Patterns of amyloid accumulation in amyloid negative cases

ABSTRACT


INTRODUCTION
Amyloid-β (Aβ) and hyperphosphorylated tau are core pathological hallmarks of Alzheimer's disease (AD) (Jack et al., 2018). Early accumulation of these proteins can be detected 10 to 20 years J o u r n a l P r e -p r o o f before the onset of dementia (Bateman et al., 2012;Rowe et al., 2010;Villemagne et al., 2013).
Until recently, the presence of AD pathology could be detected only post-mortem through autopsy studies. Currently, the development of specific PET tracers enables in vivo assessment of amyloid or tau deposition, providing valuable information on their topography. In clinical practice, the result of amyloid-PET imaging is usually dichotomized into positive or negative based on visual assessment performed by a nuclear medicine physician. Amyloid-PET scans contain more information than the one summarized by a dichotomous classification, as shown by the use of global and regional semi-quantitative measures, but their use is commonly restricted to research settings (Pemberton et al., 2022). Importantly, cortical amyloid staging models have shown that regional abnormality occurs before global positivity, associated with a different individual risk of cognitive decline based on the extent of amyloid burden.
Previous PET studies developed models for staging cortical amyloid deposition. They have consistently found that amyloid deposition in cortical regions precedes deposition in the striatum and that the accumulation in medial regions precedes that in the lateral ones (Fantoni et al., 2020).
However, the chronological order of the regions accumulating cortical amyloid was variable across studies. For example, Grothe et al 2017, has shown that amyloid accumulation begins in temporobasal and frontomedial areas, followed by the remaining associative neocortex, primary sensorymotor areas and the medial temporal lobe, and finally the striatum (Grothe et al., 2017). Moreover, Collij et al. (2020) found that SUVr abnormality was most frequently observed in cingulate, followed by orbitofrontal, precuneal, and insular cortices and then the associative, temporal, and occipital regions (Collij et al., 2020). These studies assumed that the trajectory of amyloid spread is homogeneous across individuals. However, a large clinico-pathological variability is observed in AD and could suggest the existence of different accumulation patterns (Frisoni et al., 2022).
To the best of our knowledge, only one study has explored possible subtypes of Aβ accumulation using a data driven approach, resulting in three subtypes associated with different demographic, J o u r n a l P r e -p r o o f genetic and biomarkers profiles: frontal, parietal and occipital (Collij et al., 2021). In this study, a substantial proportion of participants (i.e. 36%) was amyloid positive and guided the definition of subtypes across the continuum from negativity to positivity.
The aim of the present study is to assess whether distinct patterns of amyloid accumulation can be detected (by using a data-driven approach) investigating exclusively patients with a sub-threshold amyloid burden (centiloid < 12), i.e. in the negative range for a global assessment and to evaluate the association with different clinical, imaging features and cognitive trajectories.

