Resting-state functional MRI (rsfMRI), offering insight into the brain's functional organization and connectivity, has emerged as a powerful non-invasive tool in Alzheimer's disease (AD) research. Insights obtained using this imaging technique have shown potential for early detection of AD, as it identified brain network alterations and abnormalities before the onset of clinical signs (Sheline and Raichle, 2013). RsfMRI records spontaneous low-frequency fluctuations in the blood-oxygenation-level-dependent (BOLD) signals as an indirect measure of neuronal activity at rest, such as during quiet wakefulness or while lying still with eyes closed. It allows for the mapping of brain-wide functional connectivity (FC), defined as the strength of the temporal correlation of BOLD activity between spatially distant regions, and in turn unveils resting-state networks (RSNs) constituted by brain regions with correlated activity. Predominantly active during rest and associated with introspection and memory, the Default Mode Network (DMN) is a particularly relevant RSN in AD (Badhwar et al., 2017), as its reduced connectivity is indicative of the disease's progression (Greicius et al., 2004; van den Heuvel and Sporns, 2013; Alves et al., 2019; Ramzan et al., 2019). Furthermore, the DMN's anti-correlation with the Task-Positive Network (TPN), which is engaged during cognitive tasks, is diminished in AD, thus reducing the segregation between DMN and TPN (Belloy et al., 2018b). Understanding changes in FC within and between the DMN, TPN, and other RSNs has shown to be highly relevant to elucidate AD's impact on brain function and potential for early detection and monitoring of disease progression (Agosta et al., 2012; Ibrahim et al., 2021).

Animal models of AD provide critical insights into the disease's pathophysiology and enable the exploration of especially early impact of AD pathology. The short lifespan allows for monitoring the evolution of the disease phenotype in a relatively brief period, even before onset of the behavioral phenotype. One such model, the TgF344-AD rat, is characterized by the presence of human APP_swe and PS1_ΔE9 mutations, leading to the progressive build-up of amyloid beta (Aβ) deposits starting at 6 months of age (Cohen et al., 2013; Joo et al., 2017), accompanied by accumulation of phosphorylated tau (pTau) (Joo et al., 2017; Rorabaugh et al., 2017), gliosis, neuronal loss and cognitive impairments similar to those observed in human AD (Braak et al., 2011; Cohen et al., 2013; Munoz-Moreno et al., 2018). Impairments in spatial reference learning, spatial navigation and working memory have been observed starting at 6 months of age (Berkowitz et al., 2018; Tournier et al., 2021), with anxiety-like behavior preceding already at the age of 4 months (Pentkowski et al., 2018).

RsfMRI in rodents shows similar features as in humans; in particular analogs of RSNs have been identified in mice and rats (Lu et al., 2012). In previous work using different MRI modalities including rsfMRI in TgF344-AD female rats (Anckaerts et al., 2019) we have shown that changes in FC preceded structural changes. A small decrease in FC at the age of 6 months (in connections involving the hippocampus, cingulate, retrosplenial and sensorimotor cortices) was followed by widespread hypoconnectivity at 10 months. Another study (Tudela et al., 2019) showed lowered DMN overall mean connectivity at 5 months and decreased anterior-posterior DMN connectivity at 15 months. More evidence of network-level alterations preceding plaque formation was found in regional FC alterations at 5 months, located in the right insular cortex, amygdala and hypothalamus (Munoz-Moreno et al., 2018, 2020), The combined findings of these studies emphasize the relevance of this model in characterizing not only AD's pathology and behavioral phenotype, but also its impact, especially at early stages, on brain function.

