Elevated neuroinflammation contributes to the deleterious impact of iron overload on brain function in aging

Intracellular iron is essential for many neurobiological mechanisms. However, at high concentrations, iron may induce oxidative stress and inflammation. Brain iron overload has been shown in various neurodegenerative disorders and in normal aging. Elevated brain iron in old age may trigger brain dysfunction and concomitant cognitive decline. However, the exact mechanism underlying the deleterious impact of iron on brain function in aging is unknown. Here, we investigated the role of iron on brain function across the adult lifespan from 187 healthy participants (20-79 years old, 99 women) who underwent fMRI scanning while performing a working-memory n-back task. Iron content was quantified using R2* relaxometry, whereas neuroinflammation was estimated using myo-inositol measured by magnetic resonance spectroscopy. Striatal iron increased non-linearly with age, with linear increases at both ends of adulthood. Whereas higher frontostriatal activity was related to better memory performance independent of age, the link between brain activity and iron differed across age groups. Higher striatal iron was linked to greater frontostriatal activity in younger, but reduced activity in older adults. Further mediation analysis revealed that, after age 40, iron provided unique and shared contributions with neuroinflammation to brain activations, such that neuroinflammation partly mediated brain-iron associations. These findings promote a novel mechanistic understanding of how iron may exert deleterious effects on brain function and cognition with advancing age.


Introduction
Aging is associated with brain deterioration ( Cabeza et al., 2016 ;Grady, 2012 ;Nyberg et al., 2012 ) and cognitive decline ( Gorbach et al., 2016 ;Rönnlund et al., 2005 ;Tucker-Drob et al., 2019 ). Intracellular non-heme iron is the most abundant metal in the brain and seems to have an ambivalent role across the human lifespan Kalpouzos, 2018 ). On the one hand, brain iron is involved in numerous fundamental biological processes, such as neurotransmitter synthesis, synaptic plasticity, and myelination, which are essential for early brain development and remain crucial throughout the lifespan ( Mills et al., 2010 ;Pinero and Connor, 2000 ). On the other hand, elevated brain iron in old age is deleterious for cells, because it may trigger inflammation and oxidative stress via the Fenton reaction ( Haider et al., 2014 ; The latter age-related increase may reflect disruption in iron homeostasis . In line with this, accumulation of iron over time was linked to striatal atrophy in middle-aged and older adults ( Daugherty and Raz, 2016 ). Moreover, it was recently shown that striatal iron was negatively associated with frontostriatal brain activation during a motor imagery task ( Kalpouzos et al., 2017 ), with reduced modulation to difficulty within the frontostriatal regions during a working memory task ( Rodrigue et al., 2020 ), as well as with striatal functional connectivity during resting state ( Salami et al., 2018a ). We hypothesized that the observed link between iron and brain activity may be due to altered functioning of astrocytes, which are involved in neurovascular coupling, synaptic activity, and iron buffering. In these studies, the associations between iron and neural responses were primarily driven by older adults.
A likely mechanism underlying the deleterious effects of iron overload in normal aging and neurodegenerative disorders may be related to neuroinflammation. Data from rat models of Parkinson's disease indicated that neuroinflammation may be a consequence of accumulated brain iron, suggesting a process to scavenge toxic iron and provide neuroprotection ( Olmedo-Diaz et al., 2017 ). So far, however, no study in humans has demonstrated a direct link between brain iron load and neuroinflammation, particularly in relation to age-related differences in cognition and the underlying neural response.
Covering the adult age span ( n = 187, 20-79 years old), participants were scanned with functional magnetic resonance imaging (fMRI) while engaged in a numeric working memory (WM) task ( Salami et al., 2019 ). Iron content was quantified using relaxometry R2 * , a reliable marker of iron bound to ferritin ( Langkammer et al., 2010 ). Moreover, myoinositol (MI), a marker of neuroinflammation, was estimated using magnetic resonance (MR) spectroscopy. Increased MI concentrations have been associated with activated glia ( Chang et al., 2013 ;Woodcock et al., 2019 ). We focused on the striatum due to its involvement in many cognitive functions, including WM ( Eriksson et al., 2015 ) and because it is a preferential site for aging-related iron accumulation . We first tested whether iron differentially affected taskrelated blood oxygenation level-dependent (BOLD) responses across different age groups. More specifically, we hypothesized that higher striatal iron would be beneficial for brain function and cognition in younger age. In contrast, higher striatal iron load would be detrimental in older adults with a negative link to BOLD response and cognition. To identify iron-BOLD associations during three WM load conditions, we applied multivariate partial-least-squares (PLS ( McIntosh and Lobaugh, 2004 )). PLS accommodates the simultaneous analysis of all task conditions, iron content, and age groups, which facilitates the manifestation of similar and/or opposing iron-BOLD associations across adulthood in a datadriven fashion. Second, we investigated unique and shared contributions of iron and neuroinflammation on brain function, specifically exploring whether neuroinflammation mediated the deleterious effect of iron on brain function.

