., 2021). It should be noted, however, that the mechanisms by which Aβ and tau modulate synaptic transmission, as well as the synergistic impact of these two pathologies, are complex and continue to be an area of active investigation (Crimins et al., 2013; Manczak & Reddy, 2013; Palop & Mucke, 2010, 2016; Puzzo et al., 2017; Shipton et al., 2011).

FIGURE 2.

Effects of Aβ pathology across multiple features of the circuit in the APP/PS1 mouse model. (Left) APP/PS1 mice have truncated dendritic trees, deficits in LTP induction, degraded receptive fields and deficits in learning and memory. (Right) Network analysis of the circuit level changes in APP/PS1 animals reveals that functional connectivity is reduced as measured by correlated activity. This leads to a reduction in the combinatorial patterns generated by populations of neurons in APP/PS1 animals in CA1. The reduction in combinatorial patterns could lead to either representations being lost (the loss of patterns corresponding to features in the sensory world or episodes of memory) or degenerated (patterns come to represent multiple features or episodes of memory). Both types of changes in neural representations make specific predictions about how learning and memory are impacted in APP/PS1 animals

These results point to the critical role Aβ and tau pathology play in synaptic organization and function (Fitzjohn et al., 2010; Gengler et al., 2010; Gureviciene et al., 2004), but they also raise important questions. In some cases, Aβ pathology alters the induction of LTP in older mice (Gengler et al., 2010), suggesting that the mechanisms associated with the initial stages of plasticity are impacted in AD. In other examples, the effect seems to be on the decay kinetics of LTP (Gureviciene et al., 2004), suggesting that the effect of Aβ pathology is on the maintenance of recent plasticity events in which synaptic strengths were altered. Under what conditions the effects of AD pathology influence the initiation and maintenance of synaptic plasticity, and the extent to which these effects vary in cortical vs. hippocampal circuits remains an open question.

All findings of the effects of AD pathology on synapses are largely agnostic to the sites of that plasticity along the dendrites. As the location of a synapse on a dendrite affects how the relative timing of pre- and post-synaptic spikes will either strengthen or weaken it (Froemke et al., 2005), changes in the morphology and location of spines seen in Aβ animal models (SantaCruz et al., 2005; Šišková et al., 2014; Tsai et al., 2004; Yoshiyama et al., 2007) warrant further exploration of the parameters of STDP as they pertain to AD. The impact of AD pathology may therefore be contingent not only on how it disrupts synaptic learning but also on the precise location of that pathology on the dendritic tree.

3 |. NEURONS AND AD

Over a century ago, Santiago Ramón y Cajal drawings revealed the immense anatomical diversity of individual neurons (y Cajal, 1911). In later studies, computational models showed how these different cellular architectures, particularly the morphological variations in the dendritic tree, shape the pattern of neuronal firing (adapting, non-adapting and bursting) (Mainen & Sejnowski, 1996). In addition, the particular complement of ion channels carried by different neurons further shapes their biophysics (Gjorgjieva et al., 2016; Padmanabhan & Urban, 2010; Schulz et al., 2006).

The association between Aβ pathology and dysmorphic neurons is well documented. Studies of transgenic mouse lines that overexpress mutant isoforms of APP have shown decreased synapse density, ectopic axon sprouting, thickened boutons and degradation of dendritic spines (Hsia et al., 1999; Phinney et al., 1999; Tsai et al., 2004). Confocal microscopy studies of post-mortem brain tissue from APP mouse models (Figure 2, left) as well as from human patients with AD, showed that the dendrites inside an Aβ plaque were narrower and had a higher curvature than those outside the plaques (Knowles et al., 1999). Though some studies have reported overt neuronal loss in APP mouse models (Hsia et al., 1999), paralleling what has been observed in humans (Gómez-Isla et al., 1997; Vogt et al., 1992; West et al., 1994), others have shown conflicting results (Irizarry et al., 1997). Similar morphological changes have been reported in the context of tau pathology, especially when the tau aggregates into intraneuronal tangles. For example, a study of post-mortem brain tissue from patients with AD showed that neurons that contained neurofibrillary tangles had fewer dendritic spines and more dendritic atrophy than other neurons from the same patients that did not contain tangles (Merino-Serrais et al., 2013; Mijalkov et al., 2021). Synapse loss and brain atrophy have also been reported in mouse models of tauopathy, sometimes even before the accumulation of tangles (SantaCruz et al., 2005; Yoshiyama et al., 2007).

