Progressive circuit changes during learning and disease

Abstract

A critical step towards understanding cognition, learning, and brain dysfunction will be identification of the underlying cellular computations that occur in and across discrete brain areas, as well as how they are progressively altered by experience or disease. These computations will be revealed by targeted analyses of the neurons that perform these calculations, defined not only by their firing properties but also by their molecular identity and how they are wired within the local and broad-scale network of the brain. New studies that take advantage of sophisticated genetic tools for cell-type specific identification and control are revealing how learning and neurological disorders initiate and successively change the properties of defined neural circuits. Understanding the temporal sequence of adaptive or pathological synaptic changes across multiple synapses within a network will shed light into how small-scale neural circuits contribute to higher cognitive functions during learning and disease.

Introduction

There’s a joke amongst some neuroscientists: how many neurons does it take to make a decision? The answer is none – single-celled protists and even bacteria can orient toward light or a food source. What we mean by a decision, of course, is much more complex. Creating an account of the neural basis of cognition – from sensation to decision to motor output – remains one of the great scientific challenges of biology.

A satisfying explanation for sensation, cognition, and behavior varies according to the neuroscientist you ask and the kinds of measurements they can make. The opportunity to image and record from ever-increasing populations of neurons in awake and behaving animals tempts us into believing that we are close to understanding how cells and synapses construct mental processes, generate complex behaviors, and adapt to new experiences. This is likely not the case. Just as solving a mathematical equation requires the identification of different terms and variables, neural computations can be revealed by observing the interactions of fundamental units of the brain at the level of cells and synapses. Because the nervous system evolved within the biological constraints of cellular interactions, neural computations must be interpreted and bounded by this context – who is synapsing with whom, what information do cells receive and how do they transform that input into output. This is not new, of course, as neuroscientists have been investigating the properties and responses of individual neurons in order to infer their computation for decades. However, advances in technologies for molecular cell identification, led by pioneering efforts by Josh Huang, the Allen Brain Institute, the GENSAT project, and others (Gerfen et al., 2013; Hippenmeyer et al., 2005; Madisen et al., 2010; Nakashiba et al., 2008; Oliva et al., 2000; Taniguchi et al., 2011) have opened the door for defining the algorithms of the brain with cellular resolution.

Here we will review advances in the field that have begun to address how neurons interact with each other across multiple synapses, identifying and isolating circuit modules that can reveal fundamental principles of brain function. Such an approach contrasts to the statistical analysis of the activity of hundreds or thousands of unidentified neurons in both anatomical and temporal scales, as it necessarily focuses on how inputs are transformed into outputs by local computations within circumscribed brain areas or defined networks. We focus on the well-characterized cortical column and the hippocampal circuit as examples of small-scale computations that can be assembled into larger functional units, where specific synaptic interactions can be isolated to identify sequential and dependent changes. We will discuss how a systematic analysis of how this approach can reveal principles for information processing that are critical for both normal experience-dependent plasticity in the neocortex and also drive maladaptive circuit changes in progressive neurological disorders, centered on the hippocampal circuit in Alzheimer’s Disease (AD). Of particular interest are studies which investigate the synaptic interactions of multiple and molecularly-defined classes of neurons to address the input-output transformation of these local networks, and how they sequentially change during learning and disease.

The distinction between inferring potential neural interactions based upon neural firing during sensation or behavior, versus directly testing them with targeted recordings, synaptic analysis, and interpretations that are constrained by anatomical properties, is critical in developing a principled understanding of brain function. As we move into an era where the properties of individual circuit components can be integrated into a complex network, there is a new opportunity to define how neural circuits can be progressively changed by experience and disease. Such an approach will take into account specific sequences of cellular interactions that are spread out over space and time. Tangible outcomes from this type of systematic dissection of both the neural algorithm and its perturbation will shed light onto how learning and disease adaptively or pathologically reshape brain function.

Plasticity: a common feature of neocortical circuits

The neocortex is a highly-conserved 6-layered circuit found across all mammals, and its cellular and synaptic organization are thought to enable generalized computations that generates an impressive array of sensory and motor functions as well as higher cognitive abilities (Cohen, 2014; Douglas and Martin, 2007; Harris and Mrsic-Flogel, 2013; Lubke and Feldmeyer, 2007; Rowland and Moser, 2014). Cortical neurons are roughly organized into columns, a basic module of cortical circuitry wherein thousands of neurons differentiated by morphology, electrophysiological properties, and synaptic connectivity are arranged across 6 layers into repeating anatomical units. Plasticity within and across cortical areas is a canonical feature of neocortical circuits: it is rapidly engaged in response to experience, has been well-documented across many modalities, and is observed across the lifespan.

