Age-related iron accumulation and demyelination in the basal ganglia are closely related to verbal memory and executive functioning

Age-related cognitive decline has been linked to alterations of the dopaminergic system and its subcortical trajectories. Recent work suggests a critical role of iron accumulation within the basal ganglia (BG) in verbal memory performance, and increased iron levels have been related to demyelination. However, the specificity of age-related iron increases with respect to cognitive functions remains unclear. Therefore, we investigated the interplay of age, cognitive performance, and structural integrity of the BG. In total, 79 healthy older participants underwent a broad cognitive assessment (fluid and crystallized intelligence, verbal and numeric memory, processing speed, executive functions) and structural MRI. As expected, performance in most cognitive tests had a negative relationship with age. Moreover, BG grey matter volume and magnetization transfer (MT, indicative of myelin) decreased, and R2* (indicative of iron) increased with age. Importantly, R2* and demyelination negatively correlated with verbal memory and executive functions. Within the SN/VTA, age correlated negatively with MT, but there was no clear evidence in favor of a relationship between behavior and R2* or MT. Our results suggest that age-related increases in iron and demyelination within the BG, which are part of a fronto-striatal network, not only impact on verbal memory but also executive functions.


Materials and methods
Experimental design, procedure, and participants. All participants were part of a cognitive training study, which included a test at baseline (t1), a four-week cognitive training, and a follow up examination (t2) at the University of Lübeck. At both time points, all participants received a detailed neuropsychological assessment (see below) and a structural MRI scan (see below). For the current study, only baseline data (t1) were analyzed. The findings of the cognitive training study are reported elsewhere 39 . Note that the methods sections of cognitive assessment, image acquisition, and VBM/VBQ processing are adapted from 39 .
In total, 92 healthy, right-handed, German speaking older adults were recruited. However, nine participants were excluded due to a history of neurological, psychological or other severe physical disorders, drug abuse, CNS affecting medication intake (less than 2 weeks before testing), non-removable metal implants or claustrophobia. Four additional participants had to be excluded due to technical issues or structural abnormalities observed in the MRI data. Moreover, participants were excluded with > 5 points in the Geriatric Depression Scale (GDS, max. 15 points, > 5 points indicates mild depression; 40 ) and < 22 points in the Montreal Cognitive Assessment (MoCA, max. 30 points; 41,42 ). A value of 22 was chosen based on a study by 42 , suggesting that it might be an appropriate cut-off for mild cognitive impairment (MCI). Finally, 79 participants (age range 50-80 years, mean age 63.54 ± 8.48, 39 females) completed the MRI sessions and could, therefore, be included in further analyses. Within this group, the following numbers of participants per cohort were included: 50-60 years n = 32, 61-70 years n = 28, 71-80 years n = 19.
Participants were recruited through announcements in the local newspaper or the database of the University of Lübeck 43 . All participants signed a written informed consent and received monetary compensation. The study was approved by the local ethical committee of the University of Lübeck, Germany, and in accordance with the Declaration of Helsinki.
Intellectual functioning (i.e., fluid intelligence; Gf) was measured by the LPS 50 + short version (for people aged 50-90 years), which includes a battery of time restricted paper pencil tasks (duration ~ 30 min). For crystallized intelligence (Gc), the MWT was applied; it provides 37 rows, each containing four pseudo-words and one correct word, which has to be identified (with no time restriction).
