Synaptic density patterns in early Alzheimer’s disease assessed by independent component analysis

Abstract Synaptic loss is a primary pathology in Alzheimer’s disease and correlates best with cognitive impairment as found in post-mortem studies. Previously, we observed in vivo reductions of synaptic density with [11C]UCB-J PET (radiotracer for synaptic vesicle protein 2A) throughout the neocortex and medial temporal brain regions in early Alzheimer’s disease. In this study, we applied independent component analysis to synaptic vesicle protein 2A-PET data to identify brain networks associated with cognitive deficits in Alzheimer’s disease in a blinded data-driven manner. [11C]UCB-J binding to synaptic vesicle protein 2A was measured in 38 Alzheimer’s disease (24 mild Alzheimer’s disease dementia and 14 mild cognitive impairment) and 19 cognitively normal participants. [11C]UCB-J distribution volume ratio values were calculated with a whole cerebellum reference region. Principal components analysis was first used to extract 18 independent components to which independent component analysis was then applied. Subject loading weights per pattern were compared between groups using Kruskal–Wallis tests. Spearman’s rank correlations were used to assess relationships between loading weights and measures of cognitive and functional performance: Logical Memory II, Rey Auditory Verbal Learning Test—long delay, Clinical Dementia Rating sum of boxes and Mini-Mental State Examination. We observed significant differences in loading weights among cognitively normal, mild cognitive impairment and mild Alzheimer’s disease dementia groups in 5 of the 18 independent components, as determined by Kruskal–Wallis tests. Only Patterns 1 and 2 demonstrated significant differences in group loading weights after correction for multiple comparisons. Excluding the cognitively normal group, we observed significant correlations between the loading weights for Pattern 1 (left temporal cortex and the cingulate gyrus) and Clinical Dementia Rating sum of boxes (r = −0.54, P = 0.0019), Mini-Mental State Examination (r = 0.48, P = 0.0055) and Logical Memory II score (r = 0.44, P = 0.013). For Pattern 2 (temporal cortices), significant associations were demonstrated between its loading weights and Logical Memory II score (r = 0.34, P = 0.0384). Following false discovery rate correction, only the relationship between the Pattern 1 loading weights with Clinical Dementia Rating sum of boxes (r = −0.54, P = 0.0019) and Mini-Mental State Examination (r = 0.48, P = 0.0055) remained statistically significant. We demonstrated that independent component analysis could define coherent spatial patterns of synaptic density. Furthermore, commonly used measures of cognitive performance correlated significantly with loading weights for two patterns within only the mild cognitive impairment/mild Alzheimer’s disease dementia group. This study leverages data-centric approaches to augment the conventional region-of-interest–based methods, revealing distinct patterns that differentiate between mild cognitive impairment and mild Alzheimer’s disease dementia, marking a significant advancement in the field.


Introduction
2][3][4] However, these findings were largely derived from autopsy studies conducted during later stages of disease.][7][8][9][10] SV2A is ubiquitously expressed in all pre-synaptic axon terminals, making it a potential in vivo marker for synaptic density. 7,11,12Previously, [ 11 C]UCB-J PET studies reported decreased SV2A binding in medial temporal and neocortical regions in early Alzheimer's disease compared with cognitively normal (CN) participants. 13,144][15] A study by Mecca et al. 16 demonstrated a significant positive association between global synaptic density and global cognition and performance in five individual cognitive domains using pre-defined regions of interest (ROIs) within Alzheimer's disease participants.
Conventional PET image analysis involves either a prioriestablished ROIs typically based on anatomical features or voxel-based analysis whereby spatially standardized images are analysed on the voxel level.However, more sophisticated analytical approaches exist (e.g.cluster analysis), which may be more sensitive and thus more appropriate, especially in Alzheimer's disease.The method applied here, source-based morphometry, is a biomedical image analysis approach, which uses a data-driven method, spatial independent component analysis (ICA), to decompose mixed signal image information into maximally independent sources (images) of common variance.This method determines these variance patterns and subject-specific loading weights without information about subject group membership.Statistical analysis can then be used to identify which sources distinguish healthy controls from those with pathological conditions. 17This has previously been used to investigate spatial patterns of 18 F-fluorodeoxyglucose (FDG) PET data, 18,19 grey matter atrophy related to healthy ageing, 20 schizophrenia 17 and multiple sclerosis, 21 as well as to identify independent sources of synaptic density in healthy controls. 22e aimed to investigate whether changes in synaptic density in Alzheimer's disease follow specific patterns using this datadriven approach.Additionally, we investigated whether these patterns were associated with clinical and cognitive measures.

