Fornix white matter glia damage causes hippocampal gray matter damage during age-dependent limbic decline

Aging leads to gray and white matter decline but their causation remains unclear. We explored two classes of models of age and dementia risk related brain changes. The first class of models emphasises the importance of gray matter: age and risk-related processes cause neurodegeneration and this causes damage in associated white matter tracts. The second class of models reverses the direction of causation: aging and risk factors cause white matter damage and this leads to gray matter damage. We compared these models with linear mediation analysis and quantitative MRI indices (from diffusion, quantitative magnetization transfer and relaxometry imaging) of tissue properties in two limbic structures implicated in age-related memory decline: the hippocampus and the fornix in 166 asymptomatic individuals (aged 38–71 years). Aging was associated with apparent glia but not neurite density damage in the fornix and the hippocampus. Mediation analysis supported white matter damage causing gray matter decline; controlling for fornix glia damage, the correlations between age and hippocampal damage disappear, but not vice versa. Fornix and hippocampal differences were both associated with reductions in episodic memory performance. These results suggest that fornix white matter glia damage may cause hippocampal gray matter damage during age-dependent limbic decline.

class of model reverses the direction of causality, stating that the normal aging process in interaction with genetic and lifestyle risk factors of LOAD will cause neuroglia damage, resulting in impaired myelination, reduced microglia-mediated clearance and neuroinflammation, which in turn instigates pathological processes that lead to abnormal metabolism in key proteins and neuronal death [13][14][15][16] . The two classes of models predict opposite patterns of age and risk-related changes in white and gray matter. While in neurodegenerative models, white matter damage follows gray matter neuronal loss, the neuroglia model predicts that white matter neuroglia damage causes gray matter tissue loss.
Currently, the nature of age-related white and gray matter tissue changes and the causal direction between them remain unknown. This study therefore investigated the impact of age and risk on quantitative MRI indices of apparent neurite and glia properties and discriminated between the two types of models using linear mediation analysis [17][18][19] . More specifically, we investigated the effects of aging and risk factors on tissue properties of two important gray and white matter structures involved in age-related memory decline, the hippocampus 20 and the fornix 21 , in 166 asymptomatic individuals (39-71 years old) ( Table 1). We then tested two mediation models displayed in Fig. 1: Model A assessed whether the inclusion of hippocampal variables mediated the direct effect of age on the fornix, whilst Model B tested whether fornix variables mediated age effects on the hippocampus. If full mediation of age effects was observed in Model A but not in Model B, then this would suggest that hippocampal gray matter differences were causing white matter fornix differences, consistent with neurodegenerative models of aging. However, if full mediation occurred in Model B but not Model A, then this would suggest that age-related fornix differences were causing hippocampal gray matter differences, consistent with neuroglia models of aging.
NODDI yielded the intracellular signal fraction (ICSF), an index of apparent neurite density, the neurite orientation dispersion index (ODI), and the isotropic signal fraction (ISOSF), an estimate of free water contribution to the diffusion signal. Diffusion MRI metrics are well-known measures of white matter microstructure [31][32][33] and NODDI ICSF and ODI have been proposed to be particularly sensitive to axon density and dispersion 34,35 . In addition, we used qMT, a technique that allows the quantification of differences in brain macromolecular density, and provides indices with improved white matter glia specificity compared to diffusion MRI [27][28][29][30] . In white matter, magnetization transfer is dominated by myelin 36,37 , and is sensitive to microglia-mediated inflammation 28,38,39 . We measured the macromolecular proton fraction (MPF) as an index of white matter neuroglia that will largely   (A) MRI modalities and maps acquired from dual-shell high angular resolution diffusion imaging (HARDI), quantitative magnetisation transfer (qMT) imaging and from T 1 and T 2 relaxometry. HARDI data were modelled with neurite orientation dispersion and density imaging (NODDI), yielding maps of intracellular signal fraction (ICSF), isotropic signal fraction (ISOSF) and orientation dispersion index (ODI). The qMT maps were the macromolecular proton fraction (MPF) and the