Verbal memory was examined using the VLMT. Here, a word list of 15 nonrelated items was verbally presented for five subsequent times. Each time, participants were asked to recall as many words as possible. Recalled words were noted from the examiner (total sum of correctly recalled words of all five runs refers to VLMT learning in further analysis). In a sixth run, an interference list of 15 words was verbally presented, which had to be Image acquisition. Structural MRI was performed at the University of Lübeck using a 3 T Siemens Magnetom Skyra scanner equipped with a 64-channel head coil. Whole-brain multiparameter mapping (MPM) was acquired as reported in previous studies 62,63 using multi-echo 3D fast low angle shot (FLASH) at 1 mm isotropic resolution. The volumes were acquired for T1, proton density (PD), and magnetization transfer (MT) weightings. The weightings differed in TE, TR, and flip angles. T1-weighted: six echo times (TE = 2.2, 4.7, 7.2, 9.7, 12.2, 15 ms), TR = 19 ms, flip angle = 20°; PD-weighted: eight echo times (TE = 2.2, 4.7, 7.2, 9.7, 12.2, 15, 17.5, 20 ms), TR = 24 ms, flip angle = 6°; MT-weighted: six echo times (TE: 2.2, 4.7, 7.2, 9.7, 12.2, 15 ms), TR = 37 ms, flip angle = 6°. A Gaussian MT-pulse following Siemens product sequences was applied. To shorten the scan duration, a partial Fourier 6/8 was used. Parallel imaging with a GRAPPA acceleration factor of 2 was applied. The total scanning time of the MPM protocol was approx. 20 min.
Subsequently, two runs of diffusion weighted imaging (DWI) using an EPI sequence were performed during the same scanning session (scanning time approx. 16 min). The images were later used for further analysis (results will be reported elsewhere).
MR data were further processed on the Statistical Parametric Mapping framework (SPM 12, Wellcome Trust Center for Neuroimaging, London) and MATLAB software (R2014b version, Mathworks). R2* maps were calculated through a regression of the log signal from the PD-weighted echoes. Averaging the set of echoes for each weighting increased the signal-to-noise-ratio for estimation of the MT map 64 . The semiquantitative MT map was calculated as described by 65,66 . Subsequently, to ensure uniform orientations, images were slightly manually re-orientated according to individual posture during the MRI acquisition using SPM Check Reg and Display options 67 .
Voxel-based morphometry and voxel-based quantification. GM volumes were processed and analyzed following a protocol for voxel-based morphometry using SPM's batch system (VBM; 67,68 ). Since MT maps provide increased contrast for subcortical regions 69,70 , in a first step, they were used for segmentation of the different tissue groups. Subsequently, images of GM, WM, and cerebrospinal fluid (CSF) were generated in native space 71 . Applying high dimensional warping, images were then normalized to MNI space using the diffeomorphic registration algorithm (DARTEL) implemented in SPM, scaled by the Jacobian determinants of the deformation field and smoothed with an isotropic Gaussian Kernel of 6 mm full width at half maximum (FWHM). Finally, the resulting smoothed, modulated and normalized GM images were used for statistical analysis.
Voxel-based quantification (VBQ) analysis provides sensitivity to tissue microstructure and is therefore well suited to test differences in R2* and MT, which are sensible marker for alterations in subcortical brain regions. VBQ was processed using the open source hMRI toolbox 72 embedded in the SPM framework. Using the integrated processing pipeline of the toolbox, the previously generated MT maps were further processed using the modules tissue segmentation (GM, WM, and CSF), DARTEL, creation of templates, and normalization to MNI www.nature.com/scientificreports/ space. Subsequently, tissue-weighted smoothing with a FWHM isotropic Gaussian kernel of 3 mm 28 was performed. Resulting images of R2* and MT for GM were used to indirectly test for differences in iron levels (i.e., R2*) and myelination (i.e., MT) of brain tissue (see also 27,28 ). Note that we did not analyze WM subspace, since we focused on the BG structures that typically do not contain much WM 73 . Moreover, the estimation of iron concentrations in WM regions (on the basis of quantitative susceptibility mapping) appears to be less accurate 74 .
Analyses of MTR within GM structures have been related to demyelination 75 and cognitive impairment 76 , suggesting a link between behavior and MT integrity of subcortical structures.
