Increased Between-Network Connectivity: A Risk Factor for Tau Elevation and Disease Progression

Abstract

One of the pathologic hallmarks of Alzheimer’s disease (AD) is neurofibrillary tau tangles. Despite our knowledge that tau typically initiates in the medial temporal lobe (MTL), the mechanisms driving tau to spread beyond MTL remain unclear. Emerging evidence reveals distinct patterns of functional connectivity change during aging and preclinical AD: while connectivity within-network decreases, connectivity between-network increases. Building upon increased between-network connectivity, our study hypothesizes that this increase may play a critical role in facilitating tau spread in early stages. We conducted a longitudinal study over two to three years intervals on a cohort of 46 healthy elderly participants (mean age 64.23 ± 3.15 years, 26 females). Subjects were examined clinically and utilizing advanced imaging techniques that included resting-state functional MRI (rs-fMRI), structural magnetic resonance imaging (MRI), and a second-generation positron emission tomography (PET) tau tracer, 18F-MK6240. Through unsupervised agglomerative clustering and increase in between-network connectivity, we successfully identified individuals at increased risk of future tau elevation and AD progression. Our analysis revealed that individuals with increased between-network connectivity are more likely to experience more future tau deposition, entorhinal cortex thinning, and lower selective reminding test (SRT) delayed scores. Additionally, in the limbic network, we found a strong association between tau progression and increased between-network connectivity, which was mainly driven by beta-amyloid (Aβ) positive participants. These findings provide evidence for the hypothesis that an increase in between-network connectivity predicts future tau deposition and AD progression, also enhancing our understanding of AD pathogenesis in the preclinical stages.

1. Introduction

Alzheimer’s disease (AD) is a widely prevalent neurodegenerative disease that causes irreversible decline in multiple cognitive domains [1]. Pathological accumulation of tau tangles is a hallmark of AD that results in neurodegeneration. Recent literature has proposed several mechanisms for the initiation and spread of tau pathology, such as the abnormal transmission of signals between nerve cells, excitotoxicity, and the prion-like spread of pathogenic misfolded proteins [2,3]. Imaging studies have also lent evidence to the hypothesis that tau spread beyond the medial temporal lobe (MTL) is facilitated by beta-amyloid (Aβ) [4–6]. More importantly, recent imaging studies have shown that functional connectivity is crucial in determining the patterns of tau spreading [7–9].

Most AD literature on functional connectivity has traditionally focused on alterations within specific brain networks, often describing decreased connectivity within these networks, particularly in the default mode network (DMN) [10–12]. However, recent evidence has introduced a novel perspective, suggesting that brain functional networks undergo reduced modularity and increased inter-network connectivity during the preclinical stage of AD [13–17]. Both we and others have found evidence supporting increased between-network connectivity during this preclinical phase [13–17], occurring concurrently with decreased within-network connectivity [18–23]. Notably, our recent study and others have shown that increased between-network connectivity is highly associated with Aβ accumulation in cognitively normal individuals [22,23]. We hypothesized that the increase in between-network connectivity may be a compensatory response to Aβ aggregation, with greater integration between networks serving to compensate for decreased network functioning. This allows person to continue to function normally but at the same time provides a pathway for tau to spread throughout the brain. Considering the recent evidence on relationship between functional connectivity and observed pattern of tau spread [3,7,8,18,23,24], we hypothesize that increased between-network connectivity enables tau to spread across connected brain regions.

The association between inter-network connectivity and tau progression remains largely unexplored. In this study, in the preclinical stages of AD, we tested the hypothesis that longitudinal increase in between-network connectivity using resting-state functional magnetic resonance imaging (rs-fMRI) predicts future tau progression, progressive neurodegeneration, and progressive memory dysfunction. We utilized longitudinal scans using the second-generation tau positron emission tomography (PET) tracer (18F-MK6240), structural MRI, and rs-fMRI data from 46 healthy control elderly individuals. Employing an unsupervised agglomerative clustering method, we developed an early-stage AD biomarker capable of predicting increased risk of future tau accumulation and subsequent AD progression in cognitively normal subjects at the preclinical stages. We aim to propose an early biomarker that can effectively predict future pathological tau elevations, thus enhancing our ability to identify at-risk individuals and enabling more timely and effective interventions.

