Non-invasive assessment of normal and impaired iron homeostasis in the brain

Abstract

Strict iron regulation is essential for normal brain function. The iron homeostasis, determined by the milieu of available iron compounds, is impaired in aging, neurodegenerative diseases and cancer. However, non-invasive assessment of different molecular iron environments implicating brain tissue’s iron homeostasis remains a challenge. We present a magnetic resonance imaging (MRI) technology sensitive to the iron homeostasis of the living brain (the r₁-r₂^* relaxivity). In vitro, our MRI approach reveals the distinct paramagnetic properties of ferritin, transferrin and ferrous iron ions. In the in vivo human brain, we validate our approach against ex vivo iron compounds quantification and gene expression. Our approach varies with the iron mobilization capacity across brain regions and in aging. It reveals brain tumors’ iron homeostasis, and enhances the distinction between tumor tissue and non-pathological tissue without contrast agents. Therefore, our approach may allow for non-invasive research and diagnosis of iron homeostasis in living human brains.

Introduction

Iron is the most abundant trace element in the human body¹. It participates in fundamental processes such as oxygen transport, cellular metabolism, myelin formation, and the synthesis of neurotransmitters^1,2,3,4. Therefore, strict iron regulation is essential for maintaining normal brain function. Brain tissue’s iron homeostasis could be characterized by the tissue’s molecular iron environment, determined by the specific milieu of available iron compounds, their iron binding capacities and aggregation states. Importantly, the molecular iron environment varies between cell types and across brain regions^3,5,6,7.

Disrupted iron homeostasis plays a major role in normal aging and in neurodegenerative diseases such as Parkinson’s disease (PD), Alzheimer’s disease (AD), multiple sclerosis, Friedreich’s ataxia, aceruloplasminaemia, neuroferritinopathy, Huntington’s disease, and restless legs syndrome^{1,2,5,6,7,8,9,10}. The two iron compounds most involved in iron homeostasis are transferrin and ferritin³. Transferrin, the main iron transport protein, carries iron from the blood into brain tissue, while ferritin, the main iron storage protein, stores excess iron atoms. When iron concentrations exceed the capacity of iron-binding proteins, oxidative stress, and cellular damage can occur¹⁰. For example, the ratio of transferrin to iron, which reflects iron mobilization capacity, differs between the brains of elderly controls and patients (AD and PD)⁷. In addition, reduction in neuromelanin-iron complexes in the substantia nigra and the locus coeruleus is considered a biomarker for PD and AD^11,12.

Impaired homeostasis of the molecular iron environment also have been reported in cancer cells^13,14. Tumor cell proliferation requires a modulated expression of proteins involved in iron uptake. In addition, iron may affect the immune surveillance of tumors¹⁵. Therefore, the availability of iron in the tumor cells’ microenvironment may affect their survival and growth rate, and subsequently the course of the disease. For example, meningioma brain tumors¹⁶, compared to non-pathological tissue, were shown to contain a higher concentration of ferrimagnetic particles and abnormal expression of iron-related genes^17,18. These findings suggest there are detectable differences in iron homeostasis between brain tumors and normal brain tissue.

The extensive implications of impaired iron homeostasis in normal aging, neurodegeneration and carcinogenesis suggest that assessment of iron homeostasis in the living brain would be highly valuable for diagnosis, therapeutic monitoring, and understanding pathogenesis of diseases⁴. Iron’s paramagnetic properties make magnetic resonance imaging (MRI) a perfect candidate for non-invasive estimation of iron content in brain tissue. In particular, iron is a major contributor to the longitudinal and effective transverse relaxation rates, R₁ and R₂^* respectively^19,20,21,22. These relaxation rates can be measured using quantitative MRI (qMRI) techniques^23,24,25. Indeed, in vivo studies often use these qMRI measurements as a proxy for iron presence in brain tissue^{21,26,27,28,29,30}. However, a major limitation of current MRI techniques is that they lack information regarding the state of iron homeostasis, as they do not have the sensitivity to discriminate between different molecular environments of iron in the brain⁴.

Early in vitro and postmortem works suggest that different iron environments can be distinguished by their iron relaxivity^{30,31,32,33,34}. The iron relaxivity is defined as the dependency of MR relaxation rates on the iron concentration³⁵. It was shown that iron relaxivity varies with the specific environment in which the iron resides^{30,31,32,33,34}. However, this approach requires direct estimation of the tissue iron concentration, which can only be acquired in vitro or postmortem. Due to this major limitation, until now the phenomenon of iron relaxivity could not be studied in living humans.

Here we propose an in vivo iron relaxivity approach sensitive for the state of iron homeostasis in the brain. Our approach fully relies on MRI parameters, and does not require estimation of the tissue iron concentration, thereby allowing non-invasive assessment of different molecular iron environments in the living brain (Fig. 1). We exploit the distinct iron relaxivities of the MR relaxation rates, R₁ and R₂^*, to construct a biophysical model of their linear interdependency, which we label the r₁-r₂^* relaxivity. Using the r₁-r₂^* relaxivity, we argue that the distinct iron relaxivity of different molecular iron environments can be estimated in vivo. We confirm this hypothesis based on a unique validation framework. First, we used a bottom-up strategy in which we evaluated the r₁-r₂^* relaxivity of different iron environments in vitro. Next, we used a top-down strategy in which we measured the r₁-r₂^* relaxivity in human brains in vivo and compared it to ex vivo quantification of iron compounds and gene expression, both at the group and the single-subject levels. In healthy subjects, we assessed the biological correlates of the r₁-r₂^* relaxivity, and compared it to other MR contrasts. In meningioma patients, we tested the r₁-r₂^* relaxivity contrast between pathological and non-pathological tissues, and compared MRI measurements to ex vivo iron homeostasis estimates of tumors. Therefore, we provide a well-validated MRI framework with promising implications for the non-invasive research and diagnosis of normal and impaired iron homeostasis in living human brains.

Fig. 1: Non-invasive assessment of normal and impaired iron homeostasis in the brain.

a MRI theory: Early in vitro and postmortem works suggest that different iron environments can be distinguished by their iron relaxivity, i.e. the dependency of MR relaxation rates on the iron concentration. This approach requires direct estimation of the tissue iron concentration, which can only be acquired in vitro or postmortem. Here we propose an in vivo iron relaxivity approach. We exploit the distinct iron relaxivities of the MR relaxation rates, R1 and R2*, to construct a biophysical model of their linear interdependency, which we label the r₁-r₂* relaxivity. Using the r₁-r₂* relaxivity, we argue that the state of iron homeostasis in the brain can be estimated in vivo. b In vitro, the r₁-r₂* relaxivity distinguishes between different synthetic iron environments (ferritin, transferrin and ferrous iron ions). c Comparison between in vivo MRI scans of the healthy human brain and published postmortem histology. The r₁-r₂* relaxivity varies with the iron mobilization capacity across brain regions and in aging. d Direct comparison between in vivo MRI scans of meningioma patients and transcriptomics and proteomics of the resected tumor tissues. The r₁-r₂* relaxivity reveals tumors’ iron homeostasis and separates tumor tissue from healthy tissue.

Results

The theoretical basis for the r₁-r₂ ^* relaxivity

The iron relaxivity is defined based on the linear relationship between the relaxation rates (R₁ and R₂^*) and the iron concentration ($\left[{IC}\right]$)³⁵:

$${{{{{{\rm{R}}}}}}}_{1}={{{{{{\rm{r}}}}}}}_{(1,{{{{{\rm{IC}}}}}})}\left[{{{{{\rm{IC}}}}}}\right]+{{{{{{\rm{c}}}}}}}_{1}\,\;\;\;\;\;\;\;\;{{{{{{\rm{R}}}}}}}_{2}^{*}={{{{{{\rm{r}}}}}}}_{(2,{{{{{\rm{IC}}}}}})}\left[{{{{{\rm{IC}}}}}}\right]+{{{{{{\rm{c}}}}}}}_{2}$$

(1)

The slopes of these linear dependencies, ${r}_{(1,{IC})}$ and ${r}_{(2,{IC})}$, represent the iron relaxivities of R₁ and R₂^*, which were shown to have different values for different iron environments^30,31,32,33. ${c}_{1}$ and ${c}_{2}$ are constants. Notably, the iron relaxivities require estimation of the iron concentration ($\left[{{{{{\rm{IC}}}}}}\right]$), thereby limiting this approach to in vitro and ex vivo studies.

Here we propose a theory which advances the relaxivity model and provides in vivo iron relaxivity measurements for identifying different iron environments in the brain. We take advantage of the fact that R₁ and R₂^* are governed by different molecular and mesoscopic mechanisms^36,37,38, and therefore each of them may have a distinct iron relaxivity in the presence of paramagnetic substances. Based on our theoretical framework (“In vivo iron relaxivity model” in Methods), the linear dependency of R₁ on R₂^* can be described by the following equation:

$${R}_{1}=\frac{\,{r}_{(1,{IC})}\,}{{r}_{(2,{IC})}}*\,{R}_{2}^{*}+c$$

(2)

The slope of this linear dependency ($\frac{\,{r}_{(1,{IC})}\,}{{r}_{(2,{IC})}}$) is defined as the r₁-r₂^* relaxivity. It represents the ratio of the iron relaxivities, ${r}_{1}$ and ${r}_{2}^{*}$, which are sensitive to the molecular environment of iron. Therefore, we hypothesize the r₁-r₂^* relaxivity reveals the distinct properties of different molecular iron environments. $c$ is the intercept (residual R₁ not explained by R₂^*). Notably, in the brain, the r₁-r₂^* relaxivity could be affected by the entire milieu of available iron compounds, their iron binding capacities and aggregation states (i.e. the molecular iron environment). For an extension of the relaxivity model in the presence of a heterogeneous iron environment and myelin see Supplementary Section 1.