These rsfMRI studies assumed static, constant FC throughout the scan. However, advances like sliding window FC analyses have revealed FC fluctuations that are missed with static FC (Hutchison et al., 2013; Calhoun et al., 2014), and changes in activity patterns over the scan's duration (Lurie et al., 2020; Maltbie et al., 2022; Xu et al., 2022). One such approach uncovers the quasi-periodic patterns (QPPs) (Majeed et al., 2011), which are recurring spatiotemporal patterns of brain activity within a predefined time window. QPPs are obtained using a sliding window approach, where repetitive patterns of brain-wide neural activity are detected and averaged across them. Depending on the window size, QPPs can range from a single spatiotemporal brain state pattern to a sequence of them. In contrast to sliding window FC approaches, which yield one FC matrix per window, QPPs facilitate further insights into the dynamics of the activity states within the window itself (Belloy et al., 2018a, 2021; Yousefi et al., 2018; Yousefi and Keilholz, 2021). In the context of AD, QPP's have been used to detect early alterations in preclinical models (Belloy et al., 2018b; van den Berg et al., 2022). Our study using QPPs in TgF344-AD rats (van den Berg et al., 2022) highlighted a specific alteration: while overall FC between the basal forebrain (BFB) and the default mode-like network (DMLN) showed no notable differences, there was a distinct reduction in their coactivation within the QPP in TgF344-AD rats already at the pre-plaque stage of 4 months of age. This finding underscores the nuanced, early changes in neural interactions and disease processes in AD, detectable through dynamic QPP analysis but overlooked by conventional FC methods. However, insights gained from QPPs are inherently limited to the recurring states, covering only parts of entire scan. Additionally, an arbitrary selection of window size for analysis can yield varying outcomes, potentially missing dynamics that occur outside or across the chosen windows.

Co-activation patterns (CAPs) have been proposed as an alternative framework for investigating the brain's dynamic states. CAPs are based on point process models (Tagliazucchi et al., 2012) and use clustering (Liu and Duyn, 2013) to divide all time frames within a scan, based on spatial similarity of their voxel-level BOLD signals, into a few (typically 6–10) functional coactivation states. They offer several advantages over QPPs, including the ability to capture and quantify inter-CAP transitions and higher statistical power due to single-frame resolution. Utilizing CAPs as an analytical tool has shown promise in classifying and differentiating properties associated with various physiological and pathological conditions (Liu et al., 2018; Gutierrez-Barragan et al., 2019, 2022; Adhikari et al., 2020, 2023; Cifre et al., 2021; Li et al., 2021; Lee et al., 2023; Vasilkovska et al., 2023; An et al., 2024). Recent studies employing CAP analysis in mouse models of neurodegenerative disorders (Gutierrez-Barragan et al., 2022; Adhikari et al., 2023) have demonstrated their predictive capacity in distinguishing model mice from the wildtype at different stages of the disease.

In this work, we performed a resting-state CAP analysis on our previously obtained longitudinal dataset in the TgF344-AD rat model of AD at the ages of 4 and 6 months, corresponding to the pre- and early-plaque stages of disease progression (van den Berg et al., 2022). Our first aim was to investigate the genotype and age effects on the spatial and temporal properties of CAPs. Then we assessed the capability of CAP metrics to accurately identify the genotypes, by investigating their classification capabilities with specific interest for the pre-plaque stage. Finally, and most importantly, we characterized inter-CAP transitions, to investigate if and how the dynamic organization of intrinsic brain functional networks states, undetectable by earlier methods, is altered at these early stages in this model of AD.

The data utilized in this study were initially obtained and published in a previous manuscript (van den Berg et al., 2022). Acquisition and rsfMRI pre-processing procedures are reiterated here for completeness.