Methods and materials
The IronAge study was approved by the Regional Ethical Review Board in Stockholm. All participants signed informed consent prior to data collection.
Participants. 232 individuals were recruited through advertisements in newspapers and dedicated student websites, and 216 underwent the full protocol. None of the participants reported any current or past neurological or psychiatric conditions, and none was taking any psychoactive medication. Nine individuals were excluded due to incidental brain abnormalities (i.e., tumors, strokes, neurodevelopmental conditions) and nine for suboptimal MR quality. Four individuals were excluded due to very poor behavioral n-back data. Furthermore, one extreme outlier as well as six multivariate outliers based on Mahalanobi´s distance, with the recommended p < .001 threshold for the chi-square value, were further excluded ( Tabachnick et al., 2006 ). Thus, the fi-nal sample for analyses exploring iron-BOLD association was 187 subjects. This sample size is comparable to previous studies in which iron was linked to brain measures and cognition across the adult lifespan Rodrigue et al., 2013 ). Moreover, two previous studies reported associations between iron and brain functions ( Kalpouzos et al., 2017 ;Salami et al., 2018a ). Considering correlations reported in these studies (37 < n < 42; ~.3 < r < 0.4), we obtained 98% power on a two-tailed t -test to find iron-BOLD associations (power calculation was performed in G * Power software).
An assumption-free general additive model (GAM) revealed a curvilinear association between striatal iron and age. In young adults, striatal iron increases until the age of ~42 years (inflection point 1), plateaus until the age of ~62 years (inflection point 2), after which it increases further. To have balanced age groups in terms of sample size and age range, we split our sample into three groups as follows: young ( n = 62, 28.35 ± 5.58 years), middle-aged ( n = 62, 50.19 ± 5.91 years), and older ( n = 63, 68.69 ± 4.86 years) adults. Table 1 shows participants' demographics, brain measures and online task performance. Notably, there were no differences with respect to sex distribution, education, and MMSE. Significant differences were seen for striatal iron (significant among groups) and myo-inositol (younger vs. middle-aged and older adults).
MRI acquisition. Participants were scanned on a Discovery MR750 3.0T scanner (General Electric, Milwaukee, Wisconsin) equipped with an 8-channel phased array receiving coil at the MR center of Karolinska hospital.
A structural T1-weighted 3D IR-SPGR image was obtained with the following parameters: repetition time (TR) = 6.96 ms, echo time (TE) = 2.62 ms, 176 axial slices of 1 mm thickness, in-plane resolution = 0.94 × 0.94 mm 2 , field of view (FOV) = 24 cm, flip angle = 12°. For brain iron quantification, a 3D multi-echo Gradient Recalled Echo (meGRE) sequence was used with the following parameters: TR = 27.2 ms, 124 axial slices of 1.2 mm thickness, in-plane resolution = 0.94 × 0.94 mm 2 , FOV = 24 cm, flip angle = 17°The first TE was 1.9 ms, and it was followed by 7 consecutive TEs with a constant interval of 3.18 ms between them.
A single-voxel 1 H MR spectroscopy sequence with placement of the voxel in the right putamen was performed right after the T1 sequence, with voxel size = [18 × 22 × 33 mm 3 ], volume = 10.86 mL. For higher consistency, the same operator (G.K.) placed the voxel for all participants on their T1-weighted image, avoiding the orbital, ventricular and insular areas. Voxel positioning and volume were chosen as a compromise between the best achievable spectral linewidth, with the lowest level of spectral artifacts. Voxel gradient recalled echo shimming was reproducible and converged to unsuppressed water spectral linewidth 10 ± 2 Hz for putamen. Shimming, frequency adjustment, and water suppression were carried out automatically before each acquisition. To ensure comparability with other studies, the conventional PRESS was chosen with the following parameters: TR/TE/TE1 = 1800/68/14 ms, rfpulse flip angle (degree) / bandwidth (kHz): 90/2.3, 135/1.3, 135/1.3 and 2.3 ppm transmitter offset, spectral bandwidth 5 kHz, 4096 timedomain data points. Water suppression was achieved by three CHESS (chemical shift selected suppression) pre-pulses. To enhance the voxel definition, six very sharp outer volume suppression RF-pulses were applied, surrounding each side of the voxel. The number of averages was 96, and 8-steps phase cycle averaging during MRS data acquisition was used.
In-scanner task. A numerical n-back task was administrated in the scanner. The details of the task were given in our previous work ( Salami et al., 2019( Salami et al., , 2018b. In short, a sequence of single numbers appeared on the screen for 1.5 s, with an ISI of 0.5 s. Responses were recorded for a maximum of 2 s. During every item presentation, participants reported whether the number currently seen on the screen was the same as that shown 1, 2, or 3 digits back. Each block was preceded by instructions regarding load. Participants responded by pressing one of two adjacent buttons with the index or middle finger to reply 'yes, it is the same number' or 'no, it is not the same number', respectively. A single fMRI run with 9 blocks for each condition (1-, 2-, and 3-back) was performed in random order, each block consisting of 10 trials that included 4 correct and six incorrect responses. The trial sequence was the same for all participants with counterbalanced blocks. Performance was calculated as number of correct responses except for the first item in each 1-back condition, and for items 1-2 and 1-3 in each 2-and 3-back condition. Thus, the maximum score for each condition was 81,72, 63. Offline WM. A 2-h battery of cognitive and psychomotor tests tasks was administered, following a standardized procedure. Here, we focus on the measure of WM only. A unit-weighted composite scores (mean of z-scores) was computed for WM based on the number of correct responses for 2-and 3-back (see above) as well as number of correct items in the binding task. Numerical n-back task . A sequence of single numbers appeared on the screen. During every item presentation, subjects indicated whether the digit on the screen was the same as the one shown 1, 2 or 3 digits back. Each digit was shown for 1.5 s, with an ISI of 0.5 s. Three blocks for each condition (1-, 2-, 3-back) were performed in sequential order (1-2-3; 1-2-3). The 1-back block had 13 items, the 2-back block had 14 items, and the 3-back block had 15 items.
Binding task . The binding task assessed the ability to associate visuospatial features in WM. Four colored uppercase letters were presented in the center of a 5 × 4 grid, accompanied with 5 colored crosses displayed randomly in the other squares of the grid. Participants were asked to remember the associations between the 4 colored letters with the location of the cross of the same color. Each trial started with a fixation cross for 2 s and then consonants were shown. Five seconds were allocated to this encoding phase, followed by a retention interval of 8 s (fixation cross). Participants had to determine whether a black lowercase letter was presented in the correct location by pressing yes or no ( Lecouvey et al., 2015 ). In total, 20 trials were administered. The binding score (number correct) was adjusted for test version.

