Multimodal investigation of neuropathology and neurometabolites in mild cognitive impairment and late-life depression with ¹¹C-PiB beta-amyloid PET and 7T magnetic resonance spectroscopy

Abstract

Positron emission tomography (PET) and magnetic resonance spectroscopy (¹H-MRS) are complementary techniques that can be applied to study how proteinopathy and neurometabolism relate to cognitive deficits in preclinical stages of Alzheimer’s disease (AD)—mild cognitive impairment (MCI) and late-life depression (LLD).

We acquired beta-amyloid (Aβ) PET and 7T ¹H-MRS measures of GABA, glutamate, glutathione, N-acetylaspartate, N-acetylaspartylglutamate, myo-inositol, choline, and lactate in the anterior and posterior cingulate cortices (ACC, PCC) in 13 MCI and 9 LLD patients, and 13 controls. We used linear regression to examine associations between metabolites, Aβ, and cognitive scores, and whether metabolites and Aβ explained cognitive scores better than Aβ alone.

In the ACC, higher Aβ was associated with lower GABA in controls but not MCI or LLD patients, but results depended upon exclusion criteria. Greater variance in California Verbal Learning Test scores was better explained by a model that combined ACC glutamate and Aβ deposition than by models that only included one of these variables. These findings identify preliminary associations between Aβ, neurometabolites, and cognition.

Graphical Abstract

1. Introduction

Alzheimer’s disease (AD) is the leading cause of dementia and is projected to affect more than 130 million people worldwide by 2050 with costs estimated at 1.3 trillion USD (Prince et al., 2015). Mild cognitive impairment (MCI) is a preclinical stage of AD, wherein individuals exhibit memory deficits without significant impact on activities of daily living (Petersen, 2004). About 50% of individuals with MCI show progression to AD. Depression in late life (LLD) is a major risk factor for MCI (Steenland et al., 2012) and for all-cause dementia (Diniz et al., 2013; Modrego and Ferrández, 2004; Van der Mussele et al., 2014).

While substantial progress has been made in biomarker development, due to the multifactorial nature of cognitive decline, developing biomarkers to identify at-risk individuals and determine the likelihood of cognitive decline involves integrating multiple imaging modalities (Beach et al., 2012; Esposito et al., 2011; Jack et al., 2013). As clinical symptoms are preceded by AD neuropathology (beta-amyloid (Aβ) and tau) by many years, understanding the neurochemical and molecular mechanisms associated with AD pathology has become an increasing focus of study (Bateman et al., 2012; Breijyeh and Karaman, 2020). The National Institute of Aging—Alzheimer’s Association research framework (Jack et al., 2016) classifies AD based on objective biomarker evidence: presence or absence (+/−) of Aβ (“A”), tau (“T”), and neurodegeneration (“N”), highlighting the importance of considering mechanisms of neurodegeneration. This research framework builds upon considerable advances in the development of positron emission tomography (PET) radiotracers to image Aβ (Klunk et al., 2004) and tau deposition (Chien et al., 2013) consistent with the regional distribution of Aβ and tau observed by postmortem studies (Ikonomovic et al., 2008; Smith et al., 2019). This research framework has been applied to MCI to understand the earliest pathology associated with cognitive decline and has made fundamental insights into relationships between Aβ, tau, and neurodegeneration. The greater risk of MCI and all-cause dementia in LLD may be due to the presence of AD pathology. The majority of studies have focused on Aβ and have found mixed results, especially in LLD patients who do not meet criteria for MCI or dementia. An increase in Aβ in LLD patients compared to healthy older adults has been reported in some studies (Kumar et al., 2011; Wu et al., 2014) while other studies have reported decreased Aβ in LLD patients (Mackin et al., 2021) or no difference between groups (De Winter et al., 2017).

Multi-modal neuroimaging methods provide an opportunity to study mechanistic links between several neurobiological processes and a potential pathway from protein accumulation to cognitive decline that may be mediated by mechanisms at the molecular level, e.g. metabolic changes. Specifically, mitochondrial dysfunction, impaired oxygen metabolism, and Aβ and tau deposition interact, resulting in increased oxidative stress (i.e., formation of reactive oxygen species) that impairs cortical and limbic neural circuits that have been implicated in such cognitive domains, including executive function and memory, and in neuropsychiatric symptoms such as depression and apathy (Braskie et al., 2010; Buckner et al., 2005; Butterfield, 2002; Devi et al., 2006; Edison et al., 2007; Lin and Beal, 2006; Reddy and Beal, 2008; Rodrigue et al., 2009; Wang et al., 2014). Proton magnetic resonance spectroscopy (¹H-MRS) is a non-invasive method for the simultaneous in-vivo measurement of endogenous levels of neurometabolites involved in these processes. MRS can measure molecules relevant to AD, including metabolites associated with oxidative stress: glutathione [GSH] (Dedeoglu et al., 2004; Gu et al., 1998; Labak et al., 2010; Ramassamy et al., 2000); anaerobic glycolysis and mitochondrial dysfunction, lactate [Lac] (Liguori et al., 2015; Parnetti et al., 2000; Weaver et al., 2015); neurotransmission: γ-Aminobutyric acid [GABA] (Bai et al., 2015; Oeltzschner et al., 2019; Riese et al., 2015), glutamate [Glu] (Fayed et al., 2011; Haris et al., 2013; Hattori et al., 2002; Oeltzschner et al., 2019; Zeydan et al., 2017), N-acetyl-aspartyl-glutamate [NAAG] (Jaarsma et al., 1994; Olszewski et al., 2017; Passani et al., 1997); neuronal integrity: N-acetylaspartate [NAA] (Adalsteinsson et al., 2000; Gao and Barker, 2014; Marjanska et al., 2005; Marjańska et al., 2014); cell membrane turnover and demyelination: total choline [tCho] (Khomenko et al., 2019; Walecki et al., 2011); and astrocytic dysfunction and microglial proliferation: myo-inositol [mI] (Kantarci et al., 2013; Waragai et al., 2017).

The most consistent MRS findings in MCI are higher mI and lower NAA in the parietal cortex compared with controls, and an association between lower concentrations of these metabolites and increased risk of progression to AD (Kantarci, 2013). Low-concentration metabolites—including GABA and GSH—are difficult to measure at clinical field strength (3T or less). These metabolites have been examined in a few studies that employed high-field MRS or spectral editing techniques (Huang et al., 2017; Mandal et al., 2012; Marjańska et al., 2019; Riese et al., 2015), including our previous study using 7T MRS, which identified lower GABA and Glu in MCI compared to healthy older adults (Oeltzschner et al., 2019). As the parietal cortex is one of the brain regions associated with early neuropathology in MCI and AD, most MRS studies have focused on the posterior cingulate cortex (PCC) and precuneus, while anterior brain regions are also implicated in cognition, mood, and AD pathology (e.g. anterior cingulate cortex, ACC). Studies of the ACC have shown lower NAA and higher GABA and mI in MCI relative to healthy controls (Mihara et al., 2006; Oeltzschner et al., 2019; Shinno et al., 2007). MRS studies in LLD patients compared to controls have shown higher mI and Glx (glutamate + glutamine) (Binesh et al., 2004), lower NAA (Chen et al., 2009; Venkatraman et al., 2009), Cho, and creatine (Venkatraman et al., 2009) in the frontal cortex, lower NAA in the PCC (Smith et al., 2021b), higher Cho and mI in the basal ganglia (Chen et al., 2009), and higher GSH in the ACC (Duffy et al., 2015).