In this study, we used resting-state CAPs to longitudinally examine properties of spontaneous whole-brain activity states in the TgF344-AD rat model of AD at the pre- and early-plaque stages of disease progression. We found six CAPs were the optimal number of states in this dataset, distinctly organized into CAP- anti-CAP pairs. DMLN activation was found in two of the CAPs, with one of them showing simultaneous BFB activation. In this DMLN-BFB CAP, BFB activation was reduced in the TG group only at the pre-plaque stage. Inter-genotypic differences in spatial activations of CAPs were diminished at the early-plaque stage. Analysis of CAP temporal properties revealed a reduction in the occurrence percentage of the DMLN CAP in the TG animals at the early-plaque stage. Cross-validated four-class classification of the animals into genotype-age-groups using spatial features of CAPs displayed significantly higher accuracy than chance. However, larger confusion between genotypes was found at the early-plaque stage which was confirmed when two genotypic groups were classified at each age; accuracy was significantly higher than the chance level at only the pre-plaque stage. Through characterization of network state transitions, we identified genotypic differences in transition probabilities between two CAP pairs at the pre-plaque stage. TG animals showed increased probability for DMLN→DMLN-BFB transitions, and reduced probability for DMLN-BFB→deacLCN transitions. Interestingly, while several inter-CAP transitions at the pre-plaque stage exhibited directional preference in both genotypes, at the early-plaque stage, this directionality was observed only in the TG animals. In WT animals at the early-plaque stage, inter-CAP transitions were equally probable in both directions in most cases.

Our results, similar to earlier findings using CAP analysis in rodents (Gutierrez-Barragan et al., 2022; Adhikari et al., 2023), indicate a partition of six clusters as an optimal number to explain variance between the clusters, with diminishing returns using larger cluster numbers. These six clusters are distinctly organized into CAP-anti-CAP pairs. Across multiple CAPs, the DMLN and LCN co-activation patterns emerge as recognizable networks, previously described in the context of AD (Ibrahim et al., 2021; Liang et al., 2021; van den Berg et al., 2022). These networks are the respective rodent analogs of the DMN and TPN in humans. The DMN especially, involved in memory consolidation tasks, is the most prominent resting-state network, with AD patients typically exhibiting disrupted within-DMN connectivity (Grieder et al., 2018).

The DMLN CAP exhibits strong co-deactivation throughout cortical LCN regions (somatosensory, motor, and auditory cortex), thus further characterizing it as a brain state resembling DMLN activity. Interestingly, it anticorrelates most strongly with the LCN-Barrel CAP, rather than the main LCN CAP itself. In the DMLN-BFB CAP, strong co-deactivation is strictly localized to the caudate putamen, extending slightly into the primary somatosensory cortex. It's anti-correlated counterpart, CPu CAP, demonstrates highly localized co-activation in the caudate putamen only, with widespread cortical co-deactivation. Interestingly, the DMLN-BFB CAP shows the strongest co-activation in orbital and frontal association cortex together with BFB nuclei, while the DMLN CAP's co-activation is more centered around the cingulate and prelimbic cortex. The separated DMLN-like CAPs we observe (DMLN-BFB and DMLN) are an example of how the single-frame resolution of CAPs provides enhanced statistical power, enabling a deeper and more granular understanding of the intricate and transient nature of brain connectivity. Another study comparing CAPs in awake and anesthetised mice (Gutierrez-Barragan et al., 2022) also found CAPs resembling DMN and LCN. Additionally, in their study, two CAPs showed high basal forebrain co-activation, one with hypothalamus and the other with auditory cortex, visual cortex, and amygdala. They also showed a CAP with high CPu co-activation, which was not isolated but co-activated mainly with cingulate, retrosplenial and prefrontal cortices. Their findings were consistent using either halothane or isoflurane-medetomidine anesthesia.

One of the main outcomes of our spatial analysis of CAPs is the reduced co-activation in the BFB in TG animals at the pre-plaque stage, situated mainly in the DMLN-BFB CAP, confirming our earlier findings using QPPs (Geula et al., 2021). Basal forebrain cholinergic neurons (BFCN) experience early and substantial degeneration in AD, with the major neurodegenerative correlate being accumulation of pTau in neurofibrillary tangles (Geula et al., 2021). Consequently, BFCN degeneration is regarded as one of the hallmarks of AD (Shekari and Fahnestock, 2021), with evidence of decreased BFB volume before Aβ cortical spreading (Hall et al., 2008; Chen and Mobley, 2019; Fernandez-Cabello et al., 2020). Given its implicated roles in modulation of cognition (through projections from the diagonal band of Broca, substantia innominata, and the nucleus basalis of Meynert to the cortex) (Li et al., 2015; Turchi et al., 2018; Espinosa et al., 2019) and its modulating function on brain activity through control on the prefrontal cortex (Li et al., 2015; Turchi et al., 2018; Espinosa et al., 2019), disruptions in BFB signaling caused by AD can lead to early disturbances in the brain's network activities.

