The default mode network exhibits circuit-specific selective vulnerability despite widespread amyloid burden in an animal model of Alzheimer’s disease

Abstract

In Alzheimer’s Disease (AD) the default mode network (DMN) exhibits selective vulnerability to functional decline and amyloid beta (Aβ) plaques in early AD. DMN vulnerability may simply be the result of Aβ accumulation. Alternately, Aβ may unmask an intrinsic susceptibility of DMN circuits. Should DMN vulnerability persist amidst global Aβ deposition, this would support the latter hypothesis. We tested this using long-term, multi-region recordings of single units in mice characterized by widespread Aβ (APP/PS1). We tracked a neuronal population within the retrosplenial cortex (a DMN region) and two monosynaptically connected target populations, one within (anterior cingulate) and the other outside the DMN (primary visual cortex). We also recorded in hippocampus and examined interactions between the in-DMN and out-DMN circuits. Our data revealed progressive dysfunction restricted to the in-DMN circuit. In contrast, communication originating within but targeting an out-DMN population was unaffected by disease. Dysfunction was pronounced in sleep.

In a mouse model of amyloid pathology, the authors find deficits in interregional communication specifically in default mode network neural circuits. This suggests these circuits are selectively vulnerable despite widespread plaque burden.

Introduction

The default mode network (DMN) is a set of distributed yet highly interconnected brain regions that are coactivated during quiescent wake¹. In addition to its role in cognition, the human DMN is noteworthy as it is selectively vulnerable to the early stages of neurodegenerative disease. Specifically, in Alzheimer’s Disease (AD), amyloid-beta (Aβ) accumulation starts in DMN hubs and propagates to secondary regions along anatomical lines^2–4. Unsurprisingly, brain regions that accumulate Aβ show disrupted activity⁵. Disrupted functional connectivity between DMN regions is a hallmark of early AD⁶.

The simplest explanation is that changes in the DMN are driven by Aβ deposition (i.e., the regions are otherwise unremarkable). This rationale supports the prediction that Aβ deposits in other circuits would render them similarly disrupted. Alternatively, early AD represents the confluence of two phenomena; specific circuits are both intrinsically vulnerable to Aβ-related damage, and it is in these circuits that Aβ initially takes hold. In this case, if Aβ deposits were widespread, DMN circuitry would still be disrupted relative to non-DMN regions. However, because Aβ deposition is restricted to the human DMN early in AD, it is difficult to separate the local impact of Aβ from an intrinsic vulnerability of these brain regions. Either possibility has powerful implications for an understanding of AD etiology and intervention strategies. Here, we combine a mouse model of widespread amyloidosis with long-term, multi-site recordings spanning in-DMN and out-DMN circuits to directly test the hypothesis that neurons in the DMN are intrinsically vulnerable independent of Aβ.

The regions comprising the DMN exhibit a high degree of reciprocal connectivity^7–9, such that the majority of DMN efferents target other DMN regions. A minority of DMN projection neurons target postsynaptic sites broadly throughout the brain¹⁰. Crucially, in both cases, cell bodies lie within the anatomical borders of the DMN, such that the circuits are divisible by postsynaptic target; in-DMN projections originate and terminate in DMN loci, while out-DMN projections terminate in non-DMN regions. If DMN neurons are intrinsically vulnerable to Aβ-related dysregulation, the dissociation of DMN and non-DMN circuitry by projection target raises an interesting possibility: the cellular impact of disease may be conditioned on axonal target, even when comparing neurons whose soma are anatomically intermingled (i.e., sharing the same extracellular Aβ environment).

The recent discovery of two neuronal populations in mouse ventral retrosplenial cortex (RSPv), a core DMN region, identifies such intermingled circuitry¹⁰. Both subpopulations are composed of excitatory projection neurons, but differ in their postsynaptic targets. One group, the in-DMN targeting, sends monosynaptic connections to the dorsal anterior cingulate (ACAd), another selectively vulnerable DMN region. The other group, the out-DMN targeting, sends monosynaptic connections to the primary visual cortex (VISp), which is outside of the DMN. This anatomy crystallizes the question of selective vulnerability, allowing for interrogation not only of region but also of connectivity.

For three reasons, this circuitry, in combination with the APP/PS1 mouse, is ideally suited to test the hypothesis that subsets of neurons may be intrinsically vulnerable during disease independent of the anatomical proximity to Aβ deposits. First, the APP/PS1 mouse overexpresses amyloid precursor protein (APP) and presenilin globally and is remarkable for progressive and indiscriminate Aβ accumulation throughout the cortex and hippocampus^11–15. This age-dependent deposition pattern thus circumvents the confounding restriction of Aβ to the human DMN in AD. Second, mouse models are amenable to multi-site recording of ensembles of single neurons, allowing for the simultaneous monitoring of multiple monosynaptically connected populations in the context of free behavior. Finally, the APP/PS1 mouse is generally accepted as a model relevant to early AD progression because disruptions in cortical physiology are not explained by neuron loss^16–18.

We collected and analyzed four-site recordings of neuronal ensembles in APP/PS1 mice to test the hypothesis that two circuits, one in-DMN (RSPv-to-ACAd: R2A) and one out-DMN (RSPv-to-VISp: R2V), are differentially vulnerable in the context of global amyloid burden. To differentiate effects of constitutive APP and PS1 overexpression from the progressive accumulation of Aβ plaques, we recorded from two groups of animals: young and old. Specifically, we conducted long-term, simultaneous recordings of extracellular spiking from ensembles of single neurons in RSPv, ACAd, VISp, and hippocampus (CA1) of freely behaving WT and APP/PS1 mice. We examined neuronal dynamics and circuit interactions as a function of brain state over a 24  h cycle containing a full range of naturally occurring behaviors. Our results reveal age-dependent, selective functional disruption only in the R2A circuit. Notably, this effect is pronounced during sleep, consistent with emerging evidence that sleep quality causally mediates disease progression^19–23. These data suggest that, even within the same DMN hub region and in the face of global Aβ plaques, intrinsic vulnerability is determined, at least partly, by projection target.

Results

Chronic monitoring of single neuron activity in four brain regions in WT and APP/PS1 mice

Mice harboring APP/PS1 transgenes exhibit progressive pathology that captures key aspects of early AD in humans^13,14,24,25. Specifically, animals exhibit increased brain Aβ by 8 months and neuritic plaques in the isocortex by 10 months of age²⁴. Importantly, this molecular burden increases over time despite constitutive overexpression of the APP/PS1 genes (Fig. 1A). While Aβ expression is constrained to the DMN in the earliest stages of human AD, we observed widespread plaques throughout the isocortex and hippocampus of a large cohort of APP/PS1 animals dedicated to quantifying Aβ (N = 271; Fig. S1A, B) as described previously¹¹. To test the hypothesis that circuits whose source and target are within the DMN are intrinsically vulnerable even in the context of global APP/Aβ, we performed continuous (10–20 days), multisite recordings of extracellular signals from ACAd, RSPv, VISp, and CA1 (Fig. 1B, C) in APP/PS1 mice and WT littermates. To disentangle the effects of constitutive APP/PS1 overexpression from progressive amyloid pathology on neural activity, we studied separate groups of mice at two time points: early adulthood (N = 4 WT age 3.3 ± 0.06 months, N = 5 APP/PS1 age 3.2 ± 0.22 months) and late life (N = 4 WT age 13.6 ± 0.85 months, N = 4 APP/PS1 age 13.8 ±  1.45 months, see Table 1 for animal-by-animal details) (Fig. 1A). (NB a circuit in this context refers to the subset of neurons within a given region whose activity is relayed via projections to a target region. Both local/recurrent and projecting neurons are encapsulated by the term circuit).

Fig. 1. Aβ overexpression increases single neuron firing rates selectively in exemplar DMN regions as a function of brain state.

