Introduction

As the global population ages, an increasing number of people over 65 will develop dementia due to late onset Alzheimer’s disease (LOAD)1. LOAD is characterized by the development of amyloid-β plaques and neurofibrillary tau tangles that spread from limbic regions to neocortical areas2,3,4. As these pathological processes are thought to accumulate over many years5, it may be possible to identify brain changes related to heightened risk in asymptomatic individuals prior to the onset of memory impairment.

Carriage of the Apolipoprotein E (APOE)-ε4 genotype is the best-established genetic risk factor of LOAD6,7. APOE is the main cholesterol carrier in the brain that supports lipid transport, myelination, synaptic repair and the regulation of amyloid-β aggregation and clearance8. Individuals who carry the APOE-ε4 isoform compared to those with APOE-ε2 and -ε3 show an earlier onset of LOAD6,9 and a larger burden of amyloid-β plaques10,11,12,13,14. Such harmful effects of APOE-ε4 are heightened in individuals with a family history of LOAD15,16, probably due to the presence of other polygenic risk variants such as those of TREM217,18. In addition, APOE-ε4 is known to combine adversely with lifestyle-related risk notably central obesity19,20. Excessive abdominal visceral fat can lead to the metabolic syndrome, type 2 diabetes, and cardiovascular disease21 and obese APOE-ε4 carriers are more likely to develop hypertension, inflammation and insulin resistance22,23.

Much attention has been devoted to characterizing APOE-ε4-related changes in medial temporal lobe regions, notably in the hippocampus and parahippocampal regions24,25,26 due to their importance for episodic memory. Hippocampal volume loss on magnetic resonance imaging (MRI) is also one of the diagnostic biomarkers of LOAD27. However, hippocampal atrophy is lacking in specificity28 and usually occurs in more advanced disease stages29. Indeed, evidence regarding hippocampal atrophy in APOE-ε4 carriers is mixed and is often thought to result from the inclusion of older participants with underlying LOAD pathology30,31. It, therefore, stands to reason that hippocampal volume loss may not be sufficiently sensitive to detect very early disease changes and it has been proposed that focusing on specific hippocampal subregions such as CA1 and subiculum may be more promising32,33. However, it is also possible that limbic regions other than the hippocampus may play an important role in the development of LOAD. Notably, it has been recognised for a while that LOAD pathology may spread through the whole of the Papez circuit and may critically involve the limbic thalamus4. For instance, neurofibrillary accumulations in the anterodorsal thalamic nucleus have been found at the same time as those in the hippocampus in LOAD brains34 and reduced thalamic MRI volume has been observed in amnestic Mild Cognitive Impairment (MCI)35, LOAD36 and presymptomatic presenilin 1 mutation carriers37. Similarly, Positron Emission Tomography (PET) studies have found APOE-ε4 state to accelerate longitudinal reductions in glucose metabolism in the thalamus and frontal, parietal, and posterior cingulate regions in MCI38. Reduced glucose metabolism in anterior and posterior cingulate cortices, retrosplenial, precuneus, parietal cortex, hippocampus and thalamus was also observed in cognitively healthy middle-aged APOE-ε4 carriers39, suggesting that metabolic tissue changes in regions beyond the hippocampus can already occur at asymptomatic stages40.

While PET imaging is sensitive to metabolic changes and can identify amyloid-β and tau burden41, it is invasive and expensive and, therefore, difficult to scale up. Recent advances in non-invasive multi-parametric quantitative MRI (qMRI) methods can reveal subtle microstructural brain changes and promise to provide alternative imaging markers that may be sensitive to early risk-related changes. Up to now qMRI measurements have primarily been studied in LOAD patients and animal models, thus evidence with regards to the effects of risk factors in asymptomatic individuals is sparse.

To address this gap in the literature, we went beyond morphological analyses by employing multi-parametric qMRI to study the effects of APOE-ε4, Family History (FH) of dementia and obesity on cortical and subcortical grey matter in 165 asymptomatic individuals from the Cardiff Ageing and Risk of Dementia Study (CARDS)42,43,44 (Table 1). More specifically we applied indices sensitive to neurite dispersion and density, free water, myelin and cell metabolism from Neurite Orientation Density and Dispersion Imaging (NODDI)45, quantitative magnetization transfer (qMT)46,47,48,49 and T1-relaxometry50 (Table 2).