The differences in pre-plaque BFB co-activations we observe in the current study are like what was earlier described with quasi-periodic patterns (QPPs) in the same dataset (van den Berg et al., 2022). In that study, two short QPPs, called the DMLN and LCN QPPs consisting of coactivation of DMLN and LCN regions respectively, were found to be prominent. Co-activation of the BFB and the DMLN and the level of BFB activation was found be reduced at the pre-plaque stage in the TG animals. This reduction was not observed in the DMLN QPPs at the early-plaque stage. In the present study, while the DMLN and DMLN-BFB CAPs exhibit co-activation in similar hub regions involved in the DMLN (i.e., cingulate gyrus, retrosplenial cortex, orbitofrontal cortex, infralimbic cortex, visual cortex, and auditory cortex), the BFB activation was selective in the DMLN-BFB CAP (Figure 2C). Hence, the observed reduction of BFB co-activation in TG animals notably occurred mainly in the specific DMLN-BFB CAP, but not in the DMLN CAP, which is overall most closely resembling the DMLN (Figures 3, 4 ^* MERGEFORMAT). In line with the earlier findings, these differences are pronounced during the pre-plaque stage, and mostly diminished at the early-plaque stage. QPPs are formed by averaging across recurring segments of scans of specific window size. As CAPs cover every frame, the overlap between QPP segments/occurrence and CAP occurrences can be calculated. We speculate that the previously described DMLN QPP involves brain states corresponding to both DMLN-BFB and DMLN CAPs, where the overlap of DMLN QPP segments with DMLN-BFB CAP would be significantly higher than that with DMLN CAP, especially at the pre-plaque stage.

In the LCN-Barrel CAP, anti-correlated to DMLN CAP and strongly resembling the LCN, we observed highly localized hyperactivations in somatosensory, auditory, and visual cortex regions in the TG animals. These are indicative of increased activation in the LCN, again corroborating previous findings (van den Berg et al., 2022). Like the other spatial differences, genotypic differences were most pronounced at the pre-plaque stage and diminished at the early-plaque stage.

Temporal properties of CAPs showed few differences between the genotypes, apart from the DMLN CAP. While mean duration of the DMLN CAP increased from the pre- to the early-plaque stage across both genotypes (Figure 5A), the occurrence percentage of the CAP increased only in the WTs, but not in the TG animals (Figure 5B). This suggests a healthy aging effect on the occurrence of the DMLN CAP which is disrupted in the AD animals. In the zQ175 DN mouse model of Huntington's disease, recent work employing a similar methodology (Adhikari et al., 2023) showed reduced durations of two anticorrelated CAPs, characterized by simultaneous co-activation of DMLN and co-deactivation of LCN and vice-versa, compared to the wild-type mice. In this work, multiple timepoints were investigated with only the final (10M) timepoint indicating significant differences in temporal metrics. In this light, we speculate that temporal metrics of the CAPs will be subject to stronger genotypic effects of AD at later ages, potentially involving other CAPs. Support for this hypothesis can be found in another study (Ma et al., 2020), where CAPs were studied in healthy elderly participants, patients with mild cognitive injury (MCI), and AD patients, showing that average dwell time in the DMN-like CAP showed a decreasing trend in AD.

