Tau‐related reduction of glucose metabolism in mild cognitive impairment occurs independently of APOE ε4 genotype and is influenced by Aβ

Abstract

INTRODUCTION

Positron emission tomography (PET) imaging studies have shown that amyloid beta (Aβ) is significantly correlated with glucose metabolism in mild cognitive impairment independently of the apolipoprotein E (APOE) ε4 genotype.

METHODS

We used a singular value decomposition (SVD) approach to pairwise cross‐correlation among tau, Aβ, and fluorodeoxyglucose PET images. The resulting SVD‐based tau and Aβ scores as well as the APOE ε4 genotype, were entered as predictors in a voxelwise general linear model for statistical assessment of their effect on FDG.

RESULTS

We found cortical regions where a reduced glucose metabolism was maximally correlated with distributed patterns of tau, accounting for the effect of Aβ and APOE ε4 genotype.

DISCUSSION

By highlighting the more significant role of tau, rather than Aβ, in the reduction of glucose metabolism, our results provide a better understanding of their combined effect in the development and progression of Alzheimer's disease.

Highlights

• This study uses a data‐driven singular value decomposition approach to the cross‐correlation matrix between tau and fluorodeoxyglucose (FDG) images, as well as between FDG and amyloid beta (Aβ) positron emission tomography (PET) images.

• From a population of mild cognitive impairment subjects, we found that spatially distributed scores of tau PET are associated with an even stronger reduction of glucose metabolism, independent of the apolipoprotein E ε4 genotype and confounded by Aβ.

• By highlighting the more significant role of tau, rather than Aβ, on the reduction of glucose metabolism, our results provide a better understanding of their combined effects in the development of Alzheimer's disease.

1. BACKGROUND

Preliminary evidence regarding the pathogenesis of Alzheimer's disease (AD) led to the so‐called “amyloid cascade hypothesis,” ¹ , ² , ³ which pointed to amyloid beta (Aβ) deposition as the triggering process for AD progression. An alternative hypothesis suggests that not only Aβ, but also the accumulation of neurofibrillary tangles of tau aggregates acting in a coordinated fashion, lead to cognitive deterioration. ⁴ , ⁵ , ⁶ As such, investigators have made substantial efforts to gain a better understanding of the underlying relationships among Aβ, tau, glucose hypometabolism, and brain atrophy. ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹

Extensive studies have been conducted to elucidate the impact of Aβ and apolipoprotein E (APOE) ε4 genotype on the reduction of glucose metabolism in mild cognitive impairment (MCI). ¹² , ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ By using a singular value decomposition (SVD) approach from Worsley et al., ²⁰ we showed ¹⁴ that Aβ scores produce a stronger relationship with glucose metabolism than APOE ε4 genotype and global measures of Aβ. In concordance with Lowe et al., ¹⁷ our findings ¹⁴ reinforce the idea that, independently of the APOE ε4 genotype, Aβ accumulation and reduction of glucose metabolism are more likely to occur simultaneously throughout the clinical spectrum of AD progression. Thus, an accepted consensus ¹³ , ¹⁶ , ¹⁷ , ²¹ is that even though the APOE ε4 genotype alone can be related to hypometabolism, when interacting with Aβ deposition, most of the reduction in metabolism is attributable to the latter rather than the former.

Much less consensus has been reached about the association between tau accumulation and glucose metabolism. Preliminary studies argued that tau could be considered a mediator between Aβ and neurodegeneration, ²² , ²³ , ²⁴ , ²⁵ despite some other results suggesting a more complex interpretation and pointing to a synergistic interaction between neocortical tau and Aβ in relation to glucose metabolism. ²⁶ , ²⁷ In turn, AD‐related processes, such as reduction of glucose metabolism or gray matter (GM) volume loss, have also been thought to play a mediating role between tau deposition and cognition deterioration. ²⁸ , ²⁹ , ³⁰

It has been shown ²⁶ that spatially concurrent Aβ and tau deposition have an interactive effect on glucose metabolism. Similarly, Whitwell et al. ²³ showed that regional hypometabolism was more strongly correlated to regional tau pathology than to regional Aβ burden in AD subjects. Likewise, Strom et al. ³¹ demonstrated that, independently of the APOE ε4 genotype effect, local tau pathology was associated with local hypometabolism in AD subjects. Hanseeuw et al. ²⁷ demonstrated that tau deposition in the entorhinal cortex was not only related to local hypometabolism but extended to nearby temporal regions and the cingulate cortex. However, they also reported ²⁷ that, for low levels of Aβ, the increase of tau in the inferior temporal regions was associated with increased glucose metabolism. This surprising opposing association between tau and glucose metabolism has been further corroborated by Adams et al. ³² in cognitively normal (CN) subjects and by Rubinski et al. ³³ in MCI subjects.

With a few exceptions from Adams et al., ³² most of the correlational analysis between tau positron emission tomography (PET) and fluorodeoxyglucose (FDG) PET have been carried out at a local spatial level and/or using a single global measure for the quantification of tau. Due to such limitations, we use a SVD approach to produce individual scores of weighted averages of tau that maximally correlate with specific spatial networks of FDG. Having the ability to produce maximally correlated components, the SVD approach can uncover patterns in the FDG–tau relationship that would remain hidden if traditional global tau measurements were used for cross‐correlational purposes. We hypothesize that SVD‐based tau scores are associated with a stronger reduction of glucose metabolism independently of the effects of the APOE ε4 genotype, and that the spatial association between tau and glucose metabolism is influenced (i.e., confounded) by Aβ.