Subjects and study design
Participants provided written informed consent prior to participation as approved by the Yale University Human Investigation Committee.The study cohort consisted of 19 CN participants and 38 participants with Alzheimer's disease [14 with mild cognitive impairment (MCI) and 24 with mild Alzheimer's disease dementia] aged 50-85 years.A full description regarding recruitment and eligibility of the cohort can be found here. 14In brief, recruited individuals with mild Alzheimer's disease dementia were required to meet diagnostic criteria for probable dementia, 23 have a global Clinical Dementia Rating (CDR) score of 0.5-1.0 and have a Mini-Mental State Examination (MMSE) score <26.Individuals with MCI were included if they met diagnostic criteria for amnestic MCI, 24 had a global CDR score of 0.5, and had an MMSE score of 24-30.Mild Alzheimer's disease dementia and MCI participants were required to demonstrate impairments in episodic memory, as measured by a Logical Memory II (LMII) score 1.5 SD below an education-adjusted norm.CN participants were required to have a CDR score of 0, an MMSE score >26 and a normal education adjusted LMII score.The Rey Auditory Verbal Learning Test (RAVLT) was also administered, and the long delay was used to generate an episodic memory score.The presence of amyloid-β accumulation was determined as previously described: 13,14 all participants received a PET scan with [ 11 C]Pittsburgh Compound B ([ 11 C]PiB) and were required to be negative for CN participants and positive for MCI/mild Alzheimer's disease dementia participants.Subjects and imaging data used in the current work were highly overlapping with the cohort described in the previous study. 14

Brain imaging
Participants underwent T 1 -weighted MRI scans on a 3-T Trio (Siemens).PET scans were acquired on the High-Resolution Research Tomograph (Siemens), which acquires 207 slices (1.2 mm slice separation) with a reconstructed image resolution (full width at half maximum) of ∼3 mm.Prior to each PET scan, a 6-min transmission scan was performed for attenuation correction.PET data were acquired in list mode.Dynamic scans were acquired for 60 min after i.v.bolus administration of [ 11 C] UCB-J (572.2 ± 188.9 MBq, min 157.6-max761.8 MBq).Dynamic PET data were reconstructed with corrections for attenuation, normalization, scatter, randoms and dead time using the motion-compensation OSEM list-mode algorithm for resolution-recovery reconstruction algorithm. 25vent-by-event motion correction was included in the reconstruction by tracking motion with a Polaris Vicra optical tracking system (NDI Systems) using reflectors mounted on a swim cap worn by subject.
For each subject, motion-corrected dynamic PET data were co-registered to an early summed PET image (0-10 min after [ 11 C]UCB-J administration) using a six-parameter mutual information algorithm (FSL-FLIRT).The summed PET image was co-registered to the individual's T 1 -weighted MR image (six-parameter rigid registration), followed by a non-linear transformation to the automated anatomical labelling template in the Montreal Neurological Institute space using BioImage Suite. 26

Kinetic modelling
For a full description of the PET kinetic analysis, see our previous study. 14In brief, for [ 11 C]UCB-J image analysis, parametric binding potential (BP ND ) images were first generated using the simplified reference tissue model-2 step with the centrum semiovale as reference region. 27Simplified reference tissue model-2 step was performed using a fixed global k ′ 2 value (reference region clearance rate constant) of 0.027 min −1 , which was estimated as a k 2 population average of the centrum semiovale obtained with one-tissue compartment modelling from a previous subject group that underwent arterial blood sampling (Yale PET Center dataset).Lastly, parametric distribution volume ratio images with a whole cerebellum reference region (DVR Cb ) were computed from BP ND images as (BP ND +1)/(BP ND, Cb + 1).The whole cerebellum has previously been validated as a reference region for [ 11 C]UCB-J in Alzheimer's disease studies, demonstrating similar values of cerebellar distribution of volume (V T ) between Alzheimer's disease and CN participants. 14