These morphological changes likely have functional consequences for the biophysical properties of neurons. One study found that the increased firing rate of CA1 pyramidal neurons in the APP/PS1 mouse model could be explained by their altered dendritic morphology (decreased branching, surface area and length) and increased intrinsic excitability (Šišková et al., 2014), both features of the neuron itself rather than the synaptic inputs it receives. The evidence for the role of tau pathology in altering these intrinsic neuronal properties is less consistent. For instance, in the p301L R5 transgenic mouse model of tauopathy, in which neurons express a mutant form of tau associated with hereditary frontotemporal dementia (Deters et al., 2008; Götz et al., 2001), similar morphological changes, such as decreased dendritic length and branching, were observed (Müller-Thomsen et al., 2020). At the biophysical level, however, this study and others in tauopathy models found decreased intrinsic excitability of neurons (Hatch et al., 2017). Further research, especially on the potential differential impact of tau on different segments of the neuron (axon, dendrite and soma) (Hatch et al., 2017) and on the interaction between Aβ and tau (Hall et al., 2015), is necessary to elucidate properties of neurons that predict which intrinsic biophysical properties and what morphological elements are disrupted in AD (Cloyd et al., 2021). The initial conditions of the neuron’s identity, including the brain area it originated from, are thus paramount.

Synaptic damage, decreased dendritic branching and altered neuronal excitability in mouse models of Aβ and tau pathology (Hatch et al., 2017; Müller-Thomsen et al., 2020; SantaCruz et al., 2005; Šišková et al., 2014; Tsai et al., 2004; Yoshiyama et al., 2007) raise questions of how such changes affect the computations performed on the dendritic tree, an effective circuit within the single neuron. Previous studies have shown that synaptic inputs and the post-synaptic processing of those inputs vary along the dendritic tree (Dudman et al., 2007; Froemke et al., 2005; Johnston & Wu, 1994; Larkum et al., 1999; Svoboda et al., 1999). How does the reduction in dendritic branching observed in Aβ models (Šišková et al., 2014) affect the proximal inputs, as opposed to the distal ones? One possibility is that distal synapses are simply lost and the presynaptic cells that formerly provided inputs to the neuron no longer do so. Another possibility is that the distal inputs are remapped to more proximal sites along the dendritic tree, perhaps as a compensatory mechanism. If this latter scenario was true, how would this remodelling affect the computations performed by that neuron and the circuit more broadly? Addressing these issues could resolve outstanding mechanistic questions that link changes in dendritic morphology with alterations in both single-neuron and network activity. For instance, morphology-specific pathology could reflect the pathway-specific disruptions in neuronal activity that depend on the activity of presynaptic populations (Cash & Yuste, 1999; Hausser et al., 2001; Larkum et al., 1999; Magee, 2000). Finally, what is the interplay between homeostatic plasticity that affects features such as neuronal excitability and synaptic plasticity in AD models (Hengen et al., 2013; Turrigiano, 2008; Turrigiano et al., 1998)?

Synaptic input, dendritic integration and neural excitability are cellular descriptors that define what causes a neuron to spike. From another perspective, what causes a neuron to spike can also be defined from the perspective of systems neuroscience perspective as the set of stimuli or behaviours that generate action potentials: the neuron’s receptive field, the region of stimulus space that is best suited to evoke a neuronal response (Hartline, 1938; Hubel & Wiesel, 1962; Sherrington, 1952). Receptive fields have been observed across multiple primary sensory modalities (Hubel & Wiesel, 1962; Knudsen & Konishi, 1978) as well as in higher order neural representations (Aronov et al., 2017; Hafting et al., 2005; Okuyama et al., 2016). Recent evidence, however, suggests that individual dendritic spines or dendritic segments on that neuron respond to features that are distinct from what drives the soma (Chen et al., 2011; Jia et al., 2010; Scholl et al., 2021; Varga et al., 2011; Wilson et al., 2018), making that answer far from simple; what drives the spine appears to be different than what drives the neuron. Some studies have shown that synapses with similar receptive fields are spatially clustered on the dendritic tree (Scholl et al., 2017) and that such clustering may help enhance the feature selectivity of neurons (Wilson et al., 2016). A consequence of this functional variation among synapses is that neurons often have subthreshold responses far more complex and diverse than can be gleamed from only looking at spiking activity (Harvey et al., 2009; Poirazi et al., 2003; Schummers et al., 2002; Volgushev et al., 2000). Seen through the lens of AD pathology, questions about what causes a spine to depolarize and what causes a neuron to spike become particularly pressing. For example, changes in the ratio of synapses tuned to the preferred stimulus of the soma and those tuned to non-preferred stimuli, as well as alterations in intrinsic excitability and morphology (Šišková et al., 2014), are all likely to degrade the information encoding abilities of neurons (Cacucci et al., 2008; Cayzac et al., 2015; Grienberger et al., 2012), but which feature contribute to what types of information loss remain unknown.