Studies combining PET molecular imaging with complementary MRS measures have suggested associations between Aβ and deficits in neurotransmitters and metabolites in MCI and AD. A structured literature search (specific procedures can be found in the Supplementary Materials) revealed 12 studies that acquired both PET and ¹H-MRS data in MCI, AD, or LLD patients. Six of those studies employed an Aβ PET radiotracer, specifically, four used [¹¹C]-PiB (Chen et al., 2022; Riese et al., 2015; Sheikh-Bahaei et al., 2018; Zeydan et al., 2017) and two used [¹⁸F]-flutemetamol (Spotorno et al., 2022; Voevodskaya et al., 2016). All studies focused on the PCC and included MCI or AD, not LLD patients. All MRS studies were conducted in MR systems operating at 3T or less, with most (8/12) utilizing conventional (not edited or multi-dimensional) MRS, limiting the number of accurately quantifiable signals to tNAA, tCho, mI, Glx, and tCr. Of those studies that specifically reported metabolite-Aβ relationships in MCI/AD, two studies found a negative association between NAA and Aβ (Chen et al., 2022; Sheikh-Bahaei et al., 2018), but two studies reported no significant associations (Riese et al., 2015; Voevodskaya et al., 2016). Positive correlations between mI and Aβ were reported in the posterior cingulate cortex in three papers: in the first study, in a combined group of controls (A−T−N−), early AD (A+T+N−), and late AD (A+T+N+) (Chen et al., 2022; Sheikh-Bahaei et al., 2018); in the second study, across controls, amnestic MCI (Albert et al., 2013) and probable AD (McKhann et al., 2011); in the third study, specifically in Aβ-positive cognitively normal controls (Voevodskaya et al., 2016). No significant relationships were found between PCC Aβ and GABA and Glx (Riese et al., 2015), but Glu/mI was related to Aβ (Zeydan et al., 2017). A more recent study—not captured in our review and outside our focus on MCI and AD—reported APOE4-dependent associations between Aβ burden and grey matter concentrations of Glu and GABA in cognitively healthy older adults (Schreiner et al., 2024).

The goal of the present retrospective, exploratory, cross-sectional study was to investigate relationships between brain metabolites (measured in ACC and PCC with 7T MRS), Aβ deposition (measured by PET and the [¹¹C]-PiB distribution volume ratio, DVR, in these regions), and cognitive measures in MCI and LLD patients, and healthy older adults. DVR measurements of Aβ deposition were extracted from the same VOI as the MRS scans using a new analysis pipeline implemented in Osprey (Oeltzschner et al., 2020). We determined relationships between Aβ and the selected metabolites, specifically GABA, Glu, GSH, NAAG, NAA, tNAA, mI, and tCho, co-registering the MRS voxel to the Aβ PET maps, allowing co-localized, voxel-specific comparisons. Finally, we compared statistical models of associations between Aβ, selected metabolites, and cognitive test scores, and determined whether Aβ PET and MRS measures together were more strongly associated with cognitive performance. We hypothesized that reduced MRS-measured tCr-referenced levels of GABA and Glu in the ACC and PCC would correlate with increased Aβ deposition in the MRS volumes of interest (VOI). We further hypothesized that more variance in cognitive measures would be better explained by multimodal statistical models that included metabolite and Aβ measures than by models that only included Aβ measures.

2. Materials & methods

2.1. Participants

Thirty-five participants were included in this study: 13 healthy controls (7 female, aged 63.6 ± 7.8 years), 13 multiple-domain amnestic MCI participants (3 female, aged 69.6 ± 7.7 years), and 9 LLD patients (4 female, aged 69.7 ± 7.1 years). Table 1 contains the characteristics of the sample population.

Table 1:

Participant sample characteristics for the three groups (Controls, MCI, and LLD). Columns represent the age at time of first scan (years), participant sex (M, F), the time between the PET and MRS scans (months), Mini-mental state exam (MMSE) score, Hamilton depression rating scale (HDRS) score, grey matter fraction of the ACC, and grey matter fraction of the PCC.

	Controls	MCI	LLD
Age (years)	64 ± 8	70 ± 8	70 ± 7
Sex (M‚F)	6‚ 7	10‚ 3	5‚ 4
Time between scans (months)	9 ± 14	3 ± 9	5 ± 3 [days]
Years of Education	16 ± 3	15 ± 3	15 ± 2
MMSE score	29 ± 1	28 ± 2	29 ± 1
HDRS	1 ± 1	4 ± 4	17 ± 4
GM fraction (ACC)	0.8 ± 0.1	0.8 ± 0.0	0.8 ± 0.1
GM Fraction (PCC)	0.8 ± 0.1	0.7 ± 0.0	0.7 ± 0.1

As described previously, participants were recruited from the community or the Johns Hopkins University Alzheimer’s Disease Research Center (ADRC) and were evaluated by a geriatric psychiatrist (Oeltzschner et al., 2019; Smith et al., 2021b). All participants completed psychiatric and cognitive evaluations, including a Structured Clinical Interview for DSM-IV (First, 1997), Clinical Dementia Rating (CDR) scale (Morris, 1993), and Mini-Mental State Examination (MMSE) (Folstein et al., 1975). MCI participants had a CDR global score of 0.5 (mild cognitive impairment), whereas controls and the LLD patients enrolled in this study had a score of 0 (normal). All LLD patients had a DSM-IV diagnosis of major depressive disorder with no antidepressant treatment in the past year. Control participants did not meet DSM-IV criteria for a psychiatric disorder. All participants underwent physical and neurological examinations, laboratory testing, and toxicology screening. Participants were excluded from enrollment who 1) had a history of or active neurological or Axis I psychiatric disorder (including dementia), except for a diagnosis of a current major depressive episode (non-bipolar, non-psychotic) in the LLD patients or mild neurocognitive disorder in the MCI participants; 2) were not medically stable (i.e. poorly controlled medical conditions including hypertension and/or diabetes); 3) had a positive toxicology screening and use of psychotropic drugs or medications with central nervous system effects (e.g. antihistamines, cold medications) within two weeks prior to enrollment; 4) had contraindications for undergoing MRI scans (e.g. pacemaker, metal implants, aneurism clips).