MRI processing
Iron quantification. We performed relaxometry on the magnitude images of the meGRE sequence to obtain transverse relaxation rates R2 * , reflecting iron content (the higher the R2 * , the higher the iron con-tent). We fitted a monoexponential model to the square of the signal at each voxel of the data. R2 * estimation was conducted using the formula S 2 = S 0 2 . exp( − 2 . TE . R2 * ), where S is the measured signal magnitude, and S 0 is the signal amplitude corresponding to TE = 0 (see Garzón et al., 2017 ). Freesurfer version 6.0 was used on the T1 images to segment the nuclei of the basal ganglia. First, we spatially aligned the R2 * images on their respective T1 images. Next, mean R2 * values were extracted from the striatum (left and right caudate and putamen), eroding the outer line of each nucleus by 1 mm and removing the 15% most extreme values from each region, to exclude high signal arising from neighboring vessels, edge effects, and possible microbleeds.
Myo-inositol quantification. MRS data were first preprocessed in Matlab. This preprocessing included the signal-to-noise-ratio-weighted MRS signal coil, combining frequency and phase correction for every trace, and automatic "bad MRS traces " removal before final coherent averaging. Then, MRS data quantification was performed in the LCModel TM software (version 6.3-1 K, s-provencher.com). The basis set required for LCModel was simulated by quantum mechanical density matrix formalism in Matlab using the actual timing parameters used in PRESS pulse sequence and the chemical shifts and J-coupling constants published and available elsewhere ( Govind et al., 2015 ;Govindaraju et al., 2000 ). Overall, 20 metabolites were assessed, from which we were only interested in myo-inositol (MI) as a proxy for neuroinflammation related to glial activation. The basis set was calibrated using MRS phantom (BRAINO, GE Healthcare) and human data. The macromolecules and lipids were simulated as standard Lorenz-Gaussian shapes, and the baseline splinewidth was limited to 0.15 ppm. Thereafter, the fitted MRS spectra were visually inspected for artifacts. Seven participants failed to pass the quality control inspection, and were therefore excluded from the analyses involving MI data. The relative concentration of the MI complex was quantified using the ratio to total Creatine (tCr, 7 mM/kg assumed). Water-referenced values of tCr did not show any variation according to age ( r = 0.058; p = .454, n = 168). Therefore, tCr could be used as a standard to compute relative values for MI.
Functional MRI analyses : Pre-processing of the fMRI data included slice-timing correction and motion correction by unwarping and realignment to the first image of each volume. The fMRI volumes were then normalized to a sample-specific template generated using DAR-TEL ( Ashburner, 2007 ), affine alignment to MNI standard space, and spatial smoothing with an 6-mm FWHM Gaussian kernel (voxel size = 2 × 2 × 2 mm 3 ).
The preprocessed fMRI data were analyzed with spatiotemporal PLS McIntosh and Lobaugh, 2004 ) to assess commonalities and differences in the iron-BOLD association across experi-mental conditions (1-back, 2-back, and 3-back), and age groups (young, middle-aged, and older adults). PLS determines time-varying distributed patterns of neural activity as a function of experimental variables and striatal iron. An identified pattern reflects association changes across all regions of the brain simultaneously, that is, in a network fashion, rather than assemblies of independent regions, thus ruling out the need for multiple comparison correction. A detailed description of spatiotemporal PLS analysis for fMRI data has been given in previous reports ( Garrett et al., 2010 ;Grady and Garrett, 2014 ;Salami et al., 2010Salami et al., , 2012Salami et al., , 2013. Two types of PLS analyses were carried out. First, task PLS was the primary analysis to investigate group differences in brain networks associated with the experimental conditions (1-/2-/3-back) in the WM task. Second, the behavioral PLS (with striatal iron as a variable of interest) aimed to investigate (i) the presence of a multivariate spatial pattern of task-related BOLD response dependent on striatal iron content, (ii) whether this pattern varied across the three age groups In brief, the onset of each stimulus within each block of images (1back, 2-back, and 3-back) was averaged across blocks for each condition within the three age groups. A cross-block correlation matrix was computed as the correlation between neural activity across experimental conditions and striatal R2 * across different regions. Then, the correlation matrix was decomposed using singular value decomposition (SVD), to identify a set of orthogonal latent variables (LVs) representing linear combinations of the original variables.