Table 1 Summary of demographic, genetic, and lifestyle risk information of CARDS participants.
Full size table
Table 2 Overview of the quantitative microstructural indices and their interpretation in grey matter.
Full size table

NODDI fits a three-compartment biophysical tissue model to diffusion-weighted data acquired with a two-shell (b-values of 1200 s/mm2 and 2400 s/mm2) High Angular Resolution Diffusion Imaging (HARDI)51 protocol to separate isotropic from intra- and extracellular diffusion compartments45. This allows the calculation of the isotropic signal fraction (ISOSF), an estimate of free water, and the intracellular signal fraction (ICSF), i.e. the fraction of the tissue comprised of neurites. In addition, NODDI yields the orientation dispersion index (ODI) that reflects the spatial configuration of neurite structures (Table 2). Recent studies reported ICSF and ODI reductions in grey and white matter of patients with MCI, LOAD and young onset AD52,53,54. For instance, Fu et al. (2019) found decreased ICSF and ODI in the corpus callosum in MCI and LOAD patients, while Colgan et al.55 reported positive correlations between ICSF and histological measurements of hyperphosphorylated tau protein in the hippocampus of rTg4510 mice.

The qMT method models the exchange rate between macromolecular protons and protons in surrounding free water when macromolecular protons are selectively saturated by a radiofrequency pulse with a frequency that is off-resonance for protons in free water46,47,48,49. This allows the quantification of a number of parameters including the macromolecular proton fraction (MPF) and the magnetization transfer exchange rate kf 49. In combined neuroimaging and histology studies of Shiverer mice and puppies56,57,58, MPF has been shown to be highly sensitive to the myelin content in white matter such that MPF increases with the amount of myelin. MPF in the anterior hippocampus was also found to distinguish healthy controls from MCI and LOAD patients59. Furthermore, MCI and LOAD patients exhibit a reduced rate of magnetization transfer kf in grey and white matter59,60,61 suggesting reduced cell metabolism60. Finally, indices from relaxometry imaging such as the longitudinal relaxation rate R1 have been proposed as non-invasive biomarkers of LOAD62. R1 values are influenced by microstructural characteristics such as tissue density, macromolecular, protein and lipid composition, and paramagnetic atoms. A number of patient and preclinical studies have reported increases in R1 that may reflect LOAD pathology, although the precise mechanisms underpinning these changes remain unknown (see for review62).

Here, we characterised age and risk-related differences in mean values of ICSF, ISOSF, ODI, MPF, kfand R1 across cortical and subcortical grey matter regions that were segmented from T1—weighted images with the FreeSurfer image analysis suite (version 5.3)63. Microstructural changes were compared with differences in standard morphological metrics of cortical thickness and subcortical volumes. We expected to see risk effects in brain regions known to be early affected in LOAD including limbic regions of the hippocampus, parahippocampus, entorhinal cortex, posterior cingulate cortex as well as thalamus2,4,34,64. We hypothesised that APOE-ε4, a positive FH, and central obesity [measured with the Waist-Hip-Ratio (WHR)] would be associated with reduced ICSF, R1, MPF and kf as well as with increased ISOSF and ODI but with no differences in cortical thickness and/or subcortical volume. In addition, we expected to see the largest differences in those individuals at greatest risk, i.e. in obese APOE-ε4 carriers with a positive FH.

Results

Microstructural and morphological dependent variables were fitted to a general linear model in SPSS version 2665. All data were examined for outliers defined as above or below three times of the interquartile range (75th percentile value–25th percentile value). This led to an exclusion of 0.6% of the microstructural but no exclusions of the morphological data.

Separate multivariate analyses of covariance (MANCOVA) were carried out to test for the effects of APOE genotype (ε4 + , ε4-), FH (FH + , FH-) and WHR (WHR + , WHR-) on brain morphology (cortical thickness and subcortical volume measures) and on each of the microstructural indices (MPF, kf, R1, ISOSF, ICSF, ODI) across 68 cortical and 14 subcortical regions of interest, whilst controlling for age, sex, and IQ estimates from the revised National Adult Reading Test (NART-R)66. Significant omnibus effects were further investigated with post-hoc comparisons across all outcome measures. All first and post-hoc models were corrected for multiple comparisons with a False Discovery Rate (FDR) of 5% using the Benjamini–Hochberg procedure67 (pBHadj). As the aim of the study was to explore microstructural indices that could potentially provide novel biomarkers of dementia risk in future studies, a false positive rate of below 5% was regarded as an acceptable threshold to control for false positives while minimising the risk of missing any true risk-related microstructural differences. Information about effects sizes was provided with the partial eta squared index ηp2 for MANCOVA analyses, Cohen’s dz for group comparisons and Pearson’s r for correlational analyses.

MANCOVAs of microstructural qMT metrics

MPF omnibus effects

There were main effects of sex [F(78,46) = 2.2, pBHadj = 0.015, ηp2 = 0.8] and of APOE genotype [F(78,46) = 2.6, pBHadj < 0.001, ηp2 = 0.8] but not of FH (pBHadj = 0.137), WHR (pBHadj = 0.348), age (pBHadj = 0.385) or NART-IQ (pBHadj = 0.497). There were no interaction effects between APOE and FH (pBHadj = 1.000), APOE and WHR (pBHadj = 0.974), FH and WHR (pBHadj = 1.000) or APOE, FH and WHR (pBHadj = 0.935).