2.2. MRS acquisition

All MR data were acquired using a 7T Philips Achieva scanner (Philips Healthcare, Best, The Netherlands) with a 32-channel Tx/Rx head coil (Nova Medical, Wilmington, MA), as described (Oeltzschner et al., 2019). After collecting a structural T1-weighted image (MPRAGE), MRS voxels with dimensions of 28 mm (anterior-posterior) x 16 mm (left-right) x 20 mm (caudal-cranial) were placed in the dorsal ACC and PCC, centered on the midline with the caudal-cranial edges perpendicular to the body of the corpus callosum (Figure 1A). These regions were chosen because they have high concentrations of Aβ in normal aging and MCI and these are regions that have been targeted in MRS studies of depression (Arnone et al., 2015; Luykx et al., 2012; Maddock and Buonocore, 2012) and MCI/AD, as described above. MRS data were acquired using STEAM localization (TR = 3000 ms; TM = 25 ms; TE = 14 ms (32 voxels), TE = 15ms (28 voxels), TE = 20 ms (10 voxels); number of transients N_av = 96; number of spectral points N_pts = 2048; spectral width = 3 kHz; 2^nd-order pencil beam shim; VAPOR (Tkáč et al., 1999) water suppression; 4 water-unsuppressed transients) (Figure 1B).

Figure 1:

(A) Example voxel placement in one participant for the ACC (green) and PCC (blue) voxels overlaid on their T1-weighted image. (B) Example MR spectra for ACC (green) and PCC (blue) from one participant. (C) Multiple slices of a single subject’s T1-weighted structural scan with distribution volume ratio (DVR) image intensities overlaid. The two MRS voxels are shown in blue and green. (D) Localized PET image intensity histograms with 5-parameter Gaussian fit (distribution center, amplitude, full-width at half-maximum, linear baseline). The center of the fit to the DVR distribution is used as the Aβ deposition metric for further analysis.

2.3. PET acquisition

PET scans were performed to measure Aβ deposition using the radiotracer [¹¹C]-PiB (Klunk et al., 2004) with a second-generation High-Resolution Research Tomograph scanner (HRRT, Siemens Healthcare, Knoxville, TN), a cerium-doped lutetium oxyorthosilicate (Lu25i05[Ce] or LSO) based, dedicated brain PET scanner (Sossi et al., 2005). Dynamic scanning began immediately upon a 15 mCi ± 10% radiotracer injection and lasted for 90 minutes. Voxel-wise distribution volume ratio (DVR) values of [¹¹C]-PiB were calculated using the multilinear reference tissue method with 2 parameters (Ichise et al., 2002) using the cerebellum as the reference region for the input function (Figure 1C), as described previously (Smith et al., 2021c). The DVR measure calculated from dynamic PET scans was used instead of other outcome measures (e.g. the standardized uptake value ratio from a static scan) because of considerations including over/underestimation and test-retest variability, as previously reviewed (Ossenkoppele et al., 2013).

A separate MPRAGE was acquired within a week before the PET scans with a Philips 3 T Achieva MRI instrument with an 8-channel head coil (Philips Medical Systems, Best, Netherlands) and used for the analysis of the Aβ PET scans, as described previously (Smith et al., 2021a, 2017). Furthermore, the MPRAGE acquisitions from the PET and MRS sessions facilitated the co-localization of the MRS voxel and Aβ PET maps for region-specific analyses.

2.4. Cognitive testing

As previously described (Smith et al., 2021a, 2017; Stevens et al., 2022), a comprehensive neuropsychological assessment battery was performed, including measures of executive function, and memory. In the present analyses, we focused on six measures that showed significant differences between MCI and healthy older controls: total number of words recalled (Trials 1–5) and long delay free recall from the California Verbal Learning Test (CVLT) (Delis et al., 2000), total number of words recalled (Trials 1–3) and delayed recall scores from the Brief Visuospatial Memory Test (BVMT) (Benedict et al., 1996), and total number of words recalled on the letter and category fluency from the Delis-Kaplan Executive Function System (D-KEFS) (Delis et al., 2012).

2.5. MRS analysis

2.5.1. Calculation of MRS metabolite levels

Osprey (Oeltzschner et al., 2020) was used for loading and pre-processing of the MR spectra, including a pre-phasing step, and co-registration with the Aβ maps. Data had been exported in SDAT/SPAR format, which was already coil-combined and averaged on the Philips scanner (without alignment of individual transients). Modeling was performed with the built-in LCModel (Provencher, 2001, 1993) binary v6.3 (rather than the native Osprey linear-combination module) to more closely reproduce the methods of the previous study—LCModel v6.3 (Oeltzschner et al., 2019)—with default baseline stiffness (DKNTMN = 0.15). None of the fits exhibited gross structured residuals upon visual inspection. TE-specific simulated basis sets included: alanine (Ala), aspartate (Asp), Cr, GABA, glucose (Glc), Glu, glutamine (Gln), GSH, glycerophosphocholine (GPC), glycine (Gly), Lac, mI, NAA, NAAG, phosphocholine (PCh), phosphocreatine (PCr), phosphoethanolamine (PE), serine (Ser), scyllo-inositol (sI), taurine (Tau), and default-parametrized resonances from lipids (Lip09, Lip13a-d, Lip20) and macromolecules (MM09, MM12, MM14, MM17, MM20), as internally simulated by LCModel, with baseline parameter DKNMNT = 0.15 ppm. In our statistical analyses, we used metabolite level estimates of GABA, GSH, Glu, Lac, NAA, NAAG, tNAA (= NAA + NAAG), mI, and tCho (= GPC + PCh), each stated relative to the internal signal from total creatine (tCr = PCr + Cr). We opted for this reference—rather than unsuppressed water referencing—due to concerns over the influence of potential group differences in ACC/PCC tissue fractions and the assumed water concentrations in those compartments. To check for potential biases in our reference signal, we compared the tCr ratios to unsuppressed water for our three groups. Supplementary Figure A1 shows the resulting box plots. A three-way ANOVA revealed no statistically significant differences across the groups for either region. In addition to the MRS measures evaluated in the prior analysis of the dataset, Lac, tCho, and tNAA are also reported in the current study (Oeltzschner et al., 2020; Smith et al., 2021b).

2.5.2. Determination of Aβ deposition in the MRS VOI

The Osprey processing pipeline automatically generates a binary mask representing the MRS voxel volume, co-registered to the anatomical T₁-weighted MPRAGE image from the 7T MRS session. For this study, we implemented a modified workflow that incorporated the PET images and a second, corresponding, T₁-weighted structural MR image previously acquired at 3T and provided in the same space i.e., already co-registered to the respective PET image. The T₁-weighted images acquired during the 7T MRS session were registered to the corresponding PET-image-aligned T₁-weighted 3T images. The resulting affine transformation matrix was then applied to the Osprey-generated MRS voxel mask, allowing us to mask the PET images and acquire Aβ DVR values from the same VOI as the MRS acquisition.

To measure Aβ deposition in this VOI, we first created a histogram of the individual Aβ DVR inside the MRS voxel mask. We then thresholded this histogram above an image intensity of >0.2 to exclude the pixels with close-to-zero values. We fit a 5-parameter Gaussian model to the remaining intensity distribution with the center, width, and amplitude of the Gaussian as free parameters, as well as a linear baseline (Figure 1D). The center parameter of the Gaussian model was used as the ACC or PCC Aβ deposition DVR for statistical analysis.