Despite recent findings challenging the requirement of cortical circuits for learned behaviors (Hong et al., 2018; Kawai et al., 2015), it is clear that learning itself requires the neocortex and that these circuits themselves are highly plastic. Motor and sensory learning have been shown to change in neocortical circuits at many levels, with changes in fMRI signals (Shibata et al., 2016; Summerfield et al., 2006), topographic input organization (retino-, tono- or somatotopy; (Harris et al., 2001; Kilgard and Merzenich, 1998; Schwartz et al., 2002)), feature-selective responses and increased spike output to previously undetectable stimuli (Glazewski and Barth, 2015; Karni and Sagi, 1991; Marlin et al., 2015), increased synaptic strength (Biane et al., 2016; Cheetham et al., 2014; Clem et al., 2008; Rioult-Pedotti et al., 2000), and alterations in inhibition (Chen et al., 2015; Jasinska et al., 2010; Kuhlman et al., 2013).

Many of these changes have been observed in primary sensory cortex, indicating that even lower-level cortical areas possess the machinery to be rewired in response to experience. For example, in the rodent somatosensory cortex – a brain area of concentrated analysis because of the ease of targeting the representation of a specific sensory input – whisker stimulation can increase sensory-evoked spiking in superficial layers (Diamond et al., 1994; Glazewski and Barth, 2015; Glazewski and Fox, 1996), increase density of inhibitory synapses in layer 4 (Knott et al., 2002), trigger excitatory synaptic strengthening from layer 4 to layer 3 (L4 to L3) (Clem and Barth, 2006; Clem et al., 2008; Wen et al., 2013), and alter the dynamics of dendritic spines on layer 5 neurons (Kuhlman et al., 2014; Lendvai et al., 2000; Trachtenberg et al., 2002). However, though there is abundant evidence that experience can generate stable changes in the structure of neocortical circuits, observations have most often been piecemeal, identifying isolated sites of change without integration and analysis of how these changes influence local computations within the circuit.

A critical step towards developing a comprehensive account for how experience alters the brain will be the analysis of specific synaptic interactions defined by cell-type and how they dynamically evolve to enable long-lasting change. New studies are breaking ground in their ability to simultaneously identify multiple interacting elements of the cortical circuit and relate these small computations to large-scale alterations of cortical output. An important subset of these studies have focused on inhibitory neurons (Abs et al., 2018; Chen et al., 2015; Khan et al., 2018; Kuhlman et al., 2013; Yaeger et al., 2019), although there is an increasing appreciation regarding the diversity and discrete computational properties of different subsets of pyramidal neurons (Audette et al., 2018; Harris and Shepherd, 2015; Lu et al., 2017).

E:I balance – algorithm or phenomenon?

An algorithm is a set of rules that underlie problem-solving calculations. For neocortical circuits, these rules must be instantiated by molecularly identified neurons that are connected by synapses. How does the organization and connectivity of neurons – the cortical algorithm – facilitate circuit plasticity in response to salient experience? In a general sense, the balance of excitation and inhibition onto a neuron (E:I balance) – as well as ways that it can be modified to drive synaptic plasticity and learning - has been widely discussed for several decades as a highly conserved property of neural circuits that can be altered by neural activity and experience (Froemke, 2015; Tao et al., 2014). Importantly, E:I balance is not by itself a computation and does not relate a response transformation to the specific cellular architecture of the brain. Thus, an important modification of our understanding of E:I balance has been to recognize that inhibition comes from multiple sources and at very different times, and can reveal very different circuit functions (He and Cline, 2019).