SV D CORR = USV ′
This decomposition produces a left singular vector of striatal R2 * weights (U), a right singular vector of BOLD weights (V), and a diagonal matrix of singular values (S). In other words, this analysis produces orthogonal LVs that optimally represent relations between striatal R2 * and BOLD. Each LV contains a spatial pattern exhibiting the brain regions whose activity shows the strongest relation to striatal R2 * . To obtain a summary measure of each participant's expression of a particular LV pattern, subject-specific "brain scores " are computed by multiplying each voxel's (i) weight (V) from each LV (j) by the BOLD value in that voxel for person (m), and summing over all (n) brain voxels: Taken together, a brain score represents the degree to which each subject contributes to the multivariate spatial pattern captured by a given R2 * -driven latent variable.
The statistical significance of each LV was assessed using permutation testing. This procedure involved reshuffling the rows of the data matrix and recalculating the LVs of the reshuffled matrix using the same SVD approach. The number of times a permuted singular value exceeds the original singular value yields the probability of significance of the original LV ( McIntosh et al., 1996 ). In the present study, 1000 permutations were performed. In addition, the stability of voxel saliencies contributing to each LV was determined with bootstrap estimation of standard errors (SEs), using 1000 bootstrap samples ( Efron and Tibshirani, 1986 ). The Bootstrap Ratio (BSR; the ratio between voxel saliences and the estimated SEs) was computed and voxels with BSR > 2.8 (similar to a Z-score of 2.8, corresponding to p = .005) were considered reliable. All reliable clusters comprised contiguous voxels, with a distance between clusters of at least 10 mm. Moreover, the upper and lower percentiles of the bootstrap distribution were used to generate 95% confidence intervals (CIs) around the correlation scores to facilitate interpretation ( McIntosh and Lobaugh, 2004 ). For instance, a significant difference between correlation scores in different conditions is indicated by non-overlapping CIs. Similarly, brain or correlation scores were considered unreliable when CIs overlapped with zero.
PLS allows the analysis of relation between R2 * and BOLD in a single model, without the need of restricting the analyses to given regions-ofinterest. PLS uses all conditions, groups and behavioral measures in an experiment at once, thus offering an additional dimension by simulta-neously considering both similarities and differences across the experimental variables. If the R2 * -BOLD association is expressed in similar brain regions, but in different magnitudes across different conditions or age groups, PLS should reveal a single pattern with quantitative differences across load conditions and groups (e.g. possibly with divergent R2 * -BOLD associations across different groups). Alternatively, if the R2 * -BOLD association is expressed in different brain regions in a load-dependent manner, we would expect at least two distinct patterns.