Studies of changes in the information content of neuronal spiking patterns in AD have largely focused on the hippocampus, especially place cells (Figure 2, left). These are neurons that preferentially increase their firing rate when an animal visits a particular region of its environment, denoted as its place field (O’Keefe & Dostrovsky, 1971; O’Keefe & Nadel, 1978). Multiple studies have shown that place fields are larger, less stable from trial to trial, and less informative about an animal’s location in mouse models of Aβ or tau pathology than in age-matched non-transgenic mice (Cacucci et al., 2008; Cayzac et al., 2015; Cheng & Ji, 2013; Jun et al., 2020; Mably et al., 2017). Moreover, one such study reported that the degree of place field degradation in a given animal was proportional to its Aβ plaque load (Cacucci et al., 2008). These neural coding disruptions extend beyond the hippocampus. For example, although orientation and direction tuning in primary visual cortex neurons were diminished with age in non-transgenic mice, the magnitude of this degradation in tuning was significantly worse in APP23/PS45 mice, relative to age-matched controls (Grienberger et al., 2012). Linking these changes in the neuron’s receptive field back to the changes in the synapse and the dendrite will help illuminate how and why pathology complicates the story of what causes a neuron to spike.

4 |. ENSEMBLE CODING AND ACTIVITY IN AD

Studying the response properties of single neurons can provide some insight into the impact of AD on information coding in the brain (Cacucci et al., 2008; Cayzac et al., 2015; Grienberger et al., 2012). However, individual neurons function as part of a larger network, and it is the coordinated activity of populations of neurons that ultimately shapes perception, behaviour and cognition. It is thus critical to understand the effect of AD pathology on neural circuits and ensembles, which are defined as groups of neurons critical for carrying and representing information within a brain region.

First, even within a single brain area, individual neuronal responses vary across complex stimuli (Okuyama et al., 2016; Quiroga et al., 2005) and different neurons exhibit different kinds of variation to stimuli. It is the response diversity, both across stimuli in a single neuron and across neurons to a single stimulus, that is a necessary prerequisite for the brain to encode the enormous space of possible sensory stimuli (Barlow, 1972). The study of how neurons encode information, be it the way a population of neurons in the cortex represents visual sensory inputs via patterns of activity, or the way in which groups of cells in the hippocampus encode an animal’s position in space or a memory, is a major area of research in systems neuroscience. Consider, for example, the place cells of the CA1 region of the hippocampus (O’Keefe & Dostrovsky, 1971; O’Keefe & Recce, 1993; Skaggs et al., 1993; Wilson & McNaughton, 1993). Place cells are diverse, varying in location and size; neurons in ventral CA1 have place fields that are larger and less informative about the animal’s position than those in dorsal CA1 (Chockanathan & Padmanabhan, 2021; Jung et al., 1994; Keinath et al., 2014; Kjelstrup et al., 2008). Collectively, the ensemble of place cells ‘tile’ the animal’s environment, with the place field of each neuron covering a certain portion of the space. It is thus believed that place cells constitute a key component of the brain’s internal map of the world (McNaughton et al., 2006; O’Keefe & Dostrovsky, 1971). Ensemble representations of place are more stable and robust than single-neuron representations (Meshulam et al., 2017; Ziv et al., 2013). Thus, the animal’s position is defined by the state of the population, patterns of activity that are constrained by the correlations between neurons, including place and non-place cells (Meshulam et al., 2017; Stefanini et al., 2020). Each neuron’s representation is therefore not an island, but part of a larger system, whose structure is determined by the interplay of inputs to the population of neurons and the connections they form with one another. Further evidence for the population code comes from other studies that show that multiple place codes can exist for a single environment and that the population can spontaneously remap from one place to another over the course of months (Low et al., 2021; Sheintuch et al., 2020; Ziv et al., 2013). To decode the position of the animal in such a dynamic and distributed coding scheme, the activity of any individual place cell must be considered in the context of the other neurons in the population.