Analysis of synaptic density patterns: source-based morphometry
Parametric [ 11 C]UCB-J DVR Cb images in the Montreal Neurological Institute template space were smoothed using a Gaussian kernel with a full width at half maximum of 8 mm prior to further analysis.Source-based morphometry was performed on the spatially normalized and smoothed parametric DVR Cb maps using the GIFT toolbox (GIFTv4.0b;trendscenter.org/software/gift/).For each subject, voxel-wise DVR Cb values were concatenated into a single row vector, and the vectors of all subjects were concatenated into a 2D matrix (number of subjects × number of voxels).Analyses were constrained to a voxel mask containing the whole brain, and the average subject-wise DVR Cb value was removed from the data by subtracting the mean from each row of this matrix.Prior to decomposition, principal component analysis was performed to reduce dimensionality of the data.In this study, the number of components was set at 18 based on previous analyses in order to extract 18 independent components (ICs). 22Spatial ICA was then performed using the Infomax algorithm, 28 which decomposed the data into a mixing matrix (number of subjects × number of components) and a source matrix (number of components × number of voxels).Each row of the mixing matrix contains a subject's loading weights (unitless) of corresponding synaptic density ([ 11 C]UCB-J DVR Cb ) patterns (i.e.sources) and can be interpreted as the extent to which a pattern is present in an individual subject.Source matrix row values represent the corresponding synaptic density patterns.For visualization, each source was scaled to unit standard deviation (Z map) and thresholded at a value of |Z| > 3.
Larger loading weights for a subject indicate a stronger presence of the corresponding synaptic density pattern.If loading weights are lower for MCI/mild Alzheimer's disease dementia participants compared with CN individuals, this can be interpreted as indicating a greater loss of synaptic density in the MCI/mild Alzheimer's disease dementia group relative to global synaptic density.

Statistical analysis
Statistical analyses were performed in R 4.0.2.Demographic and clinical characteristics comparisons were made using χ 2 test for categorical variables and unpaired t-tests for continuous variables.Kolmogorov-Smirnov tests and visual inspection of the histogram were used to assess normality of the variables.Variables were not normally distributed; thus, the Kruskal-Wallis test was used.The Benjamini-Hochberg procedure was performed to control the false discovery rate (FDR) for multiple comparisons [18 comparisons for ICs between CN, MCI, and mild Alzheimer's disease dementia diagnostic groups].The relationship between loading weights and episodic memory or global memory function was assessed using Spearman's rank correlation, controlling the FDR for multiple comparisons using the Benjamini-Hochberg procedure in two analyses: (i) in the MCI and mild Alzheimer's disease dementia groups only and (ii) in the full cohort including the CN.Tests were two-tailed, and P < 0.05 was set as the threshold for significance.

Demographics and characteristics
A total of 57 participants were included, consisting of 14 with amnestic MCI due to Alzheimer's disease, 24 with mild Alzheimer's disease dementia and 19 CN participants.Diagnostic groups were well balanced for sex (χ 2 = 2.55, P = 0.28) and age [F(2,54) = 0.30, P = 0.74], though CN participants had more years of education than participants with mild dementia (CN: 17.7 ± 2.1, mild dementia: 15.8 ± 2.4, P = 0.02; Supplementary Table 1).Both the MCI and the mild dementia participants displayed typical clinical characteristics of cognitive and functional impairment [MMSE: MCI: 26.3 ± 2.9, mild dementia: 21.5 ± 3.0; CDR sum of boxes (CDR-sb): MCI: 2.3 ± 1.0, mild dementia: 5.3 ± 1.5], and CN participants demonstrated no cognitive deficits (MMSE: 29.2 ± 1.1, CDR-sb: 0 ± 0; Supplementary Table 1).Further demographic and clinical information are presented in Supplementary Table 1.1].The clusters encompassing the patterns are reported in Table 1, ordered by hemisphere and maximum Z-value, as well as the group-wise comparisons of the corresponding loading weights.Most of these patterns show distinct localization as well as symmetry between hemispheres.However, in a few patterns, there are major clusters present in only one hemisphere, e.g. the largest positive cluster in Pattern 1 is primarily located in the temporal lobe of the left hemisphere, as well as displaying a large negative cluster in the right frontal lobe.Participants with mild Alzheimer's disease dementia showed lower loading weights compared with CN participants for Patterns 1, 2, 3 and 4 (Fig. 1, Table 1).Additionally, loading weights for mild Alzheimer's disease dementia participants were significantly lower compared with MCI participants for Patterns 1 and 2. Lastly, Pattern 5 displayed lower bindings for MCI participants compared with CN.