Population
For the present study, we included 151 participants recruited consecutively at the Geneva Memory Center (GMC) and 62 Zurich. The GMC cohort consists of individuals ranging from cognitively unimpaired participants (CU, either healthy volunteers or subjective cognitive decline) to patients with mild cognitive impairment (MCI) and dementia. From the ongoing Zurich cohort 62 individuals completed baseline assessments with available 18-F-Flutemetamol and 18F-Flortaucipir PET were included. The Zurich cohort is recruited via newspaper ads and consists of healthy cognitively unimpaired individuals and participants with mild cognitive impairment. Only baseline evaluations available at the cut-off date for the analysis were transferred to Geneva for this analysis. We selected only the participants with negative Aβ scan (centiloid value lower than 12) (La Joie et al., 2019).
The inclusion criteria for the present investigation were: (i) availability of negative Aβ PET (18F-Florbetapir or 18F-Flutemetamol); (ii) T1 MRI acquisitions within 1 year from the PET scan and (ii) (iii) clinical and neuropsychological assessments within 1 year from the PET scan.
J o u r n a l P r e -p r o o f A subset of individuals underwent also 18F-Flortaucipir PET (total N=123; Geneva Florbetapir N=32, Geneva Flutemetamol N=29, Zurich cohort N=62) within 18 months from the Aβ PET and 65 of them underwent also longitudinal neuropsychological evaluation (average follow-up 2.7 year ± 1.1).
The studies were approved by local ethic committees and have been conducted in accordance with the principles of the Declaration of Helsinki and the International Conference on Harmonization Good Clinical Practice. Each participant provided a voluntary written informed consent for the participation in the retrospective study.
PET images were acquired using 18F-Flutemetamol, 18F-Florbetapir and 18F-Flortaucipir at the Nuclear medicine and molecular imaging division of the Geneva University Hospitals using a Siemens Biograph mCT or Vision PET scanner, using tracer specific protocols. In particular, 18Fflorbetapir images were acquired 50 min after the intravenous administration of approxima- MBq of 18F-AV1451 was administered and images acquired 80 minutes after injection for 40 minutes. For all PET images, time-of-flight algorithms with the necessary corrections were applied.

PET and T1 preprocessing and SUVr extraction
T1-MRIs images were segmented using FreeSurfer 6.0.0 (Fischl et al., 2004) and single-subject regional grey matter was extracted from the Desikian Killany (DKT) parcellation (Desikan et al., 2006). Aβ and Tau PET images were linearly co-registered to their corresponding T1-MRI scan, and spatially normalized to the Montreal Neurological Institute space using ANTSRegistration (Avants et al., 2011). Aβ and Tau PET images were intensity normalized using whole cerebellum and cerebellar grey matter as reference regions, respectively.

J o u r n a l P r e -p r o o f
In evaluating SUVr we considered the partial volume effect (PVE), as it could generate artifacts in regional quantification. In particular, due to the limited PET resolution, an ambiguity could arise in evaluating the uptake of ROI's border voxels, since it is not possible to discriminate whether the uptake comes from the ROI itself or from the voxels just outside the ROI's border. Then, instead of simply calculating the average uptake on each ROI, we calculated the regional quantification using a weighted average: each regional mask was smoothed with a three-dimensional isotropic Gaussian (FWMH 3.5 mm). Due to the Gaussian smoothing, voxels on the border had a lower weight in SUVr calculation than the innermost voxels, so the PV effect has been minimized.
Global Aβ SUVr was computed as the volume-weighted average of the 33 ROIs of the DKT atlas.
Global Tau SUVr was computed as the volume-weighted average of the entorhinal cortex, amygdala, lateral occipital cortex, and inferior temporal cortex (Mishra et al., 2017).

Data Driven Analysis
All the analyses described in the next sections were performed separately on the three subgroups: Geneva Florbetapir (GFP), Geneva Flutemetamol (GFM) and Zurich Flutemetamol (ZFM).