2. METHODS

2.1. Subjects and image acquisition

Data used in the preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (http://adni.loni.usc.edu). The ADNI was launched in 2003 by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, the Food and Drug Administration, private pharmaceutical companies, and non‐profit organizations, as a $60 million, 5 year public–private partnership, which has since been extended. ADNI is the result of the efforts of many co‐investigators from a broad range of academic institutions and private corporations, and subjects have been recruited from > 55 sites across the United States and Canada. To date, the ADNI, AND‐GO, ADNI‐2, and ADNI‐3 protocols have recruited > 1500 adults, ages 55 to 90, to participate in the research, consisting of CN older individuals, people with early or late MCI, and people with early AD. For up‐to‐date information, see www.adni‐info.org.

The subjects of this cross‐sectional study consisted of 119 subjects, ranging from early to late MCI, from the ADNI study who had available 2‐[18F]fluoro‐2‐deoxyglucose (FDG), [18F]flortaucipir PET, and 3D T1‐weighted anatomical magnetic resonance imaging (MRI) within a time frame of 6 months. Early MCI subjects had Mini‐Mental State Examination (MMSE) scores between 24 and 30 inclusively, a Clinical Dementia Rating of 0.5, a reported subjective memory concern, an absence of dementia, an objective memory loss measured by education‐adjusted scores on delayed recall of one paragraph from the Wechsler Memory Scale Logical Memory (WMSLM) II, essentially preserved activities of daily living, and no impairment in other cognitive domains. Late MCI subjects had the same inclusion criteria, except for objective memory loss measured by education‐adjusted scores on delayed recall of one paragraph from WMSLM II.

A detailed description of the ADNI MRI and PET image acquisition protocols can be found at http://adni.loni.usc.edu/methods. ADNI studies are conducted in accordance with the Good Clinical Practice guidelines, the Declaration of Helsinki, and US 21 CFR Part 50 (Protection of Human Subjects) and Part 56 (Institutional Review Boards), with informed written consent obtained from all participants at each site.

Out of the whole sample of 119 subjects, 116 subjects also had concurrent [18F]florbetapir PET scans within a time frame of 6 months. Those 116 [18F]florbetapir PET scans were classified into high (Aβ_H) and low (Aβ_L) amyloid subjects according to their standardized uptake value ratio (SUVR) values in an optimal target region of interest (Stat‐ROI) including areas of the posterior cingulate cortex, precuneus, and medial frontal cortex, with an associated cutoff of SUVR = 1.24. That optimal target ROI for the segregation of subjects into low and high amyloid has been previously derived from our data‐driven approach developed by Carbonell et al. ³⁴ The associated cutoff value of 1.24 was obtained from the Stat‐ROI measurements taken from our current sample of MCI subjects via receiver operating characteristic analysis, as described by Carbonell et al. ³⁴

2.2. . Image processing

MR and PET images were processed using the PIANO software package (Biospective Inc.). T1‐weighted MRI volumes underwent image non‐uniformity correction using the N3 algorithm from Sled et al., ³⁵ brain masking, linear spatial normalization using a 9‐parameter affine transformation, and non‐linear spatial normalization to map individual images from native coordinate space to Montreal Neurological Institute (MNI) reference space using a customized, anatomical MRI template derived from ADNI subjects. The resulting image volumes were segmented into GM, white matter, and cerebrospinal fluid using an artificial neural network classifier from Zijdenbos et al. ³⁶ and partial volume estimation (PVE) from Tohka et al. ³⁷

RESEARCH IN CONTEXT

1. Systematic review: The authors performed a standard literature review using traditional sources including preprints and published articles. They found that the relationship between tau and glucose metabolism at different stages of Alzheimer's disease (AD) has not been fully understood.

2. Interpretation: This study uses a data‐driven singular value decomposition approach to the cross‐correlation matrix between tau and fluorodeoxyglucose (FDG) images, as well as between FDG and amyloid beta (Aβ) positron emission tomography (PET) images. The authors found that spatially distributed scores of tau PET are associated with an even stronger reduction of glucose metabolism, independent of the apolipoprotein E ε4 genotype and confounded by Aβ.

3. Future directions: Highlighting the more significant role of tau, rather than Aβ, on the reduction of glucose metabolism, our results provide a better understanding of their combined effects in the development of AD. Future research will focus on the association of tau, Aβ and cortical atroyphy as a more accurate biomarker of neurodegeneration.

The [18F]FDG, [18F]florbetapir, and [18F]flortaucipir PET images underwent several pre‐processing steps, including frame‐to‐frame linear motion correction, smoothing with scanner‐specific blurring kernels to achieve 8  mm full width half‐maximum per Carbonell et al., ³⁸ and averaging of dynamic frames into a static image. The resulting smoothed PET volumes were linearly registered to the subject's T1‐weighted MRI and, subsequently, spatially normalized to reference space using the linear and non‐linear transformations derived from the anatomical MRI registration. The GM density map for each subject was transformed to the same final spatial resolution (i.e., re‐sampled to the same voxel size) as the PET data to account for the confounding effects of atrophy in the group‐level statistical model. SUVR maps of the PET images were generated from [18F]florbetapir and [18F]flortaucipir PET using the GM‐masked full cerebellum as a reference region. The FDG SUVR maps were generated with the pons as a reference region.

To minimize PET off‐target binding in non‐GM regions, we have restricted our analysis to those voxels in the brain cortex that have been previously masked by a GM mask in the non‐linear template space. Hence, all our voxelwise PET maps and statistical maps were projected onto the cortical surface for visualization purposes only. The GM mask in template space was generated by thresholding a template‐space GM map at 0.5, which, in turn, was obtained as the voxel‐wise average of the subject‐level GM PVE maps, each mapped into template space.