MPF post-hoc effects

APOE-ε4 carriers relative to non-carriers had lower MPF in the left thalamus (Table 3) (Fig. 1). Women had higher MPF than men in the left and right rostral middle frontal cortices, in the left superior temporal cortex and the right transverse temporal cortex (Table 3) (Fig. 2).

Table 3 Post-hoc effects of APOE genotype and sex on the macromolecular proton fraction (MPF).
Full size table
Figure 1
figure 1

Violin plots with overlaid box plots of the difference in the macromolecular proton fraction (MPF) in the left thalamus between APOE-ε4 carriers (n = 57) and non-carriers (n = 97) (pBHadj = 0.026). Boxplots display the median and the interquartile range and violin plots the kernel probability density, i.e. the width of the yellow area represents the proportion of the data located there.

Full size image
Figure 2
figure 2

displays the effects of sex on cortical thickness (CT), subcortical volume (corrected for intracranial volume), isotropic signal fraction (ISOSF) and macromolecular proton fraction (MPF) across 34 cortical regions per hemisphere parcellated with the Desikan–Killiany atlas121 and seven subcortical regions per hemisphere (hippocampus, amygdala, thalamus, caudate, putamen, globus pallidus, nucleus accumbens). Region of interest segmentations were performed with FreeSurfer (version 5.3). Regions are colour-coded according to effect sizes indicated by Cohen’s d126. Warm colours indicate positive and blue colours negative correlations. L = Left, R = Right.

Full size image

R1 omnibus effects

A significant omnibus effect was only observed for APOE genotype [F(82,43) = 2.1, pBHadj = 0.040, ηp2 = 0.08]. No main effects were present for FH (pBHadj = 0.215), WHR (pBHadj = 0.167), age (pBHadj = 0.085) sex (pBHadj = 0.060) or NART-IQ (pBHadj = 0.866) and no interaction effects between APOE and FH (pBHadj = 0.256), APOE and WHR (pBHadj = 0.582), FH and WHR (pBHadj = 0.782) or APOE, FH and WHR (pBHadj = 0.548) were observed.

R1 post-hoc effects

No APOE post-hoc effects survived FDR correction (see Supplementary Table 1).

kf omnibus effects

There were no significant main effects of APOE (pBHadj = 0.813), FH (pBHadj = 0.908), WHR (pBHadj = 1.000), age (pBHadj = 0.075), sex (pBHadj = 0.975) or NART-IQ (pBHadj = 0.870) and no interaction effects between APOE and FH (pBHadj = 0.888), APOE and WHR (pBHadj = 0.840), FH and WHR (pBHadj = 0.090) or APOE, FH and WHR (pBHadj = 0.436).

MANCOVAs of microstructural NODDI metrics

ISOSF omnibus effects

There were main effects for age [F(78,42) = 2.0, pBHadj = 0.03, ηp2 = 0.8], sex [F(78,42) = 3.4, pBHadj < 0.001, ηp2 = 0.9], and NART-IQ [F(78,42) = 2.2, pBHadj = 0.020, ηp2 = 0.8]. No main effects were present for the risk factors of APOE (pBHadj = 1.000), FH (pBHadj = 0.060) or WHR (pBHadj = 0.717) and no interaction effects between APOE and FH (pBHadj = 0.374), APOE and WHR (pBHadj = 0.551), FH and WHR (pBHadj = 0.986) or APOE, FH and WHR (pBHadj = 0.678) were observed.

ISOSF post-hoc effects

Ageing was associated with bilateral increases in ISOSF in medial regions including the cingulate, precuneus and cuneus cortices and in lateral regions including superior temporal, supramarginal, postcentral, pars opercularis and insula cortices. Age-related increases in ISOSF were also observed in left middle temporal and pars triangularis regions as well as in subcortical hippocampi, thalami, nuclei accumbens and right putamen (Table 4) (Fig. 3). Men relative to women had higher ISOSF in widespread frontal, temporal, parietal and cingulate cortices and in caudate nuclei, hippocampi, thalami and right nucleus accumbens (Table 4) (Fig. 2). In addition, NART-IQ correlated positively with ISOSF in the superior temporal sulci (left: r = 0.253, pBHadj = 0.008; right: r = 0.241, pBHadj = 0.006), left superior parietal (r = 0.227, pBHadj = 0.006), and right lingual (r = 0.182, pBHadj = 0.026) cortices (Table 4). After partialling out of age only correlations on the left hemisphere remained significant [superior parietal cortex [(r = 0.206, pBHadj = 0.048), superior temporal sulcus (r = 0.197, pBHadj = 0.032)] but those on the right did not [superior temporal sulcus (pBHadj = 0.053), lingual (pBHadj = 0.08)].