2. Method

2.1. Participants

In this study, we included 46 healthy elderly control individuals with longitudinal data (mean age 64.23 ± 3.15 years, 26 females) that included follow-up intervals ranging from 2 to 3 years from the Northern Manhattan Study of Metabolism and Mind (NOMEM), as part of the ongoing study at Columbia University Irving Medical Center [25]. All participants underwent clinical and neuropsychological exams, rs-fMRI, tau-PET (18F-MK6240), and structural MRI at baseline and follow-up. All individuals were considered cognitively normal based on these examinations and had normal activities of daily living. We then ran a statistical analysis [26] on memory performance based on the selective reminding test (SRT) delayed recall test to identify those with significant memory deficits. The SRT has been validated for its accuracy in detecting subtle memory deficits, such as in verbal memory, and has been used as a sensitive longitudinal measure of changes in memory function [27–29]. We used norms, adjusted for age, sex, and education, based on 360 participants with cross-sectional data from the NOMEM. Out of 46 individuals in this study, we identified two whose SRT delayed recall scores were more than 1.5 standard deviations below these adjusted norms. We received ethical approval for this study through the the institutional review board (IRB) of Columbia University Irving Medical Center.

2.2. Neuroimage acquisition protocols

In baseline and follow-up visits, structural and rs-fMRI scans were performed using the GE Discovery MR750 3-Tesla MRI scanner. The structural MRI was acquired with a TR/TE (time of repetition/time of echo) of 7/2.6 ms, a flip angle of 12°, a matrix size of 256×256, and 176 slices with a thickness of 1 mm. Rs-fMRI scans were performed with a TR/TE of 2000/23 ms, flip angle of 77°, matrix size of 128×128, voxel size of 1.5×1.5×3 mm, and 40 axial slices, each lasting 10 minutes.

In baseline and follow-up visits of Tau-PET scans, participants received the 18F-MK6240 tracer via an intravenous catheter in the arm, with an injection of 185 MBq (5 mCi) ± 20% (maximum volume 10 mL), administered as a single IV bolus within 60 seconds or less (equivalent to 6 secs/mL max). Brain images were acquired starting 90 minutes after the tracer injection, with six 5-minute frames over a duration of 30 minutes.

At the baseline visit to obtain the Aβ-PET scan, participants received the 18F-Florbetaben tracer via an intravenous catheter in the arm, with an injection of 300 MBq (8.1 mCi) ± 20% (maximum volume 10 mL), administered as a single IV bolus within 60 seconds or less (equivalent to 6 secs/mL max). Brain images were acquired starting 50 minutes after the tracer injection with four 5-minute frames over a duration of 20 minutes.

2.3. Neuroimage quantification process

The structural MRI scans were reconstructed using the FreeSurfer [30] (http://surfer.nmr.mgh.harvard.edu) automated segmentation and cortical parcellation software package. FreeSurfer segments the cortex into 33 different gyri/sulci-based regions in each hemisphere according to the Desikan-Killiany atlas [31], with subcortical segmentation and calculates the cortical thickness at each region and voxel/vertex. FreeSurfer longitudinal image processing framework was used for our longitudinal analyses to estimate change in cortical thickness [32]. The change in regional cortical thickness in MTL subregions (entorhinal cortex and parahippocampus) was assessed by subtracting the follow-up cortical thickness from the baseline cortical thickness.