Manual segmentation of the SN/VTA. The SN/VTA is a fairly small subcortical structure and interindividual differences of the SN/VTA boundaries can account for inaccuracy when applying normalized anatomical masks. Therefore, in the present study, a voxel-based region of interest (ROI) analysis in native space was conducted. Following previous studies 22,77 and using MRIcron tools, individual SN/VTAs were segmented based on the intense contrast change between the bright grey color and the dark grey color of the adjacent tissue in the MT-weighted image (Fig. 1). The upper limit of the SN/VTA-ROI was selected at a level of the superior colliculi. The anterior part of the SN/VTA-ROI was limited by the interpeduncular fossa and posterior borders were limited by the lateral side of the cerebral peduncle. The lower limit of the SN/VTA-ROI was identified as the last even grey colored cross-sectional area. The total rostrocaudal extension of the ROI included approx. 10 slices depending on the individual size of the SN/VTA. All segmentations were performed twice by one person. Only the latter segmentations were used for further analyses. Subsequently, the ROIs were extracted and projected as an overlay on the corresponding MT and R2* images to obtain the mean values (see also 22,78 ). Reliability of the segmentation was tested with an intra-class correlation coefficient (ICC; 79 ) using the ICC command within R statistical software 80 (see http:// perso nality-proje ct. org/r/ html/ ICC. html). There were significant correlations between the first and the second manual segmentation regarding values for MT (ICC = 0.98, p < 0.001) and R2* (ICC = 0.99, p < 0.001), suggesting high intra-rater reliability. To test inter-rater reliability, a second person performed the SN/VTA segmentation. Subsequently, segmentations of both examiners were correlated showing high inter-rater reliability for MT (ICC = 0.98, p < 0.001) and R2* (ICC = 0.99, p < 0.001). Note, that the SN/VTA had to be manually segmented since the Harvard-Oxford-Atlas does not provide it as ROI.
Statistical analyses. Participants with more than 3 SD above or below the overall mean of a specific neuropsychological test were excluded from the respective analysis. For analyses of the relationship between microstructure and cognitive functioning, multiple regression analyses for VBM and VBQ were calculated with SPM using Matlab software. Covariates were age, LPS 50 + , MWT, VLMT learning, VLMT cons, VLMT free recall, digit span, and d2-R. Due to one missing value in the VLMT recognition test and two outliers (> 3 SD) in the TMT, two additional linear regression analyses of VBM and VBQ were conducted for age and VLMT recognition, and age and TMT.
Effects within subcortical regions of interest (ROI) were investigated using a small volume correction (SVC) with a BG mask (i.e., caudate, pallidum, putamen, nucleus accumbens). The mask was taken from the Harvard-Oxford-Atlas (50% probability mask), implemented in the FMRIB Software Library (FSL; 81 ). Clusters with at least 50 voxels and a p < 0.05 after familywise error correction (FWE) at the cluster level (p < 0.001 uncorrected at peak-voxel level) were regarded as significant. For post hoc analyses and depictions of effects, MT and R2* values were extracted from significant brain regions (see "Results"). To further examine effects of age on cognition, a post hoc partial correlation for age was applied for significant results of the VBM and VBQ analyses. In addition to the SVC analysis, whole brain analysis (cerebellum excluded) was performed for all regression models on GM subspace (FWE, p < 0.05, k = 50 voxel), to test whether relationships between structure and cognition are specific to the BG.
