INTRODUCTION: Dynamic functional network connectivity (dFNC) has provided more potential to characterize different connectivity states of intrinsic networks in brain than static functional network connectivity (sFNC) and is emerging as a novel biomarker of disease. However, the mechanism of lipid metabolism-related factors impacting the dFNC across the Alzheimer’s disease (AD) spectrum remains unclear.
METHODS: We included 242 subjects, including 181 subjects with AD spectrum and 61 cognitively normal subjects. The data of lipid metabolism-related genes, serum lipid metabolites, cerebrospinal fluid (CSF) core biomarkers, structural and resting-state functional MRI, and cognitive assessment were obtained for all subjects. We used sliding-window method and k-means algorithm to obtain the dFNC matrix of all subjects and conducted cluster analysis. Spearman correlation analysis was employed to analyze the associations of dFNC temporal properties to CSF core biomarkers, cognitive scores and lipid metabolites. Polygenic score (PGS) were constructed based on the odds ratio (OR) value of each gene locus, and the lipid composite score was constructed based on the significant correlative lipid components. We also tested dynamic network difference and dFNC associations with CSF and behavioral phenotyping in lipid-related subgroups. To explore the casual relationship among lipid-related indicators, dFNC, CSF pathology and behavioral phenotype, we applied the mediation effect analysis. Finally, the abnormal dFNC was used as an objective diagnostic biomarker to classify the AD spectrum populations by support vector machine (SVM) model.
RESULTS: Two stable reoccurring connectivity states were identified: frequent but weak-connected state I, and less frequent but strong-connected state II. The AD patients showed strong connectivity strength in both states. They represented significantly differential connectivity and temporal properties compared to non-AD participants in each state. The SVM analysis showed that the dFNC, especially state I connectivity, were better in differentiating AD from other groups of subjects. Interestingly, they altered brain preference to dynamic connectivity state, in that they tended to dwell longer in the densely connectivity state II, whereas the non-AD populations were prone to reside in the sparsely state I. Then, the temporal properties of two states, including fractional windows and mean dwell time, were anti-correlated with CSF core biomarkers, cognitive scores, and lipid metabolites. More importantly, lipid-related subgroup analysis found that the effect of lipid composite score over dynamic connectivity differences was mainly in state I, whereas APOE ε4 showed a wide effect on state II network differences. The effects of PGS on the two states were not significantly different, but when APOE was removed, the effect of polygenes on state I network differences were more involved in within-network connectivity. In addition, the state II differential connections between APOE ε4 genotype subgroups mediates the effects of APOE genotype on CSF and cognitive phenotypes.
DISCUSSION: Subjects spent more time in weakly-connected state I, where new connections were more easily established to promote higher network flexibility ¹. However, less frequent state II with strong connectivity might compensate for the loss of intertwined brain function under network destruction. Herein, AD patients reversed the brain’s preference for state and took a long time with higher frequency in state II. Contrary to popular belief ², the connectivity strength of AD patients reached the highest point at both states, probably because they were still at the early stages of AD and had restricted compensatory capacity. There was a significant group-level network difference between association cortex system and sensory cortex system, similar to previous findings ^{3, 4}. Network disruption between the two systems might characterize biological and behavioral phenotyping in AD. Herein, temporal variability was lower in AD than in healthy subjects, consistent with previous studies ^{4, 5}.Moreover, the temporal properties and molecular, clinical phenotyping had an anti-correlation pattern between the two states. For instance, frequent occurrences and staying longer in state I had a protective role, thus alleviating tau pathology and improving cognition, while state II showed the opposite effect (destructive effect), causing the biological and behavioral underpinnings due to dynamic temporal characterization.The SVM analysis also showed that dFNC had better predictive power than sFNC in distinguishing AD and non-AD groups, suggesting the essentiality of extracting distinct dynamic states from the conventional sFNC. The different lipid architectures mainly influenced dFNC in state I, while the APOE ε4 mainly impacted dFNC in state II, suggesting that lipid profile and APOE have different impacts on dFNC. The PGS showed more within-network connectivity in state I after removing APOE ε4, indicating that APOE ε4 guided the direction of polygenes to destroy between-network connectivity. The differential state II connections had an indirect effect on the relationship between APOE and CSF, cognitive indicators. The long-range connections of salience and dorsal attention network interacted with visual and cerebellum network are probably related to top-down attentional control.
CONCLUSION: This study is the first to explore the temporal-spatial patterns of the whole brain dynamic network connectivity and its temporal properties across the entire cognitive spectrum ranging from normal cognition to AD, and reveals the effect of lipid biological factors over the dynamic large-scale brain networks, providing a new insight into the lipid biological basis of the reorganization of dynamic brain networks for AD.