We employed our in-house pipeline for processing rs-fMRI [33], which integrates techniques from FSL (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/) with custom-developed methods. Initially, three first volumes were removed, followed by slice timing correction to address slice acquisition time differences in the raw rs-fMRI time series. Motion parameters were then estimated using rigid-body registration with respect to the first volume. These estimated parameters, along with a geometric distortion field, were combined and applied to the slice-timing-corrected rs-fMRI data, resulting in realigned and distortion-corrected images. To further mitigate residual motion artifacts, scrubbing was performed to exclude frames with high motion (defined as exceeding 0.5 mm framewise displacement). Next, the images went through a band-pass filtering (0.01–0.08 Hz). Physiological artifacts were minimized by regressing out the whole-brain signal, as well as signals from cerebrospinal fluid (CSF) and white matter, which served as nuisance covariates to reduce non-neuronal BOLD fluctuations. Finally, brain parcellation into 200 distinct regions was carried out using the Schaefer atlas, generating a connectivity matrix [34]. This matrix was established by computing the correlation between the time series of fMRI signals for all pairs of brain regions using Pearson’s correlation coefficient.

For PET scans, we used our in-house PET processing pipeline [23,35,36]. First, subjects’ structural MRI image in FreeSurfer space was registered to the same subject’s PET (6 frames for tau and 4 frames for Aβ) composite image using normalized mutual information and six degrees of freedom, resulting in a rigid-body transformation matrix. This matrix transferred all Schaefer atlas and FreeSurfer regional masks to the static PET image space. The standardized uptake value (SUV) was calculated for selected regions and then normalized to the cerebellar gray matter to derive the standardized uptake value ratio (SUVR).

For our analyses, we calculated the mean tau SUVR within each of the seven Schaefer atlas networks in baseline and follow-up scans. These seven networks are visual, somatomotor, dorsal attention, salience/ventral attention, control, limbic, and default mode networks. To account for differences in acquisition time across subjects during scanning, we computed changes in tau SUVR on an annual basis for all subjects.

For Aβ-positivity determination using the baseline scans, we computed a global SUVR cut point of 1.25 for the frontal, parietal, temporal, anterior, posterior cingulate, and precuneus regions [23,35,36].

2.4. Regional tau cut point

Tracking pathologic tau deposition on PET raises challenges when employing a global cut-point. This is due to the heterogeneous progression of tau pathology [37–40], characterized by varying distribution across different regions. Additionally, the 18F-MK6240 tracer, known for its meningeal off-target binding [41,42], distributes non-uniformly throughout the brain. So, we employed our method [43] for regional cut-point determination using normative data from young healthy subjects. We first calculated the normal distributions of the 47 normative healthy young (mean age 29.36 ± 4.73 years, 27 females) subjects’ regional tau SUVR using the Shaefer atlas with 200 brain regions [43]. Normative young subjects are expected to have no pathological tau, allowing us to distinguish between normal and abnormal levels of tau deposition with regional resolution. Then, for each region, we determined the 95th percentile of the fitted normal distribution as regional cut points. This approach allows for region-wise and more precise assessment of tau pathology by accounting for regional differences in tau distribution.

2.5. Between-network connectivity using longitudinal rs-fMRI

As mentioned previously, we utilized the Schaefer atlas to divide the brain into 200 distinct regions, organized into seven primary functional networks. First, the connectivity matrix was constructed by correlating a time series of rs-fMRI signals for all pairs of regions using Pearson’s correlation coefficient. We assessed functional connectivity by utilizing Fisher-z transformed Pearson-moment correlations for every possible pair of regions [44]. Autocorrelations were set to zero, and to capture changes in both correlation and anti-correlation [22] in the connectivity matrix, both positive and negative values were retained [44]. Leveraging the longitudinal dataset, we generated functional connectivity maps at the subject level between these seven networks for baseline and follow-up. Following this, we analyzed subject-specific changes in connectivity patterns over time by determining the difference between follow-up and baseline connectivity maps. To account for differences in acquisition time across subjects during scanning, we computed changes in connectivity on an annual basis for all subjects. Subsequently, within each subject, we assessed the statistical significance of differences in the whole-brain connectivity map between baseline and follow-up using 10,000 random permutations [45]. To do this within each subject we randomly permuted the follow-up connectivity maps and calculated the differences. Then, a threshold of 95% of the fitted normal distribution was selected (p-values < 0.05) to identify significant differences. In the final step, we focused on each network and measured only the average of significant differences between functional networks. Consequently, we obtained seven between-network connectivity metrics as a measure of integration based on seven networks.