2.6. Statistical analysis

2.6.1. Quality control of MRS data

For the MRS data, we implemented a multi-stage outlier rejection approach. First, metabolite fits were examined group-wise, and metabolites that exhibited a median relative CRLB > 20% across any group were excluded from further analysis, a procedure suggested as an alternative to traditional single-datapoint CRLB cutoffs (Joers et al., 2018; Öz et al., 2021). Secondly, metabolite amplitude estimates were tested for significant correlations using the LCModel ‘r’ value, with |r| > 0.7 deemed to be ‘highly correlated’, and therefore not independently reportable. Finally, as previously described (Oeltzschner et al., 2019), we removed outliers for each remaining metabolite using Cook’s mean distance across all participants and brain regions. Individual data points with > 5 times the Cook’s distance were discarded.

2.6.1.1. Alternative quality control approach

In our previous publication (Oeltzschner et al., 2019), we excluded individual data points based on a widely used approach using a relative CRLB cut-off value (15%). We additionally conducted all subsequent analyses with this original CRLB quality control approach to maintain consistency with our previous work and explore analytic variability, i.e., the sensitivity of findings to the choice of statistical methods (Botvinik-Nezer and Wager, 2024; Del Giudice and Gangestad, 2021; Steegen et al., 2016).

2.6.2. Linear regression models

We designed linear regression models to identify group differences in relationships between Aβ deposition, metabolite measures, and cognitive scores. MRS measures of metabolite levels frequently exhibit considerable effects of age, most notably anterior/posterior cingulate GABA (Gao et al., 2013; Porges et al., 2021), and posterior cingulate tCho and mI (Gong et al., 2022), sensorimotor GSH (Hupfeld et al., 2021), among others (Cleeland et al., 2019; Haga et al., 2009). Sex effects of MRS levels are less well established (García Santos et al., 2010; Grachev and Apkarian, 2000), but have been reported specifically in older adults (tCho, NAA in the cingulate) (Sijens et al., 1999). Therefore, to account for their potential influence as confounding factors, we iteratively added fixed effects for age and sex—a common approach in the combined PETMRS literature, in particular when groups were not balanced in terms of age or sex (Chen et al., 2022; Matthews et al., 2021; Sheikh-Bahaei et al., 2018; Spotorno et al., 2022; Voevodskaya et al., 2016). The inclusion of age previously improved the goodness of model fit in our 7T-MRS study (Oeltzschner et al., 2019).

2.6.2.1. Group comparisons

First, we confirmed that the analysis performed with our new Osprey processing software matched the analysis performed with the LCModel software in the same dataset as in the previous study (Oeltzschner et al., 2019) (Supplementary Figure A2), which used a different LCModel version and computational setup. Group comparisons of metabolite estimates were performed as in our two previous studies (Oeltzschner et al., 2020; Smith et al., 2021b), including patient group, brain region, and age as fixed effects. We then ran the same statistical models to examine group differences in regional (ACC, PCC) Aβ deposition.

2.6.2.2. Associations between metabolite levels, amyloid deposition, and cognitive measures

We then examined relationships between MRS metabolite measures, Aβ deposition, and cognitive measures, iteratively adding fixed effects for age and sex as explained above.

First, we tested associations between Aβ deposition and MRS metabolites for each brain region. The ACC and PCC data were independently fit with several models, iteratively including the following fixed effects into the model: (1) no additional fixed effects; (2) group, plus an interaction term between Aβ deposition and group to investigate whether metabolite-Aβ relationships differed between the control and patient groups; (3) age; (4) age and sex; (5) sex.

Next, we separately examined the relationships between cognitive measures (D-KEFS, CVLT, BVMT) and MRS metabolite estimates (GABA, Glu, GSH, NAAG, NAA, mI, tCho, and Lac, as well as between cognitive measures and MRS-voxel-specific Aβ deposition. For the statistical models, the following fixed effects were added iteratively: (1) no additional effects (only cognitive scores vs. MRS measures or MRS-voxel-specific Aβ, respectively); (2) group; (3) group and age; (4) group, age, and sex; (5) group and sex; (6) group plus a metabolite-group (or Aβ-group) interaction term; (7) age; (8) age and sex; (9) sex. This analysis was performed separately for each metabolite and brain region.

Finally, we investigated whether including specific metabolites in a model testing the association between Aβ deposition and cognitive measures would improve the model fit. To limit the number of models, we selected only metabolites that had significant associations with cognitive measures in the previous statistical analysis step—namely, Glu and, under the alternative quality control approach, GABA—and cognitive measures that were associated with both metabolites and Aβ, specifically the CVLT scores.

2.6.2.3. Model performance evaluation

To establish the models that most effectively described the data, we compared several model evaluation metrics—including R² adjusted for the number of model parameters (‘adj. R²’), likelihood-ratio test (LRT), Akaike’s Information Criterion (AIC), and Bayesian Information Criterion (BIC)—in a sample of models comparing metabolite group differences and metabolite-Aβ correlations. An example, comparing PCC GABA models of Aβ, is shown in Supplementary Figure A3. The AIC, BIC, log-likelihood, and adj. R² metrics were highly correlated in the example subset, providing similar appraisals of candidate models. Log-likelihood scores provided the added benefit of significance testing via LRTs, however, their use is typically limited to nested models. Thus, for brevity and flexibility, adj. R² was used as the sole indicator of model performance in this study.

All statistical analysis was performed using R(V4.2.0) (R Core Team, 2016) within RStudio (RStudio Team, 2020), using the LME4 package (Bates et al., 2015) for linear regression analyses. The resulting p-values were considered statistically significant at a single-test alpha level of 0.05. No correction for multiple comparisons was performed. Correlation visualizations were produced using the ggpredict R package (Lüdecke, 2018) to represent conditional effects (predicted values), whereby other factors are held constant at reference levels.

3. Results

3.1. Quality control of MRS data

The individual and mean MR spectra for each patient group are shown in Figure 2A. The spectral quality of the averaged spectra was consistently good across all participants (SNR = 81.1 ± 18.9; Cr FWHM = 15.1 ± 4.2 Hz; water FWHM = 14.4 ± 4.8 Hz) and well within consensus recommendations (Lin et al., 2021). The MRS fit quality was also generally high—an example linear-combination modeling result is shown in Figure 2B. Supplementary Figure A4 shows the CRLB distributions for all considered metabolites, separated by participant group. The only metabolite that exceeded the threshold for quantifiability was Lac, and as such, was excluded from further analysis. Only one pair of individual basis functions exhibited a mean r value greater than the suggested 0.7 cutoff (PCr-Cr: r = –0.79 ± 0.05). The typically-overlapping metabolites that we report were far below this threshold so were considered independent in our analysis (Glu-Gln: 0.10 ± 0.05, NAA-NAAG: –0.29 ± 0.11). Finally, for the remaining metabolites, extreme individual outliers were identified and removed using Cook’s mean distance criterion, resulting in the following number of exclusions: GABA, 1; Glu, 2; GSH, 3; NAA, 2; NAAG, 2; tCho, 3; mI, 2.

Figure 2.