Upon finding genotypic differences in both spatial and temporal metrics of the CAPs, we aimed to investigate their predictive power in identifying the genotypic status of the animals. In line with the earlier findings using this approach in Huntington's disease mice (Adhikari et al., 2023), temporal properties of the CAPs did not perform classification better than chance in the current study. In contrast, classification using the spatial properties resulted in a significantly more accurate prediction of the genotype compared to the chance level. Interestingly, when stratifying the classifier per age, we found that accuracy of prediction reaches 65% using spatial features of the CAPs at the pre-plaque stage, but no significant improvement over the chance level at the early-plaque stage. Since earlier research using CAP analysis in AD mice at late-stage of AD development (18M) demonstrated near-perfect classification accuracy using CAP spatial features (Adhikari et al., 2020), we speculate that also in the TgF344-AD rat model, accuracy will further increase at later ages. We suspect accuracy of 4-class classification is lowered due to our timepoints being close (4M and 6M), combined with the earlier observation that genotypic spatial differences were less pronounced at the early-plaque than at pre-plaque stage.

One of the key advantages of CAP analysis over other dynamic FC methods is the additional dimension to characterization of the dynamics, by not only revealing transient brain states and investigating their spatial and temporal properties, but by also characterizing and quantifying the transitions between them. One finding of our CAP transition analysis is the consistent role of the DMLN CAP as distinct attractor for both the DMLN-BFB and deacLCN CAPs, across both genotypes and at both ages. These transitions do not only show significant probabilities; other transitions into the DMLN CAP have near-zero probabilities in each of the subgroups. Interpretation of this finding is two-fold. First, tying in with the earlier characterization of the DMLN-like CAPs (DMLN and DMLN-BFB), involvement of the BFB with the DMLN seems to be tied to a specific brain state that is separate from pure DMLN activity alone, but both states have a high affinity toward each other. Second, pure DMLN activation state rarely occurs directly out of LCN activity, but rather through more intermediate states; either the pronounced deactivation of the LCN (deacLCN) or the DMLN state with high involvement of the BFB (DMLN-BFB).

At the pre-plaque stage, the transition from DMLN to DMLN-BFB occurs significantly in the WTs (Figure 7A) but not in the TG animals (Figure 7B), in which its probability is significantly lower (Figure 7C). Moreover, while DMLN→DMLN-BFB is the only significant outgoing transition from the DMLN in pre-plaque WTs, the main outgoing transition from DMLN in TGs is directly to the LCN. This builds on observations of altered DMLN FC in AD patients (Greicius et al., 2004; Wang et al., 2013) and reduced segregation between DMN and TPN (Belloy et al., 2018b), with the current CAP analysis indicating an AD-induced shift in the flow of brain states involving the DMLN already at the pre-plaque stage. Additionally, we observe an increased probability of the DMLN-BFB→deacLCN transition in the TG animals.

At the early-plaque stage, transition probabilities are no longer different between the genotypes. Here, however, we observe a striking difference in the directionality preference of transitions. In the WT animals, besides the DMLN→CPu transition (1/11), none of the significant transitions exhibit directionality preferences. The TG animals exhibit directionality preference in seven out of their 11 significant transitions. This suggests an altered rigidity of brain state flows, where TG animals seemingly exhibit more rigid transition patterns compared to the more flexible flow of brain states in the WTs. A modeling study, based on activity flow over altered intrinsic functional connections in AD, has previously reliably predicted task activations and related dysfunction in individuals at risk for AD (Mill et al., 2020). They proposed a mechanism where alterations in FC associated with AD would disrupt the flow of activations between regions, in turn leading to aberrant task activations. Our findings support the hypothesis that the resting-state flow of activations is altered, already at the very early stages of AD.