forward exchange rate k f . Maps from relaxometry were the longitudinal relaxation rate R 1 (1/T 1 ) and R 2 (1/T 2 ). (B) Mean indices of the metrics were extracted from left and right hippocampi, fornix and parahippocampal cinguli (PHC) tracts. Hippocampi were segmented from T 1 -weighted images with FreeSurfer version 5.3 and fornix and PHC were reconstructed with damped-Richardson Lucy spherical deconvolution (dRL) based deterministic tractography on colour coded principal direction maps (RGBλ). (C) Cross-correlation matrix between white and gray matter microstructural indices that correlated with age. reflect its myelin content and the forward exchange rate k f , an index of the rate of the magnetization transfer process 27 , that has been proposed to reflect metabolic efficiency of mitochondrial function 40 ( Fig. 2A). Finally, the longitudinal relaxation rates R 1 (1/T 1 ) and R 2 (1/T 2 ) were obtained as additional estimates of the relative contribution of water, myelin and iron content of gray and white matter tissue [41][42][43][44] (Fig. 2A). Mean values of all MRI indices were extracted from the fornix and the PHC tracts, that were reconstructed with deterministic tractography, and from FreeSurfer segmentations of the whole hippocampi 45,46 (Fig. 2B). The hippocampal regions included areas of the presubiculum, subiculum, cornu ammonis subfields 1-4, dentate gyrus, hippocampal tail and fissure but excluded cortical regions such as the entorhinal cortex 47,48 .
To study the relationships between age-related differences in MRI metrics and genetic and lifestyle risk factors of LOAD as well as differences in episodic memory performance, we also acquired the following information (Table 1): Genetic risk was assessed by carriage of the Apolipoprotein-E (APOE) ε4 allele 49,50 and separately by a positive family history of a first grade relative with dementia of the Alzheimer's, Lewy body or vascular type. We predicted more pronounced white matter glia reductions in APOE ε4 compared to ε2 and/or ε3 carriers, as cholesterol transport for myelin repair in the brain has been proposed to be less efficient in ε4 carriers [50][51][52][53] . In addition, we assessed lifestyle-related risk factors associated with the metabolic syndrome, notably central obesity, hypertension, alcohol consumption and sedentary lifestyle 54,55 (Table 1). We were particularly interested in the effects of central adiposity, as obesity is globally on the rise, and is associated with chronic inflammation, insulin resistance and vascular problems 54,55 as well as with accelerated aging in brain regions that include limbic white matter 56,57 . Finally, episodic memory performance was assessed with standard neuropsychological tests of verbal and non-verbal recall 58,59 (Table 1).
Neuroglia models predict that aging and risk factors will primarily reduce glia sensitive metrics of MPF, k f , R 1 and R 2 whilst apparent neurite density (ICSF) and dispersion (ODI) should be less affected. In contrast, neurodegenerative models predict impairments in ICSF and ODI with relatively preserved glia-metrics. Age and risk-related increases in ISOSF, reflective of increased free water content in brain tissue, may be expected by both models. Importantly though, neurodegenerative models predict that hippocampal differences will mediate age-related white matter microstructural differences in the fornix (Model A), whilst neuroglia models predict that age-related differences in glia sensitive white matter metrics of the fornix will mediate age-related hippocampal differences (Model B).
Given the low numbers of diabetics, smokers, and individuals on statins amongst our sample (Table 1), these variables were not included in the analyses. This analysis revealed significant omnibus effects of age [F (35,96)  Post-hoc multivariate covariance analyses, controlling for ICV, then tested for the effects of age on each of the hippocampal, fornix and PHC MRI indices separately.  (Fig. 3).
Age correlated negatively with fornix MPF, R 1 , k f , left PHC R 2 , right PHC R 1 , left hippocampal k f and R 1 , and positively with fornix and hippocampal ISOSF and ODI (Fig. 3). No age correlations were present with ICSF in the fornix or the hippocampi and these 'null' correlations differed significantly from the smallest age correlations in the fornix (z ODI = 2.5, p = 0.01), left hippocampus (z R1 = 2.78, p = 0.005) and right hippocampus (z ISOSF = 3.18, p = 0.002) (Fig. 3). Figure 2C displays the Pearson correlation coefficient matrix between the white and hippocampal gray matter metrics. The same metrics measured in white and gray matter correlated positively with each other with the exception of fornix and left hippocampal ODI (r = −0.12, p = 0.13). MPF, k f , R 1 and R 2 correlated positively with each other and so did ISOSF and ODI, whilst MPF, k f , R 1 and R 2 were inversely correlated with ISOSF and ODI.