Hebb (1949) proposed that these ‘cell assemblies’, groups of neurons that repeatedly coactivated, could form the neural substrate for a sensory percept or a memory. Hopfield (1982) developed this idea in a network model, revealing how memory stability, pattern completion and error correction could arise emergently from the dynamics of these assemblies. Over the past decade, technological advancements have enabled the recording and manipulation of large neuronal population, thus allowing these frameworks to be explored experimentally. For example, the technique of two-photon microscopy coupled with holographic optogenetic stimulation (Packer et al., 2015; Rickgauer et al., 2014) allows for specific subpopulations of neurons to be coactivated with high spatio-temporal precision while simultaneously imaging all the activity of a large population of neurons. In these experiments, stimulating subsets of cells is sufficient to repeatedly recruit and activate larger ensembles of neurons, illustrating the interconnectedness of the component cells that cortical ensembles (Carrillo-Reid et al., 2016, 2019). These findings are not limited to the visual system; optogenetic stimulation experiments in the prefrontal cortex and hippocampus have shown the presence of ensembles that code for location, conspecific identity and fear (Kitamura et al., 2017; Liu et al., 2012; Okuyama et al., 2016; Steve et al., 2013), multiple lines of evidence that illustrate the criticality of populations of neurons in determining perception and behaviour.

An essential component of the level of analysis between the single neuron (in an APP/PS1 animal model for instance [Cacucci et al., 2008; Cayzac et al., 2015; Grienberger et al., 2012]) and behaviour in AD is the activity of the population. In vivo studies using two-photon imaging of calcium transients in multiple mouse models of Aβ pathology, including the APP/PS1 model, found populations of both hyperactive and hypoactive neurons in frontal cortex, with hyperactive neurons clustered near plaques, whereas hypoactive neurons were distributed more uniformly (Busche et al., 2008). Follow-up studies showed complex changes in neuronal activity in visual cortex (Grienberger et al., 2012) and dorsal CA1 hippocampus (Busche et al., 2012). In the former, significant changes in activity occurred even before the deposition of Aβ plaques. By contrast, experiments performed in the context of tau pathology revealed drastically different results. In the rTg4510 mouse model, which expresses a form of human tau (P301L) containing a mutation associated with frontotemporal dementia and parkinsonism, high levels of tau pathology were observed in the fore-brain (Ramsden et al., 2005). In this model, widespread suppression of neuronal activity was observed (Busche et al., 2019). Furthermore, in mice harbouring APP/PS1 mutations for Aβ pathology as well P301L mutations for tauopathy, the network activity was silenced, similar to the result observed in the rTg4510 model (Busche et al., 2019). These results suggest that, at the level of neuronal populations, the effects of tau may dominate those of Aβ, though more research is needed to elucidate the mechanisms by which these pathologies interact and the resulting alterations in network activity.