Cell-type specific connectivity and information processing in the neocortex

Progress to identify how E:I ratios are generated by the interaction of specific subtypes of neurons across multiple synapses has been rapid, particularly in the neocortex. A highly-conserved circuit motif common across neocortical areas is the strong feedforward excitation from the thalamus to both excitatory neurons and GABAergic, parvalbumin-expressing (PV) neurons in the granular and deep layers of primary sensory cortex (Audette et al., 2018; Cruikshank et al., 2007; Swadlow et al., 2002) (Figure 1A). Thalamic drive to the cortex triggers both excitatory neural firing as well as rapid PV-mediated inhibition that suppress recurrent excitation, filtering incoming information so that high-frequency afferent stimulation can only be transmitted after decay of this disynaptic inhibitory process. PV neurons are also densely connected into the local excitatory network, receiving both strong local excitation (~50% of nearby Pyr neurons will be synaptically connected to a given PV neuron) and providing broad inhibition (~50% of nearby Pyr neurons will receive inhibition from a given PV neuron) (Barth et al., 2016; Fino et al., 2013; Jiang et al., 2015). Thus, PV neurons can be potently activated in two ways, first through fast thalamic input and second from the synchronous activation of local Pyr neurons. Under both scenarios, PV neurons are poised to potently suppress high-frequency input as well as recurrent activity in the local network.

Figure 1. Rewiring of neocortical excitatory connections during learning.

A) Schematic of thalamic inputs (blue) to L2 (top), L4 (center), and L5 (bottom) excitatory neurons. B) Higher-order thalamic inputs to L5 Pyr neurons change rapidly (Audette et al, 2019; Biane et al, 2019). It is unknown how thalamic drive or feedforward inhibition from PV neurons changes during the early stages of sensory or motor learning (grey). C) Higher-order thalamic inputs to L2 Pyr neurons change later, and synaptic strength between L2 Pyr neurons is also increased (Audette et al, 2019). The changes in PV inhibition during early sensory learning have not been well-characterized.

A canonical excitatory circuit in primary sensory cortex is represented by fast thalamic drive to L4, with ascending excitation to L2/3 that subsequently activates L5 (Harris and Shepherd, 2015). This model has been revised in recent years as it has become clear that fast thalamic drive independently arrives in L5b (Constantinople and Bruno, 2013) as well as in superficial layers (see for example (Sun et al., 2016; Wimmer et al., 2010), and also that higher-order thalamic nuclei (such as POm in S1 or the pulvinar in V1) independently synapse onto neurons within L2 and L5a (Audette et al., 2018; Bureau et al., 2006) (Figure 1A). There appears to be a high degree of reciprocal connectivity for excitatory neurons within a given layer, indicating that signal amplification may be an important property of neocortical circuits under some conditions (Lefort et al., 2009; Song et al., 2005; Urban-Ciecko et al., 2015). Finally, reciprocal synaptic connectivity between L2 and L5 pyramidal (Pyr) neurons may also provide an important source of recurrent excitation that can facilitate associative memory and integrate reward and punishment signals that occur with some delay to a sensory stimulus (Lefort et al., 2009).

A ubiquitous inhibitory circuit motif found across multiple neocortical areas is composed of vasoactive intestinal peptide (VIP) inhibitory interneurons that synapse strongly onto somatostatin (SST) but weakly onto PV inhibitory neurons (Pfeffer et al., 2013), and generally avoid pyramidal neurons (Jiang et al., 2015; Kuljis et al., 2018). Consequently, the activation of VIP neurons is associated with the suppression of SST neuron activity (Lee et al., 2013), whose elevated spontaneous firing both directly and indirectly inhibits the network (Urban-Ciecko and Barth, 2016) (Figure 2A).

Figure 2. Inhibitory synaptic motifs in the neocortex.

A) Synaptic connectivity between VIP, SST, and PV neurons. Left, Serial inhibition from VIP to SST to PV, so that VIP and PV activity are positively coupled and the major effect is Pyr inhibition. Right, Parallel inhibition of Pyr from both SST and PV neurons suggests that VIP activity may reduce SST without altering PV inhibition. B) Potential scenarios for state-dependent disinhibition during learning. Left, the primary effect of increased VIP firing during learning is suppression of SST activity and enhanced PV inhibition onto Pyr neurons. Right, the primary effect of increased VIP firing during learning is reduced SST inhibition onto Pyr neurons. C) Anatomical and electrophysiological data indicates sensory experience decreases PV inhibition onto Pyr neurons.