A Experimental overview. Neural activity was recorded from WT and APP/PS1 mice in a young cohort (~3 months; prior to significant plaque formation in APP/PS1) and an old cohort (>1 year; late-stage amyloidosis in APP/PS1). The vertical offset of dots (indicating individual mice) in the timeline is for visualization. B Sixty-four ch tetrode bundles were implanted in four brain regions: two in default mode network hubs (DMN; RSPv, ACAd), one outside of the DMN (VISp), and one in CA1. C Illustration of the monosynaptic projection connecting populations in RSPv to ACAd (in-DMN) and populations in RSPv to VISp (out-DMN). D Example raw data from two tetrodes, one in RSPv and the other in ACAd. E (Left) Four single-unit mean waveforms recorded on one tetrode and (right) corresponding interspike interval histograms. The vertical red line indicates the minimum possible refractory period for a single unit. Histogram colors match the unit waveforms on the left. F Twenty-four hours of data from a single animal. Rows show (top to bottom): sleep/wake states, mean ensemble firing rate by region, animal locomotion (Mvmt), and environmental conditions (light/dark). G Single-unit firing rates by genotype, region, and brain state. Gray points are mean firing rates of individual neurons; box plots show the median and 25th–75th percentiles, and whiskers extend to the minimum and maximum non-outlier values. Firing rates were analyzed with a linear mixed-effects model: log(firing rate) ~ genotype × state × region × age group + (1∣ animal/recording epoch/neuron). Post hoc APP/PS1–WT contrasts were two-sided and Tukey-adjusted emmeans. Significant contrasts in ACAd were observed during AW ($\mathrm{\Delta }\log \left(\mathrm{FR}\right)=0.575$, 95% CI: 0.199--0.952, z = 3.004, p = 0.0027) and NREM ($\mathrm{\Delta }\log \left(\mathrm{FR}\right)=0.466$, 95% CI: 0.092--0.841, z = 2.439, p = 0.0147). Biological replicates were N = 4 young WT, N = 5 young APP/PS1, N = 4 old WT, and N = 4 old APP/PS1 mice. Counts of neurons and epochs are provided in Table 2. VISp: primary visual cortex; RSPv: ventral retrosplenial cortex; ACAd: dorsal anterior cingulate cortex; CA1: cornu ammonis hippocampus; WT: wild-type; APP: amyloid precursor protein and mutant human presenilin 1; AW: active wake; QW: quiet wake; NREM: non-rapid eye movement sleep; REM: rapid eye movement sleep. Source data are provided as a Source Data file.

Table 1.

Animal demographics and implantation timing

Animal	Genotype	Sex	Age at implantation
			PND	Months
BTH90	WT	F	90	3.0
BTH95	WT	F	94	3.1
BTH100	WT	M	92	3.0
ZBR106	WT	M	83	2.7
ZBR113	APPS1	M	58	1.9
ZBR114	APPS1	F	63	2.1
ZBR117	APPS1	M	76	2.5
ZBR118	APPS1	F	76	2.5
ZBR119	APPS1	M	107	3.5
CAF69	WT	F	374	12.5
CAF71	APPS1	M	488	16.2
CAF73	APPS1	F	374	12.5
CAF74	APPS1	M	412	13.7
CAF75	APPS1	M	383	12.8
CAF77	WT	F	393	13.1
CAF81	WT	F	432	14.4
CAF82	WT	M	435	14.5

Single units were extracted with MountainSort-based clustering²⁶ and labeled automatically by an XG Boosted decision tree²⁷ trained on measures of cluster quality and isolation, such as L-ratio, Mahalanobis distance, and the absolute refractory period²³ (Fig. 1D, E). Regular spiking units (RSUs, ~90% pyramidal) were classified based on waveform^23,28–30. In addition, video (30 FPS) was collected and processed with a convolutional neural network to track animal locomotion³¹. For sleep scoring, this was combined with 60 Hz low-passed electrophysiological data as described previously^23,32–34 (Fig. S2). We analyzed ~24 h of data from each animal, collected at least 1 week after surgery. Analyses are restricted to well-isolated RSUs detectable for the entire period of analysis (n = 3141 units with a mean presence ratio of 99.71 ± 0.04% SEM). Each animal’s RSPv ensembles contained single units with short-latency interactions (significant single-neuron pairwise correlations at <10 msec) with single units in both ACAd and VISp, consistent with monosynaptic excitatory connectivity (Fig. S3B, C). These results are consistent with monosynaptic projections from RSPv to both ACAd and VISp, as described recently in a study of mesoscale connectomics¹⁰ (Table 2).

Table 2.

Recording data and epoch counts across sleep-wake states

Animal	Rec. time (h)	# Units recorded	# Epochs
			AW	QW	NREM	REM
BTH90	24	286	37	81	65	83
BTH95	24	302	23	90	71	92
BTH100	24	174	48	77	34	49
ZBR106	24	245	22	74	68	58
ZBR113	13	38	17	16	15	22
ZBR114	24	189	44	82	57	89
ZBR117	12	160	20	33	27	20
ZBR118	24	259	15	63	58	49
ZBR119	19	229	23	34	31	43
CAF69	35.75	156	165	161	58	40
CAF71	20.84	121	180	235	104	7
CAF73	24	216	113	112	49	16
CAF74	24	163	192	281	98	5
CAF75	24	205	242	321	85	6
CAF77	23.84	58	176	206	78	34
CAF81	24	228	269	300	69	43
CAF82	24	112	243	259	69	54
Total	388.43	3141	1829	2425	1036	710

Bold values show totals for each column.

To confirm that amyloid burden in APP/PS1 animals is progressive and widespread, we quantified plaque density in an independent cohort of APP/PS1 animals (N = 143) across two age groups using serial two-photon tomography with methoxy-X04 labeling (Fig. S1). Plaques were sparse at 4 months and increased dramatically by old age (12–18 months) across all recorded regions (emmeans contrast averaged over regions, estimate = 2.89 log units, p < 0.0001), with region-specific contrasts confirming age-dependent progression in each of the four recorded sites independently (ACAd, RSPv, CA1, VISp; all p < 0.0001). In old animals, plaques were broadly distributed across all 43 isocortical regions and the hippocampus. We decomposed variance in old animals: animal ID was the dominant source of variance (90%), with regional differences explaining only 3% and residual noise accounting for 7%. Sex was a significant predictor of plaque density in old animals (broad isocortex: F = 10.01, p < 0.0001; 4 ROIs: F = 8.60, p < 0.0001), with females tending to exhibit higher plaque burden than males, consistent with prior reports in this line^35,36. The magnitude of the sex effect was small relative to inter-animal variability, and sex did not interact with genotype, age, or brain region in any measure of neural activity reported here (Table 3).

Table 3.

Sex-related effects across primary analyses

Analysis	Sex	Sex:age	Sex:geno	Sex:age:geno
Mean firing rate (Fig. 1)	F = 2.50	F = 0.0003	F = 0.33	F = 0.39
	p = 0.150	p = 0.986	p = 0.583	p = 0.550
Cross-region spike timing (Fig. 2)	F = 0.60	F = 0.0017	F = 0.69	F = 0.47
	p = 0.458	p = 0.968	p = 0.426	p = 0.510
SWR-locked peak timing (Fig. 3)	F = 0.79	F = 0.34	F = 1.32	F = 0.08
	p = 0.407	p = 0.581	p = 0.293	p = 0.786
Communication subspace dim. (Figs. 4 and 5)	F = 0.65	F = 1.26	F = 0.74	F = 0.03
	p = 0.438	p = 0.290	p = 0.411	p = 0.863

For each analysis, the linear mixed-effects model was refit with sex × age group × genotype added as fixed effects. F statistics and p-values are reported from Type III ANOVA with Satterthwaite’s method and were two-sided. No sex-related term reached significance in any analysis.