Table 4 Post-hoc effects of age, sex and NART-IQ on the isotropic signal fraction (ISOSF).
Full size table
Figure 3
figure 3

displays the effects of age on cortical thickness (CT), subcortical volume (corrected for intracranial volume), isotropic signal fraction (ISOSF) and orientation dispersion index (ODI) across 34 cortical regions per hemisphere parcellated with the Desikan–Killiany atlas121 and seven subcortical regions per hemisphere (hippocampus, amygdala, thalamus, caudate, putamen, globus pallidus, nucleus accumbens). Region of interest segmentations were performed with FreeSurfer (version 5.3). Regions are colour-coded according to the size of the age effect indicated by Pearson correlation coefficient r. Warm colours indicate positive and blue colours negative correlations.

Full size image

ODI omnibus effects

There was a significant main effect of age [F(78,51) = 2.0, pBHadj = 0.040, ηp2 = 0.8] and a significant interaction effect between FH and WHR [F(78,51) = 2.3, pBHadj = 0.010, ηp2 = 0.8] but no main effects for sex (pBHadj = 0.270), NART-IQ (pBHadj = 0.497), APOE (pBHadj = 0.153), FH (pBHadj = 0.520) or WHR (pBHadj = 0.330) and no interaction effects between APOE and FH (pBHadj = 0.436), APOE and WHR (pBHadj = 0.295) or APOE, FH and WHR (pBHadj = 0.228) were observed.

ODI post-hoc effects

Age-related increases in ODI were observed in left hippocampus, amygdala, caudate and right transverse temporal cortex (Table 5) (Fig. 3).

Table 5 Post-hoc effects of age on the orientation dispersion index (ODI).
Full size table

Post-hoc effects for the interaction between FH and WHR did not survive 5% FDR correction (Supplementary Table 2).

ICSF effects

There were no significant main or interaction effects on ICSF [age (pBHadj = 0.170), sex (pBHadj = 0.130), NART-IQ (pBHadj = 0.451), APOE (pBHadj = 0.324), FH (pBHadj = 0.342), WHR (pBHadj = 0.517), APOE × FH (pBHadj = 0.541), APOE × WHR(pBHadj = 0.236) , FH × WHR (pBHadj = 0.883), APOE × FH × WHR (pBHadj = 0.912)].

MANCOVA on cortical thickness and subcortical volume (ICV corrected)

Omnibus effects

There were main effects for age [F(82,68) = 1.8, pBHadj = 0.035, ηp2 = 0.7] and sex [F(82,68) = 1.9, pBHadj = 0.040, ηp2 = 0.7]. No main effects were observed for APOE (pBHadj = 0.597), FH (pBHadj = 0.144), WHR (pBHadj = 0.152) or NART-IQ (pBHadj = 0.651). No interaction effects between APOE and FH (pBHadj = 0.844), APOE and WHR (pBHadj = 0.978), FH and WHR (pBHadj = 0.053) or APOE, FH and WHR (pBHadj = 0.123) were observed.

Post-hoc effects

Ageing was associated with widespread thinning in bilateral frontal, temporal, and parietal cortical regions as well as with volume loss in subcortical structures, i.e. in the left hippocampus, left nucleus accumbens, bilateral thalami and putamen (Table 6) (Fig. 3). Women relative to men had larger volumes in left hippocampus, left nucleus accumbens, left putamen, right caudate and right pallidum. They also had larger cortical thickness in the right isthmus cingulate but lower cortical thickness in the left insula (Table 6) (Fig. 2).

Table 6 Post-hoc effects of age and sex on cortical thickness and subcortical volume measures.
Full size table

Exploring interaction effects between APOE, age and sex

Potential interaction effects between APOE, age and sex on left thalamus MPF were explored. Univariate analysis of variance revealed an effect of APOE [F(1,141) = 5.7, p = 0.018] and age [F(2,141) = 3.7, p = 0.027] but no interaction effects between APOE and age (p = 0.700) or APOE and sex (p = 0.900).

Exploring moderator effects of blood pressure and markers of inflammation

We then explored with two separate analyses of covariances whether controlling for differences in (i) systolic and diastolic blood pressure (BP) and (ii) inflammation-related measures of C-Reactive Protein (CRP), Interleukin-8 (IL-8) and leptin/adiponectin ratio (LAR) would account for the effect of APOE on left thalamus MPF.

While no covariate showed a main effect [systolic BP (p = 0.680), diastolic BP (p = 0.750), CRP (p = 0.150), IL-8 (p = 0.400), LAR (p = 0.500)], the APOE effect on the left thalamus MPF remained significant [F(1,149) = 6.7, pBHadj = 0.030] after accounting for BP measures, but was not significant anymore after controlling for CRP, IL-8 and LAR (p = 0.060).

Discussion

Here, we investigated whether qMRI indices of apparent neurite density and dispersion, free water, myelin, and cell metabolism were sensitive to grey matter differences related to LOAD risk in cognitively healthy individuals. Such microstructural measurements hold the potential for novel imaging biomarkers to identify asymptomatic individuals at heightened risk of developing LOAD. As such they may provide non-invasive and cheaper alternatives to PET and cerebrospinal fluid (CSF)-based biomarkers, that are currently employed in clinical trials, in the future.