An improved ability to identify and analyze different inhibitory neuronal cell types and their interactions, particularly in intact and active networks, has made it possible to gain more sophisticated insights into the specific placement of inhibitory neurons within the cortical algorithm. Because activation of VIP neurons can powerfully inhibit the activity of other GABAergic neurons particularly SST neurons, it has been suggested that their primary function is disinhibition (Figure 2B), a conclusion supported by both detailed electrophysiological recordings from in vivo recordings and correlation analysis (Lee et al., 2013; Pfeffer et al., 2013; Pi et al., 2013). Indeed, experimental evidence indicates that increased VIP or suppressed SST activity is required for the increase in visual responses following monocular deprivation in adults (Fu et al., 2015), although the role of reinforcement or cholinergic signaling in this form of passive circuit plasticity is unclear. New analysis of a select group of inhibitory Ndnf1+ neurons in Layer 1 (L1) suggests that these neurons may function similarly to VIP neurons driving SST inhibition, and they may be important for disinhibition during learning (Abs et al., 2018). An important unanswered question is whether there is long-lasting plasticity at synapses onto and from these inhibitory neurons themselves.

Given the patterns of input distribution to specific types of inhibitory neurons and their output to specific cellular targets, we can begin to extrapolate how these small computations can be leveraged and adapted during learning. However, it should be noted that a complication interpreting the effects of even these two simplified motifs is the fact that SST neurons are synaptically connected to PV neurons and will suppress PV-mediated inhibition (Pfeffer et al., 2013), and indeed, PV neurons may inhibit SST neurons as well (Jiang et al., 2015). Thus, even small circuits that have been identified pose complications in extrapolating their function within a larger circuit, and may have laminar-specific properties, something that is a subject of intense investigation (Munoz et al., 2017; Naka et al., 2019).

Disinhibition during learning

Despite these complexities, there is a new opportunity to examine how experience can influence circuit response transformations in the neocortex not just at a given instant in time using in vivo recordings of identified neurons (Abs et al., 2018; Chen et al., 2015; Khan et al., 2018; Kuhlman et al., 2013; Peters et al., 2017; Yaeger et al., 2019), but over a trajectory of circuit modifications during learning or disease (Audette et al., 2019; Chen et al., 2015). Early synaptic change at one node of the circuit is likely to permit progressive alterations across the network, and the rules for plasticity at discrete types of synapses are likely to be different (Barth, 2002). The next step is understanding how these small circuits play a role in facilitating larger-scale synaptic reorganization.

Work in this area has grown increasingly sophisticated. For example, pioneering studies by Letzkus and Luthi identified a role for inhibitory neurons in L1 in facilitating excitatory synaptic plasticity in auditory cortex via cholinergic disinhibition (Letzkus et al., 2011), experiments that presciently took into account the detailed pharmacology and synaptic circuitry of neocortical neurons as it was known at the time. However, an improved ability to identify and analyze different inhibitory neuronal cell types and their interactions, particularly in intact and active networks, has made it possible to gain more sophisticated insights into the specific placement of inhibitory neurons within the cortical algorithm and how the cellular source of inhibition can selectively influence stable excitatory plasticity during learning.

An important idea that relates changes in inhibition to excitatory synaptic plasticity is the disinhibition hypothesis, whereby transient (state-dependent) or longer-lasting reductions in inhibition may enable plasticity at excitatory synapses to occur (Hattori et al., 2017; Letzkus et al., 2015; Mohler and Rudolph, 2017; Ressler and Maren, 2019). Synaptic plasticity at and from inhibitory neurons has been characterized in both reduced experimental preparations (Castillo et al., 2011; Kullmann et al., 2012), and by in vivo manipulations particularly with deprivation in visual cortex (Kuhlman et al., 2013; Maffei et al., 2006; Nahmani and Turrigiano, 2014), but also with somatosensory (Cybulska-Klosowicz et al., 2013; House et al., 2011; Jiao et al., 2006; Knott et al., 2002) or auditory manipulations (Sarro et al., 2015). Although most investigations have focused on plasticity surrounding PV neurons, alterations in SST neurons, particularly in their output, have also begun to be implicated in experience-dependent cortical plasticity.