Comparing the current work to a previous study investigating persistence and transition probabilities in awake and anesthetised mice (Gutierrez-Barragan et al., 2022), we observe a few differences with our findings, which we can partly explain by the higher percentage of persistent transitions we detect. We observe that about 85% of our transitions are persistent, reflected in the longer durations of our CAPs while Gutierrez-Barragan et al. (2022) find that about 60% of transitions are persistent. They also find a much larger variability in the values for persistence probabilities across CAPs compared to our study. Like our earlier findings and those of other groups (Anckaerts et al., 2019; Tudela et al., 2019; van den Berg et al., 2022), most of the spatial and temporal differences we can detect are mostly present at the pre-plaque stage but diminished in the early-plaque stage. This overall trend remains true for the transition probabilities; however, we can further nuance this finding with the observed alterations in directionality preference at the early-plaque stage. This nuanced characterization of brain connectivity dynamics, inherent to CAP transition analysis, elucidates new insights into early alterations of intrinsic brain connectivity in AD.

Our study, while providing interesting insights, has several limitations. Our analysis was confined to male subjects only, which may limit the generalisability of our findings across genders. We used males only to limit variability and the number of animals required, however, sex differences in disease progression have been shown in humans (Webber et al., 2005) and AD models (Mielke, 2018), along with differences in pathological severity and behavioral deficits (Chaudry et al., 2022). We scanned under isoflurane and medetomidine anesthesia, which is known to reduce FC in in subcortical areas. Previous research indicates that FC under this anesthesia protocol closely resembles the awake state (Paasonen et al., 2018) and results in the most specific FC data compared to other protocols (Grandjean et al., 2023). However, a recent study comparing CAPs in awake and anesthetised mice brains has revealed significant alterations in topography of coactivations, their occurrences and temporal trajectories (Gutierrez-Barragan et al., 2022). Co-activation of the BFB specifically was shown to be present in the anesthetized state, but less strong compared with the awake state. Another study investigating the effects of anasthesia in mice indicated differences both in static and dynamic FC (Tsurugizawa and Yoshimaru, 2021). Additionally, while our findings provide important insights into early-stage AD, the absence of direct behavioral outcome measures limits our ability to connect these neural changes to specific cognitive or behavioral symptoms. Finally, the use of BOLD signal as indirect measure of neural activity does not capture the entire scope and interplay of neural and hemodynamic alterations in AD. These limitations underscore the need for ongoing studies to further elucidate the complex relationship between hemodynamic, neural changes, and behavioral outcomes in AD pathology.

In conclusion, our study utilizing resting-state CAPs in the TgF344-AD rat model offers novel insights into altered brain functional dynamics at the early stages of Alzheimer's disease (AD). We have validated key differences in the activation patterns and transitions of brain networks, particularly the DMLN and its interaction with the BFB, which are indicative of early disruptions in network dynamics associated with AD. Our findings confirm that these disruptions are most pronounced during the pre-plaque stage and tend to diminish as the disease progresses to the early-plaque stage. The DMLN-BFB and DMLN CAPs provide a deeper understanding of the interplay between different brain regions and their role in disease progression. Furthermore, our results underscore the potential of CAP analysis in predicting genotypic status based on spatial features, especially in early disease stages. This study not only corroborates previous findings on altered functional connectivity in AD but also advances our understanding of the nuanced changes in brain state transitions, offering a more detailed picture of the neurodegenerative process at the onset of plaque formation. Our approach, therefore, holds promise for early detection and better understanding of functional impact of the pathophysiology of AD, that may be relevant to assess the effects of potential therapeutic/disease-modifying interventions.

The data analyzed in this study is subject to the licenses/restrictions. Requests to access the complete MRI dataset generated during this study and the Matlab scripts for CAP analysis of the rsfMRI data should be directed to the corresponding authors.

The animal study was approved by Committee on Animal Care and Use at the University of Antwerp, Belgium (approval number: 2019–06). The study was conducted in accordance with the local legislation and institutional requirements.

SDW: Formal analysis, Methodology, Investigation, Writing – original draft, Writing – review & editing. MvdB: Investigation, Writing – review & editing. GAK: Resources. MHA: Conceptualization, Methodology, Formal analysis, Investigation, Supervision, Writing – original draft, writing – review & editing. MV: Conceptualization, supervision, Funding acquisition, Project administration, Writing – review & editing.