Increased spike frequency in APP/PS1 mice is age-dependent and restricted to DMN regions

To connect our recordings to prior work, we sought to replicate two previously described effects of APP-driven amyloidosis on neural activity. First, it is difficult to disentangle the effects of Aβ accumulation from the effects of constitutive overexpression of APP and PS1. A key difference is that Aβ accumulation is progressive, such that its downstream effects change with age^37,38, while gene overexpression is constant. Second, amyloidosis (and APP overexpression) disrupt neuronal firing rates, albeit with conflicting evidence of direction: calcium transients across frontal cortex and CA1 suggest hyperexcitability^39,40, while extracellular recordings in frontal cortex indicate hypoexcitability⁴¹. Experimental constraints, such as head-fixation and relatively short recordings, make it difficult to reconcile these findings. To address each of these points, we examined single neuron activity as a function of brain region, arousal state, genotype, and age group. Our approach of continuous, long-term recording allowed us to follow the activity of the same single units across behavioral and environmental conditions. As a result, a linear mixed model (LMER⁴²) can evaluate the interaction of circuit, state, genotype, and age while quantifying variance explained by individual animals and neurons. We also took advantage of the fact that animals were sampled in distinct age cohorts relevant to disease progression (Table 1). The resultant linear mixed model was $\log \left(\text{firing rate}\right)~\text{genotype}\times \text{state}\times \text{region}\times \text{age group}+\left(1\mid \text{animal/recording epoch/cell ID}\right)$, where animal ID, recording epoch, and cell ID are included as random effects.

Broadly, there was no main effect of genotype on single neuron firing rate (p = 0.113). However, there was a significant effect of age (p = 0.011) and a significant three-way interaction between age group, genotype, and brain region (p < 2.2e−16). To explore this interaction, we conducted post-hoc tests using EMMeans (R). Consistent with prior evidence of hyperexcitability, firing rates were significantly higher in the two DMN regions (ACAd and RSPv) of old APP/PS1 compared to old WT mice (p = 0.007 and 0.033, respectively). Despite near-global plaque coverage (Fig. S1A, B), this effect was circuit-specific: there was no difference between old APP/PS1 and WT firing rates in CA1 or VISp (p = 0.199 and 0.455, respectively) (Fig. S3A).

Additionally, there was a significant interaction between genotype, brain region, and age group when brain state was considered as a fixed effect (p = 6.818e−08). This was driven by state-dependent differences between old WT and old APP in ACAd (Fig. 1G; significant comparisons between genotypes: ACAd active wake, p = 0.0018; ACAd NREM, p = 0.0192). Collectively, altered firing rates are unlikely to arise from gross Aβ-related damage; single-neuron yield did not differ by genotype in old mice (Fig. S1C, D), consistent with a lack of cell loss at 1 year. Neuron counts in representative histology were similar in each recorded region (Fig. S1C). Taken together, these results suggest that neuronal activity levels in RSPv and ACAd are selectively sensitive to Aβ, in contrast to VISp and CA1, and that age contributes to firing rate dysregulation in DMN regions of APP/PS1 mice.

Summarily, the effect of genotype on firing rate was age-dependent and selectively evident only in the exemplar DMN regions recorded here. Increased spike rates in old APP/PS1 mice are consistent with prior evidence³⁸ and more easily explained by progressive amyloid pathology than constitutive transgene expression.

Timing within the R2A but not R2V circuit is disrupted in old APP/PS1 mice

While ensemble mean firing rates were elevated in each DMN region in old APP/PS1 animals, these circuits are complex and contain information that projects both locally and to diverse post-synaptic targets, some of which are contained within the DMN and some of which lie outside of the DMN. Population mean firing rates provide no insight into these distinct computational processes.

It is possible that functionally distinct yet anatomically intertwined subcircuits are differentially vulnerable to pathology. We reasoned that we could gain insight into the integrity of functional subsets by examining the strength and timing of population-level interactions. For example, neurons within RSPv process and then send information along monosynaptic projections to ACAd. Neurons in ACAd receive this transmission and, through additional local interactions, generate stereotyped population responses. As a result, the activity of a subset of neurons in RSPv should be related to the activity of a subset of neurons in ACAd within low latency (tens of ms), even if they (the RSPv neurons) are not projection neurons. Conceptually, this is similar to studying lagged functional connectivity in imaging or LFP data, albeit with single-unit resolution^43,44. Alterations in population-level interactions could arise from a variety of mechanisms, such as disruption of local interneuron timing or degradation of anatomical lines. Thus, studying population interactions with single-neuron resolution is well positioned to capture the summary effect of disrupted communication between brain regions.

Multi-site recordings allow us to ask whether the vulnerability of RSPv, a DMN hub, is a general feature of the region, or if vulnerability is specific to functional sub-circuits. To address this, we performed lagged pairwise spike correlation between all pairs of single units in the RSPv (source region) and each of the two target regions (ACAd and VISp) (Fig. 2A). We analyzed these population-wide interactions as a function of each epoch of NREM and waking. Crucially, the same population (RSPv) serves as the source in each of these circuits. We measured the directionality (delay) (Fig. 2B, C) and the magnitude (Fig. S5A, B) of the source/target interaction as a function of genotype, age group, and arousal state (1829 active wake epochs, 2425 quiet wake epochs, 1036 NREM epochs, linear mixed models described in figure legends).

Fig. 2. Inter-regional spike timing during sleep is disrupted in the R2A circuit in old APP/PS1 animals.

A Cartoon illustrating source/target lagged correlation analysis. In each behavioral epoch, single neuron activity in RSPv (source) is correlated with single neuron activity in each of two targets. Epochs of the same type are then averaged together. B The timing of R2A synchrony varies as a function of brain state in old APP/PS1 mice (bottom row) but not in young APP/PS1 mice relative to WT (top row). Lines and shading show the mean z-scored correlations ± SEM; dashed lines indicate genotype-specific peak lags. C Quantification of RSPv-ACAd peak timing. Points represent individual epochs; boxen plots show the median and nested quantile ranges extending toward the distribution tails. Peak timing was analyzed using a linear mixed-effects model: peak timing ~ genotype × state × circuit × age group + (1∣ animal), followed by two-sided asymptotic z-tests of estimated marginal means with Tukey adjustment. No significant differences were detected in young mice (p ≥ 0.5041 for AW, QW, and NREM). In old mice, peak timing was delayed in APP/PS1 mice during QW (Δ = 3.57 ms, 95% CI = 0.30–6.85 ms, z = 2.134, p = 0.0329) and NREM (Δ = 6.07 ms, 95% CI = 3.15–8.99 ms, z = 4.080, p < 0.0001), but not AW (p = 0.0777). n = 4330 neuron pairs and N = 17 mice. Source data are provided as a Source Data file.

When considering timing of inter-regional communication, the three-way interaction of brain state, genotype, and circuit was significant (p = 0.0034). RSPv to VISp (R2V; out-DMN) interaction timing was not different between APP/PS1 and WT mice in any state (active waking, quiet waking, NREM) in either young or old animals (Fig. S4B, C). In contrast, the RSPv to ACAd (R2A; in-DMN) interaction was lagged by 6 msec in old APP/PS1 mice compared to WT mice during NREM sleep (p < 0.0001) (Fig. 2B, C). In contrast, there were no genotype effects in the R2A timing in young animals. Further, there were no significant differences between genotypes in the magnitudes of R2A or R2V peaks (Fig. S5A, B). As expected, the same analyses run on shuffled spike times yielded flat lines with no peaks (Fig. S5C, D), suggesting that correlation analyses capture biologically meaningful structure.

We broke the waking state into two substates, active (AW) and quiet (QW), as meaningful differences in neuromodulatory tone and neural dynamics exist between these conditions. During QW, the R2A interaction in old APP mice was lagged by 5 msec compared to old WT animals (p = 0.033). This was not true in AW epochs (p = 0.078) (Fig. 2B, C). The timing of R2V interaction did not significantly differ between genotypes in either AW or QW (Fig. S4B, C; AW: p = 0.808; QW: p = 0.351). There were no differences between genotypes in the young cohort.

Because each circuit shared the same source neurons, these results suggest that selective vulnerability in APP/PS1 mice may not simply reflect anatomical proximity to plaques, but instead depends on circuit details, such as connectivity. Further, given the state-specificity of these results, our data suggest that disruptions in communication are unlikely to be the result of amyloid-driven anatomical damage, as this would be expected to manifest across states. A plausible hypothesis is that amyloid plaques alter the impact of neuromodulators on a subset of cortical interneurons that are over-represented in the processing and transmission of information along a specific set of projections, such as in-DMN circuits^45–47.