Additional statistical analyses
Behavioral data were analyses with repeated measures ANOVA with age group (younger, middle-aged and older adults) as between-subjects and WM load as within-subjects (1-back, 2-back, 3-back) factors. Moreover, planned comparisons and t -test were used to follow-up on significant mean differences. Correlational analyses (Pearson's r and partial correlations) were conducted to investigate differences between R2 * and behavioral performance in the different age groups.
To investigate the mediating effect of MI on the link between R2 * and brain activity, we conducted correlation analyses for different age groups, adjusting for the effects of neuroinflammation. MI values with very high variability (SD > 20%; ( Bottomley and Griffiths, 2016 )) were excluded and, in addition, they were adjusted for potential difference in variability of MI (i.e.,% SD or Cramér-Rao bounds) to account for a potential bias of low MI values, which are more variable ( r = − 0.489, p < 0.001). To estimate changes in effect size, we calculated% change in effect size after mediation using the following formula: ( r 2 after mediation-r 2 before mediation)/ r 2 before mediation. Multivariate outliers within and across groups were determined using Mahalanobi´s distance, with the recommended threshold ( p = .001) for the 2 value ( Tabachnick et al., 2006 ). For all analyses, the alpha level was set to p < .05.

Striatal iron content across the adult lifespan
Striatal iron content increased with advancing age ( r = 0.592, p < .001). An assumption-free GAM of striatal R2 * as a function of age revealed a non-linear association (degree of freedom (df) = 3.34, p < .001; i.e. significant df suggests a nonlinear fit), such that iron content increased in young, reached a plateau in middle age, and finally further increased in older adults ( Fig. 1 ). Given the non-linear association between iron and age, three age groups (Younger: n = 62, 20-39 yrs; Middle-aged: n = 62, 40-59 yrs, Older: n = 63, 60-79 yrs) were specified using a combination of the reflection points (~42 and 62 years old) in GAM, balanced number of subjects, and age range across groups.
Further correlational analyses between R2 * and WM accuracy did not reveal any significant associations in younger (0.01 < r < 0.13; p s Fig. 1. A: Averaged R2 * maps for healthy younger (20-39 yrs), middle-aged (40-59 yrs), and older adults (60-79 yrs). In the younger group, iron content is higher in the pallidum, and after middle-age, iron content is higher in pallidum, putamen, and caudate. B: Increasing striatal iron content using assumption-free general additive model (GAM) with a contour reflecting 95% confidence interval. An assumption free GAM as a function of age was fitted to accurately describe differences in iron content across the adult lifespan.

WM BOLD response and aging
We used task PLS to explore group differences in brain networks associated with load conditions during n-back task, independent of the iron content. Two significant latent variables (LVs) were identified. Given that only the first LV represented the canonical WM network and was deemed to be the most age-sensitive network, we present this LV here (see Supplementary Fig. 1 for LV2). LV1 ( p < .0001; 69.9% of cross-block covariance) identified brain regions showing load-dependent BOLD response (3-back > 2-back > 1-back) and was similarly expressed across the three age groups ( Fig. 2 A and B). In addition to similarity in the overall pattern, there were some subtle group differences. More specifically, the confidence intervals (CIs) revealed that the WM network was significantly less engaged during 3-back in the older compared to both middle-aged and younger groups (nonoverlapping CIs). Moreover, the middle-aged group showed significantly lower BOLD response at 3-back compared to the younger group (nonoverlapping CIs). Finally, the dynamic range (i.e., level of activation between conditions) of the BOLD response significantly declined from young to middle-aged to older adults. The pattern of peak voxels from this LV replicates the wellknown frontostriato-parietal task-positive network which was engaged to a larger extent at 2-and 3-back than 1-back ( Fig. 2 A; Supplementary  Table 1).