Mediation analyses between hippocampal and fornix metrics. To avoid any bias in the comparison
of Model A and Model B, the number of mediator variables, that showed significant age effects and correlated with each other in white and gray matter, was kept constant. Mediator variables in Model A were the left and right hippocampal ISOSF, left hippocampal R 1 and left hippocampal k f , and in Model B the four white matter metrics that showed the largest age effects, i.e., fornix ISOSF, MPF, k f , and R 1 . There were no significant differences between the age effect sizes in the fornix and the hippocampus (r max:age-fornix R1 versus r min:age-left hippocampal R1 : z = 1.8, p = 0.07).
Scientific RepoRts | (2019) 9:1060 | https://doi.org/10.1038/s41598-018-37658-5 In Model B, fornix mediator variables fully mediated the age effects on right hippocampal ISOSF, left hippocampal k f and left hippocampal R 1 (Fig. 4, highlighted in red) but hippocampal mediators in model A, although showing significant indirect effects on the direct effects of age on fornix R 1 and k f , did not remove the age effects on fornix MPF, R 1 and k f . A bi-directional relationship was observed between right hippocampal ISOSF and fornix ISOSF, with both variables fully mediating each other's age effect. Finally, fornix ISOSF contributed but did not fully mediate the correlation between age and left hippocampal ISOSF, and left hippocampal ISOSF only marginally contributed to the age correlation in fornix R 1 . To summarise, full mediation of age effects on glia metrics was observed in Model B but not in Model A (Fig. 1), such that age-related glia differences in the fornix mediated age differences in the hippocampus but not vice versa.

effects of genetic and lifestyle risk factors on age-mediator variables.
We further explored whether genetic and lifestyle risk factors had an effect on mediator variables that fully accounted for age effects with hierarchical regression analysis. First age, ICV, sex and years of education were entered into the regression model followed by the stepwise inclusion of all genetic and lifestyle risk variables. Table 3 summarises the results of these regression analyses. Besides age, sex was a significant predictor for differences in fornix and hippocampal ISOSF. WHR contributed significantly to fornix MPF and R 1 and alcohol consumption to fornix R 1 . There were trends (significant at the uncorrected level) for a contribution of family history of dementia and diastolic BP to differences in right hippocampal ISOSF.
Post-hoc comparisons revealed that centrally obese individuals compared with individuals with a normal WHR showed lower MPF [t(160) = 2.8, p = 0.005] and R 1 [t(160) = 3.35, p = 0.0009] in the fornix. Men compared with women showed higher ISOSF in the right hippocampus [t(162) = 6.5, p < 0.00001] and in the fornix [t(164) = 6.7, p < 0.00001]. Participants with a positive family history exhibited higher ISOSF in the right hippocampus [t(160) = 2.55, p = 0.012] and there was a positive correlation between diastolic BP and ISOSF in the right hippocampus [r(164) = 0.18, p = 0.02]. However, there was no difference in fornix R 1 between individuals that consumed alcohol units above the weekly-recommended limit compared with those within the UK recommended guidelines (p = 0.9). CSF partial volume effects. To test whether age effects on fornix and hippocampal qMT metrics were driven by free water partial volume effects, ISOSF metrics were used as mediator variables. Fornix ISOSF contributed significantly to all regression models (p < 0.0001) but did not remove the direct age effect on fornix MPF (t = −3.88, p = 0.0002), R 1 (t = −3.69, p = 0.0003), and k f (t = −3.88, p = 0.0002). Right but not left hippocampal ISOSF contributed significantly to left hippocampal R 1 (p = 0.003) and k f (p = 0.03) and removed the direct age effect on left hippocampal R 1 (p = 0.05) but not on left hippocampal k f (p = 0.0013).