2.3. Statistical analysis

The main component of the methodology used here is the SVD of the multi‐modality cross‐correlation matrix between the tau PET and FDG PET images demonstrated by Carbonell et al. ¹⁴ Briefly, the SVD of the cross‐correlation matrix C between the tau and the FDG dataset can be expressed as $C=UW{V}^{\prime }$, where U and V are orthonormal matrices, whose columns are the so‐called eigenimages or spatial loadings, and W is a diagonal matrix of ordered eigenvalues (see details in Carbonell et al. ¹⁴ ). For practical reasons and an easier interpretation, C is typically approximated by the first few components, ordered according to the values of the weights in W. Hence, the cross‐product of the first spatial loadings would produce the largest additive component of the full voxels x voxels matrix C, without the need for reconstructing it or performing the cross‐product operation for extracting significant information about the underlying cross‐correlation patterns.

Correspondingly, the SVD analysis also provides individual scores for each PET modality that can be computed by simply projecting the spatial loadings onto the corresponding PET dataset. As previously explained by Carbonell et al., ¹⁴ we used a leave‐one‐out technique to produce the individual SVD‐based scores. Thus, leaving each sample out one at a time, the SVD and corresponding eigenimages are produced from the rest of the sample. Then, the individual scores of SUVR_SVD for the left‐out sample are computed by mapping them onto the space of the orthogonal eigenimages corresponding to the rest of the sample. Such a leave‐one‐out approach overcomes any possible circularity effect produced by the computation of the SVD components and their subsequent inclusion in a generalized linear model (GLM).

Indeed, for a more precise statistical inference, those individual scores can be included as regressors in a voxelwise GLM. In particular, we used GLM for statistical assessment of the effect of SVD‐based tau scores on FDG PET SUVR.

Several models were assessed in our analysis. The first model was intended to evaluate the effect of the SVD‐based tau score and APOE ε4 status on FDG signal:

$$Y_{\mathrm{FDG}}=b_0+b_{Cov}{X}_{Cov}+b_{Tau}{X}_{Tau}+b_{Apo}ApoE\epsilon 4+b_{Int}{X}_{Tau}\times ApoE\epsilon 4$$ (1)

where Y_FDG denotes the FDG SUVR as a predicting variable; $X_{Cov}$ includes age, sex, and cognitive measurements (MMSE and Alzheimer's Disease Assessment Scale Cognitive subscale [ADAS‐Cog]) as global confounding covariates; GM density as a voxelwise confounding covariate; and SVD‐based tau scores ($X_{Tau}$), APOE ε4 status, and tau x APOE ε4 status interaction as predictors of interest. The inclusions of such global covariates aim to control for potential differences in glucose metabolism that might be caused by the sample heterogeneity regarding age, sex, and especially cognitive stages. In fact, it has been early proven ³⁹ that during the MCI stage of AD, FDG PET hypometabolism is strongly associated with cognition.

A second model was assessed to evaluate the interactive effect of Aβ and tau on FDG:

$$Y_{FDG}=b_0+b_{Cov}{X}_{Cov}+b_{Tau}{X}_{Tau}+b_{Amyloid}{X}_{Amyloid}+b_{Int}{X}_{Tau}\times X_{Amyloid}$$ (2)

where, in this case, $X_{Cov}$ also includes the APOE ε4 status and $X_{Amyloid}$ denotes SVD‐based amyloid scores resulting from an independent SVD analysis between FDG and amyloid PET images. Notice also that the previous model can be rewritten as:

$$Y_{FDG}=b_{Cov}{X}_{Cov}+\left(b_0+b_{Amyloid}{X}_{Amyloid}\right)+\left(b_{Tau}+b_{Int}{X}_{Amyloid}\right)\times X_{Tau},$$

which can be used to assess how the relationship between FDG and tau can be continuously modulated by amyloid. Therefore, when there is a true interaction (e.g., assessed by $b_{Int}$) between $X_{Amyloid}$ and $X_{Tau}$, it can be inferred that the main effect of tau over FDG (i.e., $b_{Tau}+b_{Int}{X}_{Amyloid}$) depends on, or is conditional on, the amyloid value $X_{Amyloid}$. Similar to our previous metabolic connectivity analysis by Carbonell et al., ¹² we estimate the strength of the association between FDG and tau any particular value s of amyloid by using the t test associated with the “slope” $b_{\mathrm{Tau}}+b_{Int}{X}_{Amyloid}$ at $X_{Amyloid}=s$:

$$t\left(s\right)=\frac{b_{Tau}+b_{Int}s}{\sqrt{{\sigma }_{Tau}^2+2s{\sigma }_{Tau,Int}+{\sigma }_{Int}^2{s}^2}},$$

where ${\sigma }_{Tau}$, ${\sigma }_{Int}$ denote the standard deviation of $b_{Tau}$ and $b_{Int}$, respectively, and ${\sigma }_{Tau,Int}$ denotes the covariance between $b_{Tau}$ and $b_{Int}$.

A third model included local tau measurements ($XSee{d}_{Tau}$) taken from 5 mm seeds centered on areas highly contributing (e.g., local maxima) to the first SVD‐based tau eigenimage:

$$\begin{array}{ccc}\hfill Y_{FDG}& =& b_0+b_{Cov}{X}_{Cov}+b_{Tau}XSee{d}_{Tau}+b_{Apo}ApoE\epsilon 4+b_{Int}XSee{d}_{Tau}\hfill \\ & & \times ApoE\epsilon 4\hfill \end{array}$$ (3)

Our intention here was to reveal local‐to‐distributed patterns on the relationship between (local at the seed level) tau and (distributed) glucose metabolism, compared to the distributed tau measurements produced by the SVD analysis.

Post hoc, two‐tailed Student t tests were performed to assess the main effects of interest and interaction terms in each of the previous models. The voxelwise statistical analysis was performed using an in‐house Python version of the SurfStat toolbox (http://www.math.mcgill.ca/keith/surfstat), in which statistical maps were projected onto the cortical surface for visualization purposes only. The t statistic maps corresponding to each effect of interest were thresholded using the false discovery rate (FDR) procedure (α = 0.05) from Genovese et al. ⁴⁰ to control for multiple comparisons.