The ability of VIP and L1 interneurons to inhibit SST neurons can lead to state-dependent disinhibition in the local circuit (Figure 2B), opening a window for long-lasting plasticity at behaviorally-relevant excitatory synapses onto Pyr neurons that can ultimately influence evoked firing and initiate synaptic potentiation. Analysis of these small circuit modules in the context of learning is an important step forward in thinking about cortical plasticity as a function of the pattern of highly-specified and interconnected cell types across the cortical column. Paradoxically, because SST neurons synapse strongly onto PV neurons, it might be expected that the overall effect of L1/VIP interneuron activation could be enhanced inhibition from PV neurons (Figure 2B). This complicates the disinhibition hypothesis, and it remains an open question how these different inhibitory systems interact both during normal sensation and learning.

PV neuron plasticity: themes and variations

Importantly, plasticity of PV neurons, both in their inputs and outputs, has been described in multiple experimental preparations in the neocortex (Hengen et al., 2013; Kuhlman et al., 2013; Maffei et al., 2006; Miska et al., 2018) and more generally across the brain (see for example (Donato et al., 2013; Lucas et al., 2016)). PV plasticity is diverse across cortical layers, preparations, and developmental timepoints, complicating efforts to integrate diverse observations. For example, depending on animal age and cortical layer, visual deprivation can reorganize synapses onto and from inhibitory neurons. During an early critical period in primary visual cortex, monocular deprivation in mice can drive a reduction in excitatory input to PV neurons and a decrease in PV inhibition to Pyr neurons in superficial layers, both processes reducing overall levels of inhibition (Kannan et al., 2016; Kuhlman et al., 2013) Figure 2C). The same manipulation increases inhibition from putative PV neurons onto L4 excitatory neurons in adult animals, having the net effect of reducing feedforward drive onto neurons in L2/3 (Nahmani and Turrigiano, 2014).

However, it remains unclear how PV neuron activity can be modified, particularly during learning. Do modifications occur through feed-forward inhibitory circuits from the thalamus, through altered SST inhibition (Figure 2B), or by altering the powerful role of PV neurons in feedback inhibition in the local circuit? These tantalizing results beg the question of how synaptic changes to and from PV neurons alter not just Pyr spike rates but also the local computation of the cortical circuit. The applicability of these findings to learning – not just alterations (typically deprivation) of sensory input – remains an important question, as deprivation may invoke a more radical alteration of network structure than behaviorally-relevant sensory experience.

Motor learning and sequential circuit change

Motor learning has been a powerful way to identify a sequence of synaptic changes that occur in inhibitory circuits (Peters et al., 2017) to facilitate subsequent plasticity. Motor training is relatively easy to carry out in rodents (where tools for cell-type specific analysis are more sophisticated), learning outcomes can be measured, and changes in excitatory synaptic strength and dendritic spines have been well-documented (Biane et al., 2016, 2019; Rioult-Pedotti et al., 2000; Rioult-Pedotti et al., 1998; Yang et al., 2009). Recent studies show that learning a cued lever-press task is associated with a stable reduction in SST but a transient increase in PV boutons, evaluated by longitudinal in vivo imaging in superficial layers of primary motor cortex (Chen et al., 2015). Such a finding is consistent with the general model of reinforcement cue activation of L1/VIP neurons with reduced SST activity and subsequent synaptic depression of SST inputs to Pyr neurons as well as concurrent SST-mediated inhibition of PV cells, potentially increasing PV activity and altering their inputs to Pyr neurons.

The finding that inhibitory circuits – both those targeting other GABAergic neurons as well as those targeting Pyr neurons – undergo experience-dependent reweighting indicates that the local rules for information processing are altered during the course of learning and not just dynamically at the moment of sensory stimulation. Although in vivo recordings from individual subtypes of inhibitory neurons may reveal how small-scale circuit computations are altered during the course of learning, analysis in more carefully controlled experimental preparations like the acute brain slice may be critical to provide new insights and testable hypotheses. The possibility that there are changes in inhibition that are stably manifested in synaptic strength for hours to days suggests that a more fine-scale mechanistic analysis will be possible.

The sequence of excitatory synaptic change across the cortical column

It is hard to believe that broad-scale changes in inhibition, even at in specific inhibitory subtypes, are specific enough to underlie the multitude of specific and distinct memories and representations that reside in the neocortex. In addition, changes in inhibition – particularly with PV neurons – appear to be transient (Hengen et al., 2013; Kuhlman et al., 2013). Thus, long-lasting changes in excitatory synapses that generate memory-specific neuronal ensembles have been a target of experimental focus, both because they can be generated under controlled experimental conditions (Gambino et al., 2014; Kim et al., 2016) and also observed following in vivo experience (Audette et al., 2019; Biane et al., 2016, 2019; Cheetham et al., 2007; Choi et al., 2018). Short-term (hours to days) changes in specific inhibitory circuits may be required to initiate stable (days to months) plasticity at excitatory synapses, leaving a lasting footprint in neocortical circuits that generates persistent alterations in sensory and motor responses.