In vitro validation for the sensitivity of the iron relaxivity to molecular iron environments

Before implementing this approach in the living human brain, we validated our theory by manufacturing in vitro samples of different iron compounds in a synthetic cellular membrane environment. These samples were scanned in the MRI to verify that different iron environments have different iron relaxivities. We then tested whether our r₁-r₂^* relaxivity theory could reveal these different relaxivities.

We prepared samples of transferrin, ferritin and ferrous iron ions in different cellular-like environments (water, liposomal, or protein environments; To achieve physiological iron concentrations the transferrin concentrations are higher than the ones measured in vivo^3,4). These highly controlled synthetic iron environments were scanned for R₁ and R₂^* mapping. We found that both R₁ and R₂^* increased with the concentration of iron compounds (Fig. 2a, b). The rate of this increase, defined as the iron relaxivity, was different for different iron environments (Fig. 2c, p(one-sided ANCOVA) = 1.5 × 10⁻⁷⁴, F-statistics(5) = 475 for R1, p(one-sided ANCOVA) = 2.6 × 10⁻⁷³, F-statistics(5) = 450 for R2*). We show that R₁ and R₂^* change both with the type and concentration of iron, thereby making it impossible to distinguish between iron environments with these measurements. For example, R₁ increased with the ferritin concentration, but also was higher for ferritin compared to transferrin (Fig. 2a, b, Supplementary Fig. 1a, b). Consequently, similar R₁ values can be obtained for ferritin and transferrin, depending on their concentrations (Supplementary Fig. 1b). This ambiguity can be resolved by the iron relaxivity, which differentiated the iron environments, and was consistent when computed over samples with higher or lower concentrations (Supplementary Fig. 1c). Therefore, we find that the iron relaxivity changes with the molecular environment of iron and is independent of the iron concentration.

Fig. 2: In vitro validation of the non-invasive framework for assessing the iron environments.

a, b The dependency of R1 and R2* on the iron compound concentrations for samples of different iron environments: free ferritin (N = 20), liposomal ferritin (N = 36), bovine serum albumin (BSA)-ferritin mixture (N = 22), free transferrin (N = 6), liposomal transferrin (N = 22) and liposomal ferrous iron ions (N = 20). Data points represent median values of biologically independent samples with varying concentrations relative to the water fraction ([mg/wet ml]). The linear relationships between relaxation rates and iron compounds concentrations are marked by solid lines. We define the slopes of these lines as the iron relaxivities. Dashed lines represent extrapolation of the linear fit. Shaded areas represent the 95% confidence bounds. c The iron relaxivity of R1 and R2* is different for different iron environments (p(one-sided ANCOVA) = 1.5 × 10⁻⁷⁴ and 2.6 × 10⁻⁷³ respectively). Iron relaxivities are calculated by taking the slopes of the linear relationships shown in (a, b), and are measured in [sec-1/(mg/wet ml)]. For each box, the central mark is the iron relaxivity (slope); the box shows the 95% confidence bounds of the linear fit. For the R1-iron relaxivity, the inset shows a zoom-out of the main figure, presenting the entire range of measured values. d The dependency of R1 on R2* for different iron environments. Data points represent median values of biologically independent samples with varying iron compound concentrations. The linear relationships of R1 and R2* are marked by lines. The slopes of these lines are the r₁-r₂* relaxivities, which do not require iron concentration estimation and therefore can be estimated in vivo. Dashed lines represent extrapolation of the linear fit. Shaded areas represent the 95% confidence bounds. The inset shows a zoom-out of the main figure, presenting the entire range of measured values (e) The r₁-r₂* relaxivities are different for different iron environments (p(one-sided ANCOVA) = 1.2 × 10⁻⁴⁰). For each box, the central mark is the r₁-r₂* relaxivity, and the box shows the 95% confidence bounds of the linear fit. Red dots indicate the successful prediction of the experimental r₁-r₂* relaxivity from the ratio between the iron relaxivities of R1 and R2* ($\frac{\,{r}_{(1,{IC})}\,}{{r}_{(2,{IC})}}$, shown in c, see Eq. 2). This validates our theoretical in vivo relaxivity model.

Ferritin binds thousands iron ions more than transferrin³. Thus, we wanted to exclude the possibility that the relaxivity differences were being driven by the different iron ion concentration. We estimated the iron ion concentrations for the different molecular iron environments, and verified that ferritin, transferrin and ferrous iron ions indeed have different iron relaxivities even when accounting for the discrepancies in iron concentrations (Supplementary Section 2).

The r₁-r₂ ^* relaxivity reveals the distinct iron relaxivities of different iron environments

In agreement with previous findings^30,31,32,33, our in vitro experiments indicate that iron relaxivity is sensitive to different iron environments. The iron relaxivity represents the dependency of relaxation on the iron concentration, which cannot be estimated in vivo. The r₁-r₂^* relaxivity is defined as the dependency of R₁ on R₂^*, and thus only relies on MRI measurements that can be estimated in vivo. Based on our theory, we argue that two iron environments with different iron relaxivities are also likely to differ in their r₁-r₂^* relaxivities. We validated this hypothesis using synthetic iron-containing samples. As predicted by our theoretical model, iron environments with different iron-relaxivities had different r₁-r₂^* relaxivities (Fig. 2d, e, p(one-sided ANCOVA) = 1.2 × 10⁻⁴⁰, F-statistics(5) = 103). Notably, as suggested by our theoretical formulation, the r₁-r₂^* relaxivity provides a good MRI approximation for the ratio between the iron relaxivities of R₁ and R₂^* ($\frac{\,{r}_{(1,{IC})}\,}{{r}_{(2,{IC})}}$, see Eq. (2), Fig. 2e). For liposomal transferrin and ferrous iron ions, the model’s prediction is slightly outside the confidence interval of the experimental results. This seems to be related to R₂^* estimations, as the prediction improves when replacing R₂^* with R₂ (Supplementary Section 3). Similar to the iron relaxivities, the r₁-r₂^* relaxivity is consistent across iron concentrations (Supplementary Fig. 1d). Hence, the r₁-r₂^* relaxivity is more sensitive to the molecular iron environment than R₁ and R₂^* by themselves. In addition, we validated that the r₁-r₂^* relaxivity is sensitive to the paramagnetic properties of iron-binding proteins. We found that apo-transferrin (transferrin which is not bound to iron) has a much smaller r₁-r₂^* relaxivity compared to iron-bound transferrin (p(one-sided ANCOVA) = 5.6 × 10⁻⁸, F-statistics(1) = 44.38; Supplementary Fig. 5). This implies that it is paramagnetic properties that induce the r₁-r₂^* relaxivity that we measure. Taken together, these results validate our theory, indicating that the r₁-r₂^* relaxivity can be used to measure iron relaxivity in vivo for exposing the distinct paramagnetic properties of different molecular iron environments.

Brain tissue includes a complex milieu of iron compounds. Therefore, we tested in vitro the r₁-r₂^* relaxivity of a heterogenous molecular iron environment of ferritin-transferrin liposomal mixtures (Supplementary Section 4). We found that changing the transferrin-ferritin ratio leads to considerable changes in the r₁-r₂^* relaxivity, even in mixtures with low ratio of transferrin compared to ferritin as in the brain³. Importantly, these changes were above the detection limit of the in vitro r₁-r₂^* relaxivity measurement (Supplementary Fig. 9). Hemoglobin affects the R₁ and R₂^* relaxivities of blood³⁹. In vitro, we found that ferritin and transferrin have distinct r₁-r₂^* relaxivities even in the presence of hemoglobin (Supplementary Section 5). Myelin is a major contributor to R₁ and R₂^* in the brain^{4,19,24,28,40,41,42,43,44,45}. Since myelin is composed mainly of lipids, we tested the effect of the myelin fraction on the iron relaxivity by varying the liposomal fractions in our in vitro experiments. We found that the r₁-r₂^* relaxivities are stable for different liposomal fractions and lipid types (Supplementary Section 6). These results suggest that the r₁-r₂^* relaxivity is less sensitive to the lipid concentration and composition compared to R₁ and R₂^*.