Epochs of NREM can last for thousands of seconds and contain diverse substates. Collapsing across this heterogeneity reveals state- and circuit-specific effects of Aβ, but does not provide insight into the possible disruption of functionally relevant neurophysiological events within NREM sleep.

The timing of sharp-wave ripple-driven spiking in RSPv is aberrant in old APP/PS1 animals

We reasoned that if altered communication is specific to the R2A circuit, neurophysiological events known to drive fluctuations in the DMN might be particularly vulnerable to the impact of Aβ. An ideal candidate for such an event is the sharp-wave ripple (SWR); a brief large deflection and oscillation (140–220 Hz) in the local field potential of the hippocampus that has been linked to memory consolidation between the hippocampus and isocortex^48,49. Notably, SWRs drive subsequent DMN activity⁵⁰. We next asked whether SWRs are consistent with a locus of selective vulnerability in the R2A circuit.

Arising from the dendritic layer of CA1, SWRs are observed primarily during NREM sleep but also quiet waking. The discrete, aperiodically repeating nature of SWRs presents an opportunity to approach long-term recordings of freely behaving animals with precise, trial-structured analyses. To understand intra-regional communication during SWRs, we first identified ripples in broadband CA1 data during NREM epochs using standard algorithms⁵¹ (Fig. 3A). We observed no significant effect of genotype or age on SWR peak frequency, amplitude, or duration, although there was noteworthy inter-animal variability in all three measures, particularly in WT animals (Zhurakovskaya et al.⁵², but see Cushing et al.⁵³) (Fig. S6A–C ripple density, duration, and amplitude p > 0.05).

Fig. 3. NREM sharp-wave ripple-driven activity is disrupted in ACAd of old APP/PS1 mice.

A (Left) Example sharp-wave ripples (SWRs) shown in: (top) raw local field potential (LFP) recorded in CA1, (middle) 150–250 Hz band-passed data, and (bottom) CA1 single-unit activity. (Right, top) A single SWR in raw data, and (right, bottom) band-passed data. B Raster plot of the response of a single unit in RSPv to 80 SWRs. C The average response of an RSPv ensemble (30 single units) to SWRs. NREM SWR-aligned single-unit activity in VISp (D), RSPv (E), and ACAd (F). Lines and shading show the mean ± SEM across n = 1244 neurons and N = 17 mice. Insets show boxplots of peak offsets of individual neurons. Peak timing was analyzed using a linear mixed-effects model: peak timing ~ genotype × region × age group + (1∣ animal), followed by two-sided, Tukey-adjusted emmeans contrasts. Peak timing did not differ by genotype in VISp (young: p = 0.223; old: p = 1.000) or RSPv (young: p = 0.223; old: p = 0.408). In ACAd, peak timing was delayed in old APP/PS1 mice relative to WT mice (Δ = 35.75 ms, t = 4.674, p = 0.005), but did not differ in young mice (p = 0.999). Source data are provided as a Source Data file.

One of the major inputs to RSP is a monosynaptic projection from CA1⁵⁴, and SWR-driven spiking in RSP has been described previously⁵⁵. Accordingly, in our data, individual neurons in RSP displayed increased firing around the time window of hippocampal SWRs (Fig. 3B, C).

The hippocampus is sometimes considered part of the DMN⁵⁶. As a result, we hypothesized that the timing of SWR activation in DMN regions should be disrupted in old APP/PS1 animals. Analyses revealed a significant interaction between genotype, age, and circuit (p = 4.728e−12, linear mixed model where animal ID was treated as a random effect). The timing of SWR-related activity in VISp did not differ by genotype in either age group (Fig. 3D; Median peak time: young WT vs. young APP p = 0.223; old WT vs. old APP p = 1.000). The timing of SWR-related activity in RSPv also did not differ by genotype in either age group (Fig. 3E; Median peak time: young WT vs. young APP p = 0.223; old WT vs. old APP p = 0.408). In contrast, we observed a significant delay in SWR-associated spiking in ACAd in old APP/PS1 mice relative to WTs (Fig. 3F; Median peak time: young WT vs. young APP p = 0.999; old WT vs. old APP p = 0.005).

These results are specific to SWR-driven activity; no differences between genotypes or ages were observed when the same analyses were run on randomly selected times outside of SWRs (Fig. S7A–D). In addition, while responses to SWRs during wake were observed in all three regions, they differed by neither genotype nor age in either circuit (Fig. S8). We also examined responses to NREM SWRs in fast-spiking (FS) interneurons and found no significant differences between genotypes in timing or peak magnitudes at either age range (Fig. S9).

Taken together, these results suggest that advanced amyloidosis enhances the response of local populations of putative excitatory neurons to SWRs in the RSPv, and that the timing of local SWR-driven computation is disrupted in an age-dependent manner in R2A but not R2V ensembles.

Aβ drives altered communication dimensionality in a circuit- and state-dependent fashion

While these data demonstrate that a subset of hippocampal interactions can be impacted by Aβ, it is unclear whether altered timing implies a disruption in the information content carried along a circuit. It is possible that, despite being delayed or accelerated by a few milliseconds, information transmission is otherwise normal.

To assess the impact of Aβ on the content of inter-regional communication, we measured the dimensionality of the communication subspace between each pair of regions. To do this, we performed reduced rank regression between source and target regions, first in the R2A and R2V circuits⁵⁷. Regression performance, or the degree to which source activity predicted target activity, was measured as a function of the rank of the coefficient matrix (i.e., the number of predictive dimensions). When comparing genotypes this way, a consistent shift in the number of source dimensions necessary to capture a given percentage of the target activity indicates a change in the dimensionality of the interaction of the two regions (Fig. 4A). In simple terms, complex patterns of activity require more dimensions to reproduce than simple patterns. Applied to two (or more) interacting brain regions, this analysis measures subspace dimensionality⁵⁷: a subspace of the activity in RSPv predicts activity in VISp, and another subspace of RSPv activity predicts ACAd activity. It is important to note that subspaces are latent activity patterns that often call on overlapping sets of individual cells, presumably comprising both local and projecting neurons.

Fig. 4. Communication in the R2A circuit exhibits an age-dependent reduction in dimensionality during NREM sharp-wave ripples.

A (Top) Conceptual diagram of the analysis. The activity of an ensemble of single units in a source region is used to predict activity in a target region. Source dimensionality is progressively reduced, and the resulting loss of predictive performance is measured. (Bottom) Diagram of the R2A (in-DMN) and R2V (out-DMN) circuits, both originating in RSPv. B Communication-subspace dimensionality in the R2V (left) and R2A (right) circuits during NREM sharp-wave ripples (SWRs) in young mice. Predictive performance is scaled relative to full-rank prediction. Points represent individual bootstraps; lines show mean scaled performance at each dimension, and shading indicates bootstrapped 95% confidence intervals. Insets show bootstrap distributions of optimal dimensionality, defined as the number of dimensions required to explain 90% of full-rank performance. Optimal dimensionality was analyzed using a linear mixed-effects model: optimal dimension ~ genotype × circuit × condition × age group + (1∣ animal/date/epoch/bootstrap), followed by two-sided, Tukey-adjusted asymptotic z-tests of estimated marginal means. No genotype differences were detected in young mice in either R2V (p = 0.7094) or R2A (p = 0.9229). C Same as B for old mice. Optimal dimensionality did not differ by genotype in R2V (p = 0.5548), but was reduced in APP/PS1 mice in R2A (APP/PS1–WT estimate = − 0.624, 95% CI = − 1.24 to −0.01, z = − 1.980, p = 0.0477). N = 17 mice were biological replicates; n = 10,042 bootstraps were nested resampled observations. *p < 0.05; ns, not significant. Source data are provided as a Source Data file.