Patterns of synaptic density
After applying the Benjamini-Hochberg procedure to the Kruskal-Wallis comparison, only two patterns (Patterns 1 and 2) had loading weights that remained statistically significant between diagnostic groups.For Pattern 1, participants with mild Alzheimer's disease dementia had significantly lower loading weights compared with both MCI (P = 0.0022) and CN participants (P < 0.0001).Similarly, for Pattern 2, participants with mild Alzheimer's disease dementia had significantly lower loading weights than MCI (P = 0.031) and CN participants (P < 0.0001; Fig. 1).

Association between cognitive performance and synaptic density patterns in MCI and mild dementia
We investigated the association of the loading weights of each synaptic density pattern with functional and cognitive measures, including CDR-sb, LMII, MMSE and Rey Auditory Verbal Learning Test (RAVLT)-long delay scores using Spearman's rank correlations.To assess the relationship with severity of cognitive impairment within disease state, we excluded CN participants and subsequently analysed only the MCI and mild Alzheimer's disease dementia subject loading weights.Within this symptomatic cohort, statistically significant correlations were found between the loading weights for Pattern 1 and CDR-sb (r = −0.54,ICA in early Alzheimer's disease BRAIN COMMUNICATIONS 2024, fcae107 | 5 P < 0.001), MMSE (r = 0.48, P = 0.009) and LMII score (r = 0.44, P = 0.034; Fig. 2).In addition, statistically significant associations were observed between the loading weights for Pattern 2 and LMII score (r = 0.34, P = 0.038; Fig. 3).Following FDR correction, only the relationship between the loading weights for Pattern 1 with CDR-sb (r = 0.54, P = 0.003) and with MMSE (r = 0.48, P = 0.036) remained statistically significant.

Discussion
We used the ICA of [ 11 C]UCB-J PET to identify brain patterns of synaptic density and to explore the relationship between measures of cognitive performance and pattern strength in individuals with normal cognition, MCI and mild Alzheimer's disease dementia.We focused on the brain patterns of covariance in synaptic density with subject loading weights that showed group-wise differences between diagnostic groups.Furthermore, we observed significant associations between measures of cognitive and functional performance and subject loading weights for two of these patterns with only the mild Alzheimer's disease dementia and MCI diagnostic groups included.