The r₁-r₂ ^* relaxivity provides a unique MRI contrast in the in vivo human brain

Following the in vitro validation, we measured the r₁-r₂^* relaxivity in the living human brain. For this aim we calculated the linear dependency of R₁ on R₂^* across voxels of different anatomically-defined regions of interest (ROIs, see “r₁-r₂^* relaxivity computation for ROIs in the human brain” in Methods). We found distinct r₁-r₂^* relaxivities for different brain regions (Fig. 3a). The heterogeneous distribution of the in vivo r₁-r₂^* relaxivity in the brain was consistent across healthy subjects (age 27 ± 2 years, N  =  21, Fig. 3b) and was reproducible in scan-rescan experiments (Supplementary Fig. 14). Our in vitro results indicated that the r₁-r₂^* relaxivity provides different information compared to R₁ and R₂^*. In agreement, in the in vivo human brain we found that the r₁-r₂^* relaxivity produces a unique contrast, statistically different from R₁ and R₂^* (Fig. 3b–d; p = 0.031 and test statistics = 0.13 for the two-sample Kolmogorov–Smirnov test comparing the r₁-r₂^* relaxivity distribution to R₁, and p = 2.5 × 10⁻⁴ and test statistics = 0.19 comparing it to R₂^*). While the r₁-r₂^* relaxivity represents the slope of the R₁-R₂^* linear fit, we also find that the intercept varies across the brain (Supplementary Section 7).

Fig. 3: The in vivo r₁-r₂* relaxivity provides a unique contrast in the brain.

a The dependency of R1 on R2* in four representative brain regions; occipital white matter (WM-occipital), occipital cortex (CTX-occipital), Thalamus & Putamen of a single subject. R2* and R1 measurements across voxels were binned (dots represent the median; shaded areas represent the mean absolute deviation), and a linear fit was calculated. The slopes of the linear fit represent the dependency of R1 on R2* (r₁-r₂* relaxivity) and vary across brain regions. b The r₁-r₂* relaxivity across the brain. Left: the r₁-r₂* relaxivity in different brain regions, the variation in each region is across normal subjects (age 27 ± 2, N  =  21). Within-region, this measurement is stable across subjects. It shows clear difference between regions, thus indicating its reliability. The 25th, 50th, and 75th percentiles and extreme data points are shown for each box. Right: the contrast of the r₁-r₂* relaxivity across the brain. Red, yellow and gray distributions represent the values of the r₁-r₂* relaxivities in sub-cortical (sub-CTX), white-matter (WM) and cortical (CTX) brain regions, respectively. c, d Similar analyses for R1 and R2* values, in which the gray-matter vs. white-matter contrast is much more dominant compared to the r₁-r₂* relaxivity. Hence, the r₁-r₂* relaxivity provides distinct information compared to R1 and R2*, beyond the WM-GM separation. Results in this entire figure are for ROIs in the left hemisphere.

The r₁-r₂^* relaxivity is calculated for an anatomically-defined ROI in the brain, in which the fit is performed across voxels. A voxel-wise r₁-r₂^* relaxivity visualization based on each voxel’s local neighborhood, as well as comparison to the R₁ and R₂^* contrasts, is demonstrated in Supplementary Section 8.

The effect of myelin on the r₁-r₂ ^* relaxivity

The sensitivity of R₁ and R₂^* to the myelin content is known to produce contrasts that are governed mainly by the differences between white-matter and gray-matter tissues^{4,19,24,28,40,41,42,43,44}.

As expected, we find a strong distinction between gray-matter and white-matter regions in R₁ and R₂^* values (Fig. 3c, d). However, the contrast of the r₁-r₂^* relaxivity across the brain shows a unique spatial pattern and reveals differences between brain regions beyond the typical white matter—gray matter differentiation (Fig. 3b). For example, we found the temporal, parietal and occipital white-matter regions to be indistinguishable in terms of their R₁ and R₂^* values (p(one-way ANOVA) = 0.99, statistics(2) = 0.01 for R2* and p(one-way ANOVA) = 0.48, statistics(2) = 0.74 for R1, Supplementary Fig. 20), but these regions were separable based on their different r₁-r₂^* relaxivities (p(one-way ANOVA) = 5 × 10⁻¹¹, statistics(2) = 36, Supplementary Fig. 20).

To further establish that the r₁-r₂^* relaxivity is less sensitive to the myelin content relative to R₁ and R₂^* we compared it to several in vivo myelin markers. The qMRI measurements of the macromolecular tissue volume (MTV)⁴⁶,magnetization transfer saturation (MTsat)⁴⁷, and mean diffusivity (MD) were shown to approximate the myelin content and characteristics^{48,49,50,51,52,53}. These myelin-sensitive markers were highly correlated with R₁ and R₂^* but were not significantly correlated with the r₁-r₂^* relaxivity (Fig. 4a, Supplementary Figs. 21, 22). In a biophysical model of the r₁-r₂^* relaxivity, that accounts for the presence of myelin and iron compounds, we find that the variability of myelin within an ROI can affect the r₁-r₂^* relaxivity measurement (Supplementary Section 9). Yet, an in vivo estimate of this myelin characteristic revealed it explains only 30% of the variability in the in vivo r₁-r₂^* relaxivity across the brain (Supplementary Fig. 28). We also performed a set of numerical simulations in which we consider the contributions of multiple brain tissue components to the relaxivity measurement (Supplementary Section 9). As in the in vivo brain, we found that changes in the myelin concentration substantially affect the simulated measurements of R₁, R₂^*. However, myelin-related changes were not the main component governing the simulated measurement of the r₁-r₂^* relaxivity, and in simulations of physiological conditions they could not explain the variability in the r₁-r₂^* relaxivity across the brain. We also find that T1w/T2w contrast, which serves as a semi-quantitative myelin marker⁵⁴, is different from the r₁-r₂^* relaxivity (Supplementary Section 10). Importantly, the R₁/R₂^* ratio may be more similar to the r₁-r₂* relaxivity than the individual measurements of R1 and R2*. Yet, in vivo estimation and numerical simulation of the R₁/R₂^* ratio show it is uncorrelated with the r₁-r₂* relaxivity and has a different biophysical interpretation (Supplementary Section 11).

Fig. 4: The r₁-r₂* relaxivity has unique biological correlates compared to R1 and R2* across the brain and in aging.

R1, R2* and the r₁-r₂* relaxivity were measured in vivo across younger (aged 23–63 years, N = 26) and older (aged 65–77 years, N = 13) subjects (different marker shapes) in 11 brain regions (different colors, WM white matter, CTX cortex). Each row presents the correlations of these qMRI measurements with a different in vivo or ex vivo histological feature (fitted model and 95% confidence bounds are presented for significant correlations): (a) qMRI vs. the macromolecular tissue volume (MTV), an in vivo myelin-sensitive marker, measured for younger (aged 23–63 years, N = 26) and older (aged 65–77 years, N = 13) subjects. Unlike R1 and R2*, the r₁-r₂* relaxivity is not linearly related to this myelin-sensitive marker (See Supplementary Figs. 21, 22 for additional in vivo myelin-sensitive markers). b qMRI vs. the iron concentration (postmortem, from the literature^{5, 7}) measured for younger (aged 27–64 years, N > = 7) and older (aged 65−88 years, N > = 8) subjects (for precise sample sizes see Table 2). Notably, the r₁-r₂* relaxivity is not significantly correlated with the iron content when excluding the outlier values in the globus pallidum while the R2* correlation survives this exclusion (Supplementary Section 12). c qMRI vs. the iron mobilization capacity (transferrin (TF) /iron ratio), an iron homeostasis marker (postmortem, from the literature^{5, 7}, same subjects as in b). Only the r₁-r₂* relaxivity is significantly correlated with the iron mobilization capacity, implying for its sensitivity to the iron homeostasis across brain regions and in aging. For all panels, data points show mean values and error bars show the mean absolute deviation across subjects, p-values are for one-sided F-test corrected for multiple comparisons (FDR), full statistics are in supplementary table 2.

The r₁-r₂ ^* relaxivity correlates with the iron mobilization capacity across the brain and in aging

Next, we tested the sensitivity of the r₁-r₂^* relaxivity to the state of iron homeostasis across the normal brain and in aging. We aggregated previously reported postmortem histological data describing iron, ferritin and transferrin concentrations in different brain regions of young (aged 27–64 years, N > = 7) and older (aged 65–88 years, N > = 8) adults^5,7,9 (see “Group-level comparison of qMRI parameters and histological measurements” in Methods). We performed a group-level comparison between these postmortem findings and in vivo MRI parameters, which we measured in the same brain regions and age groups (healthy young subjects aged 23–63 years, N = 26; older subjects aged 65–77 years, N = 13). As expected, R₂^* was significantly correlated with iron concentration (R² = 0.56, p(one-sided F-test corrected for FDR) = 2.8 × 10⁻⁴, statistics(20) = 25.4; Fig. 4b, full statistics are in Supplementary table 2). Importantly, this result validates the agreement between the in vivo and postmortem datasets, thus allowing to further examine the biological correlates of the r₁-r₂^* relaxivity. For this aim we estimated a feature of the iron homeostasis, the iron mobilization capacity (transferrin/iron⁷), available in the postmortem dataset. This measure was not correlated with R₂^* or R₁ (Fig. 4c). However, the iron mobilization capacity^5,7 was significantly correlated with the r₁-r₂^* relaxivity across brain regions and age groups (R² = 0.62, p(one-sided F-test corrected for FDR) = 7.3 × 10⁻⁵, statistics(20) = 32.7, Fig. 4c). Hence, the r₁-r₂^* relaxivity, unlike R₁ and R₂^*, is sensitive to the obtained ex vivo measurements of the iron mobilization capacity across the brain and can capture the effect of aging on this iron homeostasis marker.