The only significant difference between asymptomatic APOE-ε4 carriers relative to non-carriers was in the qMT measure MPF in the left thalamus with APOE-ε4 related reductions in MPF (Fig. 1). This effect was observed independently of age, sex, and verbal intelligence. Reduced MPF may arise from processes that lead to an increase in free water and/or a reduction in the macromolecular content of grey matter including changes in myelin, proteins, and and/or iron concentrations68,69. Such changes may be consistent with the presence of inflammatory processes leading to tissue swelling associated with glia activation70 and/or with a deficit in cholesterol transport in APOE-ε4 carriers 70,71,72. Consistent with this interpretation we observed that the effect of APOE genotype on left thalamus MPF was moderated by plasma markers of inflammation (CRP, IL-8, LAR). Furthermore, evidence suggests that APOE-ε4 carriage may increase susceptibility to inflammation22,23 and that inflammatory processes contribute significantly to the pathogenesis of LOAD73,74,75.

Notably these APOE-ε4-related differences in MPF were only observed in the left thalamus but not in any other cortical or subcortical region. The limbic thalamic nuclei maintain dense reciprocal connections with the hippocampal formation and the retrosplenial cortex76,77, which, together with the fornix, mamillary bodies and posterior cingulate cortex, comprise the Papez circuit important for episodic memory function78. As outlined above it is increasingly recognised that the Papez circuit, including the anterior thalamus, can be affected early in LOAD4. Neurofibrillary accumulations are found in the anterodorsal thalamic nucleus at the same time as those in the hippocampus in LOAD brains34 and neuroimaging studies have revealed reduced thalamic volume in both amnestic MCI35 and LOAD36. Furthermore, studies into the effects of APOE in middle-aged asymptomatic adults found reduced glucose metabolism in the thalamus, hippocampus and cingulate cortex39 as well as increased metabolism in bilateral thalami and superior temporal gyrus in amyloid-β positive APOE-ε4 carriers with a maternal history of LOAD79. Cacciaglia et al.80 studied the effects of APOE on grey matter volume in over 500 middle-aged asymptomatic individuals and identified reduced hippocampus, caudate, precentral gyrus, and cerebellum volumes but increased volumes in the thalamus, superior frontal and middle occipital gyri in APOE-ε4 carriers. While it remains unknown why APOE-ε4 may be related to increased thalamic volume it was suggested that this could reflect brain swelling associated with glial activation in response to larger amyloid-β burden81. As mentioned above, the here observed pattern of APOE-ε4-related reductions in MPF in the left thalamus is consistent with this interpretation56,82. One other study investigated the impact of APOE-ε4 on qMT white matter metrics in young adults and did not find any differences83. This suggests that such risk-related glial dysfunction may accumulate with age and may only become apparent from midlife onwards.

The question arises why we did not observe any risk-related effects in brain regions that have previously been reported to be affected by LOAD risk factors10,84,85. Reports with regards to the impact of APOE-ε4 on grey matter structures in healthy young and middle-aged adults have been mixed10,84, with some studies reporting no changes in hippocampal grey matter volume in APOE-ε4 carriers31,86. Studies assessing the impact of APOE-ε4 on tissue microstructure have primarily focused on diffusion tensor imaging (DTI) of white matter. While some reported widespread white matter differences in DTI measures83,87,88, this has not been replicated in all studies30,89. These discrepancies may arise due to DTI indices not being sufficiently sensitive and/or specific to detect early risk-related tissue abnormalities90. Direct comparisons between DTI and NODDI indices revealed that although fractional anisotropy (FA) was sensitive to white matter differences between healthy controls and patients with metabolic disease, FA was less anatomically specific and did not identify all brain regions that were captured by ICSF and ODI91. Thus we employed NODDI and qMT measurements to study risk effects on grey matter here and on white matter in a previous CARDS analysis92. In the previously published white matter analysis92 we did not observe any main effects of risk but found that individuals with the highest genetic risk (obese FH + and APOE-ε4) exhibited obesity-related reductions in MPF and ICSF in the right parahippocampal cingulum.

Taken together, our previous and here reported findings demonstrate that MPF from qMT can identify risk-related microstructural differences in limbic grey and white matter that were not apparent in conventional volumetric or cortical thickness measurements. We propose that these differences may reflect subtle changes related to neuroglia activation and that limbic structures including the thalamus are particularly susceptible to adverse effects of APOE-ε4 on glia cells. Inconsistencies in previous studies may have arisen from standard morphological and DTI measurements not being sensitive and/or specific enough to detect such glia-related changes.