(A) Individual MR spectra (light grey) for both ACC and PCC, mean spectra per group (solid-colored lines), and standard deviation across group (color-shaded ribbons). Visualizations are color-coded for controls (red), MCI (green), and LLD (blue). (B) Example LCModel fit (orange) of a processed MR spectrum (black) with model residual in the top row. The individual metabolite constituent contributions are plotted

3.2. Reproduction of key MRS findings

Following the previous analyses of the MCI and LLD patients (Oeltzschner et al., 2019; Smith et al., 2021b), we compared controls to the patient groups using a linear model that included age (but not sex) as a covariate and did not correct for multiple comparisons. In MCI participants compared to controls, we found lower Glu (estimate = −0.07, p = 0.048) and NAA (estimate = −0.09, p = 0.019) in PCC, and higher mI in ACC (estimate = 0.10, p = 0.008). In LLD patients compared to controls, we found lower NAA in PCC (estimate = −0.16, p < 0.001) and higher mI in ACC (estimate = 0.15, p = 0.001). In contrast to the previous studies—which had excluded individual points based on CRLB—the GABA group differences were non-significant in this analysis. No significant group effects were found for GSH, consistent with our previous findings, or for tCho, which we had not considered in our previous study. Full results of the linear regression models are tabulated in Supplementary Table A1.

3.3. Aβ deposition in the MRS VOIs

Aβ deposition in the ACC and PCC MRS VOIs was, as expected, significantly higher in MCI participants compared to controls regardless of the statistical model (ACC: estimate = 0.31, p = 0.033; PCC: estimate = 0.34, p = 0.013), and was significantly higher for males in all PCC models (estimate = 0.24, p < 0.031). No significant differences in Aβ deposition were found in the ACC and PCC regions between LLD patients and controls. ACC and PCC Aβ deposition distributions are shown in Figure 3. In our study of the larger LLD patient sample, higher Aβ deposition in LLD patients compared to controls was observed in the left superior and inferior parietal lobule, but not the PCC (Smith et al., 2021a). Results of the linear regression models for Aβ deposition are tabulated in Supplementary Table A1.

Figure 3:

Higher Aβ deposition in the ACC and PCC MRS volumes in MCI, but not LLD, compared with controls. Distribution plots of the Aβ distribution volume ratio (DVR) within the MRS volumes are shown for each region. (A) ACC data (B) PCC data. Each panel shows individual data points, whisker plots (representing mean +/− within-group standard deviation), and raincloud plots visualizing the distributions. All visualizations are color-coded for controls (red), LLD (blue), and MCI (green).

3.4. Relationships between metabolites and Aβ

Associations between Aβ deposition and metabolite measures were generally sparse, region-specific, and sensitive to the choice of outlier removal procedure and confounder variables that were iteratively included as additional effects. Modeling results are fully reported in Supplementary Table A2.

3.4.1. GABA

We found no significant associations between GABA and Aβ in the ACC or PCC. However, several significant relationships were observed under the alternative quality control approach (see Section 3.8).

3.4.2. Glutamate

The relationship between Glu and Aβ was complex. We first noted a well-fitting PCC model with a strong Aβ*group interaction effect on Glu for the LLD patients (estimate = −0.55, p = 0.075, adj. R² = 0.171), which was not present in the ACC (estimate = −0.09, p = 0.770), suggesting a specific Aβ-Glu relationship for the LLD patients. Upon closer inspection, with a post-hoc model only including LLD data in the PCC, we indeed identified a strong—albeit non-significant—negative Aβ-Glu correlation (estimate = −0.47, p = 0.162, adj. R² = 0.180) (Figure 4, Supplementary Table A3). Across all groups, the Aβ-Glu relationships were generally non-significant. The only exception was a single weak correlation in the PCC (estimate = −0.11, p = 0.027, adj. R² = 0.099) when sex, but not age was included as an effect.

Figure 4:

Group-specific associations between Glu/tCr and Aβ in the PCC, displayed as predicted metabolite values (conditional effects). The association between Glu/tCr and Aβ deposition suggests a different Glu-Aβ relationship for the LLD group compared with the control and MCI groups. Color-coding: controls (red), LLD (blue), and MCI (green). Shaded regions indicate 95% confidence intervals. The region- and group-specific adjusted-R² and p-values for the Aβ deposition correlations are similarly colored, noted with an asterisk when statistically significant.

3.4.3. Glutathione

Associations between Aβ deposition and GSH were generally weak (all adj. R² < 0.07). We identified a single significant positive GSH correlation with Aβ deposition in the PCC-specific model that included group interaction effects (estimate = 0.09, p = 0.047), although the group effects were not significant in this case. This suggested a GSH-Aβ correlation in controls but not in the patient groups; however, in control-specific post-hoc testing, this correlation was not significant (estimate = 0.09, p = 0.138).

3.4.4. NAA/NAAG/tNAA

No significant results were found for NAA, NAAG, or tNAA.

3.4.5. Choline

All models involving tCho and Aβ deposition exhibited poor overall goodness of fit (adj. R² ≤ 0.21), although a single PCC-specific model hinted at a positive correlation between tCho and Aβ deposition (estimate = 0.03, p = 0.046), this relationship disappeared when sex effects were included—ACC (estimate = 0.03, p = 0.004) and PCC (estimate = 0.02, p = 0.059)—and sex was also the only significant effect in the ACC models for tCho vs. Aβ.

3.4.6. Myo-inositol

No significant correlations with Aβ deposition were found for mI in any of the models. The goodness of fit for the mI models was generally poor (adj. R² ≤ 0.071), except for ACC-specific models which incorporated patient group or a significant but small age effect (estimate = −0.00, p = 0.025 and estimate = −0.01, p = 0.014).

3.5. Relationships between metabolite levels and cognitive measures

We found several significant region-specific correlations between metabolite estimates and cognitive measures. Importantly, however, they were only significant when evaluated across all participants without fixed effects for diagnosis group. They were not significant in models that included fixed group effects and not significant if calculated within a group. Thus, any observed relationships between metabolite levels and cognitive measures appeared to be driven by the effects of the MCI group, and there was no evidence that metabolite levels strengthened associations with cognition. The group-specific means and standard deviations of cognitive scores are reported in Supplementary Table A4 and the results of models including cognitive measures and metabolites are summarized in Supplementary Table A5.

Most notably, we found significant positive correlations between Glu and CVLT long delay free recall in group-pooled ACC models regardless of the inclusion of age and sex (estimate = 15.25–18.77, p ≤ 0.05), and in PCC group-pooled models when age was not included as a covariate (estimate = 14.79–14.96, p ≤ 0.038). A similar pattern is present for Glu and CVLT total recall, albeit less robust to cofounds. However, when diagnosis group effects were included in the models, the effect of MCI diagnosis was highly significant (poorer delayed recall for MCI compared with controls, p = 0.009–0.011) and improved model fit (adj. R² > 0.42), and the correlations between CVLT scores and Glu disappeared (p > 0.06). While ACC Glu was the only instance where effects persisted across all group-pooled models, we observed a similar pattern of group-effect-driven correlation across test scores for tCho in ACC and PCC, and mI in ACC and PCC. However, again, upon the inclusion of group effects (highly significant for MCI), the metabolite vs. cognitive score terms were not significant. Interaction terms between metabolite levels and group did not change that, i.e., relationships between metabolite levels and cognition were not significantly different between groups. We observed no significant correlations between cognitive measures and metabolite estimates in any model for the other metabolites (GABA, GSH, NAA, NAAG, tNAA, mI).