The r₁-r₂^* relaxivity was much less correlated with the absolute ferritin, transferrin or iron concentrations (Fig. 4b, Supplementary Fig. 38). Moreover, the correlation with the iron concentration was driven mostly by outliers (the globus pallidum), unlike the sensitivity of the r₁-r₂^* relaxivity to the iron homeostasis marker (Supplementary Section 12). Therefore, the r₁-r₂^* relaxivity is less sensitive to absolute changes in the iron concentrations. In return, these results imply that the r₁-r₂^* relaxivity enhances the sensitivity of MRI to the iron homeostasis. We corroborated these findings with numerical simulations of the r₁-r₂^* relaxivity (Supplementary Section 9).

The r₁-r₂ ^* relaxivity enhances the distinction between tumor tissues and non-pathological tissue

While the r₁-r₂^* relaxivity forms a unique pattern of changes across the brain, it needs to be established that this contrast contains meaningful clinical information, that can complement the contrasts of R₁ and R₂^*. For this aim, we evaluated the MRI contrast between pathological and normal-appearing tissues of patients with meningioma brain tumors (N = 18, Fig. 5a, b, full statistics are in Supplementary Table 3). The diagnosis of brain tumors and their delineation from the surrounding non-pathological tissue is routinely performed using contrast-enhanced MRI, which requires the injection of an external gadolinium (Gd)-based contrast agent with paramagnetic properties⁵⁵. As expected, when using Gd-based contrast, tumor tissue was distinct from white-matter and gray-matter tissues (Fig. 5c, Cohen’s d = 1.4, p < 10⁻⁴, statistics(16) = −5.8 for tumor-gray matter; Cohen’s d = 1.18, p < 10⁻³ statistics(16) = −4.9 for tumor-white matter). Recently renewed concerns about the long-term safety of Gd-based agents^56,57, highlight the need for Gd-free MRI techniques that can serve as safe alternatives⁵⁸. However, without Gd-agent injection, both for R₁ and R₂^* values the biggest effect size was observed between white-matter and gray-matter tissues, with no significant difference between gray-matter and tumor tissues (Cohen’s d = 0.43, p = 0.08, statistics(17) = −1.8 for R₁, Cohen’s d = 0.10, p = 0.67, statistics(17) = 0.43 for R₂^*, Fig. 5d, e). This demonstrates the poor performances of R₁ and R₂^* in Gd-free tumor tissue delineation. Importantly, the r₁-r₂^* relaxivity greatly enhanced the contrast between tumor tissue and non-pathological tissue, without contrast agent injection (Fig. 5f, Cohen’s d = 1.5, p = 7 × 10⁻⁶, statistics(17) = −6.4 for tumor-gray matter; Cohen’s d = 4.32, p = 1.2 × 10⁻¹², statistics(17) = −18.3 for tumor-white matter). This Gd-free enhancement was comparable in size to the effect of Gd-based contrast. Moreover, we provide an example for a voxel-wise visualization of the r₁-r₂^* relaxivity approach in a representative meningioma patient, which demonstrates the Gd-free tumor enhancement (Supplementary Fig. 18). These results emphasize the improved sensitivity of the r₁-r₂^* relaxivity to the unique tumor microenvironment, which may have wide clinical implications as a safe alternative for contrast agents’ injections.

Fig. 5: Application of the r₁-r₂* relaxivity on meningioma brain tumors.

a From top to bottom: Gd-enhanced T1-weighted image, R1 map and R2* map in a representative subject with a meningioma brain tumor (white arrow). b The dependency of R1 on R2* (the r₁-r₂* relaxivity) for the white-matter (WM, frontal), gray-matter (GM, frontal) and tumor tissue of the same subject. R2* and R1 measurements across voxels were binned (dots represent the median; shaded areas represent the mean absolute deviation), and a linear fit was calculated. The slopes of the linear fit represent the r₁-r₂* relaxivity. Tumor tissue exhibits distinct r₁-r₂* relaxivity relative to non-pathological tissue. The r₁-r₂* relaxivity is calculated across voxels (R1 and R2* values shown in a) for each ROI. A voxel-wise visualization of the r₁-r₂* relaxivity available in Supplementary section 8. c–f The contrast between the white-matter (WM), gray-matter (GM) and tumor tissues across patients for the Gd-enhanced contrast (inverted and scaled relative to the maximal value across subjects for visualization, [a.u.], N = 17) (c), R2* (N = 18) (d), R1 (N = 18) (e) and the r₁-r₂* relaxivity (N = 18) (f). Only the r₁-r₂* relaxivity produces significant differences between tumor and GM tissues without contrast agents. Left: boxes present the variation in the contrasts across patients. The 25th, 50th, and 75th percentiles and extreme data points are shown. The d-values represent the effect size (Cohen’s d) of the differences between tissue types, and the significance level is based on a paired-sample t test (two-sided). Gray lines extend between values of the same patient. Right: the distribution of the values between WM, GM and tumor tissue across patients. Estimates in non-pathological tissues are for the tumor-free hemisphere. p  <  0.05; **p  <  0.01; ***p  <  0.001 (full statistics are in supplementary table 3).

The r₁-r₂ ^* relaxivity is associated with unique biological pathways and gene expression profiles

We further examined how the tumor characteristics obtained by the r₁-r₂^* relaxivity differ from the information contained in R₁ and R₂^*. For this aim, we examined the associations of in vivo MRI measurements with underlying gene-expression profiles for the same tissue. We analyzed cases in which the MRI scans of the meningioma patients were followed by surgical interventions, to obtain matching resected tumor tissue samples that we profiled by bulk RNA-sequencing (Fig. 6a). For these tumor samples (N = 17), we performed an unbiased analysis to identify genes and molecular pathways that could be linked with the in vivo measured MRI parameters. For each gene we calculated the correlation between the expression level and the in vivo MRI measurements (r₁-r₂^* relaxivity, R₁ and R₂^*) across patients. We then performed gene set enrichment analysis (GSEA)^59,60 to identify molecular functions that are significantly associated with each MRI measurement. In total, we found 9, 55 and 59 significantly enriched gene sets for R₁, R₂^* and the r₁-r₂^* relaxivity, respectively (p(two-sided permutation test) < 0.01 after familywise error rate (FWER) correction; Supplementary Table 4 and Supplementary Fig. 40). These gene sets define genes linked to a specific biological pathway. Almost half of the significant gene sets were exclusively associated with the r₁-r₂^* relaxivity, and not with R₁ or R₂^* (Supplementary Fig. 40). The enrichment score represents the degree to which the genes within a set were positively or negatively correlated with MRI measurements. In examining the associations of MRI measures to biological pathways, as reflected in the enrichment score, we found that the r₁-r₂^* relaxivity clustered separately from R₁ and R₂^* (Fig. 6b). The clustering results were replicated when performed on the p-value of the enrichment, or on the subset of genes within the top enrichment pathways. This implies that the r₁-r₂^* relaxivity reflects unique cellular and molecular properties, undetectable by the separate analysis of R₁ and R₂^*. Therefore, the in vivo r₁-r₂^* relaxivity provides a unique dimension for measuring microstructure and gene expression features across the brain. The gene enrichment analysis that we performed on resected brain tumors (Fig. 6b) can provide insights into the biological pathways associated with the r₁-r₂^* relaxivity. The two most enriched pathways for the r₁-r₂^* relaxivity were “immunoglobulin complex” (normalized enrichment score (NES)= −3.62, FWER p-value < 0.001; Supplementary Fig. 41a) and “scavenging of heme from plasma” (NES = −3.27, FWER p-value < 0.001; Supplementary Fig. 41b). While the former may relate to the response of the immune system to the cancerous process^61,62, the latter involves the absorption of free heme, a source of redox-active iron⁶³. This iron-related pathway was not significantly associated with R₁ (FWER p-value > 0.05) and was less significantly associated with R₂^* compared to the r₁-r₂^* relaxivity (FWER p-value = 0.038). Moreover, we examined the main genes involved in iron regulation: transferrin receptor (TFRC), ferritin heavy-chain polypeptide 1 (FTH1) and ferritin light-chain polypeptide (FTL)^64,65. Both TFRC and FTH1 were included in the subset of genes within the top enrichment pathways for the r₁-r₂^* relaxivity, but were not found to be associated with R₁ or R₂^*. These findings provide evidence at the level of gene expression for the sensitivity of the r₁-r₂^* relaxivity to iron homeostasis.