Brain patterns of synaptic density
In this study, we identified brain patterns of co-varying synaptic density in a cohort of CN, MCI and mild Alzheimer's disease dementia participants.Previously, we identified and validated similar brain patterns of co-varying synaptic density in a study consisting of CN participants only, with a wider age range. 21While there is some overlap of CN participants, overall, we obtained consistent and similar brain patterns between these two studies.Specifically, out of the original 13 identified and validated ICs in the CN study, 9 ICs were replicated (i.e.visually similar and encompassed same brain regions) in this study (ICs 01-06, IC 08, IC 09 and IC 13).Given that significant pathology occurs in the brains of participants with Alzheimer's disease, it was unclear whether these brain patterns would be replicated.Brain patterns with subject loading weights that differed between diagnostic groups consisted of the left lateral temporal, left precuneus and bilateral posterior cingulate gyrus (Pattern 1); right temporal cortices (Pattern 2), lateral/medial occipital, cerebellar and right pre-central (Pattern 3); lingual, right medial orbital frontal and cerebellar (Pattern 4); and a medial temporal pattern extending into the parahippocampal (i.e.entorhinal and perirhinal cortices, Pattern 5).Interestingly, the brain pattern consisting of the entorhinal and perirhinal regions (Pattern 5) had loading weights that showed a group-wise reduction in the MCI group versus both the CN and the mild Alzheimer's disease dementia group.The regions detected altogether in these brain patterns (1-5) are largely in line with previous MRI studies.Taken together, the identified patterns encompass the medial and inferior temporal lobes including the hippocampus and entorhinal cortices.These regions are known to be some of the earliest affected in the progression of Alzheimer's disease pathology, 29 and ROI-based studies have shown that medial temporal measurements can provide the best discrimination between Alzheimer's disease and CN groups. 30,31Pattern 1 also includes the cingulate gyrus, a region which is intimately connected to the hippocampus and entorhinal cortices. 32nitial SV2A PET studies in Alzheimer's disease reported significant reductions in hippocampal synaptic density. 13owever, a follow-up study demonstrated significant decreases in synaptic density in the entorhinal cortex, hippocampus, amygdala, lateral temporal, pre-frontal, lateral parietal and pericentral regions in Alzheimer's disease participants. 14hese findings are largely in line with our current study, which uses a highly overlapping dataset.However, we did not observe the frontal cortex or parietal cortex within our significant brain patterns.A possible explanation for this may be that the changes in synaptic density in these regions did not occur in a consistent manner such that it did not register as a covariance within our brain patterns.

Data-driven algorithms help improve diagnostic accuracy
A major novelty of this study is the data-driven nature of the applied analysis method, where we make no a priori assumptions on the data e.g.subject info/demographics or use of pre-defined ROIs.Nonetheless, brain covariance patterns were identified that are spatially similar to early Braak staging of Alzheimer's disease.Similarly, there have been previous studies applying comparable data-driven automated image-based algorithms to identify spatial patterns of [ 18 F] FDG PET to discriminate between subtypes of parkinsonism (idiopathic Parkinson's disease, multiple systems atrophy and progressive supranuclear palsy), improving diagnostic accuracy (80%) compared with conventional clinical diagnosis (66%). 33his cohort, which has been previously studied using pre-defined template regions, 16 showed aberrant synaptic patterns in similar brain regions, namely temporal and cingulate cortices, but not frontal or occipital cortices.A parallel study using principal component analysis (PCA) identified distinct patterns across ROIs, with significant contributions from subcortical and parieto-occipital cortical regions. 34These PCA findings, particularly the positive correlation of component scores with cognitive domains and the negative correlation with global amyloid deposition, provide a complementary perspective to our ICA results. 34Of major interest is that the current study was able to discriminate between MCI and mild Alzheimer's disease dementia groups, as subject loading weights were significantly different between MCI and mild Alzheimer's disease dementia groups for Patterns 1 (mild Alzheimer's disease dementia versus MCI, P = 0.0022) and 2 (P = 0.031; Fig. 1).Comparatively, no such discriminative effect was accomplished using pre-defined (composite) ROIs, 14 suggesting that this data-driven method may be more sensitive to early synapse-related pathology in Alzheimer's disease.