3.6. Relationships between Aβ deposition measures and cognitive measures

Across all participants, greater Aβ deposition in the ACC was significantly associated with lower BVMT total score (estimate = −10.04 – −8.74, adj. R² = 0.143–0.222, p = 0.022–0.045), lower BVMT delayed recall (estimate = −4.58 – −4.17, adj. R² = 0.181–0.294, p = 0.004–0.019), lower D-KEFS category fluency (estimate = −12.36 – −11.54, adj. R² = 0.130–0.210, p = 0.011–0.034), lower CVLT total recall scores in both the ACC and PCC (estimate = −19.90 – −15.17, adj. R² = 0.16–0.265, p = 0.003–0.034). However, like the observations made for Glu, these relationships became non-significant when a fixed group effect was added. That is, again, MCI diagnosis was significantly associated with scores (p < 0.002 for all models) and strongly improved model performance (adj. R2 > 0.449 for all models). In other words, Aβ deposition did little to strengthen associations with cognitive performance in this sample once the main group effects on cognition were accounted for.

Aβ deposition did not show any significant relationship with CVLT long delay free recall in this sample, although including group effects, again, improved the model fit. D-KEFS letter fluency scores were generally very weakly associated with any Aβ model, regardless of whether group, age, or sex effects were included (adj. R² < 0.146 for all models).

All modeling results for Aβ deposition vs. cognitive measures are tabulated in Supplementary Table A6.

3.7. CVLT model fit improvement when combining Aβ and GABA or Glu levels

Since most of the significant Aβ deposition correlations with cognitive measures concerned the CVLT—which also had significant associations with Glu and GABA (with CRLB filtering; see 3.8)—we added these metabolites as additional fixed effects to region-specific Aβ deposition models of CVLT to see if they improved the model fit.

Goodness-of-fit estimates for models with Aβ-only, Aβ deposition and GABA (with and without CRLB filtering), and Aβ deposition and Glu are shown in Supplementary Table A7. We calculated adjusted R² ratios to assess the improvement of model fit (ratio > 1).

We found that Aβ deposition and Glu together almost universally improved model fits for both immediate and delayed CVLT memory measures compared to the respective Aβ-only models. This improvement was generally more pronounced in the ACC, for the CVLT long delay free recall score, and for models that did not include an additional diagnosis group effect. However, even when diagnosis group effects were already included, adding Glu to the ACC model of CVLT long delay free recall still improved the adjusted R² by almost 20%.

3.8. Effects of different quality control procedures

3.8.1. Excluded datasets under alternative quality control approach

When using the individual-CRLB-based exclusion criterion we had used in our previous work, a greater number of datasets were excluded:

GABA had 5 additional exclusions (6 vs. 1), NAAG had 28 additional exclusions (30 vs. 2), and Lac results were entirely excluded here but previously had 11 exclusions. Results for Glu, GSH, NAA, tNAA, tCho, and mI were unchanged; while results for GABA, NAAG, and Lac depended on the choice of the quality control procedure. There were no relevant significant effects of NAAG or Lac in any analysis.

3.8.2. Group comparisons

For GABA, this analysis reproduced our originally observed significant group effects of MCI, i.e., lower GABA in ACC (estimate = −0.08, p = 0.006) and PCC (estimate = −0.05, p = 0.040), as well as of LLD, i.e., lower GABA in ACC (estimate = −0.07, p = 0.023).

3.8.3. Relationships between metabolites and amyloid

The most notable finding occurred for the ACC model of GABA which also included Aβ deposition and Aβ*group. This model had a significant Aβ*MCI interaction (estimate = 0.25, p = 0.042, adj. R² = 0.361), indicating that the relationship between Aβ deposition and GABA in the ACC differed for MCI participants compared with controls. Indeed, when examining each group separately, there was a significant negative correlation between Aβ deposition and GABA in controls (estimate = −0.19, p = 0.016, adj. R² = 0.589), but not in MCI (estimate = 0.06, p = 0.126, adj. R² = 0.159) or LLD (estimate = −0.15, p = 0.309, adj. R² = 0.033) (Figure 4A, Supplementary Table A3). A negative Aβ-GABA relationship in the PCC was significant only if sex (but not diagnostic group or age) was included as the sole additional fixed effect (estimate = −0.06, p = 0.016; adj. R² of overall model = 0.147). All other models for Aβ deposition vs. GABA had very poor goodness-of-fit for the overall model (all adj. R² < |0.09|). In summary, in the ACC, higher Aβ is associated with lower GABA in controls but not MCI or LLD patients.

3.8.4. Relationships between metabolite levels and cognitive measures

Most notably, we found significant positive correlations between CVLT long delay free recall and GABA in the group-pooled ACC model (estimate = 22.63–24.77, adj. R² = 0.141–0.187, p < 0.05; regardless of whether age and/or sex were included as effects); that is, higher ACC GABA correlated with better delayed-recall performance in the CVLT. However, when diagnosis group effects were added, the effect of MCI was highly significant (poorer delayed recall for MCI compared with controls, p = 0.002–0.011) and improved model fit (adj. R² > 0.39), while the correlations between CVLT long delay free recall and GABA disappeared (p > 0.34).

3.8.5. CVLT model fit improvement when combining amyloid and GABA or Glu levels

In contrast to the universal improvement when including Glu, the inclusion of GABA tended to improve only those models without diagnosis group effects, but worsened model fits for those with group effects. In other words, adding GABA did not explain more variance in the CVLT better than Aβ-only models when MCI vs. control diagnosis was already accounted for.

4. Discussion

In this study, we investigated relationships between brain metabolites and Aβ deposition in healthy older adults and two patient groups with increased risk of developing dementia—mild cognitive impairment (MCI) and late-life depression (LLD). This study differs from previous studies based on the use of 7T MRS, quantifying several low-concentration metabolites (e.g. GABA), and targeting the ACC in addition to PCC.

4.1. Metabolites and Aβ deposition in MCI and LLD

Group comparisons for metabolite levels and Aβ deposition largely reproduced our previous analyses (Oeltzschner et al., 2019; Smith et al., 2021b; Stevens et al., 2022) and are in line with other published findings. Lower NAA and higher mI are consistently found in MCI participants compared with healthy older adults (Gao and Barker, 2014; Kantarci et al., 2013), to the extent that the NAA/mI ratio has been suggested as a potential early biomarker of MCI (Liu et al., 2021; Waragai et al., 2017). We also identified lower Glu for MCI participants (Hattori et al., 2002; Zeydan et al., 2017) and lower NAA for LLD patients compared with controls (Auer et al., 2000; Elderkin-Thompson et al., 2004). Crucially, we observed decreased GABA levels in MCI participants (ACC, PCC) and LLD patients (ACC) only when we applied the conservative single-datapoint CRLB quality control we had used previously; but this was not significant (despite considerable effect size) when we applied a more liberal criterion (using the median CRLB to exclude individual metabolites).