What excitatory synapses across the cortical column are susceptible to modification during learning, and how do they relate to the underlying circuit? In general, neural firing plasticity has been observed mainly in L2/3 and L5 and avoiding L4, something that has been particularly well-resolved in somatosensory cortex (Benedetti et al., 2009; Cheetham et al., 2008; Diamond et al., 1994; Glazewski and Fox, 1996; Jacob et al., 2012; Knott et al., 2002; Kuhlman et al., 2014). Consistent with this, synapses onto L2/3 Pyr or the apical tuft of L5 Pyr neurons have revealed long-lasting changes in the number and stability of dendritic spines with altered sensory input (Kuhlman et al., 2014; Trachtenberg et al., 2002; Yang et al., 2009). It is important to note, however, that synapses within superficial layers of the neocortex have been the subject of the majority of studies, in part because of the ease of imaging. Thus, it remains unclear whether excitatory synaptic plasticity is concentrated in superficial layers. In addition, few studies have been able to link changes in postsynaptic dendritic spines with the specific presynaptic input that might be altered. In L1, for example, spine plasticity could be linked to changes in thalamic inputs, cortico-cortical inputs, or even in inhibitory inputs from SST neurons or L1 interneurons. Although it is clear that plasticity at excitatory synapses is induced by sensory and motor experience, the way that this plasticity alters the canonical circuit regulating the flow of excitation across the cortical column remains ill-defined.

Recent studies from our lab show that in one form of sensory learning, there is a progressive alteration of excitatory synaptic strength across the cortical column that begins with higher-order thalamic inputs to L5 Pyr neurons, and proceeds to change thalamic inputs to L2 Pyr neurons and increase excitatory synaptic strength between L2 Pyr neurons (Audette et al., 2019); Figure 1B,C). These sequential changes provide a foothold by which to investigate the role of state-dependent inhibition (Figure 2B) and also longer-lasting changes in inhibitory synaptic strength in gating excitatory synaptic plasticity across the cortical column.

The differential time course of inhibitory synaptic changes (transient increase in PV and a sustained decrease in SST inputs to L2 Pyr neurons) described by Komiyama (Chen et al., 2015) suggests a complex interaction of inhibitory networks with the induction and stabilization of changes at excitatory synapses. It is important to keep in mind that salient experience likely drives a sequence of changes in both the intrinsic membrane and synaptic properties of neurons (see for example (Gainey et al., 2018; Kuhlman et al., 2013; Miska et al., 2018)). It is also likely that alterations are not only cumulative at one type of synapse (i.e., stronger and stronger synapses between thalamic inputs to L5 Pyr neurons (Audette et al., 2019), but gate plasticity at downstream nodes of the circuit.

Critical to cell-type specific analyses is the ability to capture different stages of circuit reorganization: experience-dependent plasticity must occur rapidly enough that it can be reliably and reproducibly generated, but slowly enough so that discrete stages can be isolated and mechanistically dissociated. Tools for reversible synapse-selective alterations (for example, where higher-order thalamic inputs onto target cortical neurons are specifically augmented or silenced, with fine-scale temporal control) will enable a rigorous analysis of dependent interactions that sequentially reorganize the cortical circuit during learning.

Neurodegenerative diseases: gradual adjustments and ultimate collapse

The notion that brain circuits reorganize in a sequence of dependent changes can be instructive not only for understanding the cellular and synaptic basis of learning, but also for cognitive and motor impairment in progressive neurological disorders. It is increasingly recognized that early alterations of neural function can be masked by network accommodations that involve defined area- and cell-type specific interactions prior to clinical symptoms. This is the case for multiple disorders including Huntington’s disease, amyotrophic lateral sclerosis, Parkinson’s disease, and AD (Eisen et al., 2014; Jack et al., 2013; Noyce et al., 2016; Tang and Feigin, 2012).