We first examined subspace dimensionality during NREM SWRs (Fig. 4). The relationship between the number of source dimensions and the predictability of target activity did not differ by genotype in either age group in the R2V circuit (RSPv to VISp; out-DMN) (Fig. 4B, C left). We defined the optimal dimensionality of circuit interactions as the number of dimensions necessary to capture 90% of the full-rank prediction (Fig. 4B, C inset; Young mice: p = 0.709, old mice: p = 0.555). Linear mixed model: optimal dimension ~ genotype ×  circuit × age × cond + animal/date/epoch/bootstrap. The last term represents nested random effects with random intercepts. In contrast, optimal dimensionality was significantly reduced during SWRs in the R2A circuit (RSPv to ACAd; in-DMN) in old APP/PS1 animals relative to old WT (Fig. 4B, C, right; Young mice: p = 0.923, old mice: p = 0.048). Succinctly, the complexity of circuit communication during NREM SWRs was reduced in APP/PS1 animals, but only in the R2A circuit and only in old mice.

While inset histograms display the raw distributions of optimal dimensionality in each condition (Fig. 4B, C), the same data are displayed after accounting for random effects (such as animal ID) in Fig. S10.

One possibility is that the reduction of dimensionality in the R2A circuit is general and not specific to NREM SWRs. To test this, we ran the same analysis in two additional states: (1) centered on periods of NREM sleep outside of SWRs (i.e., periods of NREM sleep dominated by slow-wave activity (delta NREM)), and (2) centered on SWRs occurring during waking⁴⁹. During delta NREM intervals, we observed no significant age-dependent differences in optimal dimensionality between genotypes in either the R2V circuit (Fig. S11A, B left; Young mice: p = 0.3853, old mice: p = 0.4316) or the R2A circuit (Fig. S11A, B right; Young mice: p = 0.6009, old mice: p = 0.9361). During SWRs occurring during wake, we again did not observe any significant, age-dependent differences in optimal dimensionality between genotypes in either the R2V circuit (Fig. S11C, D left; Young mice: p = 0.2948, old mice: p = 0.6204) or the R2A circuit (Fig. S11C, D right; Young mice: p = 0.0510, old mice: p = 0.4761).

Summarily, age-dependent reduction in communication dimensionality only occurred during NREM SWRs in the R2A circuit. This is consistent with our hypothesis that the RSPv, a DMN hub, exhibits age-, state-, and circuit-dependent vulnerability to Aβ such that R2A is degraded while R2V is not. However, there is a simple alternate explanation: our results could be trivially explained by a failure in ACAd. To address this, we took advantage of the directionality of our analyses. We inverted each source and target, allowing us to examine a second set of circuits: V2R (VISp to RSPv) and A2R (ACAd to RSPv) (Fig. 5A). In contrast to the results presented above, the A2R circuit did not exhibit an age-dependent reduction in dimensionality during NREM SWRs in either genotype (Fig. 5B, C right; Young mice: p = 0.8361, old mice: p = 0.2481). Interestingly, in the V2R circuit, there was a significant difference in dimensionality between genotypes in young mice but not in old mice (Fig. 5B, C left; Young mice: p = 0.0165, old mice: p = 0.8792). In contrast to the other results described here, this effect is unlikely to be driven by amyloid pathology.

Fig. 5. Switching source and target reveals no age-dependent reduction in communication subspace dimensionality during NREM sharp-wave ripples.

Data were analyzed as in Fig. 4, but source and target regions were reversed to generate the V2R and A2R circuits. A (Top) Conceptual diagram of the analysis. The activity of an ensemble of single units in a source region is used to predict activity in a target region. Source dimensionality is progressively reduced, and the resulting loss of predictive performance is measured. (Bottom) Diagram of the V2R and A2R circuits, both targeting RSPv. B Communication-subspace dimensionality in the V2R (left) and A2R (right) circuits during NREM sharp-wave ripples (SWRs) in young mice. Predictive performance is scaled relative to full-rank prediction. Points represent individual bootstraps; lines show mean scaled performance at each dimension, and shading indicates bootstrapped 95% confidence intervals. Insets show bootstrap distributions of optimal dimensionality, defined as the number of dimensions required to explain 90% of full-rank performance. Optimal dimensionality was analyzed using a linear mixed-effects model: optimal dimension ~ genotype × circuit × condition × age group + (1∣ animal/date/epoch/bootstrap), followed by two-sided comparisons of estimated marginal means. Genotype-comparison p-values were p = 0.0165 for V2R and p = 0.8361 for A2R. C Same as B for old mice. Genotype-comparison p- values were p = 0.8792 for V2R and p = 0.2481 for A2R. N = 17 mice were biological replicates; n = 11,521 bootstraps were nested resampled observations. Source data are provided as a Source Data file.

We observed a similar effect when analyzing delta NREM intervals. Specifically, there were significant differences in optimal dimensionality between genotypes in the A2R circuit but only in young mice (Fig. S12A, B left; Young mice: p = 0.0062, old mice: p = 0.9415). There were no significant differences in any comparisons within the V2R circuit (Fig. S12A, B right; Young mice: p = 0.9255, old mice: p = 0.0871). Likewise, during waking SWRs, there were significant differences in dimensionality between genotypes in the V2R circuit (Fig. S12C, D left; Young mice: p = 0.5885, old mice: p = 0.0138) as well as the A2R circuit (Fig. S12C, D right; Young mice: p = 0.0073, old mice: p = 0.0334). These results highlight a genotype difference in the young cohort; transgene overexpression is a more parsimonious explanation than progressive Aβ accumulation. Intriguingly, these genotype differences were due to a higher dimensionality in APP/PS1 mice relative to WTs rather than an age-dependent reduction in dimensionality described above. In other words, to the extent that transgene overexpression impacts circuit dynamics, it increases complexity, while Aβ does the opposite.

Taken together, these data rule out ACAd as the locus of failure. If ACAd were intrinsically dysfunctional, one would expect the A2R circuit (in which ACAd acts as source) to also show degraded dimensionality in old APP/PS1 mice. It does not; A2R shows no age-dependent reduction during NREM SWRs (old mice: p = 0.2481). ACAd, when considered as a source, appears intact. The deficit therefore implicates a disruption in how RSPv organizes and transmits information specifically toward ACAd: a directional, projection-specific vulnerability in a computational subspace of RSPv. This suggests that the failure resides in RSPv’s output to ACAd, not in ACAd itself.

Across four circuits, the only incidence of age-dependent degradation in dimensionality occurred during NREM SWRs in the R2A circuit. This demonstrates that the functional consequence of Aβ-related vulnerability may not be strictly reducible to structural and anatomical effects. Structural/anatomical changes do not vary on the timescale of hundreds of ms (e.g., SWRs). Taken in the context of global Aβ, global APP overexpression, and global presenilin overexpression, these data suggest that selective vulnerability is not a product of proximity to disease-related molecules (although it is unmasked by those elements). This interpretation is further supported by a direct comparison of plaque burden across recorded regions: in old APP/PS1 animals, while all examined regions exhibited extensive plaque pathology, VISp exhibited significantly higher plaque density than either RSPv or ACAd (all p < 0.0001; Fig. S1). However, VISp showed no age-dependent physiological disruption in any measure reported here. Computationally defined subsets of key circuits are disrupted in a dynamic and state-dependent fashion, consonant with substrates related to the fast patterning and routing of information, such as neuromodulatory systems and recurrent inhibition^58,59.

Discussion

The brain regions comprising the DMN are selectively vulnerable to Aβ deposition and Alzheimer’s disease (AD)-related disruption. This vulnerability has been widely described, particularly in human clinical populations, yet it has been challenging to separate two explanations of this phenomenon. The simplest case is that the pattern of Aβ deposition early in AD determines the pattern of affected circuits. In other words, Aβ plaques are toxic to nearby neurons, and plaque expression is initially limited to the DMN. Alternatively, early AD could represent the confluence of two factors. Specific circuits are intrinsically vulnerable to insult, and it is in these circuits that Aβ takes hold. To parse these explanations, we made four-site recordings of neuronal ensembles in individual APP/PS1 mice, which are characterized by a widespread and progressive Aβ burden across the isocortex and hippocampus. We recorded two key circuits, one previously defined as in-DMN and one previously defined as out-DMN¹⁰. Each of these shares the same source (RSPv, a DMN hub) but differs in their targets (ACAd, a DMN region, and VISp, a non-DMN region). We recorded from two age groups: young and old. Our data reveal age-dependent, circuit level-differences in the impact of Aβ, despite both widespread Aβ load (and constitutive, pan-neuronal APP/PS1 overexpression) and the fact that the same ensembles comprised the source of each circuit. Not only do our data offer a unique resolution of intraregional circuitry in an animal model of amyloid pathology, but our recordings span 24 h of unrestrained, natural behavior. Disruptions of the R2A (in-DMN) circuit were exaggerated throughout NREM sleep and showed robust effects during sharp-wave ripples (SWRs), which are involved in memory consolidation.