WM BOLD response and iron content
The second PLS was carried out to investigate (i) the presence of multivariate spatial pattern of task-related BOLD response dependent on striatal iron content, and (ii) whether this pattern of iron-BOLD association varied across different age groups. The analysis revealed only one significant LV ( p = .03; 22.3% of cross-block correlation: Fig. 3 A and  B). This WM network demonstrated a positive and reliable correlation with striatal iron across all WM conditions in the younger group; that is, higher striatal iron was associated with higher BOLD signal. In contrast, a negative correlation between iron content and BOLD response across WM conditions was observed in the middle-aged and older groups; that is, higher iron was related to lower BOLD response. Note that iron-BOLD

Fig. 2.
Multivariate relationships of BOLD response to the load conditions as a function of age groups identified by task PLS (LV1). A: Red color is associated with working memory regions exhibiting greater activity during 3-back than 1and 2-back in a load-dependent fashion (i.e. 3-back > 2-back > 1-back). These regions mostly correspond to the fronto-parietal network and the striatal circuit. Blue color indicates brain regions reflecting greater activity during 1-back than 2-back and 3-back, and mostly maps into the default mode network. B: Brain scores for LV1. Error bars denote 95% confidence intervals, and thus nonoverlapping CIs reflect significant differences. associations were significantly different between younger and middleaged groups, but also between younger and older groups (nonoverlapping CIs; Fig. 3 B). No significant load-dependent modulation of the iron-BOLD association was observed in any groups (overlapping CIs). The network demonstrating opposing BOLD-iron associations across age groups engaged primarily frontal-parietal and striatal regions ( Fig. 3 A,  Supplementary Table 2). Importantly, some of the iron-sensitive frontal regions were part of the WM network (for overlap see Fig. 3 C) showing load-dependent BOLD modulation identified by primary task PLS analyses ( Fig. 2 A). Overall, our findings suggest that although iron impacts the regions which exhibit load-dependent modulation, the association itself is not load-dependent. Fig. 3. Multivariate relationships of BOLD response to striatal R2 * as a function of age groups. A: Red regions indicate positive R2 * -BOLD associations in the young, but negative R2 * -BOLD associations in the middle-aged and older groups. B: R2 * -BOLD associations across the three age groups. C: Yellow and green areas are the iron-sensitive regions overlaid on the regions showing load-dependent increases in BOLD response (shown in red) across the three age groups ( Fig. 2 A).
In terms of BOLD-performance links, the brain score during 3-back was related to 3-back accuracy after adjusting for age across the sample ( r = 0.197, p = .007), but there was no relationship for the two easier conditions (1-back: r = 0.009, p = .901; 2-back: r = 0.012, p = .876). This pattern suggests that the network showing opposing BOLD-iron association is beneficial for WM performance across the adult lifespan.

Associations of inflammation with BOLD and iron content
Older age was associated with higher neuroinflammation (MI; r = 0.169, p = 0.030; n = 166). Also, higher MI was correlated with higher striatal R2 * ( r = 0.314, p < 0.001, adjusting for age). Given similar patterns observed in the PLS analysis (in relation to iron-BOLD association; see Fig. 3 ) for middle-aged and older adults compared to younger adults, we investigated whether WM-related brain activations might be differently linked to neuroinflammation in younger compared to middleaged and older adults. In younger adults, individual differences in MI were not related to BOLD response (1-back: r = 0.000, p = 0.997; 2-back: r = 0.083, p = 0.544; 3-back: r = 0.186, p = 0.170) or R2 * ( r = 0.189, p = 0.163). By contrast, in middle-aged and older adults, higher MI was related to lower BOLD response during 2-back at trend level ( r = − 0.179, p = 0.062) and significantly to 3-back ( r = − 0.224, p = 0.018), as well as to higher R2 * ( r = 0.298, p = 0.001; Fig. 4 B). There was no relationship between MI and BOLD during 1-back ( r = − 0.134, p = 0.162). Given similar patterns, we collapsed across 2-and 3-back conditions to inves-tigate whether MI mediated the relationship between R2 * and BOLD in middle-aged and older adults ( Fig. 4 A). The relationship between R2 * and BOLD ( r = − 0.294, p = 0.002) was attenuated, but remained significant, after adjusting for individual differences in MI ( r = − 0.232, p = 0.014; about 38% drop in r 2 ), suggesting a partial mediation effect.