To obtain a clearer picture about the age-structure-behavior relationship, a mediation analysis was performed for significant structure-behavior correlates with age as predictor, structure as mediator, and cognitive www.nature.com/scientificreports/ performance as dependent variable. The mediation analysis was performed with the TMT-diff score (instead of TMT-sum), since it provides a purer measure of executive functioning compared to the TMT-sum score. Moreover, in post-hoc calculations of the MRI regression analysis only TMT-diff but not TMT-ratio became significant; therefore, no further analyses with TMT-ratio were performed here. The analysis was performed using the medmod module (see https:// cran.r-proje ct. org/ web/ packa ges/ medmod/ medmod. pdf) within R statistical software yielding indirect, direct, and total effects. Z-transformation was performed. Finally, SN/VTA values (MT and R2*) and their relationship to cognitive performance were investigated using a correlation analysis in SPSS. MT and R2* values were averaged across voxels for each individual SN/VTA-ROI (see above). The correlation analysis of the structural parameter (MT, R2*), age and the neuropsychological tests was Benjamini-Hochberg (FDR) corrected. To test whether significant correlations with age or cognitive tests within the BG are different from those within the SN/VTA, a comparison of correlation coefficients for dependent groups with one overlapping variable (i.e. age, cognitive tests) was performed 82 . Analysis was computed using the cocor command (cocor.dep.groups.overlap) within R statistical software (see https:// CRAN.R-proje ct. org/ packa ge= cocor; 83 ).
Not all data were normally distributed; however, due to the large sample size (79 participants), our inferential statistics are supposed to be robust against a violation of the assumption of normality (i.e., central limit theorem; 84 ).

VBM analysis.
A multiple regression analysis was computed, including the covariates age, LPS 50 + , MWT, VLMT learning, VLMT cons, VLMT free recall, digit span, and d2-R. Due to outliers/missing values in the VLMT recognition and TMT, two additional regression analyses were conducted for age and VLMT recognition, as well as age and TMT. Therefore, three separate regression analyses were performed; the effects were small volume corrected, using a BG mask (cluster level FWE, p < 0.05; see "Materials and methods").
The VBM multiple regression analyses revealed a negative relationship between age and GM volume within the right caudate, pallidum, and putamen (Table 1; Fig. 2a-b), and a positive correlation between GM volume and VLMT recognition in the caudate (Table 1). There were no further significant correlations between GM and behavior. When adding age as a covariate (i.e., post hoc partial correlation), the effect between GM volume and VLMT recognition only reached trend level (r = 0.208, p = 0.07). No significant correlates between structure and cognition could be observed on subsequent whole brain analysis (FWE, p < 0.05, k = 50 voxel). VBQ analysis on R2*. Similar to the VBM analysis, a multiple regression analysis on GM R2* was performed, including the covariates age, LPS 50 + , MWT, VLMT learning, VLMT cons, VLMT free recall, digit span, and d2-R. Due to outliers/missing values in the VLMT recognition and TMT, two additional regression analyses were conducted for age and VLMT recognition, as well as age and TMT. Therefore, three separate regression analyses were performed; the effects were small volume corrected, using a BG mask (cluster level FWE, p < 0.05; see "Materials and methods").
The analyses showed a significant positive effect between age and R2* within the left and right putamen (SVC, cluster level FWE, p < 0.05; Table 1; Fig. 3a). Post hoc, MT and R2* values were extracted from that cluster for further analyses. As expected, there was a significant positive correlation between age and GM R2* (r = 0.475; Fig. 3b), a negative correlation between age and GM MT (r = −0.656, p < 0.001; Fig. 3c), and a negative correlation between GM MT and GM R2* (r = −0.339, p = 0.002; adjusted alpha level: 0.05/3 = 0.016; Fig. 3d). However, when adding age as covariate in a subsequent partial correlation, the relationship between GM MT and GM R2* no longer remained significant (r = −0.042, p = 0.717).
Regarding behavioral performance, GM R2* correlated negatively with performance in VLMT recognition (i.e., left and right putamen) and free recall (i.e., left and right putamen/pallidum) -cluster level FWE-corrected, p < 0.05 (Table 1; Fig. 4a-b). Furthermore, GM R2* of the left and right putamen correlated positively with the TMT (higher values [s] in the TMT imply worse performance; Table 1; Fig. 4c-d). There were no other significant effects with regard to R2*. When adding age as a covariate in a post hoc partial correlation, the effects remained significant (GM R2* and VLMT recognition: r = -0.400, p < 0.001; GM R2* and TMT: r = 0.240, p = 0.041).