In the hippocampus, a study of population coding in the rTg4510 mouse model of tauopathy found that although the spatial information of individual place cells was decreased relative to non-transgenic controls, as had been reported in mouse models of Aβ pathology (Cacucci et al., 2008; Cayzac et al., 2015), ordered sequences of place cell activation were preserved (Cheng & Ji, 2013). Even though hippocampal neurons in the rTg4510 model failed to fire at the same position of the environment in each trial run, ensembles of neurons still activated in a replicable sequence, similar to the manner in which ensembles of place cells sequentially activate when an animal explores its environment. This suggests that the mechanisms that functionally bind groups of individual neurons into recurring ensembles are intact in the rTg4510 model. Interestingly, however, features of population activity tend to be weakened in other models of Aβ and tau pathology. In an electrophysiological study of large dorsal CA1 populations in awake mice, correlations between neurons were diminished in the APP/PS1 mice, relative to the controls, and this produced a reduction in the diversity of patterns generated by populations of neurons (Chockanathan et al., 2020). Reduction in the diversity of patterns or ensembles reveals that the neural ‘vocabulary’ available to populations of neurons is truncated in an Aβ model (Figure 2, right). That different studies find different effects of AD pathology on population coding is not surprising. For example, differences between Cheng and Ji (2013) and Chockanathan et al. (2020) may evidence the heterogeneous effects of pathology in balancing the influence of external stimuli (the position of the animal) and internal states and structures (the correlations between neurons) (Meshulam et al., 2017). AD pathology may decouple features of intrinsically generated hippocampal activity from external sensory inputs (Cacucci et al., 2008; Cayzac et al., 2015; Grienberger et al., 2012), thereby increasing the impact of local networks within the hippocampus on structuring activity. Additionally, changes observed at the network level at a single point in the late stages of Aβ pathology likely reflect the accumulated impact of many changes in the circuit over the progression of the disease. For example, aberrantly excitatory neural circuits in mouse models of amyloidosis show compensatory remodelling of inhibitory circuits (Palop et al., 2007), and studies that track changes in population activity over time will provide critical insights into how pathology ushers network activity into different states over the course of the disease.

These studies of population coding in the context of Aβ pathology have mostly employed transgenic models, such as the APP/PS1 line (Jankowsky et al., 2001). The advantage of these models is that they show robust plaque expression in the cortex and hippocampus (Whitesell et al., 2018), as seen in humans with AD (Thal et al., 2002; Thal, Rüb, et al., 2000). However, in addition to increasing Aβ pathology, such models also present potential confounds resulting from the insertion of the transgene and the overexpression of protein, including destruction of endogenous genes, overproduction of APP-associated species other than Aβ and non-specific endoplasmic reticulum stress (Barbero-Camps et al., 2014; Saito et al., 2014, 2016; Sasaguri et al., 2017). Knock-in mice present a way of modelling Aβ pathology while limiting these confounding effects (Drummond & Wisniewski, 2017; Saito et al., 2014). One widespread model is the APP^NL-G-F strain (Saito et al., 2014). Encouragingly, recent studies that employed this model have replicated some of the key earlier findings from transgenic lines, such as spatial memory impairments (Saito et al., 2014), LTP deficits (Latif-Hernandez et al., 2020), decreased spatial information of place cells (Jun et al., 2020) and hyperactivity of neurons that are near sites of Aβ accumulation (Takamura et al., 2021). These studies naturally raise questions about how single-neuron and synaptic disruptions in these AD models may influence population and network activity.

5 |. OSCILLATORY ACTIVITY AND AD

These mesoscopic measures of circuit structure and function, as well as their disruptions in AD, reflect global patterns of activity throughout the brain. Methods such as electroencephalogram (EEG) and functional magnetic resonance imaging (MRI) that study the organization of brain activity at these scales in patients have provided the closest links to the mesoscopic activity of circuits. A critical link between measures of neural population activity in humans and those in animal models is the oscillations in the local field potential (LFP). LFP oscillations have been shown to play critical roles in spatial processing, temporal synchronization of spiking activity and memory consolidation (Dupret et al., 2010; Foster & Wilson, 2007; Jadhav et al., 2012; O’Keefe & Recce, 1993; Skaggs et al., 1996). Three frequencies in the LFP have a particular relevance to AD: theta oscillations, gamma oscillations, and sharp waves and ripples (SWRs).

Theta waves are 6–12 Hz oscillations found most strongly in the hippocampus (Vanderwolf, 1969). Several lines of evidence implicate theta oscillations in spatial coding. First, in both animal models and humans, the frequency and power of theta oscillations have been shown to increase during locomotion and active exploration and decrease during periods of immobility (Vanderwolf, 1969; McFarland et al., 1975; Rivas et al., 1996; M. Aghajan et al., 2017). Second, during exploration of an environment, place cells fire in a stereotyped manner with respect to the phase of the local theta oscillation (Harvey et al., 2009; O’Keefe & Recce, 1993; Skaggs et al., 1996). Namely, a place cell will fire action potentials at progressively more advanced phases of the theta oscillation as an animal traverses that neuron’s place field. This phenomenon, known as phase precession, offers numerous computational benefits, from increasing the precision of the cognitive map to addressing ambiguities that arise from different trajectories through the same location (Mehta et al., 1997, 2002). In populations, which are activated as an animal explores its environment, phase precession temporally orchestrates the firing of place cell into repeated sequences that occur on the time scale of STDP (Bi & Poo, 1998); even though an animal may take several seconds to traverse multiple place fields, theta-orchestrated place cell sequences could occur within tens of milliseconds. This offers a potential mechanism to strengthen the synaptic connections between the place cells involved and, in turn, increases the reliability of the hippocampal spatial code through population-level representations of space (Meshulam et al., 2017; Stefanini et al., 2020). The strengthening of these sequences may also facilitate memory consolidation during SWRs, as well as replay and preplay events.