Discussion
Our study shows that the effects of brain iron content on neural activity and cognition strongly depend on age. Striatal iron load increased with age, indicating a non-linear pattern in line with postmortem data ( Hallgren and Sourander, 1958 ). Whereas higher frontostriatal activity was related to better memory performance independent of age, the link between brain activity and iron differed across age groups. Higher striatal iron was linked to greater frontostriatal activity in younger, but reduced activity after middle age. Finally, higher iron content in middleaged and older, but not younger, adults was specifically associated with higher MI, reflecting higher neuroinflammation, the latter being also linked to lower BOLD response. The pattern of results indicates that the interplay between iron and neuroinflammation exerts a deleterious effect on the neurofunctional correlates of WM in midlife and older age. Within the limits of the cross-sectional data, a possible interpretation would be that iron accumulation triggers neuroinflammation, leading to disrupted frontostriatal activation and WM decline in older age.
Our primary task PLS analysis, independent of iron content, revealed age-related alterations in the dynamic range of BOLD response within the frontostriato-parietal WM network. More specifically, older groups tend to show less load-dependent modulation, and thus more similarity in BOLD response across loading conditions. This finding concords with evidence showing age-related reductions in dynamic range of functional modulation at the level of large scale brain network, likely due to the reduced neural resources ( Hakun and Johnson, 2017 ;Kennedy et al., 2017 ;Nagel et al., 2009 ;Nyberg et al., 2009 ;Salami et al., 2018b ;Turner and Spreng, 2015 ).
Previous lifespan studies supported the notion that accumulation of subcortical iron is a marker for cognitive decline in aging Ghadery et al., 2015 ;Rodrigue et al., 2013 ). Similarly, elevated iron content was found to have negative consequences for BOLD response during different mental states ( Kalpouzos et al., 2017 ;Rodrigue et al., 2020 ;Salami et al., 2018a ). In contrast, some studies reporting beneficial effect of iron for cognition in childhood and younger age ( Darki et al., 2016 ;Hect et al., 2018 ). Our lifespan sample provide the first human evidence supporting positive and negative effects of iron on brain's functional integrity as a function of age. Iron-BOLD associations were significantly different between younger and middle-aged, but also between younger and older adults across all load conditions. This suggests state-invariant differences in iron-BOLD associations ( Salami et al., 2018b ), which can be readily observed during a low demanding task (i.e. 1-back). The opposing association between iron and BOLD in younger and older groups can be explained by aging-related changes in the homeostasis of iron.
Alterations in neurobiological mechanisms lead to iron dyshomeostasis in aging  due to declining efficiency of mechanisms such as storage of iron in the ferritin protein and efficient transportation of iron via transferrin ( Ward et al., 2014 ). Aging-related increase in free ferrous iron triggers oxidative stress and inflammation Ward et al., 2014 ). The R2 * signal does not allow distinction between free and bound iron, and mostly reflects ferritin, which is the only paramagnetic element that is present in sufficient quantity to enable detection by an MR scanner . However, accumulating presence of free iron triggers the production of ferritin ( Arosio et al., 2009 ;Arosio and Levi, 2002 ). It is therefore plausible that in older individuals, higher R2 * reflects higher levels of free ferrous iron. The fact that higher R2 * was associated with lower brain activity suggests iron dyshomeostasis in older adults. Histological studies showed that age-related iron accumulation primarily oc- Fig. 4. A: Path model indicating that the link between iron (R2 * ) and brain activations (BOLD) during working memory (2-back and 3-back) was partly mediated through myo-insitol (MI). The u-turn arrow indicates% change in r 2 after mediation analysis (see text). B: Higher MI was associated with higher R2 * in striatum in middle-aged and older adults. MI is measured in millimoles, but presented as z-scored values due to adjustment for% standard deviation in MI. curs in astrocytes ( Connor et al., 1990b ;Schipper, 1996Schipper, , 2004. It is reasonable to speculate that the link between higher iron content and lower BOLD signal is due to an astrocytic dysfunction in the neurovascular coupling due to iron accumulation in these glial cells ( Kalpouzos et al., 2017 ;Salami et al., 2018a ).
The analyses of relations between striatal MI and age revealed increased neuroinflammation in aging. Critically, increased striatal neuroinflammation, possibly due to astrocytic rather than microglia activation, was associated with elevated local iron content and decreased frontostriatal BOLD response in middle-aged and older adults. Our findings concord with previous studies showing that neuroinflammation is elevated in mild cognitive impairment and Alzheimer's disease ( Bradburn et al., 2019 ;Kreisl et al., 2013 ). Relatedly, age-related inflammatory changes contribute to cognitive decline ( Di Benedetto et al., 2017 ;Ownby, 2010 ;Papenberg et al., 2016 ). Moreover, given the sparse data on the link between neuroinflammation and cognition as well as its neural correlates in non-pathological conditions, our data also provide the first direct evidence for the detrimental effect of brain inflammation on neural correlates of cognition in midlife and older age.