To further break down different components of the TMT, regression analyses for TMT-A, TMT-B, TMT difference, and TMT ratio were performed. As a result, three clusters for TMT-B and one cluster for TMT difference (positive correlation) were revealed (Table 1). No significant effects were found for TMT-A and TMT ratio.
No significant correlates between structure and cognition could be observed on subsequent whole brain analysis (FWE, p < 0.05, k = 50 voxel).
VBQ analysis on MT. Similar to the analyses above, a multiple regression analysis on GM MT was performed, including the covariates age, LPS 50 + , MWT, VLMT learning, VLMT cons, VLMT free recall, digit span, and d2-R. Due to outliers/missing values in the VLMT recognition and TMT, two additional regression analyses were conducted for age and VLMT recognition, as well as age and TMT. Therefore, three separate regression analyses were performed; the effects were small volume corrected, using a BG mask (cluster level FWE, p < 0.05; see "Materials and methods").
The analyses revealed negative correlations between age and MT within the BG (SVC). Three clusters within the caudate and pallidum were identified (Table 1; Fig. 5a-b). Regarding cognitive performance, GM MT correlated positively with VLMT recognition ( Post-hoc regression analyses for the TMT revealed three clusters for TMT-A, seven clusters for TMT-B, and one cluster (negative correlation) for TMT differences (Table 1). There was no effect for TMT ratio. Finally, no significant correlates between structure and cognition could be observed on subsequent whole brain analysis (FWE, p < 0.05, k = 50 voxel).  Fig. 6a), but not between age and R2* (r = 0.031, FDR adjusted p = 0.787; Fig. 6b) or MT and R2* (r = 0.184, FDR adjusted p = 0.21; Fig. 6c). Furthermore, a positive correlation between SN/VTA R2* and TMT reached trend level (FDR adjusted p = 0.06; Table 3) -note that higher values (s) in the TMT imply worse performance. Finally, a comparison between the correlations within the BG vs. SN/VTA was computed (see methods). All comparisons reached significance (p < 0.05), except from the TMT in the R2* analysis (p = 0.838; Table 2).
Relationship between age and cognitive performance. As expected, age correlated negatively with LPS 50 + , VLMT learning, VLMT free recall, and VLMT recognition, d2-R, and TMT ( Table 3). Instead of TMT-sum, TMT-diff was included in the analysis, since it offers increased sensitivity to executive functioning; moreover, post-hoc MRI regression analysis with TMT-diff (but not TMT-ratio) became significant (Table 1).

Mediation analysis.
In a next step, a mediation analysis for significant structure-behavior correlates within the R2* and MT analysis (Fig. 7, Table 4) was performed, to further investigate the interplay between brain structure, cognitive performance, and age. In our models, age served as predictor, cognition (VLMT recognition and TMT-diff, respectively) as dependent variable, and brain structure (R2* and MT, respectively) as mediator. The    www.nature.com/scientificreports/ Figure 6. Relationship between MT, R2*, and age within the SN/VTA. We observed a significant negative correlation between MT and age (a); but no significant correlations between (b) R2* and age or (c) MT and R2* (p > 0.05). Benjamini-Hochberg (FDR) p-value adjustment was applied. www.nature.com/scientificreports/ analysis revealed that brain structure partially mediated the effects of age on cognition for both, R2* and MT. For the analysis of R2* and cognition, the indirect path mediated the effects of iron on verbal memory (Fig. 7a) to a degree of 29.80%, but failed significance for executive functioning (Fig. 7c; Table 4). The direct pathway via age explained 70.21% and 89.12%, respectively (Table 4). For the MT analysis, the indirect path mediated the effects of MT on verbal memory (Fig. 7b) to a degree of 68.50% and on executive functioning (Fig. 7d) to a degree of 64.24% (Table 4). In contrast, the direct pathways via age failed significance ( Table 4). The R2* and MT analyses with TMT-sum score revealed congruent results (see supplements).  www.nature.com/scientificreports/

Discussion