It is important to note that while we did not find any risk-related effects on brain morphology we did replicate the well-established pattern of widespread age-related thinning in frontal, temporal and parietal regions93 as well as volume loss in subcortical structures including the hippocampi and thalami (Fig. 3). The subcortical volume loss was accompanied by age-related increases in ISOSF in bilateral hippocampi and thalami but effects on cortical regions were more localised: increased ISOSF was apparent along medial regions of the cingulate and parietal cortices including the precuneus as well as in superior temporal and lateral and orbito prefrontal cortices. Age-related increases in ISOSF have been previously observed94 and most likely reflect lost tissue being replaced by CSF. Consistent with a previous study95 we also observed a positive correlation between age and ODI, an estimate of neurite dispersion, in the hippocampus and the left caudate and amygdala. In contrast to Nazari et al.95 however, we did not find any effects in cortical regions, while they reported reduced ODI with age in fronto-parietal regions. These opposing patterns in cortical and subcortical regions may reflect age-related reductions of neocortical dendritic spine density96 with accompanying compensatory increases in the dendritic extent of dentate gyrus granular cells97,98. Similar age-related increases in the dendritic tree have also been reported in the basolateral nucleus of the amygdala of rats99.

Furthermore, we observed positive correlations between ISOSF and NART-IQ in superior temporal, parietal and lingual cortices that were partly driven by age. NART requires the reading of irregularly pronounced words and older relative to younger adults tended to perform better in the NART. However, positive albeit weak correlations between NART-IQ and ISOSF remained for the left superior temporal sulcus and left superior parietal cortex. Developmental imaging studies have revealed cortical thinning during adolescence100 that may be due to increased myelination100 or synaptic pruning and dendritic arborization101,102. It may therefore be possible that childhood developmental differences in cortical maturation as well as in education may have contributed to this effect. For instance, childhood cognitive abilities have been found to account for relationships between cognitive performance and brain cortical thickness decades later in older adults from the Lothian birth cohort103.

Consistent with previous reports104 we did not observe widespread sex-differences in brain morphology measurements with the exception of larger volumes in the left hippocampus in women than men105. However, qMRI indices revealed the following pattern: Women compared to men, had lower ISOSF in widespread cortical and subcortical regions and larger MPF in frontal and temporal regions. Previously we also reported higher MPF and lower ISOSF for white matter in women than men44. Overall this pattern of sex differences suggests higher cortical myelination and lower free water signal in women as they tended to be overall in better health i.e. were less obese, had lower systolic BP, and reported drinking less alcohol than men44. All of these factors may have contributed to women showing “healthier” grey and white matter in the CARDS cohort.

Finally, some study limitations need to be considered. First of all, CARDS is a cross-sectional study that cannot answer whether the observed APOE effects on left thalamus MPF are predictive of accelerated development of LOAD pathology, cognitive, or neuronal decline. Future prospective longitudinal studies are required to address this question. We also propose that our findings require replication in larger samples that can control for possible interactions between APOE and other LOAD risk genes such as variants of TREM2 and polygenic risk hazards as the number of participants in the CARDS study was too small to do so. It is also worth mentioning that other qMRI measurements, that were not included in the current study, may prove helpful in characterising risk effects on the brain. Notably quantitative T2 and T2* measurements have been proposed to be sensitive to neurodegenerative processes. For instance, prolonged T2 relaxometry has been reported in the hippocampus of LOAD patients106 and has been proposed to increase the sensitivity and specificity of MCI and LOAD detection107. Finally, it should be noted that we only studied the thalamus as a whole structure while neuropathological evidence suggests a specific vulnerability of the anterodorsal thalamic nucleus to LOAD pathology. Future studies may investigate risk-related effects on specific subthalamic nuclei, which was beyond the scope of the current study as we were focusing on risk effects across the whole brain.

In summary, we have shown APOE-ε4 related reductions in the qMT measure MPF in the left thalamus that were moderated by peripheral markers of inflammation. This effect occurred independently of age, sex and NART-IQ and was not observed in morphological or microstructural indices from diffusion-weighted imaging. In addition, the effect was specific to the left thalamus and was not present in other cortical and subcortical grey matter regions. We propose that MPF reductions may reflect the effects of glia-mediated inflammatory and demyelination processes in APOE-ε4 carriers. As such qMT measurements hold the potential for non-invasive and cheaper biomarker alternatives to PET, that may aid our understanding of the pathological processes leading to LOAD. In addition, qMT may help with the identification of asymptomatic individuals at heightened risk of LOAD for stratification into clinical trials for future preventative therapeutics.

Materials and methods

The Cardiff Ageing and Risk of Dementia Study (CARDS) has been described previously including a detailed description of the participant sample43,92, assessment of genetic and metabolic risk factors44,92 and the acquisition and processing of the MRI data43,44,92,108. Here we provide a brief summary of the most important points. CARDS received ethical approval from the School of Psychology Research Ethics Committee at Cardiff University (EC.14.09.09.3843R2) and all participants provided written informed consent in accordance with the Declaration of Helsinki. All research methods were performed in line with Cardiff University’s Research Integrity and Governance Code of Practice and relevant data protection regulations.