While high-density, multi-region recordings of single units in freely behaving animals offer powerful insights into complex neuronal interactions during behavior, there is a tradeoff between resolution and the anatomical breadth of observation. Given the constraints of this approach, we carefully chose the specific in-DMN and out-DMN circuits, guided by prior work from our group¹¹. Specifically, we sought a pair of in- and out- DMN circuits monosynaptically connected to a common source. While our choice of circuits meets these criteria, it is currently impossible to simultaneously record single-neuron activity with spike-time resolution from all circuits originating in the DMN. As a result, our data offer support for the hypothesis that circuits originating and terminating in the DMN are intrinsically vulnerable, even in comparison to those with the same origins but out-DMN targets. However, extensive further work will be required to establish the generalizability of these findings throughout the extent of the DMN.

When using transgenic animals, inducible models are desirable as they introduce a temporal pattern to the expected results. This can demonstrate that effects are not simply the result of transgene overexpression. A similar logic applies to transgenic models of neurodegenerative diseases; the progressive nature of the pathology imposes an intrinsic temporal pattern on the results. Here, because disrupted neuronal activity was largely evident only late in life, we conclude that network-level deficits are most likely a consequence of progressive Aβ pathology; it is well known that APP/PS1 animals accumulate Aβ plaques in a time-dependent fashion^37,38. Directly manipulating Aβ plaque levels in future work would provide more evidence in support of causality. However, it is essential to consider the details of any animal model of a human disease, particularly those that are not naturally occurring within the wild-type animal population. Beyond replicating results across mechanistically distinct models of a disease, results must ultimately hold in humans for validity.

Many mouse models of amyloidosis rely on the transgenic overexpression or knock-in of humanized APP carrying familial disease mutations that increase Aβ production. Lines differ in the anatomical distribution and timing of plaque deposition^11,60, as well as the timing and extent of behavioral, functional, and cellular damage. The APP/PS1 mouse accumulates plaques in the isocortex and hippocampus with little anatomical specificity relative to humans and other models¹¹ and little to no neurodegeneration by 12–16 months of age¹⁶. Apropos this, we detected no significant differences in the yield of single units or histology as a function of genotype that might have indicated neuronal loss in old mice. The broad anatomical expression of Aβ in these mice is related to the promoter (PrP) used to drive the transgenes. This is an important caveat that is often considered undesirable for modeling natural human disease progression.

While APP overexpressing mice exhibit relatively little neurodegeneration, there is evidence of neuronal dysfunction by 12 months in the APP/PS1 line. This takes the form of frequent spontaneous epileptiform discharges on EEG as well as seizures, which we noted in our observations of behavior and physiology, exclusively in the old cohort⁶¹. In the APP/PS1 mouse, Aβ has been shown to both increase^39,40 as well as decrease neuronal activity⁶², perhaps explained by a neuron’s proximity to nearby plaques^39,41. This may also depend on brain state or context^63,64, highlighting the need to examine neuronal activity in multiple regions across a range of naturally occurring conditions. At a lower spatial resolution, APP overexpression reduces the power of gamma oscillations⁶⁵ and alters coupling between the hippocampus and cortex during sleep⁵². Our results suggest that alterations in activity are likely to be circuit-specific and may require simultaneous monitoring of multiple populations of neurons.

The major impact of pathology, especially early in disease progression, may vary throughout the circadian cycle and depend on brain state and behavior. Our data emphasize a key role for state in understanding amyloidosis; in the in-DMN circuit, dysfunction appears pronounced during NREM sleep. This is of particular interest in light of the relationship of NREM sleep to memory. Much work has linked sleep, particularly NREM, to memory consolidation⁶⁶, often involving hippocampal-isocortical interactions⁶⁷. Perhaps it is unsurprising then that early AD in humans is marked by disrupted sleep structure^68–70. Chronically disrupted sleep is understood to causally exacerbate AD in humans^19,20,22,71. Similarly, APP/PS1 mice show disrupted sleep/wake cycles⁷².

Notably, peak firing rates during SWRs shifted significantly later in ACAd in APP/PS1 mice, with a non-significant trend toward earlier shifts in RSPv. This asymmetry suggests that the impact of Aβ on higher-order features of neuronal activity, such as timing and patterning, is not a trivial consequence of hyperexcitability⁶¹. Subtle disruptions of the dynamics of specific transcriptomic cell types⁷³ and/or variable expression of compensatory plasticity^{28,32,74–76} are plausible mechanisms by which Aβ could drive complex and circuit-specific alterations in neuronal dynamics.

Disruption of SWRs in the R2A circuit, both in terms of peak timing and dimensionality, is specific to state. SWRs that occur during quiet waking appear to be impacted differently by Aβ pathology than SWRs that occur during NREM. SWRs in wake and NREM have distinct neuromodulatory contexts and cognitive implications: waking SWRs are associated with planning future actions and memory retrieval, whereas NREM SWRs are linked with memory consolidation⁷⁷. Further, the dimensionality of communication in the in-DMN circuit is significantly reduced during NREM SWRs, while the out-DMN circuit is equivalent to WT. The suggestion that SWRs are a locus of pathological effect is consistent with observations of abnormal SWR properties in mouse models of a range of neurodegenerative diseases^52,78,79. Further work is required to test the hypothesis that disruption in the balance of excitation and inhibition⁶⁵ may be restricted to brain states or events⁸⁰.

Our findings align well with the cascading network failure hypothesis in AD⁸¹. This asserts that AD pathogenesis is the result of network failures, beginning in the posterior DMN, that shift processing burden to the next step of the system, which subsequently fails as the damage propagates along lines of high connectivity (i.e., through the DMN, initially). This model predicts increases in connectivity that precede structural and functional disruption, and accounts for the prion-like spread of Aβ^25,69,82. Key to this theory is that altered functional communication between DMN regions is an inciting event. Consistent with this, our data confirm impaired communication in the in-DMN pathway, despite equivalent amyloid burden in the off-DMN pathway. Also of note, our data reveal lagged connectivity in the direction of RSPv to ACAd (posterior to anterior DMN), which is also consistent with data from humans with AD⁸¹. The consistency of our observations and key features of human pathology suggests that the mechanisms underlying the selective vulnerability of DMN circuits are phylogenetically conserved and are not tethered to the specifics of the disease process.

We propose that in AD, the selective vulnerability of the DMN is a consequence of intrinsic vulnerability to injury and is not an artifact of the anatomical specificity of amyloid accumulation. Within DMN regions, the vulnerability of individual neurons to functional disruption is likely to be determined by the cell type-specific gene expression underlying connectivity patterns^10,83. How this interacts with the increased activity in the DMN at rest is a crucial question, as neuronal activity can exacerbate Aβ pathology⁸⁴. The advent of new technologies to measure mesoscale connectivity, record neuronal activity, and resolve spatial transcriptomics will make it possible to further disentangle cellular and circuit rules by which disease spreads through the DMN.

Methods

Animal husbandry and model

All procedures involving mice were performed in accordance with protocols approved by the Washington University in Saint Louis Institutional Animal Care and Use Committee, following guidelines described in the US National Institutes of Health Guide for the Care and Use of Laboratory Animals. We used eight heterozygous APP^+/- mice and nine nontransgenic littermates (APP^-/-) from the transgenic APP/PS1 line (B6.Cg-Tg[APPswe, PSEN1dE9]85Dbo/Mmjax, MMRRC Stock No: 034832-JAX¹³). In the young mouse dataset, all mice were between 2 and 4 months of age at the time of the experiment. In the old mouse dataset, all mice were between 12 and 16 months of age at the time of the experiment. Mice were housed in an enriched environment and kept on a 12:12 h light:dark cycle with ad libitum access to food and water.