2.6. Machine learning implementation

To implement our unsupervised machine learning approach, we first used principal component analysis (PCA) to reduce the dimensionality of our feature set including seven between-network connectivity features (based on seven functional networks). We then ordered the components by the amount of variance they explained in the data, and we selected the components that collectively explained a high percentage (96%) of the variance that effectively represented the dominant patterns in the data [46]. Next, we used an unsupervised agglomerative clustering analysis [47], utilizing the first two principal components (PCs) as input features. This unsupervised clustering method starts by considering each subject as a separate cluster, and pairs of clusters are merged as one moves up the hierarchy based on the distance measure. We used complete linkage with the distance between two clusters defined as the longest distance between two points in each cluster. We incorporated an analysis comparing distance metrics (Euclidean) between subjects for our clustering algorithm (Figure S1). Finally, to identify the optimal number of clusters, we varied the number of clusters from 2 to 10 and used the silhouette criterion to confirm the proper number of clusters. The implementation was conducted using Python.

2.7. Statistical analyses

Differences in continuous measures, such as age, were analyzed using t-tests and analysis of variance (ANOVA) while categorical variables (e.g., gender) were evaluated with chi-squared tests. The clusters’ output was analyzed for significant differences in pathological, neuropsychological, and anatomical characteristics. This analysis involved conducting various statistical tests, including t-tests, ANOVA, and post-hoc Tukey HSD (honestly significant difference) tests with a family-wise error rate (FWER) of 0.05. These tests were employed to assess whether there were statistically significant differences in tau elevations, regional cortical thickness of MTL subregions, and baseline SRT scores between the identified clusters. Finally, multiple regression analysis was utilized to investigate the association between between-network connectivity of the limbic network and tau elevation in the limbic network, with adjustments made for baseline age, gender, and APOE ε4. We used the White test to check the regression analyses for homoscedasticity. All analyses were conducted using Python.

3. Results

3.1. Subject’s characteristics

Table 1 shows the complete demographic, clinical, and behavioral information for the study subjects. Our study data were comprised of 14 Aβ-positive individuals, 15 APOE-positive individuals, and 2 individuals with significant memory decline. In Figures 1a–b, we visually depict the rise in the regional probability of abnormal tau deposition among the 46 longitudinal subjects over a two-to-three-year follow-up period, utilizing our regional cut-point method. Notably, we observe a substantial increase of approximately 20% or more, not only in MTL subregions but also across several neocortical regions (e.g., middle temporal and lingual gyrus). Additionally, around 5–10% of subjects exhibit tau spread even in the frontal lobe. These findings underscore the rationale that a follow-up duration of 2–3 years, even among initially cognitively unimpaired subjects, is sufficient, as a significant proportion demonstrates a noteworthy spread of tau during this period.

Table 1.

Demographics, clinical and behavioral information for the study subjects

	Total Number = 46
Age	64.23 ± 3.15
Sex (Male / Female)	20 / 26
Ethnicity (Hispanic / Non-Hispanic Black / Non-Hispanic White)	35 / 10 / 1
Aβ status (Negative / Positive)	32 / 14
APOE ε4 status (Negative / Positive)	31 / 15
Education (Year)	11.30 ± 4.06
Memory Status (No Decline / Decline)	44 / 2
Selective Reminding Test: Delayed Recall score	5.67 ± 2.56

Figure 1.

Visual representation of a region-wise probabilistic atlas depicting the prevalence of tau pathology by region in a cohort of 46 subjects at (a) baseline and (b) after 2–3 years of follow-up.