Fig. 6: The r₁-r₂* relaxivity measured in vivo in meningioma patients agrees with iron homeostasis markers estimated ex vivo on surgical specimens of meningiomas from the same patients.

a Unique comparison between qMRI parameters measured in vivo and ex vivo iron histology of the same human tissue; MRI scans of meningioma patients (N = 17) were followed by surgical interventions, to obtain matching resected tumor tissue samples. These tissue samples were profiled by bulk RNA-sequencing to obtain gene expression profiles, and by western-blot analysis to estimate the transferrin/ferritin ratio, a marker for iron homeostasis. b Gene set enrichment analysis for the correlation of MRI with gene expression. Rows show significant biological pathways, columns represent R1, R2* and the r₁-r₂* relaxivity. The dendrogram shows hierarchical clustering of the normalized enrichment scores. The r₁-r₂* relaxivity clustered separately from R1 and R2* and is therefore enriched for unique biological pathways. The two most enriched pathways for the r₁-r₂* relaxivity are highlighted in yellow. c The r₁-r₂* relaxivity, R1 and R2* measured in vivo for tumor tissues classified as having either low or high transferrin-to-ferritin ratios (Tf/Fer). Tf/Fer ratio, an iron homeostasis marker, was estimated using western-blot analysis following surgical resection of the tissue. Data points represent independent measurements of individual patients (N = 16). The threshold between groups was set to 1 based on the median across patients (N = 8 for each group). While R1 and R2* cannot distinguish between the groups with different Tf/Fer ratios, the r₁-r₂* relaxivity is higher in tissues with a high Tf/Fer ratio, indicating its sensitivity to iron homeostasis. Each box shows the 25th, 50th, and 75th percentiles and extreme data points. p-values presented are for two-sample t tests (two-sided).

The r₁-r₂ ^* relaxivity reveals differences in iron homeostasis between tumor tissues

We further validated the sensitivity of the r₁-r₂^* relaxivity to the iron homeostasis at the proteomics level, by comparing in vivo MRI measurements to ex vivo iron homeostasis estimation on the same tissue (Fig. 6a). MRI scans of the meningioma patients were followed by surgical interventions, to obtain matching resected tumor tissue samples that we analyzed by western-blot. We compared in vivo MRI values of tumor tissue to its transferrin/ferritin ratio which serves as a proxy for iron homeostasis. Neither R₁ nor R₂^* showed significant differences between tumors with low and high transferrin/ferritin ratios (Fig. 6c, p = 0.84, statistics(14) = −0.2 for R1 and p = 0.93, statistics(14) = −0.1 for R2*). Notably, the r₁-r₂^* relaxivity was significantly higher for tumors with high transferrin/ferritin ratio compared to tumors with low transferrin/ferritin ratio (p = 0.005, statistics(14) = −3.3, Fig. 6c). Therefore, as established by both gene expression and proteomics analyses, the r₁-r₂^* relaxivity measured in vivo detects pathological disruptions in iron homeostasis which were previously only observable ex vivo.

Discussion

We present an MRI relaxivity approach, with increased sensitivity to different molecular environments of iron. First, we confirm in vitro that different iron environments induce different relaxivities, which can be estimated with MRI using the r₁-r₂^* relaxivity. When examining R₁ and R₂^* of different iron environments, we find the molecular state of iron is confounded by the strong effects of iron and myelin concentrations. In this in vitro setting, we show that the r₁-r₂^* relaxivity reveals the sensitivity of MRI to the intrinsic paramagnetic properties of different iron environments. In the healthy human brain, we show that the r₁-r₂^* relaxivity provides a unique MRI contrast. This contrast is less sensitive to myelin compared to other quantitative MRI parameters. Interestingly, it does vary with an iron homeostasis marker⁷ across brain regions and age groups. We further demonstrate that the r₁-r₂^* relaxivity contains meaningful clinical information associated with iron homeostasis, which was previously inaccessible to conventional MRI approaches. In meningioma patients, we find that the r₁-r₂^* relaxivity contrast is useful for enhancing the distinction between tumor tissue and non-pathological tissue. We substantiate this finding by associating in vivo MRI measurements with RNA sequencing and protein expression levels of the same tumor tissues. We find that the r₁-r₂^* relaxivity is associated with unique iron-related biological pathways and reveals the state of iron homeostasis in tumors.

Relaxivity is commonly employed to characterize MR contrast agents⁶⁶. While most contrast agents induce relaxation based on their paramagnetic or superparamagnetic properties, some agents elevate the R₁ relaxation rate more efficiently while others elevate R₂^*. R₁ relaxation mechanisms are affected by local molecular interactions, while R₂^* is sensitive to more global effects of extended paramagnetic interactions at the mesoscopic scale³⁶. We show that by contrasting these two different mechanisms, we gain sensitivity to the endogenous iron environment, without the injection of an external contrast agent.

The concept of iron relaxivity, and its sensitivity to the molecular environment of iron, was previously suggested by several postmortem and in vitro studies^{30,31,32,33,34}. We reproduce these results in our in vitro experiments and further demonstrate that different iron environments have different iron relaxivities. Moreover, Ogg et al. ³⁰ calculated the iron relaxivity by comparing postmortem measurements of iron concentration for different age groups to the typical R₁ values in those age groups. They found that this approximation of iron relaxivity was higher in the gray matter and white matter than in sub-cortical structures. We replicate this result in living subjects, based on our approach for estimating iron relaxivity in vivo. The theoretical derivation we propose for the r₁-r₂^* relaxivity shows that it represents the ratio of the iron relaxivities of R₁ and R₂^*. This theory was supported by our in vitro experiments. Therefore, we exploit the different relaxation rates for a biophysical model of their linear interdependency, thus allowing to approximate the iron relaxivity in the living brain.

First, we evaluated the sensitivity of the r₁-r₂^* relaxivity to the molecular environment of iron based on in vitro experiments with ferritin, transferrin, and ferrous iron ions. We show that these iron environments induce different relaxivities, even when accounting for the discrepancies in iron binding. Moreover, we show that the macromolecular environment in which the iron reside can alter the relaxivity as well. For example, ferritin induces different relaxivities in the presence of liposomes, proteins or when it is free. As implied previously³¹, this effect can be explained by the encapsulation and the spatial distribution of paramagnetic molecules. Therefore, similarly to different contrast agents, the physical mechanisms by which iron compounds interact with the surrounding water environment are inherently different. To further confirm that our results are related to paramagnetic properties and not the presence of the proteins themself, we tested MRI measurements of apo-transferrin (transferrin unbound to iron). In this case, the r₁-r₂^* relaxivity of transferrin vanished. Therefore, the presence of iron in different in vitro environments can be detected by the r₁-r₂^* relaxivity. Moreover, evaluating ferritin-transferrin mixtures, we find that in heterogenous iron environments the specific composition of the iron milieu affects the r₁-r₂^* relaxivity.

The biological interpretation of the r₁-r₂^* relaxivity measurement is more ambiguous when applied to the in vivo human brain. This is due to the fact that most qMRI parameters, including R₁ and R₂^*, are known to suffer from low biological specificity in the brain^{4,19,24,28,40,41,42}. The primary MRI contrast between gray matter and white matter usually is associated with myelin, while an additional and often correlated effect is attributed to the iron concentration⁴. A major strength of our relaxivity approach is that it captures distinct biological and pathological information, not acquired by traditional qMRI parameters such as R₁ and R₂^*. We show that the great contrast between white matter and gray matter usually observed in R₁ and R₂^* is no longer as substantial in the r₁-r₂^* relaxivity. In vitro, we show that the r₁-r₂^* relaxivity is stable across iron and liposomal concentrations. To further demonstrate the minimal effect of myelin on the r₁-r₂^* relaxivity in the brain, we employed the qMRI measurements of MTV⁴⁶, MTsat⁴⁷ and MD, which were shown to approximate the myelin content^{48,49,50,51,52,53}. These in vivo myelin markers are all highly correlated with R₁ and R₂^* but not with the r₁-r₂^* relaxivity. The semi-quantitative T1w/T2w myelin marker⁵⁴ is also not correlated with the r₁-r₂^* relaxivity. Adapting the biophysical model for the r₁-r₂^* relaxivity to account for myelin, we find that the variability in myelin within an ROI can still affect the r₁-r₂^* relaxivity measurement. Yet, in vivo estimate of the myelin variability only explains ~30% of the variation in the in vivo r₁-r₂^* relaxivity across the brain. In brain tissue numerical simulations, we show that the myelin content substantially affects the measurements of R₁ and R₂^*, but it cannot by itself explain the measured variability in the r₁-r₂^* relaxivity across the brain. The R₁/R₂^* ratio may be more similar to the r₁-r₂^* relaxivity than the individual measurements of R₁ and R₂^*. However, the R₁/R₂^* ratio is sensitive to the relative magnitudes of R₁ and R₂^*, and the r₁-r₂^* relaxivity is only affected by the shared variation of R₁ and R₂^*. We demonstrate in vivo that these measurements are different. In simulations we find that unlike the r₁-r₂^* relaxivity, the R₁/R₂^* ratio is sensitive to iron compounds concentration and myelin content. Nonetheless, it could be that the r₁-r₂^* relaxivity contains some residual contribution of the myelin content. We show that this contribution is much more limited compared to traditional qMRI parameters such as R₁ and R₂^*. In return, the MRI measurement of the r₁-r₂^* relaxivity enhances the contrast between pathological tissue and normal tissue and is associated with distinct gene expression pathways. Hence, the relaxivity framework reveals distinct biological features otherwise undetectable in standard qMRI measurements.