Quantification of amyloid plaque burden

To quantify the magnitude and variability of progressive plaque accumulation in the APP/PS1 mouse line, we quantified plaque density in ACAd, RSPv, CA1, and VISp in an independent cohort of APP/PS1 animals across two age groups. This dataset is publicly available at https://brain-map.org/our-research/connectivity/mouse-mesoscale-connectivity-in-aging-and-alzheimers-disease, and contains 25 brains previously published in Whitesell et al.¹¹ plus 246 additional unpublished brains spanning multiplnd hippocampus of a large cohort of APPe transgenic lines and ages; analyses here are restricted to the APP/PS1 subset at 4 months (N = 3; 1 male, 2 female) and 12–18 months (N = 140; 16 female, 124 male). Plaques were labeled with methoxy-X04 and imaged across the entire brain using serial two-photon tomography as described previously¹¹. Plaque density was calculated as the fraction of brain structure volume containing plaques using automated segmentation and registration to the Allen Mouse Brain Atlas CCFv3. Statistical analyses were performed using linear mixed-effects models (lme4⁴²) with structure, age group, and sex as fixed effects and individual animal as a random effect. Estimated marginal means and pairwise contrasts were computed using emmeans. For models comparing young and old animals, a small offset equal to half the minimum nonzero observed density was added prior to log-transformation to all values, ensuring zero values present in young animals could be accommodated while preserving comparability across age groups. For models restricted to old animals, no transformation or offset was necessary as all plaque density values were strictly positive and residual diagnostics confirmed approximate normality on the raw scale.

Surgical procedures

All mice underwent multisite electrode array implantation surgery. Mice were anesthetized with inhaled isoflurane (1–2% in air) and administered slow release buprenorphine (ZooPharm, 0.1 mg kg⁻¹) for analgesia. The mouse’s skull was secured in a robotic stereotaxic instrument (NeuroStar, Tubingen, Germany), and the skin and periosteum covering the dorsal surface of the skull was removed. Lambda and bregma midline positions were identified and used as reference coordinates for alignment with stereotaxic atlases using NeuroStar StereoDrive software. Four craniotomies (diameter 1–1.5 mm) were drilled over target implantation sites using the automatic drilling function of the stereotaxic robot, and the dura mater membrane was resected. Four brain regions were implanted with custom 64-channel nichrome tetrode-based arrays cut at a 45^∘ angle such that the distance between the shortest and longest tetrode was 350–500 μm. Arrays were fixed (not drivable) and separated from headstage hardware by a flex cable, thus allowing an arbitrary geometry of multiple probes. In each mouse, four 64 channel arrays were targeted so that the longest recording wire was centered on the following stereotaxic coordinates (AP/ML/DV relative to bregma and dura, in mm): CA1 (−2.52/−1.75/1.5), ACAd (0.6/−0.25/1.15), RSPv (−3.28/−0.95/1.17), and VISp (−4.03/−2.88/1.25). Electrode bundles were lowered into brain tissue at a rate of 3 mm/min using a custom built vacuum holder. Anatomical location was confirmed post hoc via histological reconstruction. Arrays were secured with dental cement (C&B-Metabond Quick! Luting Cement, Parkell Products Inc; Flow-It ALC Flowable Dental Composite, Pentron), and headstage electronics (eCube, White Matter LLC) were bundled and secured in custom 3D-printed housing. Four-module (256 channel) implants weighed approximately 2 g. Post surgery, mice were administered meloxicam (Pivetal, 5 mg kg⁻¹ day⁻¹ for 3 days) and dexamethasone (0.5 mg kg⁻¹ day⁻¹ for 3 days) and allowed to recover in the recording chamber prior to recording.

Recording procedures

Recordings were made using custom tetrode-based arrays. Sixteen tetrodes (64 channels) were soldered to a custom-designed PCB (5 mm × 5 mm × 200 μm) which stacked horizontally with a similarly sized amplifier chip (White Matter LLC, Seattle). PCB/amplifier pairs were further stacked with 3 additional pairs (total 4 modules, 256 channels). Recordings were conducted in an enriched home cage environment with social access to a litter mate through a perforated acrylic divider. Freely behaving mice were attached to a custom built cable with in-line commutation. Neuronal signals were amplified, digitized, and sampled at 25 kHz alongside synchronized 15–30 fps video using the eCube Server electrophysiology system (White Matter LLC). Recordings were conducted continuously for between 2 weeks and 3 months. Data and video were continuously monitored using Open Ephys⁸⁵ and Watch Tower (White Matter LLC). Twenty-four-hour blocks of data were identified for inclusion in these studies first by the absence of hardware and/or software problems, for example, cable disconnects or dropped video frames, respectively, and second by the maximal yield of active channels. Beyond these criteria, selection of 24 h blocks was arbitrary.

For experiments involving spike-sorted data, raw data were bandpass filtered between 350 and 7500 Hz and spike waveforms were extracted and clustered using a modified version of SpikeInterface⁸⁶ and MountainSort4²⁶ with curation turned off. A custom XGBoosted decision tree was used to identify those clusters constituting single units. Clusters identified as single units were manually inspected to confirm the presence of high-amplitude spiking, stable spike amplitude over time, consistent waveform shapes, and little to no refractory period contamination. For all downstream analyses, we excluded neurons with firing rates <0.5 Hz or presence ratio <0.8 throughout the duration of a sleep/wake epoch. RSU and FS units were classified using trough-to-peak times of spike waveforms: units with a trough-to-peak time of less than or greater than 0.4 ms were classified as FS or RSU units, respectively.

Probe localization

Following recording, mice were perfused with 4% formaldehyde (PFA), the brain was extracted, and immersion fixed for 24 h at 4 °C in PFA. Brains were then transferred to a 30% sucrose solution in PBS and stored at 4 °C until brains sank. Brains were then sectioned at 50 μm on a cryostat. Sections were rinsed in PBS prior to mounting on charged slides (SuperFrost Plus, Fisher) and stained with cresyl violet. Stained sections were aligned with the Allen Institute Mouse Brain Atlas⁸⁷, and tetrode tracks were identified under a microscope (Fig. S1A).

Sleep-wake scoring

LFP was extracted in the 0.1–60 Hz range from 5 manually selected channels for each subject. Channels were averaged together, and then spectrograms were generated in 1-h increments. Sleep scoring was performed semi-manually in custom GUI software based on spectrogram data and movement tracking of the headstage generated with DeepLabCut³¹. States were assigned to NREM, REM, or Wake (Quiet, Active) based on delta bandpower (high in NREM), theta bandpower (high in REM), and movement/broadband LFP activity (high in wake) using well established methods^34,88. A random forest trained on sleep scores from multiple human raters was used to assign sleep scores in 4 s epochs from spectrogram and movement data. These were checked manually by human raters and corrected where appropriate. For periods where sleep state could not be easily discerned by spectrogram and movement alone, temporally aligned video was consulted. For downstream analyses, we only included sleep/wake epochs of duration longer than 20 s to avoid epochs dominated by the effects of state transitions.

Single unit FR and correlation analysis

For each single unit, the average number of spikes was calculated in 1 s bins across each epoch. Then the mean firing rate was calculated across all epochs of a particular sleep state for each neuron individually. For each sleep epoch, spikes were binned in 1 ms windows for source and target regions. The Pearson correlation was calculated between the binned spike time series for each source-target pair of units at lags from −80 to +80 ms to generate a lagged correlogram. The mean of all lagged correlograms for every pair of source-target units was calculated for each epoch. Only epochs of greater than 60 s duration were included, after which correlograms were z-scored from the mean and standard deviation of the 20 ms flanks of the correlogram (−80 : −60 ms, +60 : +80 ms). Z-scored correlograms were then smoothed with 2 ms Gaussian kernel for visualization. For each epoch’s lagged correlogram, the peak position was found by identifying the time lag at which the maximum of the correlogram trace occurred on the z-scored correlogram in a 40 ms window around the center (−20 : +20 ms).