Participants

The CARDS cohort comprised 166 community-dwelling individuals between the age of 38 and 71 years who underwent cognitive and health assessment as well as MRI scanning (Table 1). Exclusion criteria were a history of neurological and/or psychiatric disease, head injury, drug/alcohol dependency, high risk cardio-embolic source, large-vessel disease or MRI incompatibility due to pacemaker, stents or other surgical implants. As a group, participants intellectual functioning was above average as assessed with the National Adult Reading Test (NART)66. All but one participant scored > 26 on the Mini Mental State Exam (MMSE)42 thus the remaining 165 participants were classified as cognitively healthy. Eight participants scored ≥ 10 in the Patient Health Questionnaire (PHQ)-9109, suggesting moderate levels of depression but no participant was severely depressed.

Assessment of risk factors

Saliva samples were collected with the Genotek Oragene-DNA kit (OG-500) and APOE genotypes ε2, ε3, and ε4 were determined with TaqMan genotyping of single nucleotide polymorphism (SNP) rs7412 and KASP genotyping of SNP rs429358. Participants self-reported their family history of dementia, i.e., whether a first-grade relative was affected by Alzheimer’s disease, vascular dementia or any other type of dementia.

Central obesity was assessed from the waist-hip ratio (WHR)44 with abdominal obesity defined as a WHR ≥ 0.9 for males and ≥ 0.85 for females. Resting systolic and diastolic blood pressure (BP) readings were taken with a digital blood pressure monitor (Model UA-631; A&D Medical, Tokyo, Japan) and the means of three readings were calculated. Participants self-reported other metabolic risk factors, including diabetes mellitus, high levels of blood cholesterol controlled with statin medication, history of smoking, and weekly alcohol intake. There were only few diabetics, smokers, and individuals on statins and, hence, these variables were not included in the analyses.

Blood plasma analysis

As previously reported44,92, venous blood samples were drawn into 9 ml heparin coated plasma tubes after 12 h overnight fasting and were centrifuged for 10 min at 2000 × g within 1 h from blood collection. Plasma samples were then transferred into 0.5 ml polypropylene microtubes and stored in a freezer at − 80 °C. Circulating levels of high-sensitivity C-Reactive Protein (CRP) in mg/dL were assayed using a human CRP Quantikine enzyme-linked immunosorbent assay (ELISA) kit (R & D Systems, Minneapolis, USA). Six individuals had a CRP value > 10 mg/ml indicative of acute infection and were, therefore, excluded from the statistical analyses testing for moderating effects of inflammation. Leptin concentrations in pg/ml were determined with the DRP300 Quantikine ELISA kit (R & D Systems) and adiponectin in ng/ml with the human total adiponectin/Acrp30 Quantitkine ELISA kit (R & D Systems). Leptin/adiponectin ratios for each participant were calculated. Interleukin IL-8 levels in pg/mL were determined using a high sensitivity CXCL8/ INTERLEUKIN-8 Quantikine ELISA kit (R & D Systems). Determination of interleukin-1β, interleukin-6 and Tumor Necrosis Factor α (TNFα) levels were trialled with high-sensitivity Quantikine ELISA kits but did not result in reliable measurements consistently above the level of detection for each assay.

MRI data acquisition

MRI data were acquired on a 3 T MAGNETOM Prisma clinical scanner (Siemens Healthcare, Erlangen, Germany) as described in43,44,92,108. T1-weighted images (1 × 1 × 1 mm voxel) were collected with a three-dimension (3D) magnetization-prepared rapid gradient-echo (MP-RAGE) sequence (256 × 256 acquisition matrix, TR = 2300 ms, TE = 3.06 ms, TI = 850 ms, flip angle θ = 9°, 176 slices, 1 mm slice thickness, FOV = 256 mm and acquisition time of ~ 6 min).

High Angular Resolution Diffusion Imaging (HARDI)51 data (2 × 2 × 2 mm voxel) were collected with a spin-echo echo-planar dual shell HARDI sequence with diffusion encoded along 90 isotropically distributed orientations110 (30 directions at b-value = 1200 s/mm2 and 60 directions at b-value = 2400 s/mm2) and six non-diffusion weighted scans with dynamic field correction and the following parameters: TR = 9400 ms, TE = 67 ms, 80 slices, 2 mm slice thickness, FOV = 256 × 256 × 160 mm, GRAPPA acceleration factor = 2 and acquisition time of ~ 15 min.