Discussion
The aim of this study was to discriminate between two classes of causal models of aging with mediation analysis. The neurodegenerative model predicted that age-related hippocampal differences would account for age effects on fornix metrics whilst the neuroglia model predicted that age differences in fornix glia would cause age-differences in the hippocampus. Our results are consistent with the neuroglia model as we found that fornix glia sensitive metrics of MPF, R 1 and k f fully mediated the effects of age on hippocampal tissue properties but not vice versa, i.e., hippocampal mediator variables, although demonstrating significant indirect effects, did not remove the age effects on fornix MPF, R 1 and k f (Fig. 4). This pattern of results is consistent with a growing body  Table 2. Summary of the effects of age on gray and white matter microstructural indices. * Controlled for intracranial volume, **5% FDR corrected. ISOSF = isotropic signal fraction, MPF = macromolecular proton fraction, ODI = orientation dispersion index, PHC = parahippocampal cingulum.
of evidence pointing towards an important role of neuroglia changes in aging and neurodegeneration 11,[13][14][15][60][61][62] . However, to the best of our knowledge, this is the first study to show that fornix white matter glia damage may cause hippocampal gray matter damage during age-dependent limbic decline in the human brain.   Both in the hippocampus and the fornix, we observed age-related increases in ISOSF and ODI and reductions in glia-sensitive metrics from qMT and relaxometry imaging, whilst age had no effects on ICSF. Overall this pattern of results suggests age-dependent i) increases in the free water content of tissue, ii) reductions in glia-sensitive indices, and iii) increases in apparent neurite dispersion in the absence of differences in apparent neurite density.
In white matter, MPF, k f , R 1 and R 2 are known to be sensitive to myelin, as well as to neuroinflammation and to a lesser degree to iron changes [27][28][29][30]38,44,63,64 . While some effects were observed on R 1 and R 2 in the PHC, the largest age-dependent reductions in MPF, k f , and R 1 were present in the fornix tract. Reductions in these indices in the absence of any effects on ICSF suggest that aging affects glia-sensitive properties of white matter, notably myelin, but not apparent axon density, i.e. not the number and the size of axons. These results provide novel in vivo neuroimaging evidence consistent with neuropathological findings of a reduction of up to 45% of the length of myelinated fibers, rather than a loss of axons per se across the lifespan [65][66][67][68][69] . Similarly, age-related reductions in fornix myelination and glia changes in the absence of any changes of the fornix cross-sectional area were also observed in non-human primates 70 . Myelin damage is closely linked to neuroinflammation and qMT metrics are sensitive to both processes 71 . Thus it may also be possible that the observed differences in MPF and k f reflect age-dependent dystrophic and reactive microglia changes 39,72 .
Similarly, in hippocampal gray matter we observed age-related reductions of R 1 and k f in the absence of differences in ICSF. Age reductions in hippocampal R 1 were accounted for by increases in hippocampal ISOSF and may, therefore, primarily reflect increases in the water content of hippocampal tissue. Increases in hippocampal ISOSF are consistent with neuropathological findings of a loss of hippocampal pyramidal cells and dentate gyrus granule cells associated with normal aging 73 . However, the absence of an effect on hippocampal ICSF suggests that apparent dendritic density was unaffected and increased ISOSF may also be compatible with glia-related changes in gray matter. Furthermore, the observed reductions in hippocampal k f , a marker of the rate of the magnetization transfer between macromolecular and free water pool, may reflect age-dependent metabolic differences. Giulietti et al. 40 found k f reductions in the hippocampus, temporal lobe, posterior cingulate and parietal cortex in LOAD, and proposed that these changes may reflect reductions in metabolic mitochondrial activity 74 . Age-related reductions of hippocampal metabolic activity have also been observed in animal studies 75,76 and age-related damage to mitochondria of microglia has been linked to sustained microglia neuroinflammation 7 .