3.2. Pathological characteristics of clustered subjects

In this study, we employed cluster analysis to identify distinct subgroups within our cohort, based on between-network connectivity features in seven functional networks. After applying the PCA to our feature set, we utilized the first two principal components, which collectively explained approximately 96% of the total variance, as inputs for the unsupervised agglomerative clustering method. Following this, we selected three clusters based on silhouette results, with a score of 0.456. The resulting clusters were comprised of 4,11, and 31 subjects in three groups. Figure S1 provides a dendrogram illustrating the hierarchical clustering of these groups. As shown in Table 2, there is no significant difference between any demographic or clinical status in the three groups comparison. However, subjects in the significant increase group have relatively higher amyloid-positive status compared to the mild increase group (with marginal significance, p<0.08). Moreover, both individuals with subtle memory deficits are included in the moderate increase group.

Table 2.

Demographics, clinical and behavioral information for the categorized subjects

	Mild increase	Moderate increase	Significant increase	Statistics (three groups comparison)
Total number	11	31	4	-
Age	65.18 ± 2.47	64.03 ± 3.46	63.25 ± 0.82	F = 0.76 , p = 0.47
Sex (Male / Female)	6 / 5	12 / 19	2 / 2	X2 = 0.9, p = 0.64
Ethnicity (Hispanic / Non-Hispanic Black / Non-Hispanic White)	11 / 0 / 0	20 / 10 / 1	4 / 0 / 0	-
Aβ status (Negative / Positive)	10 / 1	20 / 11	2 / 2	X2 = 2.87, p = 0.23
APOE ε4 status (Negative / Positive)	8 / 3	21 / 10	2 / 2	X2 = 0.69, p = 0.70
Education (Year)	10.63 ± 4.18	10.7 ± 4.02	9 ± 5.19	F = 0.30, p = 0.74
Memory Status (No Decline / Decline)	11 / 0	29 / 2	4 / 0	-

In Figure 2, we used the boxplot to characterize the tau pathological annual changes between these clustered groups based on the average annual change in tau across all networks. Through the ANOVA test, we identified a significant difference among the three subject groups (p-value<0.0001, F-value=13.27). Subsequent post-hoc Tukey HSD testing revealed that the significant increase in between-network connectivity group exhibited significantly higher tau elevation longitudinally (p-value<0.0003) compared to both the moderate and mild increase connectivity groups (p-value<0.002). Additionally, the moderate increase group displayed significantly higher tau elevation than the mild increase group (p-value< 0.03).

Figure 2.

Unsupervised clustering of our cohort identified three clusters (mild increase, moderate increase, and significant increase). These clusters showed differences in longitudinal change in tau. * Survived after FWER correction.

3.3. Anatomical and neuropsychological characteristics of clustered subjects

After characterizing the pathology of these clustered subjects, we explored their distinct anatomical and psychological characteristics. As depicted in Figure 3a, our analysis reveals that even with our small sample the significant increase group exhibits a statistically significant reduction (p-value<0.02) in the baseline SRT delayed recall scores compared to the moderate increase group. Moreover, the significant increase group demonstrates marginally significantly lower scores (p-value<0.06) compared to the mild increase group, suggesting varying degrees of cognitive impairment across these categories.

Figure 3.

Unsupervised clustering identified three clusters (mild increase, moderate increase, and significant increase). These clusters showed differences in a) baseline selective reminding test (delayed recall) score, and b) longitudinal change in cortical thickness of the entorhinal cortex. * Survived after FWER correction.

We also assessed the anatomical characteristics of these clusters by investigating the longitudinal cortical thickness change (baseline – follow-up) in MTL subregions, including the entorhinal cortex and parahippocampus. We did not annualize the cortical thickness measurements because changes in cortical thickness over a two to three-year period may not provide sufficient resolution to detect meaningful annualized differences. Although no significant differences were found in parahippocampus thickness among the clusters, as illustrated in Figure 3b, the significant increase group displayed a marginally significant rate of longitudinal cortical thinning compared to both the moderate increase and mild increase groups, with p-values<0.05 and <0.07, respectively. It is important to note that the significant increase group exhibited negative values, indicating consistent longitudinal cortical thinning in the subjects of this group. These findings underscore the potential clinical relevance of the cluster outputs, as evidenced by the observed cognitive performance differences across groups and cortical atrophy, particularly within the significant increase group.