Still, an open question remains whether the biological interpretation of the r₁-r₂^* relaxivity demonstrated in vitro is also measurable in vivo. We bring several evidence that the r₁-r₂^* relaxivity reveals the sensitivity of MRI to properties of the molecular iron environment, otherwise confounded by myelin and iron concentrations. First, in the healthy and aging brain we show that unlike R₁ and R₂^*, the r₁-r₂^* relaxivity is correlated with an iron homeostasis marker, the iron mobilization capacity (transferrin/iron). We also show that the r₁-r₂^* relaxivity is less correlated with the absolute ferritin, transferrin or iron concentrations, indicating that it enhances the sensitivity to the homeostasis between iron compounds, rather than their absolute concentrations. Importantly, we corroborate these findings based on in vivo and ex vivo analyses of meningioma patients. We find that the variability in the in vivo r₁-r₂^* relaxivity of brain tumors is explained by their transferrin-ferritin ratios and is associated with iron-related genes. These results could not be obtained for the individual measurements of R₁ and R₂^*. To further validate that the r₁-r₂^* relaxivity could be sensitive to the iron homeostasis, even when accounting for the massive effect of the myelin and iron concentrations on the MRI signal, we generated a simulation of a brain-like environment which contains multiple tissue. We provide an example for physiological changes in the iron environment, based on the ferritin-transferrin fraction, which could lead to considerable changes in the r₁-r₂^* relaxivity, well above the detection limit of this MRI measurement. Therefore, we show in vitro, in vivo, ex vivo and in numerical simulations, that the r₁-r₂^* relaxivity measurement could allow to detect changes in iron homeostasis under physiological conditions.

Our in vitro experiments are based on ferritin, transferrin and ferrous iron ions as examples for variable iron environments. Yet, in the human brain our approach is probably more broadly sensitive to the entire milieu of iron compounds, their iron binding characteristics and aggregation states. Other iron compounds that exist in the brain, such as hemoglobin, hemosiderin, neuromelanin, magnetite, ferric ion, lactoferrin and melanotransferrin⁴, might have distinct iron relaxivities as well (as we have shown for the iron relaxivity of hemoglobin). Moreover, other characteristics of the iron environment such as iron compounds’ aggregate sizes, intra-aggregate spacing, spatial distributions and iron loadings, could all have an additional effect on the iron relaxivity³⁴. For example, catecholamine neurons of the substantia nigra and locus coeruleus are rich in neuromelanin-iron complexes^4,67 which could contribute to the r₁-r₂^* relaxivity measurement in these regions (as evident by our distinct results in the pallidum). In addition, the paramagnetic properties of deoxyhemoglobin in capillaries and veins and their orientations are known to affect the R₂^* measurement^68,69,70. In vitro, we found that ferritin and transferrin have distinct r₁-r₂^* relaxivities even in the presence of hemoglobin. However, hemoglobin could have a more substantial effect in vivo. Thus, the r₁-r₂^* relaxivity might also be sensitive to the iron environment in the vascular system, which is crucial for brain iron metabolism and homeostasis⁷¹. Moreover, on the cellular scale there is substantial heterogeneity in the iron environment^72,73. R₁ and R₂^* each represent some average of different spatial and cellular compartments, but it is unclear how the r₁-r₂^* relaxivity is weighting such cellular contributions. Similar to the measurements of R₁ and R₂^*, the r₁-r₂^* relaxivity might be contaminated by biases related to magnetic field inhomogeneities²³. Yet, we show that the r₁-r₂^* relaxivity is reproducible in scan-rescan experiments (both in vitro and in vivo) and is stable across subjects, thus indicating the limited effect of such biases. Another limitation of our approach is that we estimate the dependency of R1 on R2* across voxels, and it therefore produces one relaxivity measurement per anatomically-defined ROI. To visualize the contrast of the r₁-r₂* relaxivity in the brain, we present a voxel-wise implementation based on a sliding-window approach. However, this implementation is more sensitive to partial volume and smoothing effects. To conclude, while we demonstrate our approach in vitro based on ferritin, transferrin and ferrous iron samples, we believe that in the human brain the r₁-r₂^* relaxivity provides more comprehensive information on the molecular iron milieu previously inaccessible by MRI.

In order to further model the separate contributions to the MRI signal of different iron compounds, it would be necessary to increase the dimensionality of the in vivo iron relaxivity measurement. The intercept of the R₁-R₂^* linear fit represents the residual R₁ not explained by R₂^*. We find that the relaxivity and the intercept contribute complementary information both in vivo and in vitro. Therefore, combining these measurements may allow better in vivo characterization of brain tissue. In addition to R₁ and R₂^*, other qMRI parameters known to be sensitive to iron include quantitative susceptibility mapping (QSM) and R2^3,4. In addition, it was suggested that magnetization transfer (MT) measurements are affected by neuromelanin–iron complexes^12,74. The linear interdependencies of these other iron-related MRI measurements may uncover additional features of the iron environment⁷⁵. For example, we show that the dependency of R₁ on R₂ is also useful for differentiating between iron environments. Therefore, we speculate that the concept we introduce here, of exposing the iron relaxivities in vivo based on the linear dependency of R₁ on R₂^* (the r₁-r₂^* relaxivity), can be generalized to further increase MRI’s specificity for iron using additional complementary measurements. In a previous work we implemented a different aspect of relaxivity for the detection of lipid composition, based on the linear dependency of qMRI parameters on the macromolecular tissue volume (MTV)⁴⁶. Here, we demonstrate that the r₁-r₂^* relaxivity and the dependency of R₁ on MTV provide two orthogonal microstructural axes. The dependency of R₁ on MTV changes according to the lipid types, even in the presence of iron, while the r₁-r₂^* relaxivity is insensitive to the lipid types and provides better distinctions between iron environments. Therefore, the presented framework can be generalized further to boost MRI’s specificity and support a more comprehensive in vivo histology with qMRI.

While the field of in vivo histology with MRI is rapidly growing, ground-truth validation remains a great challenge. Here we propose a cutting-edge validation strategy combining both bottom-up and top-down approaches in which we incorporate in vitro, in vivo and ex vivo analyses. For the bottom-up analysis, we developed a unique, synthetic biological system that allows us to examine the biophysical interpretation of the r₁-r₂^* relaxivity in highly controlled in vitro settings. For the top-down analysis, we tested whether our interpretation remains valid in the context of the highly complex biological tissue. We compared the r₁-r₂^* relaxivity measured in vivo to histological measurements of iron homeostasis and gene expression. This comparison was done both at the group level, based on previously reported findings, and at the single-subject level, by analyzing resected tumor tissues. To the best of our knowledge, no other study have compared qMRI parameters measured in vivo to ex vivo iron histology and gene expression of the same human tissue. Taken together, the different validation strategies all indicate that the r₁-r₂^* relaxivity increases the specificity of MRI to different molecular environments of iron, highlighting the robustness of our findings.

Our proposed approach for measuring iron homeostasis in vivo using the r₁-r₂^* relaxivity may have wide clinical and scientific implications. First, the r₁-r₂^* relaxivity provides a unique contrast for imaging the brain, which is associated with 45 distinct gene sets not associated with R₁ or R₂^* by themselves. Moreover, we show that the r₁-r₂^* relaxivity, which captures paramagnetic properties, enhances the contrast between tumor tissue and normal-appearing white-matter and gray-matter tissues. In agreement with these findings, meningioma tumors have been shown to contain a higher concentration of ferrimagnetic particles and an abnormal expression of iron-related genes compared to non-pathological brain tissue^17,18. Indeed, the contrast enhancement we saw with the r₁-r₂^* relaxivity is similar to the one observed for Gd-enhanced imaging, which is based on the altered relaxivity in the presence of paramagnetic agents⁷⁶. Concerns regarding the safety of Gd-based contrast agents raise the need for Gd-free diagnosis of brain tumors^56,57. Adjusting our approach for clinical imaging might offer safer alternatives for brain tumor diagnosis.

Finally, the sensitivity of the r₁-r₂^* relaxivity to iron homeostasis in the brain may have clinical implications for neurodegenerative diseases. Alterations in the distribution of molecular iron compounds can lead to cellular damage which is disease-specific¹⁰. In particular, the iron mobilization capacity was shown to differ between elderly controls and patients diagnosed with either Parkinson’s disease (PD) or Alzheimer’s disease (AD)⁷. We found that this iron homeostasis marker is correlated with the r₁-r₂^* relaxivity across brain regions and age groups. This result, in addition to our other validation strategies, demonstrates the sensitivity of the r₁-r₂^* relaxivity to the iron homeostasis in the brain and in the aging process. Therefore, our approach can add an important layer of information to existing in vivo PD and AD biomarkers such as neuromelanin MRI¹².