Gamma oscillations are 30–100 Hz fluctuations in the hippocampus and cortex that manifest in a range of behavioural contexts, from sensory stimulation to selective attention to active exploration (Bragin et al., 1995; Fries et al., 2001; Gray et al., 1989; Pesaran et al., 2002). Several studies have linked the generation of gamma rhythms to inhibitory interneurons in the hippocampus (Bartos et al., 2007; Bragin et al., 1995; Cobb et al., 1995; Hájos et al., 2004; Lytton & Sejnowski, 1991; Whittington et al., 1995). In particular, slow gamma oscillations (25–50 Hz) arise from networks of fast-spiking parvalbumin (PV) interneurons (Buzsáki & Vanderwolf, 1983; Cardin et al., 2009; Carlén et al., 2012; Fries, 2009; Sohal et al., 2009; Traub et al., 1996). These oscillations are important for synchronizing neural activity from different parts of the brain (Engel et al., 1991; Gray, 1994; Gray et al., 1989), whereas fast gamma oscillations (50–100 Hz) facilitate communication between entorhinal cortex and hippocampus to allow for the encoding of new sensory information into memories (Bieri et al., 2014; Zheng, Bieri, Hsiao, & Colgin, 2016; Zheng, Bieri, Hwaun, & Colgin, 2016).

Finally, the SWR is a large-amplitude high-frequency (approximately 200 Hz) oscillation observed during immobility, eating, drinking and slow-wave sleep (Buzsáki, 1986; O’Keefe & Nadel, 1978; Vanderwolf, 1969). In SWRs that occurred during rest periods, it was found that ensembles of place cells rapidly reactivate. The temporal compression of these ‘replay’ and ‘preplay’ events, which occur on the order of hundreds of milliseconds (Diba & Buzsáki, 2007; Foster & Wilson, 2006), has led some to posit that their purpose is to strengthen the synaptic connections between the involved neurons by STDP (Bi & Poo, 1998), thereby facilitating learning and memory (Dragoi & Tonegawa, 2011; Dupret et al., 2010; Gillespie et al., 2021; Girardeau & Zugaro, 2011; Heller & Bagot, 2020; Jadhav et al., 2012; Ji & Wilson, 2006; Joo & Frank, 2018; Maingret et al., 2016; Pfeiffer & Foster, 2013).

Mouse models of Aβ and tau pathology also show disruptions in the LFP, including theta and gamma oscillations as well as SWRs. In the APP/PS1 mouse model, both theta power and theta frequency were reduced, relative to age-matched non-transgenic mice, even though average running velocity was higher in the APP/PS1 group (Cayzac et al., 2015; Scott et al., 2012). Though the explicit impact of this weakened theta oscillation on phase precession has yet to be extensively studied, the relationship between the theta phase and rhythmic spiking activity in place cells is weakened in the 3xTg mouse model of Aβ and tau pathology (Mably et al., 2017). Decrease oscillatory synchrony likely erodes sequences of activation (phase precession), as well as plasticity during those windows, possibly compounded by deficits in STDP and induction of LTP in APP/PS1 hippocampal neurons, particularly when the neuron is near an Aβ plaque (Garad et al., 2021). The net effect of these changes may be the poor spatial memory and decreased place cell information content in mouse models of Aβ pathology (Figure 2, left), as well as the recent finding that place cells in APP knock-in mice fail to remap between different environments (Cacucci et al., 2008; Cayzac et al., 2015; Jun et al., 2020).