We investigated age-related differences in GM and microstructural markers of myelin (i.e., MT) and iron (i.e., R2*) levels within the BG and their link to cognitive performance. In line with our hypotheses, cognitive performance negatively correlated with age in most tests. At the neural level, we observed age-related declines of GM volume and MT, and an increase of R2* relative to the tested age range. Importantly, performance in verbal memory and executive function was predicted by MT (positive relationship) and R2* (negative relationship). Finally, a ROI-analysis of the SN/VTA revealed age-related demyelination (as indicated by MT) but no clear link between behavior and R2* or MT, respectively. As such, our findings are compatible with the role of the BG in multiple cognitive functions, and they give new insights into the specific functional consequences of age-related microstructural changes. Neural degeneration within the BG is typical for heathy aging 27,30,85 , and might reflect a loss of neuronal and dendritic architecture, rather than a loss of neurons 86,87 . Although the underlying processes are still unclear, microstructural changes and oxidative damage might be of particular relevance (see as a review 86 ). Indeed, apart from GM reductions, our findings show that MT negatively correlated with age within the caudate and pallidum (Fig. 5a-b). This presumably indicates less macromolecular content (e.g., oligodendrocytes) and might reflect demyelinating processes 26,88 . A positive correlation between age and R2* within the putamen were accompanied by a negative relationship to MT (Fig. 3). This is in accordance with previous studies 3 , and may indicate a dysfunction of myelin-forming oligodendrocytes, which are sensitive to oxidative stress 26 , possibly triggered by increased iron levels 13,24 . Importantly, the negative relationship between R2* and MT did not remain significant when controlling for age in a partial correlation, suggesting that the interplay between iron and demyelination is not independent from age. Therefore, future studies should further investigate the relationship between all three factors ideally in a longitudinal design.
With regard to behavior, we can confirm our hypothesis 2,3 by demonstrating a negative relationship between R2* and verbal memory performance within the putamen/pallidum (Table 1; Fig. 4a-b). In order to encode novel information into long-term memory, the hippocampal VTA-loop model suggests that a hippocampal novelty signal is sent to the DA neurons within the SN/VTA via a polysynaptic path including the subiculum, nucleus accumbens, and ventral pallidum before DA-neurons back-project to the hippocampus 89,90 . While iron is required for DA synthesis 14 , an accumulation of it can impair DA production 8,14 . Therefore, increased iron levels within the bilateral putamen/pallidum may account for an imbalance of the loop, leading to impairments in verbal memory.
Importantly, while previous work has already demonstrated a relationship between striatal iron and episodic memory 2,3,91 , our findings show that iron accumulation (as indicated by R2*) within the BG also impact on executive functioning (as tested by the TMT) in healthy older adults (Fig. 4c-d). Psychological models of executive functioning typically include several mental processes, such as planning, task switching, inhibition, and cognitive flexibility (see e.g., 92 ). At the neural level, executive processing is strongly related to the medial frontal lobe 93 , but there is also a body of evidence for a close link to the fronto-striatal circuit [94][95][96][97][98] . Notably, in Parkinson's disease, which is characterized by increased iron levels within the SN/VTA, red nucleus, putamen, pallidum, and caudate (e.g., 15 ), executive dysfunction is already present in early stages of the disease 99,100 . It has been suggested that executive disorders in Parkinson's disease originate from DA depletion within the striatum, resulting in a dysfunction of the fronto-striatal circuit 98,101,102 . More evidence comes from a PET study testing patients with mild Parkinson's disease and healthy controls on a typical planning task (Tower of London). While both groups showed overlapping task-related activation patterns within the PFC, only healthy controls showed activation within the right caudate. In contrast, in patients with Parkinson's disease, activation within the right hippocampus was increased, suggesting a compensatory shift towards the declarative memory network 103 . Therefore, increases in BG iron levels during healthy aging may impair fronto-striatal circuits, leading to reduced executive functioning.