Clustering
To investigate the presence of different Aβ patterns, a k-means clustering with the Aβ SUVr of the 33 selected ROI using Pearson correlation distance has been applied ( Figure 1). Based on silhouette plot, we choose an optimal number of cluster equal to 2. To find the best clustering solution, the algorithm was repeated 300 times, and the result with the lowest within-cluster variance was identified (Lisboa et al., 2013).
Once we obtained the two clusters, we aimed to verify if there were regions in which the Aβ accumulation was significantly higher or lower in one of the two clusters with respect to the other one. Let A 1r and A 2r be the medians of the amyloid SUVr of the r-th ROI in the first (labeled by 1) J o u r n a l P r e -p r o o f and second (labeled by 2) clusters respectively. We used, for each ROI, two one-tailed Mann Whitney tests, assuming as null hypothesis A 1r ≤ A 2r for the first test and A 1r ≥ A 2r for the second one ( Figure 1).
ROIs where the null hypothesis A 1r ≤ A 2r is rejected will be called positively significant (S+), while ROIs where the null hypothesis A 1r ≥ A 2r is rejected will be called negatively significant (S-). Thus, the amyloid load in S+ ROIs are greater in the first cluster with respect to the second one, and vice versa for S-ROIs.
In the present investigation, we could not test on longitudinal data whether the two clusters we obtained represent two temporal stages of the same accumulation pattern (stage hypothesis) or they describe two different accumulation patterns (pattern hypothesis). Therefore, to answer this question, we introduced the following working hypothesis, which we will call the amyloid nondecreasing constraint (Figure 1): the accumulation of amyloid never decreases over time, neither regionally nor globally. A direct consequence of the non-decreasing constraint is that, given two temporal stages of the same pattern, all the ROIs of the later stage must consistently show an accumulation greater than or equal to the ROIs of the earlier stage. This means that:  the presence of S+ or S-ROIs only is compatible with the stage hypothesis, as it does not violate the amyloid non-decreasing constraint  the simultaneous presence of S + and S-ROIs violates the amyloid non-decreasing constraint, therefore it is not compatible (p-value 0.005) with the stage hypothesis, leading us to accept the pattern hypothesis. Furthermore, the S+ ROIs define the first pattern, while the S-defines the second one.

J o u r n a l P r e -p r o o f
In order to evaluate if the S+ and S-ROIs obtained as described in section 2.3.1 were stable with respect to variations of the sample used, we performed a validation of the S+ and S-ROIs through a bootstrap analysis. From the sample considered, we randomly extracted 1000 bootstrap samples and, for each one, we clustered and identified the significant S+ and S-regions using the same methods presented in section 2.3.1. Thus, 1000 lists of S+ regions and S-regions were obtained, one for each bootstrap sample. From these lists we extracted the percentage of times that each of the 33 considered ROIs resulted to be of type S+ or S-. ROIs that at least 50% of times are S+ (respectively S-) will be considered as positively (resp. negatively) significant and stable with respect to bootstrap analysis; such ROIs will be named SB+ (resp. SB-).

Statistical Analyses
Statistical analyses were performed within the whole population. We compared demographics, cognitive, and imaging differences between the two clusters. Numerical data (age, education, Mini Mental State Examination (MMSE), Free and Cued Selective Reminding Test (FCSRT) total free recall immediate recall, FCSRT delayed free recall, FCSRT delayed total recall, hippocampal volume, Global Amyloid SUVr and Global Tau SUVr) were tested using a nonparametric two-tailed Mann-Whitney test (p-value 0.05). Categorical variables (i.e. gender) were tested using a Fisher's exact test. Furthermore, we compared regional Tau SUVr between clusters. We used an analogous approach to the one presented in section 2.3.1 to determine the S+ and S-regions. We performed, for each ROI, two one-tailed Mann-Whitney tests (p-value 0.005) using T 1r ≤ T 2r and T 1r ≥ T 2r as null hypotheses, where T 1r and T 2r are the medians of the Tau SUVr in the r-th region related to the first and second clusters respectively. ROIs where T 1r ≤ T 2r is rejected are the ones where T 1r is significantly greater than T r2 , while ROIs where T 1r ≥ T 2r is rejected are the ones where T 2r is significantly greater than T r1 . This allowed us to find any significant regional differences between amyloid clusters. To investigate the effect of the cluster on cognitive changes over time, linear J o u r n a l P r e -p r o o f mixed effect models were performed with Mini Mental State Examination (MMSE), as dependent variables and cluster, time (years), cluster*time interaction as independent variables. Random intercept and random slope were considered to account for individual differences at baseline as well as for individual change over time. The model has been adjusted by age.

