Whole-brain functional connectivity predicts regional tau PET in preclinical Alzheimer's disease

Abstract

Preclinical Alzheimer's disease, characterized by the abnormal accumulation of amyloid-β prior to cognitive symptoms, presents a critical opportunity for early intervention. Past work has described functional connectivity (FC) changes in preclinical Alzheimer's disease, yet the predictive nature between the functional connectome and Alzheimer's disease pathology during this window remains unexplored. We applied connectome-based predictive modelling to investigate the ability of resting-state whole-brain FC to predict tau (18F-flortaucipir) and amyloid-β (18F-florbetapir) PET binding in preclinical Alzheimer's disease (A4, n = 342 amyloid-β-positive, age 65-85). Separate models were developed to predict amyloid PET signal in the posterior cingulate, precuneus, and cortical composite regions, and to predict tau PET signal in each of 14 cortical regions that demonstrated meaningful tau elevation as identified through a Gaussian mixture model approach. Model performance was assessed using a Spearman's correlation between predicted and observed PET binding standard uptake value ratios. We assessed the validity of significant models by applying them to an external dataset and visualized the underlying connectivity that was positively and negatively correlated to regional tau. We found that whole-brain FC predicts regional tau PET, outperforming FC-amyloid-β PET models. The best-performing tau models were for regions affected in Braak stage IV-V regions (posterior cingulate, precuneus, lateral occipital cortex, middle temporal, inferior temporal, and banks of the superior temporal sulcus), while models for regions of earlier tau pathology (entorhinal, parahippocampal, fusiform, and amygdala) performed poorly. Importantly, FC-based models predicted tau PET signal in the Alzheimer's Disease Neuroimaging Intitative-3 dataset (amyloid-β-positive, n = 211, age 55-90) in tau-elevated but not tau-negative individuals. For the posterior cingulate tau model, the most accurate model in A4, the predictive edges positively correlated with posterior cingulate tau predominantly came from nodes within temporal, limbic, and cerebellar regions. The most predictive edges negatively associated with tau were from nodes of heteromodal association areas, particularly within the prefrontal and parietal cortices. These findings reveal that whole-brain FC meaningfully predicts tau PET in preclinical Alzheimer's disease, particularly in regions affected in advanced disease, and are relevant across the Alzheimer's disease clinical spectrum in individuals with elevated tau PET burden. This suggests that functional connectivity, likely in conjunction with other factors, may play a key role in early processes that facilitate later-stage tau spread. These models highlight the potential of the functional connectome for the early detection and monitoring of Alzheimer's disease pathology, especially in later-stage target regions.

Keywords: asymptomatic; dementia; machine learning; multimodal; neurodegeneration.

Graphical Abstract

Figure 1

Whole-brain FC predicts regional tau PET signal in preclinical Alzheimer’s disease. (A) Amyloid CPM results, representing accuracies of whole-brain FC predicting amyloid PET SUVRs. Violin plots represent distributions over 1000 model iterations of global amyloid SUVR and regional amyloid SUVR in the PCC and PreC regions. Model performance is measured by Spearman’s correlation ($r_s$) of predicted versus observed amyloid PET. n = 1490 amyloid-positive and amyloid-negative participants. *** P < 0.001, false-discovery rate (FDR)-adjusted for 3 tests (1-tailed, permutation test). (B) Regional tau CPM results, representing accuracies of whole-brain FC predicting tau PET SUVRs. Violin plots represent distribution over 1000 model iterations, with the median model performance depicted by the black dot. Model performance is measured by Spearman’s $r_s$ of predicted versus observed regional tau PET. n = 342 amyloid-positive participants. * P < 0.05 against permutation testing; FDR-adjusted for 14 tests (1-tailed, permutation test). ‡ P < 0.05 against spatially permuted model, FDR-adjusted for seven tests (1-tailed permutation test). (C) Similarity of predictive edges between regional tau models, using the Jaccard index (intersection of edges divided by the union of edges present between each pair of models).

Figure 2

Regional tau binding models generalize to an external dataset. Validation of A4 regional tau models in the ADNI-3 dataset, differentiated by tau-elevated and tau-negative groups. Model performance is measured by Spearman’s $r_s$ of predicted versus observed regional tau PET in ADNI (n = 211). *P < 0.05 against permuted model, FDR-adjusted for six tests (1-tailed permutation test).

Figure 3

PCC tau model edges positively and negatively correlated with FC. (A) Circle plots of the predictive edges for the PCC model (n = 342). Each circle represents a node from the Shen-268 atlas, grouped according to their respective region. Edges are depicted by lines between each node pair. Plot with red lines (left) shows edges that were positively correlated with PCC tau, while the plot with blue lines (right) indicates edges that were negatively correlated with PCC tau. Only edges from the top 5% most predictive nodes (by degree) are displayed, based on edges significant in ≥ 3/5-folds across ≥ 600/1000 connectome-based predictive modeling iterations. (B) The same predictive edges from circle plots A plotted onto brain surfaces.