Verbal memory and executive functioning could also be predicted on the basis of BG myelination (as indicated by MT; Fig. 5c-f). At the physiological level, age-related reduction of myelin may impair saltatory conduction at the nodes of Ranvier and thereby, decreased velocity of action potentials along the myelin sheaths 104,105 . This, in turn, could impair behavior -in our case verbal memory and executive functioning. Indeed, previous studies could demonstrate that myelin facilitates learning processes (see as a review 106 ), while demyelination is related to a deceleration of processing speed in healthy aging 107 . Moreover, in mild cognitive impairment, demyelination is found within several brain regions 108,109 .
In terms of the different components of the TMT, both subtests (TMT-A and TMT-B), the TMT-diff score and TMT-ratio were investigated (see methods). For the R2* analysis, positive associations were revealed for the TMT-B subtest and TMT-diff score, suggesting a stronger functional link to executive/cognitive control (TMT-B) compared to the working speed component (TMT-A). In other words, higher values of R2* resulted in worse performance in executive functioning, corrected for working speed. For the MT analysis, negative associations were observed for both TMT subtests (TMT-A and TMT-B), as well as for the TMT-diff score. Again, this suggests that impairments of BG integrity impact on executive functioning rather than on working speed.
In the present study, we observed a negative correlation between age and GM volume within the right caudate, pallidum, and putamen ( Fig. 2a-b), and a (weak) link between GM volume and verbal memory (i.e., VLMT recognition) in particular within the caudate nucleus. This is in line with previous work on age-related structural degeneration of the BG 30,85,110 , and the notion of a close relationship between caudate volume and associative memory 30 , intelligence 111 , cognitive flexibility 112 , and learning 113 . However, it should be noted that the relationship between GM volume and VLMT recognition was only borderline significant (p = 0.07) when adding age as covariate in a partial correlation. Again, this suggests that the interplay between BG integrity and cognitive performance (including verbal memory) is not independent from age. Indeed, in our mediation analyses (Fig. 7) the indirect effects of brain structure (MT and R2*) on cognition (VLMT recognition and TMT-diff) were statistically significant and explained additional variance on top of the direct effects via age. However, this www.nature.com/scientificreports/ was only true for age/R2*/VLMT, age/MT/VLMT and age/MT/TMT but not age/R2*/TMT where the indirect pathway was not statistically significant. While this could be due to a non-significant relationship specifically between age and R2* (Fig. 7c), it also indicates that different markers of structural brain integrity (MT vs R2*) allow different conclusions regarding the interplay of age, brain structure, and cognition. This is further supported by the fact that for the MT but not R2* analyses, only the indirect paths reached significance, suggesting a subordinated direct effect of age. Together, structural integrity of the BG mediates the direct effect of age on cognitive performance. However, the strength of this relationship, in particular the direct effect of age on cognitive performance, appears to depend on the imaging modality (R2* vs MT).
A ROI-analysis of the SN/VTA revealed a significant negative correlation between age and MT, but no significant effect between age and R2* (Fig. 6). These observations are compatible with previous studies in healthy older adults, which also indicate age-related MT decreases 22,78 but no significant changes in iron 19 . With regard to the pattern observed in the BG, this suggests that, during healthy aging, myelin decreases are typical in both regions (BG and SN/VTA), whereas iron increases are only typical within the BG but not SN/VTA. Interestingly, in patients with Parkinson's disease, iron increases have been reported in the BG and SN/VTA 15 . This suggests that BG iron increases as well as demyelination in the BG and SN/VTA are typical for healthy aging, whereas SN/VTA iron levels are specific to pathologic neurodegeneration 14,17 . However, since only older adults have been tested here (50-80 years), we cannot exclude the possibility that a group of young controls might have shown significantly lower iron levels in the SN/VTA. Evidence for a continuous increase of SN/VTA R2* over the life-span comes from a study testing people aged 18-85 years, showing a positive relationship between age and SN/VTA R2* 28 . Thus, it is likely, that a younger control group would have shown significant lower R2* values compared to our older cohort.
