Sam de Waegenaere1,2, Alya Al-Awlaqi1,2, Lori Berckmans1,2, Marleen Verhoye1,2, and Mohit H Adhikari1,2
1Bio-Imaging Lab, Biomedical Sciences, University of Antwerp, Antwerp, Belgium, 2µNEURO Research Centre of Excellence, University of Antwerp, Antwerp, Belgium
Synopsis
Keywords: Alzheimer's Disease, Alzheimer's Disease
Motivation: Resting-state fMRI studies of Alzheimer’s disease impact on brain’s function commonly use functional connectivity (FC) ignoring sensitive and dynamic readouts such as the co-activation patterns (CAPs) occurring at short timescales.
Goal(s): We aimed to assess changes in CAPs, in addition to network-level FC, in a transgenic rat model of Alzheimer’s disease longitudinally.
Approach: We acquired high temporal resolution resting-state fMRI and performed FC and CAP analysis.
Results: We found increased lateral cortical network FC that correlated with memory impairments at the plaque stage, and hyper and hypoactivation of the default-mode-like-network and hippocampal regions in two CAPs at the pre-plaque and plaque stages respectively.
Impact: Our
findings demonstrate that metrics of brain dysfunction of Alzheimer’s disease derived
from high temporal resolution resting-state fMRI not only explain behavioural
manifestations but also capture alterations preceding plaque formation further
validating their translational potential as an early, functional biomarker.
Introduction
Resting-state
functional MRI (rsfMRI) has been instrumental in characterizing brain
dysfunction due to Alzheimer’s disease (AD) from a network perspective1. In addition to traditional functional
connectivity (FC) analyses, dynamic FC states such as co-activation patterns
(CAPs)2,3 reveal functional network dynamics at short
timescales that are accurate at classifying transgenic rodents4,5 of AD from healthy controls. Here, we used
rsfMRI to evaluate AD’s functional impact at pre-plaque and plaque stages in
the TgF344-AD rat model using FC and CAPs and then correlated them with memory impairments
at the plaque stage.Methods
Fifteen
TgF344-AD (TG) and 15 wildtype (WT) rats underwent rsfMRI scans, under anesthesia
(isoflurane/medetomidine), at 4 (pre-plaque) and 10 (plaque stage) months of
age using a 9.4T MRI-system (TR 0.6s, 1000 GE-EPI volumes). Post 10-month scans,
we assessed learning and memory using a radial arm maze (RAM) consisting of a
central platform and 8 arms. After a habituation phase (5-7 days), food was
placed in four arms, and food-restricted rats were tested for working and
reference memory in the acquisition phase (10 days). We pre-processed MRI
images using an in-house MATLAB pipeline and SPM12 and calculated region of
interest (ROI), and network-level FC. We obtained CAPs by (a) clustering concatenated
pre-processed volumes of all 30 subjects at both ages into 2-20 clusters,
using spatial correlation distance, (b) identifying the optimal number of
clusters, and (c) taking the voxel-wise average across all volumes with identical
cluster membership. Group and age effects on FC, spatial, and temporal
properties of the CAPs were assessed using repeated measures ANOVA. Behavioral
learning curves were analyzed for multiple metrics of reference and working
memory and correlated with FC. Finally, cross-validated (80-20 train test split) accuracy of CAP temporal and
spatial features to classify the animals into four classes (TG 4M, TG 10M, WT 4M, WT 10M) was evaluated.Results
WT rats improved their working memory
over time while TG rats did not. TG rats committed more reference memory errors
and explored more arms in the RAM than WT rats (Fig 1A-D). ROI-FC did not show a significant genotype or age effect for any connection (P > 0.05, FDR
corrected). Lateral cortical network (LCN) FC was significantly increased in TG
(p < 0.05, FDR corrected for 25 within- & between-network FC
comparisons) at the plaque stage and correlated with reference memory errors
and time in incorrect arms (Fig 1 E-F).