Higher Aβ deposition in ACC and PCC in MCI and AD is consistently reported (Rabinovici and Jagust, 2009; Ziolko et al., 2006), and consistent with our findings. Greater Aβ deposition in LLD patients is not consistently observed. In a larger group of LLD patients, including the patients enrolled in the present study, higher Aβ deposition was observed in LLD patients in the left superior and inferior parietal lobule (Smith et al., 2021a) It is important to note that the MRS voxel placement was not informed by the regional Aβ deposition in the LLD patients or MCI participants enrolled in the study and thus, did not overlap entirely with the MRS voxels.

4.2. Associations between Aβ deposition and GABA

Like the group comparisons, the most prominent findings in our analysis of relationships between MRS metabolites and Aβ deposition were sensitive to the choice of CRLB quality criterion. With the more conservative approach we had used in our previous analysis, we made several interesting observations of GABA in the ACC that were not reproduced when we used the more liberal CRLB criterion (Joers et al., 2018; Öz et al., 2021). Specifically, group-pooled models of GABA vs. Aβ deposition were not statistically significant, but adding group interaction terms and subsequent post-hoc analyses showed a negative correlation between GABA and Aβ deposition in controls that is not present in MCI. The lack of a significant finding was surprising, given that MCI participants had significantly lower GABA and higher Aβ deposition than controls in the ACC. This constellation of findings may reflect an ongoing parallel GABA decline and Aβ deposition increase in controls, suggesting a connection between proteinopathy and declines in GABA availability and/or neurotransmission through GABAergic neurotransmission; it might suggest that this parallel decline does not continue further once MCI has become manifest. Previous work has shown declining GABA with age independent of Aβ (Hone-Blanchet et al., 2022), and one study has previously examined the relationship between GABA and Aβ deposition in MCI (Riese et al., 2015). However, this study only considered the PCC, where lower GABA was found in amnestic MCI compared with healthy older adults, but GABA was not correlated with Aβ. The absence of a robust PCC GABA-Aβ deposition correlation in our data (except when a sex effect was added) supports these findings and suggests that connections between proteinopathy and GABA levels may be region specific. Overall, the sensitivity of our findings to the quality control procedure suggests substantial analytic variability and requires further investigation, ideally with greater sample sizes.

4.3. Associations between Aβ deposition and Glu

Similarly, our data suggested group-specific interactions between Aβ deposition and Glu levels in the PCC; although non-significant in the post-hoc analysis, a negative association between Aβ deposition and Glu appeared to apply solely to the LLD patients. We found no other published data reporting interactions between Glu and Aβ deposition in LLD patients. The absence of a relationship between PCC Glu and Aβ deposition in our MCI data is in line with a 3T MEGA-PRESS report of no significant differences in Glx levels for MCI participants compared with controls (Riese et al., 2015) and another study examining cognitively unimpaired older adults (Kara et al., 2022). Another study reported a weak negative correlation between the ratio of Glu/mI and global cortical Aβ deposition across a cohort of controls and MCI participants (Zeydan et al., 2017); the authors suggested this reflected the neuronal loss accompanying increasing amyloid burden and/or specific glutamatergic impairment in MCI.

4.4. Associations between Aβ deposition and other metabolites

Prior studies that performed Aβ PET and MRS in MCI participants have found negative correlations between Aβ deposition and NAA and positive correlations between Aβ deposition and mI in the PCC (Chen et al., 2022; Sheikh-Bahaei et al., 2018). Our analysis did not provide evidence to support these findings for mI, NAA, or tNAA in either brain region. However, it has been suggested (Kantarci et al., 2007) that tNAA levels steadily decline over time from the prodromal stages of AD. In contrast to our MCI-only group, both prior PET-MRS studies included a larger cohort of (early and/or late) AD patients and did not stratify their correlation analysis by group, which therefore included a greater dynamic range of NAA, mI, and Aβ deposition compared with our analysis. Although links between protein agglomeration, neuronal activity, oxidative stress, and impaired mitochondrial glucose turnover have been proposed (Bero et al., 2011; Mosconi, 2005; Mosconi et al., 2008; Mullins et al., 2018; Weaver et al., 2015; Yamada et al., 2014), we did not find any evidence that higher Aβ deposition was associated with differences in Lac or GSH levels (which are markers related to cellular energy function and antioxidant availability).

4.5. Associations between cognitive measures and metabolites/Aβ

We previously found associations between CVLT scores and PCC Glu, PCC NAA, and mI in the ACC and PCC, as well as a DKEFS category-fluency relationship with PCC GSH for MCI (Oeltzschner et al., 2019). In the present study—which, in contrast to our prior work (Oeltzschner et al., 2019) included an LLD group as well as the MCI group—we only found associations between ACC and PCC Glu and both CVLT scores, while the GSH-DKEFS correlation narrowly missed significance. However, in general, most of the identified metabolite-cognition correlations were lost when a fixed group effect was added to the model. That is, most of these relationships appeared to be driven by the (per definition) pathological cognitive performance of the MCI cohort. We observed a similar pattern for the Aβ deposition models of cognition in both the ACC and PCC, with significant correlations largely explained by group differences, suggesting that no single MRS or Aβ measure was sufficiently associated with cognitive measures. It is likely that the sample sizes in the groups were too small to detect group-specific associations between metabolite levels and cognitive measures.

4.6. Glu improved model fits of cognitive measures compared to Aβ-only models

Interestingly, CVLT scores were more strongly associated with—i.e., achieved greater adjusted R²—a model that combined ACC Glu and Aβ deposition than by models that only included one of these variables. This held true even when a diagnosis group effect was included, despite correlations between CVLT and Glu themselves appearing to be primarily driven by diagnosis group effects, as described above (i.e. due to inherent differences between groups, see section 4.5). Decreased Glu levels have previously been demonstrated with MRS in MCI (Zeydan et al., 2017) and AD (Fayed et al., 2011). It has been suggested that Aβ-induced downregulation of postsynaptic glutamatergic receptors (e.g., α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptors) contributes to dysfunction in synaptic plasticity that may ultimately contribute to cognitive decline (Fayed et al., 2011; Liu et al., 2010). Changes in receptor sensitivity and/or density are not observable by MRS (Stagg, 2014), though, and the relationship between glutamatergic receptor expression, actual glutamatergic signaling, and MRS-derived Glu levels is unclear. Notably, including GABA did not provide a similar model fit improvement, although it was one of the few metabolites that exhibited associations with cognition, at least under the more conservative CRLB exclusion criterion. Due to the low sample size and high model complexity, these findings should be considered preliminary, warranting further investigation.

4.7. Limitations and perspectives

The small size (13 controls, 13 MCI, and 9 LLD patients) and imbalance (3/10 female MCI participants compared to 7/13 healthy controls and 4/9 LLD patients) of the sample are significant limitations due to the complexity of the statistical analysis required for a multi-modal study with many potential confounders.