3. RESULTS

3.1. Demographics

The subject characteristics analysis from Table 1 revealed no statistically significant associations between APOE ε4 status and sex (χ²  = 2.59, P = 0.11) or age (t = 0.26, P = 0.63). Neither the MMSE (t = 0.23, P = 0.72) nor the ADAS‐Cog (t = 0.89, P = 0.07) showed statistically significant differences between the APOE ε4 statuses. A contingency table analysis within the sample of 116 subjects with FDG, tau, and amyloid PET revealed statistically significant associations between APOE ε4 status and amyloid status (χ²  = 8.61, P = 0.003).

TABLE 1.

Summary of subject characteristics.

	All	APOE ε4 carrier	APOE ε4 non‐carrier	AβL	AβH
Sample size	119	21	98	50	66
Age	72.32 ± 7.81	71.91 ± 8.08	72.41 ± 7.79	70.58 ± 8.88	73.70 ± 6.93
Sex (F/M)	52/67	13/8	39/59	17/33	34/32
MMSE	28.03 ± 1.79	27.95 ± 1.74	28.05 ± 1.81	28.33 ± 1.64	27.83 ± 1.61
ADAS‐Cog	12.66 ± 3.96	11.97 ± 3.20	12.83 ± 4.12	8.62 ± 4.08	10.04 ± 4.40
Aβ (low/high)	50/66	1/19	49/47	50/0	0/66
Global tau SUVR	1.09 ± 0.12	1.13 ± 0.09	1.08 ± 0.12	1.03 ± 0.06	1.14 ± 0.13
Global Aβ SUVR	1.39 ± 0.22	1.49 ± 0.19	1.31 ± 0.21	1.14 ± 0.07	1.49 ± 0.17

Abbreviations: Aβ, amyloid beta; Aβ_H, high amyloid; Aβ_L, low amyloid; ADAS‐Cog, Alzheimer's Disease Assessment Scale Cognitive subscale; APOE, apolipoprotein E; MMSE, Mini‐Mental State Examination; SUVR, standardized uptake value ratio.

3.2. SVD analysis

For the sake of this section, any reference to amyloid data will be understood as an analysis restricted to the sample of 116 subjects with FDG, tau, and amyloid PET images.

By using a leave‐one‐out strategy, individual tau scores were computed one at a time from the SVD analysis of tau and FDG datasets composed of 115 subjects. Correspondingly, the associated spatial loadings were computed as the average of the resulting eigenimages produced during the 116 independent SVDs. The first, second, and third SVD components accounted for an average of 21.49%, 8.81%, and 7.21% of the total co‐variability between the FDG and flortaucipir PET images, respectively.

Figures 1A and B show the average (across 116 leave‐one‐out SVD decompositions) spatial representation of the corresponding spatial loadings for each PET modality. The strongest positive weights in the first tau eigenimage (Figure 1A) correspond to the entorhinal cortex, the lateral inferior temporal gyri, the fusiform gyri, the bilateral angular gyrus, as well as areas of the posterior cingulate cortex. Thus, such regions appeared to be negatively correlated to the strongest negative loads in the FDG eigenimage (Figure 1B), which correspond to the lateral inferior temporal gyri, angular gyri, and small areas within the posterior cingulate cortex. In a similar manner, Figure S1 in supporting information shows the spatial loadings for tau and amyloid corresponding to the second and third SVD components. Notice that the second component (Figures S1A and 1B) highlights the situation in which an increase in tau deposition within areas of the lateral and medial frontal lobules is associated with both a reduction of glucose metabolism in the lateral inferior temporal gyri and an increase of glucose metabolism within the medial fronto‐parietal cortex. The third SVD component reflects (Figures S1C and 1D) that the increase of tau tracer binding within the medial orbito‐frontal cortex and the parahippocampal gyri is associated with an increase of glucose metabolism within the temporo‐occipital cortex. Figure 1C shows boxplots corresponding to the individual tau SVD scores for the two classification groups according to both the APOE ε4 status and the amyloid status. The mean values for the tau SVD scores were SUVR = 1.84 ± 0.38, 1.68 ± 0.38, 1.50 ± 0.24, and 1.86 ± 0.41 for the APOE ε4 carrier, APOE ε4 non‐carrier, Aβ_L, and Aβ_H groups, respectively. Those values resulted in no statistically significant differences between APOE ε4 carrier and non‐carrier (t = 1.72, P = 0.08), but highly significant differences between Aβ_L and Aβ_H groups (t = −5.54, P < 0.001). Figure 1D shows the scatter plot for the first tau and FDG SVD‐based scores, which produced a large correlation value of r = −0.49 (P < 0.001), providing evidence for an overall linear dependency between FDG and tau. Also notice that, when segregated by Aβ, the tau and FDG scores remain strongly negatively correlated. Similarly, Figure 1E shows that the first tau scores were also significantly correlated to the Stat‐ROI measurements (r = 0.53, P < 0.001). The strength and direction of this linear association did not change when the data were segregated by the APOE ε4 status.

FIGURE 1.

The first SVD component from the cross‐correlation between FDG and tau PET images from MCI subjects accounts for 21.49% of the total co‐variability. A, The strongest positive weights in the first tau eigenimage correspond to the medial and lateral inferior temporal gyri. B, The strongest negative loads in the FDG eigenimage correspond to the lateral inferior temporal gyri, angular gyri, and small areas within the posterior cingulate cortex. C, Boxplots for the segregation of the tau scores according to APOE ε4 and global Aβ statuses shows strong statistically significant differences between Aβ_L and Aβ_H groups. D, The scatter plot for the first tau and FDG scores shows large overall and by Aβ group correlation values. E, The strong correlation between the tau scores and the Stat‐ROI amyloid measurements does not change with APOE ε4 status. Aβ, amyloid beta; Aβ_H, high amyloid; Aβ_L, low amyloid; APOE, apolipoprotein E; FDG, fluorodeoxyglucose; MCI, mild cognitive impairment; PET, positron emission tomography; Stat‐ROI, optimal target region of interest; SVD, singular value decomposition.