Currently, a causal link between age-related neuroglia changes and the development of LOAD pathology in the human brain remains speculative. However, white matter disease, characterised by a loss of myelin, oligodendrocytes, axons and reactive astrocyte gliosis 77-79 is a known feature of LOAD. Accumulating evidence also suggests that neuroinflammation and a reduction of glia mediated clearance mechanisms contribute to LOAD 11,12,80 . For instance, reactive microglia were found in the hippocampus of LOAD brains 81 and microglia derived ASC protein specks have been shown to cross-seed amyloid-β plaques in transgenic double-mutant APPSwePSEN1dE9 mice 82 . It is therefore possible that age-related neuroglia changes involving myelin and inflammation pathways may not only occur in response to neuropathology but may even trigger protein abnormalities and synaptic and neuronal loss. Thus, an intriguing interpretation of our results may be that aging is associated with glia changes including loss of myelin in the fornix, which in turn may cause tissue damage in the hippocampus, that might make this region more vulnerable to the development of LOAD pathology 81,83 . Clearly, future longitudinal prospective studies assessing the predictive value of MRI indices of white matter glia for amyloid and tau burden are needed to test this hypothesis.
We also observed a bi-directional mediation effect between right hippocampal ISOSF and fornix ISOSF, such that both variables fully mediated the age effect on each other. This result is unsurprising and reflects that unspecific tissue loss in the hippocampus is associated with tissue loss in the fornix and the other way around. In contrast, fornix and left hippocampal ODI metrics were not correlated with each other. This finding suggests that ODI indices may capture different microstructural properties and may not be directly comparable in gray and in white matter, potentially due to differences in tissue complexity and organisation.
The question arises how genetic and lifestyle risk factors of LOAD impact on medial temporal tissue properties. Consistent with our hypothesis, centrally obese individuals exhibited reductions in fornix MPF and R 1 . Obesity-related reductions in MRI metrics of proton density, R 1 , and R 2 * , have previously been observed in white matter pathways connecting frontal and limbic regions in young adults 84 . Obesity has also been associated with accelerated aging of cerebral white matter 56 and we previously reported Body Mass Index related increases in mean and axial diffusivities in the fornix 57 . Although the mechanisms underpinning these changes remain unclear, it may be possible that the observed MRI differences may reflect changes in neuroglia that may be linked with obesity-related systemic inflammation. Consistent with this interpretation, differences in white matter fractional anisotropy in obesity were reported to be associated with inflammation and to a lesser degree with glucose regulation, whilst lower blood pressure and dyslipidemia appeared to have positive effects on white matter microstructure 85 .
Unexpectedly and in contrast to previous reports 52,53,86 we did not observe a main effect of the APOE genotype on hippocampal gray matter or fornix/PHC white matter tissue properties. There is substantial evidence that APOE ε4 carrier status in older individuals (>65 years of age) is related to accelerated atrophy in the hippocampus, increased beta-amyloid burden and to an earlier onset of dementia 87,88 . However, the effect size of APOE-ε4 genotype on age-related hippocampal atrophy over the adult life course is small (e.g. β = −0.05, p = 0.04 in n = 3749 between 43-69 years of age) 89,50,[90][91][92][93] and the biological mechanisms underpinning these relationships remain poorly understood. The APOE gene is involved in many complex functions, including lipid and amyloid-β metabolism, neuroinflammation and vascular regulation 94 . It is likely that APOE genotype, rather than exhibiting straightforward main effects, interacts with age and with multiple genetic and lifestyle-related factors. As the effects of APOE genotype were not the main focus of this study, testing for potential interaction effects between APOE genotype and other demographic and risk factors was beyond the scope of this paper.
Scientific RepoRts | (2019) 9:1060 | https://doi.org/10.1038/s41598-018-37658-5 In contrast to APOE, we observed some contribution of family history of dementia and diastolic blood pressure to hippocampal ISOSF, but these effects did not survive multiple comparison correction. We also found that men compared with women showed larger ISOSF in the right hippocampus and in the fornix. These results are compatible with previous findings, indicating that both family history 95 and being male [96][97][98][99] are associated with accelerated age-related atrophy of cortical and subcortical brain regions, including the hippocampus and the fornix.
Finally, we found evidence that the observed age-dependent decline in limbic structures had some impact on individual differences in episodic memory performance, as individual differences in fornix and hippocampal ISOSF were negatively correlated with differences in delayed verbal recall performance. This finding is consistent with a large body of evidence demonstrating an important role of the hippocampal-fornix axis in mediating age-dependent decline of episodic memory function 21 .