3.4. Between-network connectivity associated with limbic network tau elevation

Finally, we present evidence from our study revealing the association between between-network connectivity and longitudinal accumulation of tau, particularly focusing on the limbic network. Our investigation aims to illustrate the underlying mechanisms driving tau pathology progression, with a particular focus on how the connectivity of the limbic network with the other brain networks facilitates the spread of tau. Tau tangles often initiate formation within regions of the limbic network, particularly the entorhinal cortex. Figure 4a illustrates a significant association (p-value<0.003, r=0.55) between between-network connectivity and longitudinal tau accumulation within the limbic network. Moreover, in Figure 4b, we compare the association between the annual change in between-network connectivity and the annual change in tau accumulation in the limbic network for Aβ-positive and Aβ-negative subjects separately. While Aβ-positive subjects exhibited a strong significant association (p-value<0.008, r=0.71), the Aβ-negative subjects showed a marginally significant association (p-value<0.048, r=0.38), which did not survive after FWER. This demonstrates the significant impact of Aβ-positivity in this relationship. We checked our linear regression models for homoscedasticity (p-value>0.942), suggesting that residuals are uniformly scattered.

Figure 4.

a) Association between the annual change of between-network connectivity and the annual change of tau uptake in the limbic network. b) Association between the annual change of between-network connectivity and the annual change of tau uptake in the limbic network for Aβ-positive and Aβ-negative subjects. * Survived after FWER correction.

It is noteworthy that we also tested the association between baseline tau uptake and between-network connectivity of the limbic network. As shown in Figure S2, our analysis reveals that higher baseline tau in the limbic areas is not significantly associated with greater between-network connectivity of the limbic network.

4. Discussion

This study aimed to investigate the relationship between alterations in between-network connectivity and the risk of future tau spread and AD progression. Initially, by using between-network connectivity as a feature and unsupervised agglomerative clustering, we demonstrated that the three groups of clustered subjects exhibited significantly different levels of longitudinal tau accumulation. Additionally, our investigation into MTL subregions’ cortical thickness and performance on the SRT revealed that significant increase subjects had significantly lower recall scores and greater entorhinal thinning. Finally, we found a strong association between elevated tau levels in the limbic network and increased between-network connectivity which was mainly driven by Aβ-positive subjects.

Previous research has reported that older adults exhibit lower network segregation, reduced modularity, decreased efficiency, and diminished hub function [48]. Several recent studies have documented an increase in between-network connectivity during the preclinical and mild stages of AD [48–50]. It has been shown that between-network connectivity is associated with several functional aspects of the brain such as memory and cognitive control [51–53] and more accurately correlates with patient symptoms at the early stages of the disease [19]. The broader exploration based on between-network connectivity is pivotal as cognitive functions often rely on the collaboration of multiple brain networks [54]. In preclinical stages the organizational shift towards increased between-network connectivity in older adults may signify a compensatory mechanism [55] of the brain in response to pathologies [20,22,56] (especially Aβ), marked by more extensive recruitment of neural resources.

The role of Aβ in accelerating the spread of tau [43] through neuronal communication pathways has been shown in previous studies [57,58]. Also, it has been hypothesized that the cortical Aβ specifically in the posteromedial cortex may drive local MTL circuit hyperexcitability and further accumulations [59]. In our previous research, we established that even sub-threshold levels of Aβ accumulation can induce an increase in cortical thickness within the MTL [35], suggesting the presence of compensatory mechanisms in the brain at preclinical stages. Our recent study also found that increased between-network connectivity is highly associated with the buildup of Aβ [23]. Supporting our findings, another study [22] demonstrated that the increase in between-network connectivity induced by Aβ was driven by a reduction in anti-correlation and increased correlation in DMN. So, Aβ may cause disruptions to the local excitatory-inhibitory balance in the DMN, which triggers hyperexcitability in the MTL and results in tau accumulation [60]. It has been also reported that increased connectivity between the posterior DMN and high-connectivity hubs (such as MTL sub-regions) is linked to amyloid accumulation [61] and leads to the amyloid-associated compensatory response in brain networks [62]. While this inter-network reorganization may initially act as a compensatory mechanism to protect neuronal function, it might inadvertently create conditions that induce the spread of tau.