Six CAPs in 3 pairs of anticorrelated patterns
(Fig 2A, C) were found to be optimal; they explained saturation-level variance
across all subjects’ volumes. No significant genotype or age effects were
observed in CAPs’ temporal properties. CAPs' spatial activation showed a significant interaction effect for several voxels/CAPs. At the pre-plaque stage,
we found functional hyperactivation in the default mode-like network (DMLN)
regions and hippocampal formation in CAP 1 (Fig 2D, panel 3 from top), and in
the cingulate region in CAP 6 (Fig 2E, panel 3). At the post-plaque stage, TG rats presented a
general decrease in activation levels, especially in hub regions within the
DMLN (Fig 2D & E, bottom panels). These changes are also reflected in significantly
higher than-chance classification accuracy when spatial features of CAPs were
used as features (Fig 3A). The confusion matrix showed the highest prediction for the
TG 4M group and confusion between the genotypic groups at the plaque stage (Fig
3B).Discussion
Correlation between significantly
higher LCN FC with memory impairments at the plaque stage demonstrates importance
of rsfMRI in understanding AD’s impact on brain’s functional architecture. Hyperactivation
found in the default mode-like network regions such as the cingulate cortex and
the hippocampus at the pre-plaque stage is in line with a widely accepted view
that abnormal neuronal activities precede the plaque formation in animal and
cell models6. During memory encoding
tasks, hippocampal hyperactivation and decreased inactivation of the DMLN hub
regions has been reported as a consistent fMRI signature of AD, particularly at
the early stages7. A pathogenic role in early hyperactivation is
further supported by observation that individuals with AD exhibit increased
susceptibility to epilepsy and seizures, particularly, in early-onset familial
AD7. Thus, our findings using
RS-CAPs demonstrate AD phenotypes that are consistent with studies using other
techniques such as electrophysiology and task-fMRI.Conclusion
Our findings demonstrate the
effectiveness of high temporal resolution rsfMRI and advanced analyses such as
co-activation patterns to tease out signatures of Alzheimer’s disease in a
transgenic rat model. In a next step, we will use CAP activations and FC
changes at pre and plaque stages to predict memory impairments found in this
study using a cross-validated machine learning approach.Acknowledgements
This work was funded by FWO-G045420N
(MV), Stichting Alzheimer Onderzoek (SAO‐FRA 2020/027 (MV). The computational resources and
services were provided by the HPC core facility CalcUA, the VSC, funded by the
Hercules Foundation, and the Flemish Government department EWI. The Bruker
Biospec 9.4T system was upgraded to AVANCE-NEO through Hercules Foundation
funding (I007120N -MV-co-promotor).References
- Badhwar, A. et al.
Resting-state network dysfunction in Alzheimer’s disease: A systematic review
and meta-analysis. Alzheimers Dement (Amst) 8, 73–85 (2017).
- Liu, X. & Duyn, J. H.
Time-varying functional network information extracted from brief instances of
spontaneous brain activity. PNAS 110, 4392–4397 (2013).
- Gutierrez-Barragan, D., Basson, M.
A., Panzeri, S. & Gozzi, A. Infraslow State Fluctuations Govern Spontaneous
fMRI Network Dynamics. Current Biology 29, 2295-2306.e5 (2019).
- Adhikari, M. H., Belloy, M. E., Van
der Linden, A., Keliris, G. A. & Verhoye, M. Resting-State Co-activation
Patterns as Promising Candidates for Prediction of Alzheimer’s Disease in Aged
Mice. Front. Neural Circuits 14, (2021).
- Adhikari, M. H. et al.
Longitudinal investigation of changes in resting-state co-activation patterns
and their predictive ability in the zQ175 DN mouse model of Huntington’s
disease. Sci Rep 13, 10194 (2023).
- Targa Dias Anastacio, H., Matosin,
N. & Ooi, L. Neuronal hyperexcitability in Alzheimer’s disease: what are
the drivers behind this aberrant phenotype? Transl Psychiatry 12,
257 (2022).
- Kazim, S. F. et al. Neuronal
Network Excitability in Alzheimer’s Disease: The Puzzle of Similar versus
Divergent Roles of Amyloid β and Tau. eNeuro 8,
ENEURO.0418-20.2020 (2021).