In conclusion, we provide novel evidence in support of the neuroglia model of aging. We propose that age-related damage to fornix glia, rather than a loss of axons, may cause hippocampal tissue changes associated with reduced metabolism, glia and neuronal loss. These results provide novel insights into the mechanisms underpinning age-dependent decline in limbic structures and suggest that MRI metrics sensitive to neuroglia may potentially provide useful biomarkers of midlife risk of LOAD.

Methods
This study was approved by the Cardiff University Psychology Research Ethics Committee (EC.14.09.09.3843R2). All participants provided written informed consent in accordance with the Declaration of Helsinki.
participants. Community-dwelling participants between 35 and 75 years of age were recruited from Cardiff University panels and notice boards and via internet and poster advertisements. Exclusion criteria included a history of neurological or psychiatric disease, head injury with loss of consciousness, drug/alcohol dependency, high-risk cardio-embolic source, large-vessel disease, and MRI contraindications. From 211 recruited volunteers, 166 underwent MRI scanning at CUBRIC. Table 1 summarises the demographic, cognitive, and dementia risk information for these 166 participants.

Assessment of genetic and lifestyle related risk factors. Genetic-related risk. Saliva samples were
collected with the Oragene-DNA (OG-500) kit (Genotek). APOE genotypes ε2, ε3 and ε4 were determined by TaqMan genotyping of single nucleotide polymorphism (SNP) rs7412 and KASP genotyping of SNP rs429358. Genotyping was successful in 165 individuals that underwent an MRI scan (Table 1). Participants also provided information about their family history of dementia, i.e. whether a first-grade relative was affected by LOAD, vascular dementia or Lewy body disease with dementia.
Lifestyle-related risk. Abdominal adiposity was assessed with the waist-hip-ratio (WHR) 100 . Central obesity was defined as WHR ≥ 0.9 for men and ≥ 0.85 for women. Systolic and diastolic blood pressure was measured with a digital blood pressure monitor (Model UA-631; A&D Medical, Tokyo, Japan). Hypertension was defined as systolic BP ≥ 140 mm Hg. Risk factors of diabetes mellitus, high levels of blood cholesterol, history of smoking and weekly alcohol intake were self-reported. Information about participants' physical activity over the previous week was collected with the International Physical Activity Questionnaire (IPAQ) 101 . Participants' intellectual and cognitive functions were assessed with the National Adult Reading Test (NART) 102  High Angular Resolution Diffusion Imaging (HARDI). Diffusion data (2 × 2 × 2 mm voxel) were collected with a spin-echo echo-planar dual shell HARDI 105 sequence with diffusion encoded along 90 isotropically distributed orientations (30 directions at b-value = 1200 s/mm 2 and 60 directions at b-value = 2400 s/mm 2 ) and six non-diffusion weighted scans with dynamic field correction, 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.
MRI data processing. The HARDI data were corrected for distortions induced by the diffusion-weighted gradients and artifacts due to head motion with reorientation of the encoding vectors 108 in ExploreDTI (Version 4.8.3) 109 . EPI-induced geometrical distortions were corrected by warping the diffusion-weighted image volumes to down-sampled T 1 -weighted images with a resolution of 1.5 × 1.5 × 1.5 mm 110 . After preprocessing, the NODDI model 22 was fitted to the dual-shell HARDI data using fast, linear model fitting algorithms of the AMICO framework 111 . Restricted non-Gaussian diffusion in the intra-axonal space was quantified with ICSF and free isotropic Gaussian diffusion with ISOSF. ODI was modelled with a Watson distribution.
MT-weighted GRE volumes for each participant 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 using Elastix 112 . The 11 MT-weighted GRE images and T1-maps were modelled by the two-pool Ramani's pulsed MT approximation 113 yielding MPF, k f , and R 1 maps. MPF maps were threshholded to an upper intensity limit of 0.3 and k f maps to an upper limit of 3 using the FMRIB's fslmaths imaging calculator to remove voxels with noise-only data.
A monoexponential decay function was fitted to the T 2 images using quimultiecho from the Quantitative Imaging Tools (QUIT) library (http://spinicist.github.io/QUIT). This fitting excluded the first acquired TE due to a non-contained stimulated echo artifact in the first echo of the multi-contrast spin echo Siemens library sequence. R 2 was calculated as 1/T 2 in second units.