Here we discuss how analysis of progressive cell-type and synapse-specific changes is changing our understanding of neurological disorders, with a focus on AD as an example of the insight that can be provided by this perspective. We focus on the hippocampus circuit because it is strongly implicated in early stages of AD, and the cellular and synaptic organization of the hippocampal circuit are well-defined making it possible to identify and interpret synapse- and cell-type specific changes. Early alterations in hippocampal synaptic function in early AD may initiate a cascade of compensatory, reactive circuit alterations that ultimately drive the devastating cognitive impairment at end stages of the disease.

Preclinical changes in circuit function

AD is a progressive neurological disorder characterized by impairments in memory and cognition, and is histologically associated with the accumulation of amyloid plaques and neurofibrillary tangles in the brains of affected individuals in multiple brain areas but concentrated in the hippocampus and associated cortical circuits. The hippocampus has been implicated in latent circuit changes prior to onset of memory deficits and shows hyperactivity in early stages in both humans and different animal models of AD (Busche et al., 2012; Kazim et al., 2017; Palop et al., 2007). It has been proposed that initial neuronal hyperactivity may trigger a positive feedback loop within this circuit (Palop and Mucke, 2016), and activity-dependent production of Aβ may eventually culminate in large-scale synapse loss, arbor shrinkage and hypoexcitability in the late stages (Zott et al., 2018). Identification of the circuit mechanisms that generate this early hyperexcitability may reveal important nodes of neuronal susceptibility associated with abnormal APP processing and may help in development of disease-modifying therapies independent of molecular correlates like Aβ or tau.

Pathway-specific changes in synaptic function in the hippocampus

How does hyperactivity initially arise in the hippocampal circuit, and how might it be related to synapse loss in later stages of the disease? New studies suggest that preclinical hippocampal hyperexcitability may be linked to Aβ-dependent neuronal hyperactivation within the hippocampus and neocortex (Busche et al., 2012; Zott et al., 2019). This hyperexcitability is observed in CA1 pyramidal neurons and could be directly tied to glutamatergic synapses (Zott et al., 2019), providing a precise locus within the hippocampal circuit to investigate specific hypotheses about early alterations in hippocampal function. Indeed, early gain of synapses across the dendritic arbor of CA1 Pyr neurons in the APP/PS1 model at 1 month of age has been observed (Megill et al., 2015), consistent with prodromal CA1 hyperexcitability.

While extensive spine loss has typically been observed in later stages the disease, a closer inspection has revealed that some spines (defined by dendritic compartment) may be more vulnerable than others to alterations, especially pathological loss during disease progression (Megill et al., 2015; Perez-Cruz et al., 2011; Reinders et al., 2016). Importantly, Aβ monomer and oligomers can be released at synapses in an activity-dependent manner (Cirrito et al., 2008; Kamenetz et al., 2003) and some studies in cultured hippocampal neurons suggest that Aβ oligomers may selectively bind to certain sites along the dendritic (Lacor et al., 2004). Thus, excitatory synaptic signaling may be strengthened early on in AD, possibly contributing to hyperactivity (Figure 3A,B). At later stages, due to extensive spine loss, neurite retraction, impairment of LTP, enhancement of long term depression (LTD), and axonopathy among others, this excitatory drive appears to be lost resulting in cells becoming hypoactive (Figure 3C (Busche and Konnerth, 2015; Spires-Jones and Knafo, 2012).

Figure 3. Progressive circuit changes in the hippocampal CA1 circuit during AD.

The schematic demonstrates gradual changes in the different synaptic inputs onto a CA1 pyramidal neuron (in black) across stages of AD pathology. Entorhinal cortex (EC, blue) and CA3 (black) represent two major excitatory inputs, while PV and SST represent two major inhibitory inputs to the CA1 Pyr. A) In the healthy condition, different excitatory inputs are balanced by a combination of inhibition ensuring normal activity levels of the CA1 pyr. B) During early stages of AD, CA1 Pyr may become hyperactive due to a combination of decreased inhibition as well as an increase in excitation mediated by low levels of Aβ hypothetically through blocking of GABA signaling, strengthening of LTP, axonal sprouting, and/or decreased glutamate reuptake may result in hyperactivity. EC inputs may be more susceptible to early Aβ-pathology even in earlier stages. C) During later stages, high Aβ levels may affect synapses globally - by blocking NMDAR signaling, impairing LTP, strengthening LTD, excessive synaptic pruning and/or axonal dystrophy - resulting in hypoactivity.