Decreased power in the slow gamma frequency band has been observed in mouse models of tauopathy (Booth et al., 2016; Mably et al., 2017) and Aβ pathology (Goutagny et al., 2013; Iaccarino et al., 2016) as well as in the ApoE4 knock-in model of late-onset AD (Gillespie et al., 2016). In many of these studies, the deficits in gamma oscillations do not occur in isolation but rather manifest as decreased coupling between theta oscillations and gamma bursts (Booth et al., 2016) or decreased SWR-triggered gamma oscillations (Iaccarino et al., 2016). Consistent with the hypothesis that fast gamma rhythms are involved in encoding of new memories, whereas slow gamma rhythms are involved in retrieval of previously stored memories (Mably & Colgin, 2018), one recent study showed that APP/PS1 mice have a deficit in the retrieval, rather than storage or encoding, of memories (Roy et al., 2016). Researchers were able to rescue the deficit in contextual fear conditioning seen in APP/PS1 mice by optogenetically reactivating the ensembles of neurons that were initially activated in a particular context. This result implied that even though recall of that memory was impaired, the neural substrate of the memory was still present in the brain. In multiple mouse models of AD, GABAergic interneuron dysfunction has been observed (Kurudenkandy et al., 2014; Leung et al., 2012), presenting a possible mechanism for disrupted slow gamma rhythms. Moreover, optogenetic stimulation of PV interneurons with a 40 Hz entrainment signal can not only restore slow gamma oscillations but also reduce Aβ pathology in 5XFAD mice (Iaccarino et al., 2016). Encouragingly, follow-up studies have suggested that non-invasive audiovisual sensory stimulation at 40 Hz can also restore slow gamma oscillations throughout the brain, reduce tau and Aβ pathology, and improve cognition in multiple mouse models (Adaikkan et al., 2019; Iaccarino et al., 2016; Martorell et al., 2019).

Many studies have shown a reduction in the amplitude, intensity, occurrence and frequency of hippocampal SWRs in mouse models of Aβ and tau pathology (Ciupek et al., 2015; Gillespie et al., 2016; Jones et al., 2019; Jura et al., 2019; Nicole et al., 2016). One study of the apoE4 mouse model, which expresses a major genetic risk factor in sporadic human AD, showed not only that SWRs were sparser in transgenic mice than the controls but also that future impairments in learning were correlated with the abundance of SWRs (Jones et al., 2019). A recent study in 5XFAD mice revealed that SWRs occur less frequently and with a shorter duration relative to wild-type controls (Prince et al., 2021). Furthermore, although place cells were less likely to be reactivated during SWRs, the activity of place cells during active exploratory behaviour was similar between the two groups, suggesting that the features of the LFP and ensemble activity may be damaged in AD even in the absence of changes to single-neuron tuning properties (Prince et al., 2021). Decreased occurrence of SWRs was also observed in the rTg4510 tauopathy model; here, the reduction of ripples actually preceded the onset of tauopathy and the degradation of place fields (Ciupek et al., 2015). In both Aβ and tauopathy models, the relationship between SWRs and neuronal class appears to vary by neuron type, with excitatory pyramidal neurons more likely to be recruited during SWRs and inhibitory interneurons less likely to fire during SWRs, relative to age-matched controls (Caccavano et al., 2020; Witton et al., 2016). Taken together, these data suggest that features of population activity, such as SWRs, may be better indicators of neuropathology and cognitive dysfunction than the firing patterns of individual neurons. Despite this well-documented association between disruptions in SWR and mouse models of tau and Aβ pathology, the impact of such pathology on phenomena such as replay and preplay is unknown.

6 |. CONCLUSION

The current landscape of AD research is one of scale and approach. Studies at the behavioural level and the molecular and cellular level reflect some of the leading-edge strategies for understanding and treating the disease (Figure 1a, green). At the junction between these approaches in the circuit (Figure 1a, yellow), broadly defined at scales from the synapse and neuron (with complex dendritic topology and principles of synaptic integration) on one end to the connections of groups of cells across multiple brain regions at the other end. As such, there remain numerous outstanding questions in AD research that may inform what cell types and regions are most affected by the cellular pathology as well as how behavioural and cognitive disruptions arise out of changes to circuits and networks in AD.