All image modality maps and region of interest masks were spatially aligned to the T 1 -weighted anatomical volume as reference image with linear affine registration (12 degrees of freedom) using FMRIB's Linear Image Registration Tool (FLIRT).
Five R 2 and four qMT data sets were missing as five participants did not complete the MRI session due to claustrophobia.
tractography. The RESDORE algorithm 114 was applied to identify outliers, followed by whole brain tractography with the damped Richardson-Lucy algorithm (dRL) 115 on the 60 direction, b = 2400 s/mm 2 HARDI data in single-subject space using in house software 114 coded in MATLAB (the MathWorks, Natick, MA). To reconstruct fibre tracts, dRL fibre orientation density functions (fODFs) were estimated at the center of each image voxel. Seed points were positioned at the vertices of a 2 × 2 × 2 mm grid superimposed over the image. The dRL algorithm interpolated local fODF estimates at each seed point and then propagated 0.5 mm along orientations of each fODF lobe above a threshold peak of 0.05. This process was repeated until the minimally subtending peak magnitude fell below 0.05 or the change of direction between successive 0.5 mm steps exceeded an angle of 45°. Tracking was then repeated in the opposite direction from the initial seed point.
The fornix and the PHC were reconstructed with an in-house automated segmentation method based on principal component analysis (PCA) of streamline shape 116 . This involved the manual reconstruction of a set of tracts from 20 randomly selected datasets, that were then used to train a PCA model of candidate streamline shape and location. The fornix and the PHC tracts were reconstructed by manually applying region of interest (ROI) gates to isolate specific tracts from the whole brain tractography data on colour-coded fiber orientation maps in ExploreDTI following previously published protocols 21,57,117,118 . The trained PCA shape models were then applied to all datasets: candidate streamlines, i.e. those bridging the gap between estimated end points of the fornix/PHC, were selected from the whole volume tractography and spurious streamlines were excluded by means of a shape comparison with the trained PCA model. All automatic tract reconstructions underwent visual quality control and any remaining spurious fibers that were not consistent with the tract anatomy were removed.
Whole Hippocampal segmentation. Volumetric segmentation of the left and right whole hippocampi from T 1 -weighted images was performed with the Freesurfer image analysis suite (version 5.3), which is documented online (https://surfer.nmr.mgh.harvard.edu/). Mean intracranial volume fractions (ICV) were extracted for each brain as estimates of individual differences in head sizes. Two datasets had to be excluded from the analysis due to motion artefacts. statistical analyses. Statistical analyses were conducted in MATLAB, SPSS version 20 119 and the PROCESS computational tool for mediation analysis 120 . Multiple comparisons were corrected with a False Discovery Rate (FDR) of 5% using the Benjamini-Hochberg procedure 121 . Partial Eta 2 (ηp 2 ) and correlation coefficients are reported as indices of effect sizes. Differences in the size of correlation coefficients were tested with the Fisher's r-to-z transformation 122 . All reported p-values are two-tailed.
Omnibus multivariate regression analysis was conducted to test for main effects of age, genetic and lifestyle risk factors, sex, education and head size simultaneously on all MRI metrics. Omnibus effects were followed up with post-hoc multivariate covariance analysis testing for effects on individual MRI metrics in all regions of interest while correcting for any confounding variables. The direction of age effects and the relationship between hippocampal and white matter microstructural metrics were investigated with Pearson correlation coefficients.
Linear mediation analysis was then used to test for the indirect effects a*b of hippocampal mediator variables on the direct effects c' of age on the fornix metrics (Fig. 1, Model A) and for indirect effects a*b of fornix mediator variables on direct age effects c' on hippocampal differences (Fig. 1, Model B). The significance of indirect and direct effects was assessed with a 95% confidence interval based on bootstrapping with 5000 replacements 120 .
Scientific RepoRts | (2019) 9:1060 | https://doi.org/10.1038/s41598-018-37658-5 The impact of genetic and lifestyle risk factors on mediator variables was explored with hierarchical linear regression analyses, and Pearson correlation coefficients between MRI mediator variables and episodic memory measures were calculated to explore brain-cognition relationships.

Data Availability
The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.