First, to gain a complete picture of possible metabolite, Aβ relationships, and confounders, we tested many different models without correcting for multiple comparisons. Second, we employed two possible quality control criteria based on MRS data modeling uncertainty (CRLB). This decision influenced our findings, particularly for GABA. Using hard cut-offs for relative CRLBs identifies (and removes) individual datasets with low measurement confidence. Although bias towards higher concentration estimates has been acknowledged as a weakness of this practice (Kreis, 2016; Öz et al., 2021), it remains widely used. The (more liberal) criterion that we designed based on group median CRLB represents a different “ethos”, opting to judge whether metabolites can be estimated reliably in a cohort rather than in an individual. This avoids the pitfalls of excluding single datapoints, albeit at the cost of potential vulnerability to including genuinely uncertain measurements, which can also precipitate as high CRLB.

Both aspects—modeling and outlier removal—exemplify sources of analytical variability, i.e., the “ability to identify a finding consistently across variation in [statistical] methods” (Botvinik-Nezer et al., 2020; Simonsohn et al., 2020). While our study identified several novel relationships between metabolites and Aβ, further study is required to substantiate them, e.g., with greater sample size or methods like multiverse analyses (Dafflon et al., 2022; Del Giudice and Gangestad, 2021; Steegen et al., 2016) which consider (and weigh) multiple admissible statistical models and isolate the most influential variables. These methods may further allow the incorporation of other factors that may influence the PET and MRS measures, for example, years of education, apoe4 carrier status, and the time between the PET and MRS scans. Finally, a logical next step would be to perform a longitudinal study to elucidate the time course of Aβ-metabolite relationships and to evaluate whether early metabolite level changes are associated with subsequent cognitive decline in MCI and LLD patients.

The MRS analysis was performed using on-scanner-averaged SDAT data. While the relatively short acquisition time will limit frequency and phase drift during acquisition, only the pre-averaged data were available, which prevented us from performing shot-to-shot drift correction.

In-vivo MRS suffers from a relatively low spatial resolution and usually requires voxels of mixed tissue composition (including both grey and white matter). Furthermore, previous work has shown grey/white matter differences in MRS-measured metabolite estimates (Hui et al., 2024; Krukowski et al., 2010; Pouwels and Frahm, 1998; Safriel et al., 2005). For Aβ imaging, grey/white-matter contrast is dependent on disease progression (Chapleau et al., 2022) and [¹¹C]-PiB, specifically, has been shown to have a binding affinity for myelin-rich white matter (de Paula Faria et al., 2014; Stankoff et al., 2011), further confounding mixed-tissue studies. Although we did explore the feasibility of a tissue-specific Aβ metric in this work, our Gaussian modeling approach was less robust in the grey/white-matter-specific analyses. To avoid further data exclusions, we opted for the mixed-tissue model at the cost of specificity. White matter “bleed-in” is therefore a potential confound of this study. Fortunately, segmentation of structural T₁-weighted images revealed no statistically significant group differences in the tissue composition of the MRS voxels, which should limit the impact of groupwise differences. Future MRS studies could utilize MRSI or smaller voxel sizes to increase tissue specificity (Schreiner et al., 2024), but this comes at the cost of reduced spectral quality and efficacy of the voxel localization. An alternative approach could be to do the opposite and increase the MRS voxel size to encompass both the PCC and the precuneus—both implicated as early sites of Aβ deposition—this might provide improved SNR to both Aβ and MRS measures if sufficient field homogeneity can be maintained.

Reliable detection of J-coupled metabolites—like Glu and GABA—necessitate specialized MRS acquisitions. GABA detection often requires spectral editing at lower field strength (≤3T), and Glu requires higher field strengths to separate from Gln (usually reported together as Glx). The increasing popularity of spectral editing at 3T, the increasing availability of 7T scanners, and the standardization of MRS methodology should improve access to these methods for future neuroimaging studies of aging.

Conclusion

In this study, we demonstrated region- and group-specific associations between low-concentration metabolites (GABA, Glu) and Aβ deposition in controls, MCI participants, and LLD patients, although findings were sensitive to the choice of quality control criteria. Furthermore, we gained evidence that incorporating Glu levels in statistical modeling improved how well cognitive outcomes could be explained, compared to Aβ deposition alone. These findings may help identify mechanistic links between AD neuropathology and cognitive decline, although larger sample sizes will be required to substantiate our findings.

Figure 5:

Group-specific associations between GABA/tCr and Aβ in the ACC, displayed as predicted metabolite values (conditional effects). (A) Without individual CRLB exclusions: Differences between groups and relationships within group are not significant (B) With individual CRLB exclusions: A significant control-specific correlation between GABA and Aβ is found and is not present in MCI. Color-coding: controls (red), LLD (blue), and MCI (green). Shaded regions indicate 95% confidence intervals. The region- and group-specific adjusted-R² and p-values for the Aβ deposition correlations are similarly colored, noted with an asterisk when statistically significant.

Highlights

• A study of mild cognitive impairment, late-life depression, and healthy controls.

• We examined beta-amyloid (Aβ) and neurometabolites using PET and 7T MRS.

• We report novel associations between brain metabolites and Aβ.

• Glu + Aβ predicts verbal learning scores better than Aβ-only models.

Verification

This work has not been published previously, is not under consideration for publication elsewhere, is approved by all authors and tacitly or explicitly by the responsible authorities where the work was carried out, and it will not be published elsewhere in the same form, in English or in any other language, including electronically without the written consent of the copyright-holder.

Abbreviations:

¹H-MRS

Proton magnetic resonance spectroscopy

ACC

anterior cingulate cortex

AD

Alzheimer’s disease

adj. R²

adjusted-R²

ADRC

Alzheimer’s Disease Research Center

Ala

alanine

Asp

aspartate

ATN

Aβ-tau-neurodegeneration

Aβ

beta-amyloid

CDR

Clinical Dementia Rating

Cr

creatine

CRLB

Cramer-Rao Lower Bound

CVLT

California Verbal Learning Test

BVMT

Brief Visuospatial Memory Test

D-KEFS

Delis-Kaplan Executive Function System

DVR

distribution volume ratio

VOI

volume of interest

FWHM

full-width at half maximum

GABA

γ-Aminobutyric acid

Glc

glucose

Gln

glutamine

Glu

glutamate

Glx

glutamate + glutamine

Gly

glycine

GPC

glycerophosphocholine

GSH

glutathione

Lac

lactate

Lip

lipids

LLD

late-life depression

MCI

mild cognitive impairment

mI

myo-inositol

MM

macromolecules

MMSE

Mini-Mental State Examination

MPRAGE

3D magnetization-prepared rapid gradient echo

NAA

N-acetyl-aspartate

NAAG

N-acetyl-aspartyl-glutamate

PCC

posterior cingulate cortex

PCh

phosphocholine

PCr

phosphocreatine

PE

phosphoetanolamine

PET

positron emission tomography

Ser

serine

sI

scyllo-inositol

SNR

signal-to-noise ratio

STEAM

stimulated echo acquisition mode

tau

tau protein

Tau

taurine

tCho

total choline GPC + PCh

tCr

total creatine Cr + PCr

tNAA

total N-acetylaspartate NAA + NAAG

VAPOR

variable pulse power and optimized relaxation delays