Similarly, the SVD analysis was carried out between the FDG and the amyloid datasets to produce SVD‐based amyloid scores. The resulting scores and spatial loadings are presented in Figure S2 in supporting information. Notice that the FDG loadings in Figures 1B and S2B show a somewhat different spatial distribution, with the negative loadings in the FDG–amyloid case not as extensive as in the FDG‐tau case. This result demonstrates that the FDG–tau and FDG–amyloid relationships carry underlying spatially distinct patterns of cross‐correlation.

3.3. Voxelwise GLM with SVD scores

The assessment of the main effects of APOE ε4, SVD‐based tau scores, and their interaction effect on FDG SUVR was carried out by statistical inference over the estimated coefficients in Model 1. Although showing an overall trend of relationship with glucose metabolism, the main effect of APOE ε4 did not produce areas of statistical significance after controlling for multiple comparisons (Figure 2A). Here, a full color spectrum for both positive and negative t statistic values in the colormap indicates that there were no areas showing statistical significance differences. Similarly, the interaction effect between APOE ε4 and tau did not produce areas of statistical significance (Figure 2C). In contrast, the increase in tau scores was significantly associated with a reduction in glucose metabolism in several regions, including the precuneus, the posterior cingulate cortex, the lateral inferior temporal areas, the entorhinal cortex, and extended regions of the parietal cortex (Figure 2B). Notice that the colormap has been thresholded to reflect only the statistical significance regions, where the areas of no significance are shown in gray.

FIGURE 2.

Statistical assessment of APOE ε4, SVD–tau scores, and APOE ε4 x SVD–tau on FDG SUVR. A, No region of statistically significant effect APOE ε4 after multiple comparisons FDR thresholding. B, Tau scores were significantly associated with a reduction in glucose metabolism in several regions, including the precuneus, the posterior cingulate cortex, the lateral inferior temporal areas, the entorhinal cortex, and extended regions of the parietal cortex. C, No region of statistically significant interaction between APOE ε4 and SVD–tau on FDG. APOE, apolipoprotein E; FDG, fluorodeoxyglucose; FDR, false discovery rate; SUVR, standardized uptake value ratio; SVD, singular value decomposition.

For comparison purposes, and to put value on the importance of the SVD‐based tau scores in revealing cross‐correlation patterns with FDG, Model 1 was also evaluated with a global tau measurement, $X_{GlbTau}$, taken as the average of the SUVR over the whole cortex, and the SVD‐based amyloid scores $X_{Amyloid}$. Figure 3A shows that, in contrast to the SVD‐based tau scores, the global tau measurement only produced small areas of statistically significant relationships with FDG within the lateral inferior temporal gyri and the parietal cortex. It is not a surprising result given that, by definition, the SVD tau scores would produce a maximal correlation with FDG among the set of all possible tau scores that could be constructed as linear combinations of the original tau measurements.

FIGURE 3.

Global tau and Aβ scores were significantly associated with FDG in small areas the lateral inferior temporal gyri, entorhinal cortex, and the parietal cortex. Aβ, amyloid beta; FDG, fluorodeoxyglucose; SVD, singular value decomposition.

Similarly, Figure 3B shows that the regions of statistically significant FDG–amyloid association were located within the lateral inferior temporal gyri, entorhinal cortex, and the parietal cortex. Notice that, in both cases, the areas of significant relationship with FDG were very localized and not as strong as the overall extent of significant regions observed for the case of the SVD‐based tau score (i.e., Figure 2B).

We subsequently evaluated Model 2 to assess the interactive effects of distributed Aβ and tau on glucose metabolism. Figure 4A shows that the SVD‐based tau scores still produced extended areas of association with the reduction of glucose metabolism after adjusting for the effect of distributed amyloid. In contrast, the main effect of Aβ did not produce areas of statistical significance on FDG SUVR (Figure 4B) after adjusting for the effect of distributed tau. Similarly, the interaction term between Aβ and tau did not show regions surviving the multiple comparisons thresholding (Figure 4C), although the tau‐related reduction of glucose metabolism in areas such as lateral inferior temporal cortex, precuneus, and parietal cortex seems to be relatively associated with an increase of Aβ. Notice that, although not statistically significant, the increase in Aβ seems to be associated with a positive relationship between FDG and tau in large areas of the cortex, particularly within the frontal cortex (Figure 4C). Similarly to Figure 2, the thresholded colormap was used only for the cases in which the t statistics maps showed regions of statistical significance. Figure S3 in supporting information shows how the relationship between glucose metabolism and the SVD‐based tau scores varies with the spatially distributed Aβ scores. Despite the fact that the interaction between amyloid and tau was not statistically significant after multiple comparisons correction, one can observe an overall trend in tau‐related reduction of glucose metabolism associated with the increase in Aβ levels.

FIGURE 4.

Assessment of the interaction Amyloid x Tau effect on FDG PET. A, The SVD‐based tau scores show large areas of association with the reduction of glucose metabolism after adjusting for the effect of Aβ. B, Amyloid did not produce areas of statistical significance on SUVR FDG after adjusting for tau. C, Interaction between Aβ and tau did not show regions surviving the multiple comparisons thresholding. Aβ, amyloid beta; FDG, fluorodeoxyglucose; PET, positron emission tomography; SUVR, standardized uptake value ratio; SVD, singular value decomposition.