The results of the mediation analysis indicated that the association between higher striatal iron and lower frontostriatal response was partly mediated through neuroinflammation ( Daugherty et al., 2019 ). That is, the direct link between iron and BOLD was still significant, albeit weaker, when neuroinflammation was included as a mediator. This finding provides the first in vivo evidence for the link between iron and neuroinflammation in the human brain ( Ndayisaba et al., 2019 ), such that the interplay between iron and MI exerts a deleterious effect on brain function in older age. Previous histopathological studies suggested that MI estimated by MRS might be indicative of astrocyte activation ( Bitsch et al., 1999 ;Woodcock et al., 2019 ). Moreover, iron accumulation in aging mostly occurs in astrocytes ( Connor et al., 1990a ;Dringen et al., 2007 ), which have a pivotal role in coupling blood flow to neural activity ( Petzold et al., 2008 ;Rossi, 2006 ). Age-related iron accumulation in astrocytes may induce elevated astrocytic inflammation which in turn disturbs the neurovascular coupling, and thus reduced BOLD signal in aging. These findings should be considered with caution, as MI has modest specificity for microglia and astrocytic activation ( Chang et al., 2013 ). Future PET studies involving radioligands specifically targeting astrocytes and microglia are called for to make firm conclusion ( Narayanaswami et al., 2018 ). Although the associations revealed by these data are cross-sectional; thus, results from the mediation analysis should be interpreted with caution. A possible mechanism of the interplay between iron, inflammation, and brain function is that increased neuroinflammation may further accelerate iron accumulation, resulting in exacerbated detrimental effects of iron on brain functioning. In concordance with this interpretation, in vitro experiments with rat cells showed that pro-inflammatory cytokines led to increment of iron content, by influencing the expression of iron transporters involved in maintaining iron homeostasis ( Urrutia et al., 2013 ;Wang et al., 2013 ).
Iron content increases tremendously during development, and reaches a supposedly optimal concentration in young adulthood ( Darki et al., 2016 ). Hence, younger adults with higher iron concentration may have an optimal cellular metabolism, resulting in optimal efficacy of specific mechanisms such as neurotransmitter synthesis and synaptic plasticity ( Hare et al., 2013 ;Ward et al., 2011 ). The fact that higher R2 * was associated with higher brain activity may suggest that R2 * reflects optimal iron homeostasis in younger adults. However, a recent study showed that younger adults with higher striatal iron load exhibited decreased brain activity with increasing task difficulty during an n-back task ( Rodrigue et al., 2020 ). Since the non-significant pattern of decreased load-dependent modulation in younger adults mimics the one reported by Rodrigue et al., an alternative interpretation of the positive iron-BOLD association in the young could be that higher neural activity reflects a compensatory response to reduced cellular metabolism, induced by higher deleterious iron load. With increasing age and iron load, such a compensatory activity may not be possible anymore. Finally, as no link was found between MI, iron and BOLD in the younger group, neuroinflammation may not be a player if this alternative interpretation is true.
In the current study, we used R2 * as a measure for iron content. Although R2 * has been widely validated as an in vivo measure of iron in the brain ( Langkammer et al., 2010 ), it is also susceptible to myelin as well as circulating blood in living tissue. The present results should be replicated with an alternative measure of brain iron, quantitative susceptibility mapping, which might be more robust in quantifying iron ( Langkammer et al., 2012 ). Moreover, we have focused on the frontostriatal circuit, but it is also important to note that other circuits (e.g., mammillothalamic tracts), and other regions are also important for memory processes ( Düzel et al., 2010 ;Murty et al., 2011 ;Vann, 2010 ).
In conclusion, our study is the first to highlight the striking differential effects of brain iron load on neural activations, with higher iron being seemingly beneficial in the young, but likely deleterious after middle age. Iron and neuroinflammation were positively associated after the age of 40 years; besides, iron has unique and shared contributions with neuroinflammation to brain dysfunction. The interplay between neuroinflammation and iron provides a mechanistic understanding of how iron content might exert opposing effects on neural function and, ultimately, cognition, possibly through a disturbance of the neurovascular coupling due to astrocytic iron-related dysfunction and inflammation. Brain iron overload has been shown in virtually all neurodegenerative disorders ( Ward et al., 2014 ). Hence, the observed findings have implications for a better understanding of the mechanisms involved in neurodegeneration.

Author credit statement
Grégoria Kalpouzos designed and performed research; Alireza Salami and G.P., Grégoria Kalpouzos, Jonas Persson, Erika J. Laukka and Rous-lan Sitnikov analyzed data; and Alireza Salami, Grégoria Kalpouzos, Goran Papenberg wrote the paper, which was edited by all authors.

Data availability statement
Data from the IronAge study are available upon request to Dr. Grégoria Kalpouzos who is the principal investigator for this project. According to the ethical application for IronAge project all requests for accessing data should be approved by the principal investigator. Conditions for data availability are to provide a short description of the project in mind as well as to agree on the general conditions for withdrawal of data as described in the agreement form specified below.

Declaration of Competing Interest
The authors declare no conflict of interest.