To conclude, we present an MRI contrast, based on the r₁-r₂^* relaxivity, which is sensitive to the iron homeostasis in the human brain. This technology can differentiate between tumor tissue and non-pathological tissue without injecting contrast agents, and can detect biological properties inaccessible to conventional MRI approaches. We validated the sensitivity of the r₁-r₂^* relaxivity to the molecular state of iron using in vitro, in vivo and ex vivo analyses. We show that our MRI technology reveals the intrinsic paramagnetic properties of different in vitro iron environments. Furthermore, the r₁-r₂^* relaxivity varies with the iron mobilization capacity across brain regions and age groups, and reveals differences in iron homeostasis and iron-related gene expression in pathological tissues. Therefore, this approach may further advance our understanding of the impaired iron homeostasis in cancer, normal aging and neurodegenerative diseases, and may open new avenues for the non-invasive research and diagnosis of the living human brain.

Methods

In vivo iron relaxivity model

The iron relaxivity model assumes a linear relationship between relaxation rates and iron concentration (Eq. (1))^30,31,32,33.

This linear relationship for two different iron environments a and b with iron concentrations [a] and [b] can be expressed using the following equations:

$${R}_{1}={r}_{(1,a)}\left[a\right]+{c}_{(1,a)}\;\;\;\;\;\;\;{R}_{1}={r}_{(1,b)}\left[b\right]+{c}_{(1,b)}$$

(3)

$${R}_{2}^{*}={r}_{(2,a)}\left[a\right]+{c}_{(2,a)}\;\;\;\;\;\;\;{R}_{2}^{*}={r}_{(2,b)}\left[b\right]+{c}_{(2,b)}$$

(4)

where ${r}_{(1/2,a/b)}$ represents the R₁-iron relaxivity or the R₂^*-iron relaxivity of the two iron environments, and ${c}_{(1/2,a/b)}$ are constants.

The two iron environments are distinguished by their iron relaxivities under the assumption:

$${r}_{(1,a)} \, \ne \, {r}_{\left(1,b\right)}\;\;\;\;\;\;\;{and}/{or}\;\;\;\;\;\;\;{r}_{\left(2,a\right)} \, \ne \, {r}_{\left(2,b\right)}$$

(5)

Rearranging Eq. (4):

$$\left[a\right]=\frac{{R}_{2}^{*}-{c}_{(2,a)}}{{r}_{(2,a)}}\;\;\;\;\;\;\;\left[b\right]=\frac{{R}_{2}^{*}-{c}_{(2,b)}}{{r}_{(2,b)}}$$

(6)

Substituting Eq. (6) in Eq. (3):

$${R}_{1}=\frac{{r}_{(1,a)}}{{r}_{(2,a)}}{R}_{2}^{*}+{c}_{a}\;\;\;\;\;\;\;{R}_{1}=\frac{{r}_{(1,b)}}{{r}_{(2,b)}}{R}_{2}^{*}+{c}_{b}$$

(7)

Where $\frac{{r}_{(1,a)}}{{r}_{(2,a)}}$ and $\frac{{r}_{(1,b)}}{{r}_{(2,b)}}$ represent the linear dependencies of R₁ on R₂^* (r₁-r₂^* relaxivities) of the two iron environments a and b (see Eq. (2)). The intercept of the linear dependency of R₁ on R₂^* can be described by the following expression: ${c}_{a/b}={c}_{(1,a/b)}-\frac{{r}_{(1,a/b)}}{{r}_{(2,a/b)}}{c}_{(2,a/b)}$. For a discussion on the biophysical meaning of the R₁-R₂^* intercept see Supplementary Section 7.

Importantly, the MRI-measured r₁-r₂^* relaxivity serves as an in vivo estimator of iron relaxivity and reveals intrinsic properties of the iron environment.

Assuming that the iron relaxivity of R₁ provides a different separation between the two iron environments a and b compared to the iron relaxivity of R₂^*:

$$\frac{{r}_{(1,a)}}{{r}_{(2,a)}}\ne \frac{{r}_{(1,b)}}{{r}_{(2,b)}}$$

These two iron environments a and b then can be distinguished by their r₁-r₂^* relaxivities (i.e., the in vivo iron relaxivities). In the brain, the r₁-r₂^* relaxivity could be affected by the entire milieu of available iron compounds, their iron binding capacities and aggregation states (i.e. the molecular iron environment). For an elaboration of this model under the assumption of a complex iron environment and in the presence of myelin see Supplementary Section 1.

Phantom samples experiments

Phantom system

We prepared samples of four different iron compounds: transferrin (holo-transferrin human, Sigma), apo-transferrin (apo-transferrin human, Sigma), ferritin (equine spleen, Sigma), and ferrous (iron (II) sulfate heptahydrate, Sigma). These samples were prepared in three different molecular environments: liposomes, 18.2 MΩcm water, and bovine serum albumin (BSA, Sigma)^77,78. For each combination of iron compound and molecular environment, we made different samples by varying both the iron compound concentration and the lipid/BSA-water fractions. For liposomal/BSA environments, the iron compounds concentrations were divided by the water-lipid or water-BSA fractions to get units of [mg/wet ml]. The liposomes were made from a mixture of soy phosphatidylcholine (PC) and egg sphingomyelin (SM) purchased from Lipoid and used without further purification. Additional results with PC and PC-cholesterol (Sigma) liposomes are presented in Supplementary Fig. 12. The lipid samples were mixed in chloroform at desired mole ratios and evaporated under reduced pressure (8 mbar) in a Buchi rotary evaporator vacuum system (Flawil, Switzerland). The resulting lipid film was resuspended in a 10 mM ammonium bicarbonate solution, lyophilized, and subsequently hydrated in the reassembly buffer. To achieve the desired lipid-protein concentration, the protein solution (~50 mg/ml in water) was diluted to the right concentration and subsequently was added to the lyophilized lipid powder. For the BSA phantoms, samples were prepared by dissolving lyophilized BSA in 18.2 MΩcm water at the desired concentrations. For the Hemoglobin phantoms, samples were prepared by dissolving lyophilized human hemoglobin powder (Sigma) in 18.2 MΩcm water at the desired concentrations. The hemoglobin concentrations in this experiment were constrained by its solubility (20 mg/mL for hemoglobin alone, and lower when it is in a mixture with other iron-binding proteins).

Samples were placed in a 2-ml glass vials glued to a glass box, which was then filled with ~1% SeaKem® LE Agarose (Ornat) and 0.1% gadolinium (Gadoteric acid (Dotarem, Guerbet)) dissolved in distilled water (DW). The purpose of the agarose with gadolinium (Agar-Gd) was to stabilize the vials and to create a smooth area in the space surrounding the samples that minimized air-sample interfaces.

MRI acquisition for phantoms

Data were collected on a 3T Siemens MAGNETOM Skyra scanner equipped with a 32-channel head receive-only coil at the ELSC Neuroimaging Unit at the Hebrew University.

Quantitative R₁ and MTV

3D Spoiled gradient echo (SPGR) images were acquired with different flip angles (FA = 4°, 8°, 16°, and 30°). The TE/TR were 4.45/18 ms. The scan resolution was 0.5 mm × 0.5 mm × 0.6 mm. For B1+ mapping, we acquired an additional spin-echo inversion recovery scan (SEIR). This scan was done on a single slice, with an adiabatic inversion pulse and inversion times of TI = 2,000, 1,200, 800, 400, and 50 ms. The TE/TR were 73/2,540 ms. The scan resolution was 1.2 mm × 1.2 mm × 2.0 mm.

Quantitative R₂ ^*

SPGR images were acquired with different flip angles (α = 4°, 8°, 16°, and 30°). The TR was 27 ms and 5 echoes were equally spaced between 4.45 and 20.85 ms. The scan resolution was 0.5 mm × 0.5 mm × 0.6 mm.

Quantitative R₂

Multi spin-echo images with 15 equally spaced spin echoes between 10.5 ms and 157.5 ms. The TR was 4.94 s. The scan resolution was 1.2 mm isotropic.

Estimation of qMRI parameters for phantoms

Quantitative R₁ and MTV mapping

R₁ and MTV estimations for the lipid samples were computed with the mrQ⁴⁶ (https://github.com/mezera/mrQ) and Vista Lab (https://github.com/vistalab/vistasoft/wiki) software packages. The mrQ software was modified to suit the phantom system⁷⁸. The modification utilizes the fact that the Agar-Gd mixture which fills the box around the vials is homogeneous, and therefore can be assumed to have a constant R₁ value. We used this gold-standard R₁ value generated from the SEIR scans to correct for the inhomogeneous excitation bias in the SPGR scans.

A mask labeling the different phantom samples was generated based on MATLAB’s “imfindcircles” function, and was filtered to remove voxels with extreme signals. Voxels were filtered based on a fixed threshold on the SPGR signal at FA = 16. In addition, we also filtered out those voxels in which the SPGR signal at FA = 16 was two median absolute deviations away from the median value. We further edited this mask manually, removing voxels with susceptibility artifacts resulting from the vials and air pockets. To fit the R₁ and proton density of each phantom sample, we calculated the median values of the SPGR signal as well as the inhomogeneous excite and receive biases across all the voxels of each sample. These median values were used in the Vista Lab function “relaxFitT1” to find the median R₁ and proton density of each sample. proton density values then were calibrated using the proton density of a water-filled vial in order to calculate the MTV values.