To remove any potential bias on the choice of the Aβ measurements, we repeated the previous analysis by using alternative amyloid scores resulting from the SVD analysis between amyloid and tau PET images (as opposed to amyloid and FDG PET images). The result of such an analysis is shown in Figure S4 in supporting information. In this figure, one can still observe extended areas of reduction in glucose metabolism associated with tau scores after controlling for the effects of the alternative Aβ scores (the one from the tau‐amyloid SVD analysis). Similarly, neither the main effect of Aβ nor the interaction between Aβ and tau produced statistically significant results.

3.4. Voxelwise GLM with local tau seeds

Several bilateral seed regions were identified as local maxima in the first average tau eigenimage: entorhinal cortex (with MNI coordinates [−26 −7 −29] and [26 −7 −29]), lateral inferior temporal gyri ([−52 −52 −12] and [52 −52 −12]), dorsal posterior cingulate cortex ([−7 −51 27] and [7 −51 27]), fusiform gyri ([−31 −42 −14] and [31 −42 −14]), or local minima in the first average FDG image: angular gyri ([−49 −57 35] and [49 −57 35]). A detailed seed‐based correlation analysis is presented for two seed regions: (1) the left fusiform gyrus (LFUSI) and (2) the right angular gyrus (RANG), which were regions that showed distinct amyloid–glucose metabolism correlation patterns in our previous study. ¹⁴ Notice that, for each of these seeds’ locations, local tau measurements were computed by taking the SUVR average over voxels within a radius of 5 mm centered on the corresponding anatomical location. Finally, Model 3 was fitted independently for each of the selected tau seed measurements.

Figure 5A and B show the spatial distribution of the tau pattern of spread, or “tau network” associated with the left fusiform and the right angular seeds, respectively. As expected, due to the mathematical properties of Pearson correlation coefficient, strong correlation values were obtained around the seed location. Notice that the seed‐based tau network relative to the LFUSI tau seed (Figure 5A) appears to be strong and spatially extended beyond the fusiform gyrus and covers large bilateral areas of the temporal‐parietal lobes and precuneus. In contrast, the RANG tau network is particularly strong within the RANG itself and within the bilateral precuneus and posterior cingulate cortex. Pearson correlation between local tau and local glucose metabolism (i.e., at the same seed location) in the LFUSI and RANG were r = −0.44 (P < 0.001) and r = −0.11 (P = 0.28), respectively. Table 2 shows the average tau, amyloid, and FDG SUVR values for the 10 seed regions as well as their (local) FDG–tau and FDG–amyloid correlation values. The RANG and LFUSI tau measurements not only show a distinctive local‐to‐local correlation pattern with glucose metabolism, but also a distinct local‐to‐distributed cross‐correlation pattern as well. Indeed, Figure 5C,D show the t statistics maps resulting from fitting Model 3 with LFUSI and RANG seed tau measurements, respectively. Although this figure shows non‐thresholded t statistics maps, the pattern observed with LFUSI tau seed (Figure 5C) is composed of strong (statistically significant) negative correlations with glucose metabolism in the lateral inferior‐temporal gyri, fusiform gyri, angular gyri, and precuneus, which resembles the significant regions obtained with the SVD‐based tau scores (i.e., Figure 2B). As shown in Figure 5D, tau in the RANG seed was weakly related to glucose metabolism and the strongest negative correlations were observed only within the RANG itself and the posterior cingulate cortex. Notice also that the local tau in the RANG seed was associated with an increase of glucose metabolism in areas spread all over the cortex. Notably, the inclusion of SVD‐based Aβ scores in the model (i.e., model 2 with $XSee{d}_{Tau}$ instead of $X_{Tau}$) decreases the strength and spatial extent of the local‐to‐distributed patterns of the local tau–glucose metabolism relationship (see Figure 6). In fact, the stronger negative correlations between local tau and glucose metabolism after controlling for the SVD‐based Aβ scores corresponded to the LFUSI seed (Figures 6A) and are only particularly strong within the left fusiform gyrus itself and the left lateral inferior temporal gyrus. On the other hand, the (weak) negative correlations for the RANG case appear restricted to the left lateral inferior temporal–parietal regions, while much larger areas of positive correlations were enhanced throughout the brain (Figures 6B).

FIGURE 5.

A, B, Spatial distribution of the tau network associated with the left fusiform and the right angular seeds. A, The seed‐based tau network relative to the LFUSI seed appears to be strong and spatially extended beyond the fusiform gyrus and covers large bilateral areas of the temporal‐parietal lobes and precuneus. B, The RANG tau network is particularly strong within the right angular gyrus itself and within the bilateral precuneus and posterior cingulate cortex. C, D, Assessment of the LFUSI and RANG seed tau measurements on FDG PET. C, The LFUSI tau seed shows strong negative correlations with glucose metabolism in the lateral inferior‐temporal gyri, fusiform gyri, angular gyri, and precuneus. D, The RANG tau seed was only weakly related to the reduction of glucose metabolism. FDG, fluorodeoxyglucose; LFUSI, left fusiform gyrus; PET, positron emission tomography; RANG, right angular gyrus.

TABLE 2.

Local tau–FDG and amyloid–FDG correlations for different seed locations.