Quantitative magnetization transfer weighted imaging (qMT) data were acquired with a prototype sequence, i.e. an optimized 3D MT-weighted gradient-recalled-echo sequence46 to obtain magnetization transfer-weighted data with the following parameters: TR = 32 ms, TE = 2.46 ms; Gaussian MT pulses, duration t = 12.8 ms; FA = 5°; FOV = 24 cm, 2.5 × 2.5 × 2.5 mm3 resolution. The following off-resonance irradiation frequencies (Θ) and their corresponding saturation pulse nominal flip angles (ΔSAT) for the 11 MT-weighted images were optimized using Cramer-Rao lower bound optimization: Θ = [1000 Hz, 1000 Hz, 2750 Hz, 2768 Hz, 2790 Hz, 2890 Hz, 1000 Hz, 1000 Hz, 12,060 Hz, 47,180 Hz, 56,360 Hz] and their corresponding ΔSAT values = [332°, 333°, 628°, 628°, 628°, 628°, 628°, 628°, 628°, 628°, 332°]. The longitudinal relaxation time, T1, of the system was estimated by acquiring three 3D gradient recalled echo sequence (GRE) volumes with three different flip angles (θ = 3°,7°,15°) using the same acquisition parameters as used in the MT-weighted sequence (TR = 32 ms, TE = 2.46 ms, FOV = 24 cm, 2.5 × 2.5 × 2.5 mm3 resolution). Data for computing the static magnetic field (B0) were collected using two 3D GRE volumes with different echo-times (TE = 4.92 ms and 7.38 ms respectively; TR = 330 ms; FOV = 240 mm; slice thickness 2.5 mm)111. The acquisition time for the complete qMT sequence including all fieldmaps was ~ 30 min.

HARDI and qMT data processing

As described in43,44,92,108, the dual-shell HARDI data were split and b = 1200 and 2400 s/mm2 data were corrected separately for distortions induced by the diffusion-weighted gradients and motion artifacts with appropriate reorientation of the encoding vectors112 in ExploreDTI (Version 4.8.3)113. EPI-induced geometrical distortions were corrected by warping the diffusion-weighted image volumes to the T1—weighted anatomical images114. After pre-processing, the NODDI model45 was fitted to the HARDI data with the fast, linear model fitting algorithms of the Accelerated Microstructure Imaging via Convex Optimization (AMICO) framework115 to gain ISOSF, ICSF, and ODI maps.

Using Elastix116, MT-weighted GRE volumes were co-registered to the MT-volume with the most contrast using a rigid body (6 degrees of freedom) registration to correct for inter-scan motion. Data from the 11 MT-weighted GRE images and T1-maps were fitted by a two-pool model using the Ramani pulsed-MT approximation117. This approximation provided MPF and kf maps. To remove voxels with noise-only data, MPF maps were thresholded to an upper intensity limit of 0.3 and kf maps to an upper limit of 3.0 using the fslmaths imaging calculator from the Functional Magnetic Resonance Imaging of the Brain (FMRIB) library (version 6).

All image modality maps were spatially aligned to the T1-weighted anatomical volume as reference image with linear affine registration (12 degrees of freedom) in within-subject space using FMRIB’s Linear Image Registration Tool (FLIRT)118,119.

Cortical and subcortical grey matter region segmentation

Grey matter cortical and subcortical regions were automatically segmented from T1—weighted images with the Freesurfer image analysis suite (version 5.3), which is documented online (https://surfer.nmr.mgh.harvard.edu/)64. The images were processed by running the “recon-all” script using the default analysis settings. In brief, the images were registered to the Montreal Neurological Institute standard space and intensity normalization was performed. This was followed by automatic skull stripping to remove extracerebral structures, the cerebellum and the brain stem, followed by segmentation into grey matter, white matter and CSF and separation of the hemispheres. Pial surfaces were obtained by tessellating the grey and white matter boundary and by surface deformation following intensity gradients for optimal placement of grey and white matter and grey matter and CSF boundaries120. Surface inflation and registration to a spherical atlas were then performed and the cerebral cortex was parcellated into 34 regions per hemisphere based on gyral and sulcal structures following the Desikan-Killiany atlas121. Cortical thickness measurements were estimated as the average shortest distance between the pial surface and the white matter boundary122. For each hemisphere, seven deep grey matter structures (hippocampus, amygdala, thalamus, caudate, putamen, pallidum, and nucleus accumbens) were automatically parcellated using a probabilistic atlas so that average volumetric measurements could be determined123,124. Mean intracranial volume fractions (ICV) were extracted for each brain as estimates of individual differences in head sizes and all volumetric measurements were adjusted for ICV by dividing each participant’s subcortical volume by their ICV.

Finally, the mean values of all microstructural indices were extracted from each participants’ cortical and subcortical region of interests. Mean measurements were taken in each participants’ native space. This was done by first converting each participants’ cortical and subcortical masks from the FreeSurfer Massachusetts General Hospital volume file format (MGZ) into the Neuroimaging Informations Technology Initiative (NIfTI) analyze-style data format and then uploading the microstructural maps onto each region of interest mask using the fslmaths command from the FMRIB library. Mean values of each index for each mask were then extracted using the FMRIB fslstats command. NODDI and qMT indices of ISOSF, ICSF, ODI, MPF and kf, could not be extracted from bilateral caudal middle frontal, left isthmus cingulate and left pericalcarine regions and R1 could not be extracted from the right postcentral region.