Despite a significant relationship between age and demyelination (as indicated by MT) of the SN/VTA, there was no clear link between structural integrity (MT or R2*) and behavior ( Table 3). The only hint was based on a negative correlation between R2* and executive function (TMT-diff, uncorrected p = 0.048), but this did not survive Benjamini-Hochberg correction (adjusted p = 0.140). This absence is in contrast to previous work on SN/VTA integrity in healthy older adults demonstrating a link between MTR and verbal memory 22 , MT and reward-related reaction times 114 as well as MT and learning in a go-no-go-task 78 . These apparently divergent findings may relate to differences in MR data-acquisition protocols or differences in cognitive/behavioral readouts and they may suggest that the relationship between SN/VTA integrity and behavior may be modulated by other factors that were not explicitly included in our analysis (e.g., personality traits).
A direct comparison of the correlations between structural integrity and cognitive performance in the BG vs. SN/VTA revealed statistically significant differences (except from the correlation of TMT with R2*, Table 2). This means that the observed associations between age and R2*, age and MT, VLMT and R2*, VLMT and MT, as well as TMT and MT were significantly more pronounced within the BG as compared to the SN/VTA. Taken together with the absence of any significant effects outside of the BG in the whole brain analyses, this finding further underscores that the associations between structural integrity (R2* and MT) and cognitive functions (verbal memory and executive functioning) within the BG were region specific effects.
Finally, our findings must be interpreted with several limitations. First, MT and R2* are only indirect makers of myelin and iron contents and therefore, should be interpreted with caution (e.g., 24 ). However, post mortem studies provide histological evidence for a strong relationship between MT and myelin 21 as well as R2* and iron 38 . Along the same lines, R2* is sensitive to both, iron and myelin 115,116 but previous work suggests that the correlation between R2* and iron concentration in GM might be higher as compared to R2* and iron concentration in WM 38 . In any case, the interpretation of R2* is not only indirect but also complex (see e.g., 74,117 ). Second, VBM might be vulnerable to include voxels of WM, CSF, or blood vessels and can, therefore, confound estimates of R2* and MT (see 118 ). Manual segmentation of the SN/VTA, on the other hand, might have been biased by iron concentration 70 . Third, we would like to point out that the current study followed a cross-sectional but no longitudinal approach. Therefore, terms like "increase" or "decrease" relate to between-subjects comparisons within the age range of our sample but they do not imply individual development over time. Moreover, separate regression models for VLMT recognition and TMT were performed due to missing data. Therefore, one could argue that the observed effects of both models are due to reduced complexity. However, for verbal memory performance, correlations were also observed for VLMT recall in the multiple regression models, which speaks for a robust finding. Finally, future studies should more thoroughly take underlying mechanisms of age-related structural changes into account. In this view, hypertension and vascular diseases may relate to the brain's microstructure and cognitive functioning. Indeed, hypertension is one of the main risk factors for stroke 119 and small vessel disease 120 , both strongly linked to dementia [121][122][123] . Moreover, education 124 and life-style factors 125 have been shown to account for inter-individual differences in healthy aging.
To summarize, age-related markers of iron levels and demyelination within the BG were correlated, which is in line with the role of iron in dysfunctional myelin synthesis. Importantly, iron levels and demyelination predicted both verbal long-term memory and executive functioning, which gives novel insights into the behavioral consequences of BG microstructural changes. From a more general perspective, our results further suggest that increased iron and demyelination within the BG are typical for healthy aging and they might be distinguished from age-related differences in SN/VTA microstructure.

Data availability
The data that support the findings of this study are available on reasonable request from the corresponding authors (DB or NB). The data are not publicly available due to data security regulations by the local ethics committee.