Strength of brain pattern encompassing temporal and cingulate cortices is related to cognitive performance
Building on our recent work, 34 we explored the relationship between brain patterns of synaptic density and clinical measures of cognitive performance in Alzheimer's disease.We found correlations between subject loading weights for several patterns and outcome measures of cognitive performance, even when excluding the CN participants.After adjusting for multiple comparisons, the pattern encompassing temporal and cingulate cortices showed associations with CDR-sb and MMSE.This suggests potential links between synaptic density in these regions and cognitive performance.Previous post-mortem electron microscopy studies have found significant decline in synaptic levels in the posterior cingulate gyrus of individuals in the early stages of Alzheimer's disease compared with CN individuals, with MCI patients exhibiting synaptic numbers that were between Alzheimer's disease and CN cohorts. 32Additionally, the cingulate gyrus is strongly connected with the hippocampal/entorhinal cortex regions, areas of demonstrated robust and early synaptic loss in AD 4,35,36 and part of the medial temporal memory circuit.Synapse levels in the temporal gyrus are also lower to a similar degree in both MCI and early Alzheimer's disease patients compared with CN. 37 It is worth noting that some earlier studies might not have had the sample size to detect these associations.
When comparing correlation strengths between regional SV2A DVR and cognitive measures as in our previous studies, there are a few key differences.In the 2020 study, strong correlations between hippocampal synaptic density and clinical measures (CDR-sb and episodic memory) were observed in a pooled analysis of Alzheimer's disease and CN participants.This was also the case for a composite ROI of Alzheimer's disease-affected regions (entorhinal, hippocampal, parahippocampal, amygdala, pre-frontal, lateral temporal, posterior cingulate cortex/precuneus, lateral parietal and lateral occipital regions).However, none of these correlations were significant within the Alzheimer's disease or CN groups. 14In a follow-up study from 2022, correlations of synaptic density (as assessed by SV2A PET) and cognitive performance were evaluated within a homogeneous Alzheimer's disease sample using a comprehensive neuropsychological testing battery. 16However, our current ICA approach is the first data-driven study to definitively distinguish between MCI and mild Alzheimer's disease dementia participants with respect to synaptic density PET-derived information.

Limitations
There are a number of limitations worth reporting.The precise role of SV2A as a pre-synaptic density marker is unconfirmed, and hence, it remains a limited indicator of synaptic density.Clinical diagnosis was made on the basis of standard clinical criteria and amyloid PET status.Biomarkers of tau pathogenesis were not included and there is no post-mortem validation, so this cohort could not be fully classified according to amyloid, tau, neurodegeneration criteria. 38This study has limited power to investigate demographic variables due to its modest sample size.It is also important to note that ICA has not been previously utilized to investigate synaptic density patterns in Alzheimer's disease patients, and future studies with larger sample sizes may be needed to further assess and validate findings of this nature.In addition, longitudinal studies would be highly valuable to determine whether the ICA loading weights can provide information on disease progression.

Conclusion
Multiple patterns of synaptic density relevant to underlying Alzheimer's disease pathology were demonstrated in SV2A PET data using data-driven methodologies.Synaptic density loss in Alzheimer's disease may occur in a systematic and network-wise manner, possibly (at least partly) according to distinct anatomical patterns.In addition, several of these patterns of synaptic density showed clear significant differences across disease stages (between MCI and mild Alzheimer's disease dementia participants), as well as strong associations with clinical measures of cognition within the symptomatic Alzheimer's disease group.This is a promising step forward for the application of data-driven algorithms in the assistance of clinical diagnosis and disease staging in Alzheimer's disease.

Figure 1
Figure 1 Significant ICA patterns and loading weights.Synaptic density spatial patterns were identified by ICA.Patterns were Z-scores transformed, and the Kruskal-Wallis comparison of the loading weights determined 5 (out of 18) patterns with significant diagnostic group differences, depicted in Rows 1-5.Each row displays co-varying spatial density patterns with |Z| > 3 and corresponding loading weights.

Figure 2
Figure 2 Pattern 1 loading weights versus cognition.Correlations of Pattern 1 subject loading weights and measures of cognitive and functional performance in MCI and mild Alzheimer's disease dementia participants.Within the disease group only, scatter plots depict significant relationships between loading weights and CDR-sb, LMII and MMSE scores.Spearman's correlations were calculated, and following FDR correction, only the relationship between the loading weights with CDR-sb and with MMSE remained statistically significant.The dashed line represents linear regression line.

Figure 3
Figure 3 Pattern 2 loading weights versus cognition.Correlations of Pattern 2 subject loading weights and measures of cognitive and functional performance in MCI and mild Alzheimer's disease dementia participants.Within the disease group only, scatter plots depict significant relationships between loading weights and LMII score.Spearman's correlations were calculated without correction for multiple comparisons.The dashed line represents linear regression line.

Table 1 Synaptic density patterns identified by ICA and group-wise comparisons of corresponding loading weights
Clusters with |Z| > 3.0 and size > 500 voxels are listed, ordered by maximum absolute Z-value.CN, control.*Non-FDR corrected indicates significant group differences in loading weights: *P < 0.05, **P < 0.01 and ****P < 0.0001.