Sample description and k-means clustering
The three cohorts studied (GFP, GFM and ZFM) consisted of 45, 44, and 62 individuals, respectively. Demographics (age, sex, years of education), cognitive (Mini Mental State Exam), and imaging features (hippocampal volume, global amyloid and tau SUVr) data are provided for all the three cohorts. Cohorts' differences in terms of demographics, cognitive and imaging features are shown in Table 1. The three cohorts do not differ in terms of age, gender and for imaging features, despite a difference in in education, MMSE and syndromic composition, with the Zurich cohort having a higher proportion of CU individuals.
As described in section 2.3.1, we performed a k-means clustering independently over each cohort, resulting in two clusters (Table 2, Figure 2) per cohort with the following sample size: GFP Cluster 1 N= 25, Cluster 2 N= 20; GFM Cluster 1 N = 23, Cluster 2 N=21 for; ZFM Cluster 1 N= 34, Cluster 2 N=28. Table 2 shows the demographics, cognitive and biomarker features of these two clusters. No differences in demographics or cognitive features were observed. However, global Tau SUVr was higher in the Temporal predominant (TP) cluster than in the Cingulate predominant (CP) (Figure 4a and 4b).

Pattern (S+ and S-ROIs type)
J o u r n a l P r e -p r o o f Using the analysis described in section 2.3.1 we obtained both S+ and S-ROIs type in each cohort.
In the GFP cohort we found 7 S+ (higher in the CP subtype) and 3 S-(higher in the TP subtype) ROIs, in the GFM we found 2 S+ and 8 S-ROIs and in the ZFM we found 2 S+ and 10 S-ROIs. The S+ and S-ROIs as well as the two accumulation subtypes in each cohort are shown in Figure 2a and 2b.
The S+ and S-ROIs are reported in table 3 for the three cohorts in the format "ROI -p-value".

ROIs validation
The bootstrap analysis (as presented in section 2.3.2) resulted in SB+ and SB-ROIs reported in table 3 (in the format "ROI -percentage of times that they are significant"). Four SB regions, namely the fusiform, inferior temporal cortex (SB-, thus higher in TP), and the posterior cingulate and transverse temporal cortex (SB+, thus higher in CP) significantly different between the two clusters across the three cohorts. Figure 3 shows the TP and CP ROIs that survived the bootstrapping.
Moreover, no inconsistent behaviors from bootstrap analysis were recognized: there are no SB+ regions of one cohort that are SB-in another cohort and vice versa; and, focusing on single cohorts, it never happens that the SB+ regions have a counterpart of SB-frequencies and vice versa.

Regional tau accumulation related to amyloid-driven clustering
The statistical comparison of the two clusters populations in both cohorts shows a significant global Tau SUVr difference, with a global deposition is significantly greater in cluster TP than in CP (p-value < 0.05, Figure 4a). When comparing regional Tau SUVr deposition between the two clusters, participants of the TP subtype showed a greater Tau load than participants of the CP subtype in the fusiform and in the inferior temporal ROIs. Regional Tau differences are shown in Figure 4b. The linear mixed model revealed a trend for a cognitive decline in the TP (N=30) compared to the CP (N=35), yet not statistically significant (TP β=-0.39 vs CP β=-0.07, p-value=0.079, Figure 5).