Seed	Tau SUVR	FDG SUVR	Amyloid SUVR	Tau–FDG correlation	Amyloid–FDG correlation
Left ENT	1.28 ± 0.25	1.15 ± 0.11	1.17 ± 0.15	−0.35	−0.22
Right ENT	1.27 ± 0.25	1.16 ± 0.11	1.17 ± 0.14	−0.38	−0.24
Left INT	1.35 ± 0.34	1.50 ± 0.18	1.50 ± 0.32	−0.45	−0.29
Right INT	1.22 ± 0.33	1.40 ± 0.20	1.19 ± 0.31	−0.27	−0.14
Left PCC	1.25 ± 0.24	1.85 ± 0.24	1.67 ± 0.36	−0.27	−0.19
Right PCC	1.24 ± 0.23	1.86 ± 0.24	1.66 ± 0.35	−0.27	−0.19
Left ANG	1.14 ± 0.28	1.42 ± 0.21	1.29 ± 0.31	−0.28	−0.14
Right ANG	1.13 ± 0.24	1.43 ± 0.23	1.30 ± 0.30	−0.21	−0.05
Left FUSI	1.28 ± 0.26	1.45 ± 0.13	1.33 ± 0.23	−0.44	−0.25
Right FUSI	1.29 ± 0.29	1.45 ± 0.14	1.36 ± 0.23	−0.42	−0.23

Abbreviations: ANG, angular gyrus; ENT, entorhinal cortex; FDG, fluorodeoxyglucose; FUSI, fusiform gyrus; INT, xxxx; PCC, posterior cingulate cortex; SUVR, standardized uptake value ratio.

FIGURE 6.

Assessment of LFUSI and RANG tau on FDG after adjustment of Aβ. A, The LFUSI tau seed shows spatially extended negative correlations between tau and glucose metabolism particularly within the left fusiform gyrus itself and the left lateral inferior temporal gyrus. B, The negative correlations for the RANG tau seed appear restricted to the left lateral inferior temporal‐parietal regions. Large areas of increase in glucose metabolism are associated with RANG tau after adjustment for the Aβ effect. Aβ, amyloid beta; FDG, fluorodeoxyglucose; LFUSI, left fusiform gyrus; RANG, right angular gyrus.

4. DISCUSSION

We have used an SVD‐based voxelwise cross‐correlation analysis to reveal whole brain associations between FDG and tau PET images from the ADNI study. Our analysis also assessed the impact of spatially distributed Aβ scores on the relationship between glucose metabolism and tau. Our analysis highlights that the set of spatially distributed regions implicated with the reduction of glucose metabolism are the areas typically demonstrating abnormal tau deposition during the initial stages of AD (e.g., medial temporal lobe and lateral inferior temporal gyri). The individual SVD‐based tau scores showed a significant association with glucose metabolism. Further, significant associations between the SVD‐based tau scores and glucose metabolism hold after accounting for the remote effect of spatially distributed Aβ. Our seed‐based correlation analysis indicated that the strong association between glucose metabolism and local tau in the left fusiform gyrus was considerably diminished after controlling for the effect of Aβ. Taken together, our findings suggest that widespread Aβ confounds the relationship between tau and glucose metabolism, with the greater impact occurring at local levels of tau compared to more spatially distributed tau beyond the medial temporal lobe.

The SVD‐based tau scores showed a spatial representation mainly characterized by positive high loading values within the entorhinal cortex and the lateral inferior temporal gyri, regions that have consistently shown prominent tau deposition in the initial stage of AD. ²⁴ , ⁴¹ Thus, the fact that the spatial loadings associated with our tau scores follow the same spatial patterns of tau deposition suggests that the initial processes of tau seeding within the entorhinal cortex and spreading to the inferior temporal cortex have an immediate influence on the reduction of glucose metabolism. In contrast, measures of distributed Aβ have a weaker impact on glucose metabolism, and it is only apparent when the associated spatial loadings are widely distributed over the cortex, indicating a more advanced stage of pathologic signs of AD.

While the SVD approach was useful for deriving maximally correlated tau and FDG scores, a confirmatory analysis using GLM was necessary to assess the voxelwise effect of those tau scores and the APOE ε4 genotype on FDG. That modeling approach showed that, in agreement with Strom et al., ³¹ there was no statistically significant effect of the APOE ε4 genotype on glucose metabolism after adjusting for tau. This finding also goes in the same direction as our previous study, in which we demonstrated that spatially distributed Aβ, rather than the APOE ε4 genotype, was related to the reduction of glucose metabolism in MCI. ¹⁴ Given that Sepulcre et al. ⁴² found the APOE ε4 genotype to play a central role in tau propagation patterns, one could envision a mechanism in which the genetic profile of individuals with MCI has only an indirect effect on glucose metabolism. It appears that the main contribution of APOE ε4 can be related to tau propagation, which ultimately has a more direct impact on the reduction of glucose metabolism. More importantly, our results showed that the SVD‐based tau scores were statistically significantly correlated with the reduction of glucose metabolism in extended areas of the cortex after adjusting for the indirect effect of the APOE ε4 genotype. As a basis for comparison, equivalent SVD‐based Aβ scores were only weakly related to glucose metabolism and restricted to focal areas of the entorhinal cortex, lateral inferior temporal gyri, and angular gyri. It is worth noting that the previously mentioned regions are precisely those AD‐vulnerable areas in which tau had the strongest impact on glucose metabolism. However, such a weak association between Aβ and glucose metabolism turned out to be not statistically significant after adjustment for the tau effect. Our results seems to be in correspondence with the causal mediation analysis presented by Bilgel et al., ¹¹ who showed that Aβ has only an indirect effect on glucose metabolism and pointed to tau in the entorhinal cortex and the inferior temporal gyri as the main drivers of neurodegeneration.

In contrast to previous studies from Adams et al. and Rubinski et al. ³² , ³³ that characterized the tau–metabolism relationship using a dichotomous global status (e.g., positive and negative) of Aβ levels, we followed a continuous modeling approach for the distributed scores of Aβ. As pointed out by Carbonell et al., ¹² the dichotomization of Aβ measurements lowers the power to detect true statistical effects and thereby limits our capacity to understand the “continuum” effect of different levels of Aβ on the tau–glucose metabolism relationship. The inclusion of the interaction term Amyloid x Tau in our formulation not only allowed us to assess the synergistic effect of tau and Aβ on glucose metabolism, but also provided a tool to estimate the strength of the tau–metabolism association at any level of Aβ in a continuum. Although not statistically significant after multiple comparisons correction, the spatial distribution of the t values in the Amyloid x Tau term suggest that a gradual deposition of Aβ may contribute to modulation of the spatial association between distributed measures of tau and glucose metabolism. Indeed, the strongest pattern of tau‐related reduction in glucose metabolism corresponded to very high levels of Aβ, which follows the line of similar findings for CN subjects ²⁷ , ³² and AD patients ²³ , ²⁶ , ³¹ with positive Aβ status.