It is important to note that the acute effects of Aβ on synaptic function have been primarily explored at excitatory synapses (typically at CA3 inputs to CA1 Pyr), and there may also be input-selective effects at other pathways. Recent studies indicate that specific inhibitory pathways are suppressed in early stages of the disease, a process that could contribute to hyperexcitability in the circuit.

In hippocampal CA3 neurons, spontaneous IPSCs as well as evoked, feedforward inhibition (from an unknown source) are reduced in the APP/PS1 strain at ~6 months of age (Viana da Silva et al., 2019). Somatostatin axons in the CA1 (stratum lacunosum moleculare) are lost by at least 5 months in the same strain (Schmid et al., 2016), suggesting that inhibition from this cell type will also be reduced onto CA1 Pyr neurons. Interestingly, this axonal loss is likely linked to reduced excitatory drive onto SST neurons, as the dendritic spine turnover and number of dendritic spines onto SST neurons are also reduced. In the Tg2576 model of AD, PV inputs to CA1 Pyr neurons are reduced by 6 months of age (Yang et al., 2018). On the other hand, PV inputs appear to be increased in hippocampal CA1 and CA3 Pyr neurons in APPPS1-21 mice at 3 months of age (Hollnagel et al., 2019). Thus, both SST and PV inputs may be modified in early AD, prior to plaque deposition (Figure 3B).

Indeed, early hyperactivity may be a common feature of early AD pathogenesis. It is not restricted to the hippocampus alone but manifest in other neocortical circuits as well. Hyperexcitability has been observed in other brain areas (Busche et al., 2012; Kazim et al., 2017; Palop et al., 2007). Reduced in inhibitory input to some Pyr neurons, particularly from PV cells, has also been observed in various cortical areas (Chen et al., 2018; Petrache et al., 2019; Verret et al., 2012). For neocortical studies in particular, it will be important to precisely identify the pre- and postsynaptic identity of neurons analyzed in order to interpret the effects of aberrant wiring on circuit transformations. Overall, both SST and PV inputs may be modified in early AD, prior to plaque deposition.

Interpretation of all these results is complicated by the use of different animal models, different brain circuits studied, and analysis of different timepoints, and changes may be dynamic, proceeding in opposite directions as the circuit adapts. Thus, the temporal sequence of these cell-type and pathway-specific alterations still remains an open question. For example, it remains unknown whether inhibitory synapse loss occurs in response to a primary effect of Aβ at glutamatergic synapses, or whether these deficiencies proceed in parallel to excitatory synapse hyperactivation. Targeted cell-type analysis across multiple timepoints is warranted to identify the sequence of events that generated this abnormal circuit structure.

While it is abundantly clear that Aβ contributes to synaptic dysfunction (Spires-Jones and Knafo, 2012; Zott et al., 2018), future studies should evaluate defined classes of synapses to determine whether presynaptic input or neurotransmitter identity confers greater susceptibility to synapse gain or loss. For example, it remains unknown whether Aβ is produced at and accumulates equally at all types of synapses, and how synaptic transmission at different types of synapses is influenced by this molecule.

Currently, treatment for AD patients remains symptomatic, and an important and unmet goal is reversal of memory loss and cognitive decline. Earlier diagnosis and interventions that slow or halt disease progression may be intermediate steps towards this ultimate objective. Careful cell-type and pathway-specific neural circuit analysis for detection of predictive, pre-symptomatic changes may enable a better understanding of disease etiology and drive new therapeutic approaches for delaying or reversing later cognitive symptoms.

Progressive rewiring of brain circuits

Understanding how neurons, defined by their distinctive patterns of gene expression and direct and indirect patterns of connectivity, interact with each other in specific sequences of activation and inhibition brings a new sophistication and mechanistic understanding into how experience can transform brain circuits. However, the identification of learning- or disease-related changes in neural circuits is complicated by the fact that neural interactions can rapidly adjust in response to altered activity, leading to confusion about cause and effect. Analysis of single types of neurons or synapses can be confounded by the question of whether changes are primary or reactive, particularly when the stimulus (learning or disease) is ongoing. New efforts to identify progressive changes in defined neural circuits can reveal important principles of both normal circuit function and learning- or disease-dependent plasticity, enabling scientists to build and test detailed models of neuron interactions that may ultimately explain cognitive processes.