DISCUSSION
In the present investigation, we have identified two different clusters of early regional amyloid accumulation. One cluster had higher amyloid deposition in medial and inferior temporal areas, named TP cluster, while another one had higher amyloid in the posterior cingulate cortex (CP cluster). Importantly, this observation was confirmed in two independent cohorts, studied in Geneva and Zurich, respectively, and two different tracers (Flutemetamol and Florbetapir).
In line with our data, previous studies have shown that the cingulate could be one of the first regions of amyloid deposition and further propagation (Collij et al., 2020;Farrell et al., 2018;Palmqvist et al., 2017). Other studies have also shown that medial frontal, parietal and temporal areas might be the trigger regions of amyloid deposition (Braak et al., 2003;Villeneuve et al., 2015). The main reason of these inconsistent results is presumably due to study design differences in terms of studied population and methodological approach.
To the best of our knowledge, only one study has assessed in vivo subtypes of amyloid accumulation with a data driven approach (Collij et al., 2021). Collij et al. have identified three subtypes, namely Frontal, Parietal and Occipital according to the earliest regions to become abnormal. Our clusters are overall consistent with these results, with our CP cluster overlapping with the Frontal and Parietal subtypes and the TP cluster matching with the Occipital subtype. In particular, the regions with the highest deposition in the CP cluster are the posterior cingulate, one of the first regions becoming abnormal in the Parietal subtype, and the superior and middle frontal regions, re-J o u r n a l P r e -p r o o f ported as early abnormal in the early phases in the Frontal subtype. The TP cluster may correspond to the Occipital subtype, characterized by the lateral occipital and temporal cortices involvement.
Our two clusters did not differ in terms of demographic, clinical and cognitive features. However, the TP cluster showed higher tau deposition than the CP cluster, in particular in the fusiform and inferior temporal regions. Tau deposition is closely associated with neurodegeneration, and both are directly linked to cognitive decline. Based on the model of Hanseeuw et al. (2019), Aβ and tau in inferior temporal neocortex interacted and potentiated tauopathy and cognitive decline (Hanseeuw et al., 2019). Indeed, the longitudinal analysis on the Geneva dataset revealed steeper cognitive decline in the TP cluster. The interaction cluster*time was not statistically significant, possibly because of the small sample size. However, we can speculate that individuals in the TP cluster have poorer prognosis than the CP. The increased 18F-Flortaucipir uptake in mesial temporal regions in the TP cluster might reflect AD-related tau pathology but also primary age-related tauopathy (PART), a common pathological finding in the elderly. In favor of an AD-related tau pathology are the neuropathological data available, showing that Flortaucipir in these regions is correlated to tau pathology and is on the contrary low or negligible in the case of PART (Kotari et al., 2023;Lowe et al., 2020), and the notion that PART is typically not associated with cognitive decline (Iida et al., 2021), while we observe a decline, albeit not significant in this small size cohort.

Limitations
The main limitations of the present study are: (i) The lack of longitudinal imaging data prevents us from studying the evolution of biomarkers and confirming whether this population is homogeneous and progressing towards AD. It is possible that this is a mixed population comprising both normal aging individuals and those with early AD pathological changes; (ii) the relatively small sample size per tracer and per center and different settings, different acquisition and reconstruction protocols across cohorts may impact the generalizability of the results to the whole popula-J o u r n a l P r e -p r o o f tion. The results presented in this work should be considered suggestive only, pending further confirmation using larger datasets and diverse methodological approaches; (iii) the missing longitudinal evaluation for the Zurich cohort; (iv) lack of APOE status that could influence the amyloid accumulation, its topography and the disease progression.

CONCLUSIONS
In our study we observed two amyloid patterns in the earliest phases of amyloid accumulation, differently prone to tau pathology. This study confirms the heterogeneity of the amyloid deposition and allows speculating that a prognostic stratification might be possible in the earliest disease phases. However, these results should be replicated in independent longitudinal cohorts. ^between Geneva Flutemetamol and Zurich.      J o u r n a l P r e -p r o o f Figure 4b shows tau mean SUVr difference by cluster in the three cohorts. Mean SUVr difference is only highlighted in ROIs where the SUVr median of the Cingulate Predominant cluster is significantly higher (red) or lower (blue) than the SUVr median of the Temporal Predominant cluster.

Tables and Figures
The colorbar represents the mean difference, thresholded at p-value <0.005. Trajectories of global cognition by clusters (TP, CP). The X-axis represents time in years. Lines represent predicted mean trajectories obtained from linear mixed models. Shaded areas indicates and 95% confidence intervals. Higher values indicate better performance. We observed a trend for a cognitive decline in the TP (N=30) compared to the CP (N=35), yet not statistically significant (TP β=-0.39 vs CP β=-0.07, p-value=0.078).