The absence of statistical significance associated with the RANG seed is somewhat of a surprising finding considering that the angular gyri turned out to be regions of prominent reduction in glucose metabolism associated with distributed tau. In combination, these findings reinforce the idea that a tau measure taken from those areas typically associated with initial seeding and spreading of tau pathology would have the greatest role in the reduction of glucose levels in remote neocortical regions that are supposed to be impacted by tau at a later stage. Hence, the tau‐related increases in glucose metabolism in areas such as the angular gyri have been typically interpreted, for example, by Hanseeuw et al., ²⁷ as a compensatory mechanism of increased neuronal activity driven by the impact of tau spreading. Even though the molecular mechanisms causing the tau‐related hypermetabolism at low Aβ levels are still largely unknown, Adams et al. ³² suggested that it might possibly represent a normal aging process. Alternatively, Rubinski et al. ³³ proposed a scenario in which tau enhances neuronal hyperactivity at lower levels of Aβ but reduces neuronal function with the gradual accumulation of Aβ.

In any case, a plausible mechanistic explanation for the uncovered tau–metabolism associations would necessarily involve multi‐factorial, integrative scenarios of the AD progression. From a pure brain connectivity perspective, one must consider the relationship between tau deposition, abnormal structural and functional connectivity, and underlying amyloid‐dependent patterns of spread. ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁶ For instance, modeling approaches by Vogel et al. ⁴⁷ support the notion that tau pathology propagates through neuronal connections, likely with seed origin in the medial temporal cortex, a process that is accelerated under the influence of a gradual increase in Aβ accumulation. On the other hand, given the well‐known strong association among multiple neuroanatomical, neurophysiological, and neurovascular systems as per Iturria‐Medina et al. ⁴⁸ in neurodegenerative disease progression, it is unlikely that pathologic alterations caused by toxic Aβ and/or tau proteins would not alter the brain's metabolic equilibrium or vice versa. For instance, according to Qosa et al., ⁴⁹ disruptions in the neurovascular integrity and neurovascular coupling can affect misfolded proteins clearance, leading to high intraparenchymal Aβ and tau protein concentration levels. Correspondingly, Iadecola ⁵⁰ demonstrated that gradual increases in Aβ and tau may induce blood–brain barrier alterations, thereby damaging endothelial cells, pericytes, vesicular transport, and calcium balance. Hence, the combined effect of the previously mentioned processes would lead to an inevitable alteration in the brain metabolism, producing a multi‐factorial and continuous degenerative cycle.

Despite the finding that our current SVD approach can uncover patterns in the FDG–tau large‐scale association, it is not without several limitations. First, SVD data exploration can only reveal linear associations among pairwise imaging datasets. Indeed, non‐linear–based techniques, such as kernel SVD and/or tensor‐based decompositions, might be used to reveal complex non‐linear associations among multiple modalities within a more theoretical framework. Second, our current approach only allows us to assess the significance of one canonical component at a time, potentially making the biological interpretations a tedious and complex process. However, we consider it acceptable compared to the almost impossible task of exploring and interpreting a full, large‐scale voxel‐by‐voxel cross‐correlation matrix. Thus, our results primarily focus on findings associated with the first principal component only. We acknowledge that a more thorough assessment of the rest of the components might also reveal interesting cross‐correlation patterns between tau and FDG. Finally, our study was designed to explore association among cross‐sectional datasets only. An analysis involving longitudinal datasets would certainly improve our understanding about the spatiotemporal evolution of the FDG–tau and FDG–amyloid associations across the progression of AD.

5. CONCLUSIONS

We have explored whole‐brain associations among glucose metabolism, tau, Aβ, and APOE ε4 genotype in a sample of MCI subjects. Our analysis also assessed the impact of Aβ on the spatial relationship between glucose metabolism and tau. By exploring the large‐scale, cross‐correlation between tau and FDG PET images with the SVD approach, our analysis revealed key findings, including: (1) reduction of glucose metabolism is not associated with APOE ε4 genotype, (2) spatially distributed tau has a much greater role in the reduction of glucose metabolism than Aβ, (3) tau‐related reduction of glucose metabolism was pronounced even after adjustment for Aβ, and (4) local tau in initial seeding regions had a prominent role the reduction of glucose levels in remote neocortical regions. To our knowledge, this study is the first to relate glucose metabolism with tau and Aβ from a whole brain network perspective that accounts for distributed‐to‐distributed and local‐to‐distributed patterns of cross‐correlation.

AUTHOR CONTRIBUTIONS

Felix Carbonell and Barry J. Bedell designed research; Felix Carbonell, Carolann McNicoll, Alex P. Zijdenbos, and Barry J. Bedell performed research; Felix Carbonell and Barry J. Bedell analyzed the data and drafted the paper.

CONFLICT OF INTEREST STATEMENT

Authors Felix Carbonell and Carolann McNicoll are employees of Biospective Inc. Authors Alex P. Zijdenbos and Barry J. Bedell are shareholders of Biospective Inc. Author disclosures are available in the supporting information.

INFORMED CONSENT

Informed consent was obtained from all individual participants included in the study.

Supporting information

Supporting Information

Supporting Information

Carbonell F, McNicoll C, Zijdenbos AP, Bedell BJ. Tau‐related reduction of glucose metabolism in mild cognitive impairment occurs independently of APOE ε4 genotype and is influenced by Aβ. Alzheimer's Dement. 2025;21:e14625. 10.1002/alz.14625