Ripple detection

For ripple detection, data from CA1 probes was downsampled to 1500 Hz and notch-filtered at harmonics of 60 Hz to eliminate 60 cycle noise (60, 120, 180, 240, and 300 Hz)⁸⁹. The four channels with the lowest contamination in the ripple band were selected by determining the lowest mean bandpower in the ripple band (150–250 Hz) in 5 min windows. Signal from the four selected channels was then bandpass filtered between 150 and 250 Hz before performing ripple detection with a publicly available Python package, github.com/Eden-Kramer-Lab/ripple_detection, implementing ripple detection from Karlsson and Frank⁵¹. Ripple amplitude was calculated as the difference between the maximum and minimum values of the ripple band-filtered LFP signal during each ripple. Only ripples with amplitude >50 μV were included in downstream analyses. Ripples less than 100 ms apart were combined.

To calculate peri-event time histograms (PETHs) around sharp-wave ripples, single units were separated by putative cell type using spike waveform characteristics (RSU and FS). The PETH was then calculated by summing all spike events for each unit in a given region in a 1 s bin around each ripple time occurring during a NREM epoch. PETHs were then averaged across ripples for each unit. Average PETHs were then z-scored using the flanks of the PETH (−30% : +30%) to enable comparison across animals/recording epochs. All z-scored PETHs from all single units of a given cell type and region were then averaged together. To ensure that only responsive cells were included in genotype comparisons, only PETHs with activity two standard deviations above PETH flank activity were used in statistical comparison.

Communication subspace analysis

For each recording epoch and animal, spike data from each brain region was aligned around instances of each of three neurophysiological phenomena: sharp-wave ripples occurring during NREM, sharp-wave ripples occurring during waking, and randomly selected intervals of NREM demonstrating slow waves (delta: 0.1–4  Hz) and lacking sharp-wave ripples. For each type of event (NREM SWRs, waking SWRs, and delta NREM), spike times were binned (100 ms) in a 1 s window centered on each event (e.g., ripple). The average peri-event time histogram (PETH) across all events was subtracted from each individual PETH. Each PETH was then z-scored to account for trial-to-trial variance, and then concatenated into one final time series for each recording epoch. In essence, we conducted standard, z-score normalized PETH analyses around SWRs or outside of SWRs, considering both NREM and wake.

From this preprocessed data, communication subspace analysis was performed using MATLAB code adapted from Semedo et al. (github.com/joao-semedo/communication-subspace)⁵⁷. Briefly, reduced rank regression was performed between source RSP activity and target activity from VISp and ACAd. If data was rank deficient, non-contributory units were found via QR decomposition and removed. Only datasets with more than 10 units were used for subsequent reduced rank regression. To ensure that source and target contained equal numbers of units, units were randomly sampled from the larger dataset of either source or target data across 25 bootstraps to match the size of the smaller dataset. Reduced rank regression was then performed across 10 predictive dimensions using 10-fold cross-validation. Optimal dimensionality was calculated as the number of predictive dimensions needed to explain 90% of the full-rank performance⁵⁷. Reduced rank regression was performed using the RidgeInit and Scaling parameters available in the Semedo MATLAB code. Normalized squared error was used for the loss function (sum of squared errors divided by the total sum of squares of the target data).

Results from each bootstrap were used in comparison of optimal dimensionality across genotypes. To ensure that analysis only contained reduced rank regression results where performance improved with increasing rank, RRR performance curves were included if performance at full rank was higher than performance at lowest rank and if performance at full rank was non-negative. The predictive dimensions- performance relationship curve was min-max scaled from 0 to 1 for each curve in order to make comparison across recording epochs and animals feasible.

Statistics and reproducibility

Unless otherwise noted, all statistical analyses involved the implementation of a linear mixed-effects model using the lmer function from the lme4 package in R⁴². This model was developed to evaluate the potential influence of different factors on the relevant response variable (e.g., firing rate, predictive power). Fixed effects in models comprised main effects, such as firing rate, genotype, circuit, age group, and state, as well as their interaction effects. These variables represent distinct biological or experimental parameters in our study.

Random effects were also incorporated into the model to account for non-independence and potential clustering in the data. For example, we modeled random effects of animal, recordings, epochs, neuron ID, and bootstrap (where applicable) as random effects. When appropriate, we included nested random effects (such as bootstrap). These random effects account for the hierarchical or nested structure of the data and control for the non-independence of observations from the same animal, epoch, date, or bootstrap. In this dataset, the mouse (N = 17) constitutes the biological unit of replication, while neurons, arousal-state epochs, and recording sessions represent non-independent, nested measurements within each animal; linear mixed-effects modeling was used specifically to account for this hierarchical structure rather than treating nested measurements as independent data points.

Model fitting was conducted using restricted maximum likelihood (REML) estimation, which provides unbiased estimates of variance and covariance parameters. After fitting the model, the fixed effect estimates were extracted to interpret the influence of each variable and their interaction effects on the response variable. We used Tukey-adjusted emmeans contrasts to assess p-values for comparing means.

Data are reported as mean ± SEM unless otherwise noted. All tests were performed with a significance threshold set at p < 0.05.

Sex as a biological variable

To account for established sex differences in the progression of amyloid pathology in APP/PS1 mice^35,36, we tested whether sex contributed to the age-dependent genotype effects reported here. For each major analysis, we refit the linear mixed-effects model with the additional fixed-effect term sex × age group ×  genotype, which adds main effects of sex, sex × age, sex × genotype, and, critically, the sex × age × genotype interaction. This directly tests whether age-dependent genotype effects differ between males and females. We then evaluated the resulting type III ANOVA (Satterthwaite’s method) for significance of the sex main effect and all sex-containing interactions. Across all measures of neural activity reported in this study, no sex-related term reached significance, and the primary genotype, age, region, and state effects were preserved (Table 3). We note that, because sex is an animal-level variable, these statistical tests involving sex are evaluated at the animal level and therefore have modest power to detect small effects. Null results regarding sex should be interpreted with this limitation in mind.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

S.J.B. assisted in recordings, performed data preprocessing, designed experiments, performed statistical analysis, and wrote the manuscript. J.N.M. performed recordings, data preprocessing, conducted analyses, wrote the manuscript, and produced figures. K.B.N. provided technical consultation and software pipelines for data preprocessing. C.A.F. performed surgeries and recordings. H.E. performed and analyzed histology. B.T.H. performed surgeries and recordings. Z.B.R. performed surgeries and recordings. D.K. conducted analyses and produced figures. E.L.D. provided mentorship and assisted in writing the manuscript. J.M.G. provided statistical assistance and helped construct figures. J.D.W. and J.A.H. contributed to the initial concept, mentorship, and consultation. K.B.H. lead and directed the project, produced figures, and wrote the manuscript.

Peer review

Peer review information

Nature Communications thanks Todd Golde and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by NIH BRAIN Initiative R01NS118442 (K.B.H.), R01AG091773 (K.B.H.), 1R01EB029852 (E.L.D. and K.B.H.), BrightFocus Foundation Standard Award A2022038S (K.B.H.), Cure Alzheimer’s Fund (K.B.H. and J.N.M.), and in part by the National Institute on Aging R01AG047589 (J.A.H.).

Data availability

The datasets generated and/or analyzed in this study constitute ~100 terabytes of raw neural (broadband) data. The data are stored in a cost-efficient manner, not immediately accessible to the internet. Raw data are available upon request. For the production of the figures. Source data are provided with this paper.

Code availability

All relevant code from our lab, including software needed to run recordings like ours, is in Python and is publicly available at github.com/hengenlab. Code necessary for the analyses used in this paper are available at github.com/hengenlab/appps1. Other groups’ code, including Open Ephys, SpikeInterface, MountainSort4, DeepLabCut, RippleDetection, and Communication Subspace, is publicly available as specified in Methods.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76952-z.