The framework we have proposed implicitly assumes the translatability of Aβ and tau models. Although the studies of Aβ and tau models described in this review have significantly increased our understanding of AD, several important caveats should be noted (Morris et al., 2014). First, although transgenic models of Aβ pathology have generally expressed forms of APP and presenilin associated with early onset autosomal dominant AD pathogenesis and progression (Goate et al., 1991; Sherrington et al., 1995; Wolfe et al., 1999), most cases of AD in humans are sporadic and not associated with mutations in these genes. Epidemiological studies have revealed certain gene mutations associated with sporadic AD, most notably the ApoE4 allele (Corder et al., 1993; Genin et al., 2011; Jonsson et al., 2012; Morris et al., 2010). Mouse strains incorporating such mutations have been developed (Sullivan et al., 1997, 2004), and as these may model the pathology of human AD more closely than traditional transgenic lines, they represent an important complementary line of investigation. Interestingly, the effect of this pathology across various levels of analysis could provide a parallel avenue for insights into how AD affects circuits. Second, more work needs to be done on the interplay of tau and Aβ, such as the induction and exacerbation of tau pathology by Aβ (Bennett et al., 2017; Bolmont et al., 2007; Lewis et al., 2001). Mouse lines that incorporate both Aβ and tau overexpression, such as the 3xTg strain (Oddo et al., 2003), represent one way to study these complex synergistic effects. Third, it should be acknowledged that Aβ and tau pathology do not occur in a vacuum but rather are imbedded within a rich immunological landscape within the brain (Akiyama et al., 2000; Heneka et al., 2015). Elucidating the role that diverse neuroimmunological cells play in regulating AD associated pathology may reveal how changes in neuronal–non-neuronal interactions percolate across various levels of the neural circuit (Lee & Landreth, 2010; Wang et al., 2015). All three of these caveats represent fertile grounds for inquiry, particularly at the level of circuits and networks.

The changes in the structure of synapses and dendrites and parallel changes in the rules of learning and memory suggest that the picture of single-neuron pathology is multiparametric, a combination of intrinsic properties such as excitability, structural features such as dendritic morphology, and network elements such as the synaptic inputs and the rules by which those synapses change. As a result, it remains unclear if different cell types (e.g., Layer 2/3 pyramidal cells vs. Layer 5 pyramidal cells), which may vary in these features, are equally affected by Aβ and tau. Here, cell ‘identity’ may be defined not only in terms of molecular and cellular features, such as excitatory vs. inhibitory or PV-positive vs. somatostatin-positive, but also in terms of the role that those cells occupy in the network. For example, the hyperactive and hypoactive cells observed in some Aβ and tau models (Busche et al., 2008, 2019) may serve as highly connected network hubs that therefore exert a disproportionate effect on the overall correlated activity of the population. Aβ and tau pathology may also have heterogeneous effects across different brain regions. For example, recent studies show that the functional connectivity across populations of neurons in the dorsal and ventral CA1 region of the hippocampus can be distinct, reflecting differences not only in the underlying architecture of neurons in these areas but also in their dynamics and this coding strategies (Chockanathan & Padmanabhan, 2021). Does Aβ or tau pathology have heterogeneous effects on the networks in these distinct hippocampal subregions? In light of these questions, it is possible that the most successful approaches to understanding and treating AD will focus on restoring network function in parallel with the elimination of Aβ plaques or tau tangles (Adaikkan et al., 2019; Canter et al., 2016; Iaccarino et al., 2016; Martorell et al., 2019).

Abbreviations:

AD

Alzheimer’s disease

AMPA

α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid

APP

amyloid precursor protein

Aβ

amyloid-β

CA1

cornu Ammonis subfield 1

EEG

electroencephalogram

EPSP

excitatory post-synaptic potential

GABA

gamma-aminobutyric acid

GSK-3β

glycogen synthase kinase-3β

LFP

local field potential

LTD

long-term depression

LTP

long-term potentiation

MRI

magnetic resonance imaging

NMDA

N-methyl-d-aspartate

PACSIN1

protein kinase C and casein kinase substrate in neurons protein 1

PSD-95

post-synaptic density protein 95

PV

parvalbumin

STDP

spike-timing-dependent plasticity

SWR

sharp wave-ripple

DATA AVAILABILITY STATEMENT

No data was generated for this manuscript.