Biologically Aβ interacts with a variety of receptors on the surface of both neuronal and glial cells, initiating signaling cascades that alter both gene transcription and protein expression [63–65], leading to neuronal hyperactivation and neurotoxicity [66,67]. As shown in Figure 5, Aβ can also lead to activated microglia [68], which can also prune synapses and remodel network circuitry [69]. Aβ can even directly bind to different components of cell membranes [67], leading to wide-scale brain network reorganization. In this milieu of neuronal hyperactivity and glial cell hyperreactivity, neurons are stimulated to secrete more tau, which can spread trans-neuronally/trans-synaptically or even transfer to adjacent neurons via the extracellular space [70,71]. Because of the Aβ-mediated increase in network connectivity, tau has a pathway along which to spread [72]. As a result, even though synaptic pruning and network reorganization may be a compensatory response to preserve neuronal function [20,22,56], these processes may set the stage for the further spread of tau. This is supported by experimental evidence that Aβ injected into the cortex and hippocampus can induce tau in synaptically connected, but distant brain regions [72–74].

Figure 5.

The proposed biological mechanism by which cortical beta-amyloid deposition (red) promotes neuronal hyperconnectivity, possibly mediated by activated microglia, creates pathways for the transneuronal spread of tau (blue).

As shown in this study, the increase in connectivity across networks in the brain may provide a pathway for the spread of tau out of the MTL. This can be potentially explained by the numerous connections between the MTL and cortical brain regions [75], leading to neuronal vulnerability of the MTL region to AD pathologies [76]. We previously investigated whether there could be associations between Aβ and tau across spatially distinct regions [36], and we found strong remote associations between early-stage tau accumulation in the MTL and cortical Aβ across different brain regions. The critical role of functional connectivity in the accumulation and spread of tau has been shown by several previous studies [7,8,24]. The variability in functional connected brain regions to tau spread could also explain the heterogeneity of observed patterns of tau spread [9,77]. As demonstrated in an in-vivo study, resting-state connectivity alone accounted for 21.8% to 39.2% (across three cohorts) of the variability in regional tau accumulations in Aβ-positive individuals [78]. More importantly, based on a recent postmortem study, a rare soluble species of high-molecular-weight (HMW) hyperphosphorylated tau is capable of intercellular spread [79] which could be triggered by neuronal connectivity. Taken together, by using the between-network connectivity, this study underscored the potential benefits associated with finding early biomarkers or risk factors for AD pathologies during preclinical stages in future clinical trials and AD research.

The main limitation of our study was the number of longitudinal samples of elderly healthy individuals who underwent imaging with the second-generation tau tracer (F18-MK6240). This led to a relatively limited sample size which may have impacted the statistical power of certain analyses, particularly the ANOVA, post-hoc, chi-square tests in anatomical, neuropsychological, and clinical characteristics. So, it is essential to replicate our findings using independent datasets with a larger longitudinal sample size. Additionally, second limitation is the inherent variability in rs-fMRI that can affect the longitudinal connectivity measurements. This variability may stem from several factors, such as physiological fluctuations, and subject-related factors such as varying levels of alertness during scanning sessions. However, to mitigate factors, the rs-fMRI scans at both baseline and follow-up were performed using the same MRI scanner, acquisition protocol, and instructions.

Significance statement.

This study characterized a mechanism by which between-network connectivity induces the spread of tau, thereby facilitating the prediction of individuals at risk of Alzheimer’s disease progression during the preclinical stages.

Data availability

The data for this project are confidential but may be obtained with Data Use Agreements with the Columbia University Irving Medical Center. It can take some weeks to negotiate data use agreements and gain access to the data. The author will assist with any reasonable replication attempts for the following publication.