Quantitative R₂ ^* mapping

We used the SPGR scans with multiple echoes to estimate R₂^*. Fitting was done by taking the median values of the SPGR signal across all the voxels of the phantom sample for each TE. To label the different samples, we used the same mask that was used to calculate R₁ and MTV. We then used an exponential fitting process to find R₂^*. As we had four SPGR scans with variable flip angles, we averaged the R₂^* values acquired from each of these scans for increased signal to noise ratio.

Quantitative R2 mapping

We used spin echo scans with multiple echoes to estimate R2. Fitting was done by taking the median values of the spin-echo signal across all the voxels of the phantom sample for each TE. To label the different samples, we used the same mask that was used to calculate R₁ and MTV. We then used an exponential fitting process to find R2.

r₁-r₂ ^* relaxivity computation for phantoms

For each iron compound in each molecular environment, we computed the linear dependency of R₁ on R₂^* across samples with varying iron-binding protein concentrations relative to the water fractions. We fitted the following linear model across samples:

$$R_1=a\cdot{R_2}^ \ast+b$$

(8)

The slope of this linear model ($a$) represents the r₁-r₂^* relaxivity Eq. (2) and (7). $b$ is the intercept (the residual R₁ not explained by R₂^*). This process was implemented in MATLAB.

Estimation of total iron content in phantoms

We estimated the iron content of our transferrin and ferritin samples using the following equation:

$${{{{{{\rm{iron}}}}}}}\,\left(\frac{{{{{{{\rm{mg}}}}}}}}{{{{{{{\rm{ml}}}}}}}}\right)=\,\frac{{{{{{{\rm{iron}}}}}}}\,{{{{{{\rm{binding}}}}}}}\,{{{{{{\rm{protein}}}}}}}\,\left(\frac{{{{{{{\rm{mg}}}}}}}}{{{{{{{\rm{ml}}}}}}}}\right)}{{{{{{{\rm{protein}}}}}}}\,{{{{{{\rm{molecular}}}}}}}\,{{{{{{\rm{weight}}}}}}}\,\left(\frac{{{{{{{\rm{mg}}}}}}}}{{{{{{{\rm{mol}}}}}}}}\right)} \cdot \frac{{{{{{{\rm{iron}}}}}}}\,{{{{{{\rm{ions}}}}}}}}{{{{{{{\rm{protein}}}}}}}} \cdot {{{{{{\rm{iron}}}}}}}\,{{{{{{\rm{molecular}}}}}}}\,{{{{{{\rm{weight}}}}}}}\,\left(\frac{{{{{{{\rm{mg}}}}}}}}{{{{{{{\rm{mol}}}}}}}}\right)$$

(9)

Transferrin contains 2 iron ions per protein molecule³, and its molecular weight was estimated as 76 × 10⁶ mg/mol (based on manufacturer information). The iron loading of ferritin was estimated as 2,250 iron ions per protein molecule (based on manufacturer information) and its molecular weight was estimated as 440 × 10⁶ mg/mol³. The molecular weight of iron was set to 55.847 × 10³ mg/mol.

This resulted in the following equation for converting iron-binding protein concentrations into iron concentrations:

1 mg/ml transferrin = 1.4 µg/ml iron

1 mg/ml ferritin = 0.29 mg/ml iron

Importantly, this model is only applicable in vitro, for controlled and homogeneous iron environments. We do not apply this molecular model to MRI values of brain tissue which is complex and heterogenous.

Human MRI datasets

Table 1 summarizes the different human MRI datasets used in the study.

Table 1 human MRI datasets used in the study

Healthy human subjects

We scanned 26 young adults (aged 27 ± 10 years, 10 females), and 13 older adults (aged 70 ± 3 years, 4 females). Healthy volunteers were recruited from the community surrounding the Hebrew University of Jerusalem. The experimental procedure was approved by the Helsinki Ethics Committee of Hadassah Hospital, Jerusalem, Israel. Written informed consent was obtained from each participant prior to the procedure. This data was first used in our previous work⁷⁷. Sex and gender were determined based on self-report were not considered in the study design as this information is irrelevant for our findings.

Meningioma patients

During the study period May 2019 to August 2020, we recruited 19 patients who had undergone surgery for the resection of brain meningiomas. All patients had preoperative qMRI scans in addition to their clinical brain MRI assessment. One subject, with a titanium cranial fixation plate adjacent to the tumor, was excluded from the study due to local disruption of the magnetic field. The final cohort included 18 patients (11 females). Imaging studies were anonymized before they were transferred for further analysis. Brain meningioma surgical specimens, available for 16 patients, were obtained from the fresh frozen tissue biobank of the Department of Neurosurgery, Shaare Zedek Medical Center, Jerusalem, Israel, and were transferred on dry ice for western-blot and gene expression analyses. The experimental procedure was approved by the Helsinki Ethics Committee of Shaare Zedek Medical Center, Jerusalem, Israel. Study participants provided informed consent according to an institutional review board, approval numbers 0299-17 SZMC and 0159-20 SZMC. Sex and gender were determined based on self-report were not considered in the study design as this information is irrelevant for our findings.

MRI acquisition for healthy human subjects

Data were collected on a 3T Siemens MAGNETOM Skyra scanner equipped with a 32-channel head receive-only coil at the ELSC Neuroimaging Unit at the Hebrew University.

Quantitative R₁, R₂ ^* and MTV mapping

SPGR echo images were acquired with different flip angles (α = 4°, 10°, 20°, and 30°). Each image included 5 equally spaced echoes (TE = 3.34–14.02 ms) and the TR was 19 ms. The scan resolution was 1 mm isotropic. Additional SPGR echo image was acquired with an MT pulse (TE = 3.34, TR = 27, α = 10, 1 mm isotropic). For B1+ mapping, we acquired an additional spin-echo inversion recovery scan with an echo-planar imaging read-out (SEIR-epi). This scan was done with a slab-inversion pulse and spatial-spectral fat suppression. For SEIR-epi, the TE/TR were 49/2,920 ms. The TIs were 200, 400, 1200, and 2400 ms. We used 2-mm in-plane resolution with a slice thickness of 3 mm. The EPI readout was performed using 2× acceleration.

Anatomical images

3D magnetization-prepared rapid gradient echo (MPRAGE) scans were acquired for 30 of the 39 healthy subjects. The scan resolution was 1 mm isotropic, the TE/TR were 2.98/2300 ms. Magnetization-prepared 2 rapid acquisition gradient echo (MP2RAGE) scans were acquired for the remaining 9 subjects. The scan resolution was 1 mm isotropic, the TE/TR were 2.98/5000 ms.

Whole-brain DTI measurements

performed using a diffusion-weighted spin-echo EPI sequence with isotropic 1.5-mm resolution. Diffusion weighting gradients were applied at 64 directions and the strength of the diffusion weighting was set to b = 2000s/mm² (TE/TR = 95.80/6000 ms, G = 45mT/m, δ = 32.25 ms, Δ=52.02 ms). The data includes eight non-diffusion-weighted images (b = 0). In addition, we collected non-diffusion-weighted images with reversed phase-encode blips. For two subjects (1 young, 1 old) we failed to acquire this correction data and they were excluded from the diffusion analysis.

MRI acquisition for meningioma patients

Data were collected on a 3T Siemens MAGNETOM Skyra scanner equipped with a 32-channel head receive-only coil at the Shaare Zedek Medical Center.

Quantitative R₁, R₂ ^* and MTV mapping

SPGR echo images were acquired with different flip angles (α = 4°, 10°, 20°, and 30°). Each image included 5 equally spaced echoes (TE = 2.85–14.02 ms) and the TR was 18 ms. The scan resolution was 1.5 mm isotropic. For calibration, we acquired an additional SEIR-epi scan. This scan was done with a slab-inversion pulse and spatial-spectral fat suppression. For SEIR-epi, the TE/TR were 49/2,920 ms. The TIs were 200, 400, 1200, and 2400 ms. We used 2-mm in-plane resolution with a slice thickness of 3 mm. The EPI readout was performed using 2× acceleration.

Gd-enhanced anatomical images

Gd-enhanced MPRAGE scans were acquired. The scan resolution was 1 mm isotropic, the TE/TR were 2.4/1800 ms. The contrast agent was either Multihance or Dotarem at a dose of 0.1 mmol/kg. Contrast agent injection and MPRAGE acquisition were done after the acquisition of the quantitative MRI protocol, or on a different day.

Estimation of qMRI parameters for human subjects

Quantitative R₁ and MTV mapping

Whole-brain MTV and R₁ maps, together with bias correction maps of B1+ and B1-, were computed using the mrQ software^46,79.

Quantitative R₂ ^* mapping

We used the SPGR scans with multiple echoes to estimate R₂^*. Fitting was performed with the Voxel Based Quantification (VBQ) toolbox (v2e) in SPM12⁸⁰. As we had four SPGR scans with variable flip angles, we averaged the R₂^* maps acquired from each of these scans for increased SNR.

Quantitative MD mapping

Diffusion analysis was done using the FDT toolbox in FSL^81,82. Susceptibility and eddy current induced distortions were corrected using the reverse phase-encode data, with the eddy and topup commands^83,84. MD maps were calculated using vistasoft (https://github.com/vistalab/vistasoft/wiki).

Quantitative MTsat mapping

MTsat maps were computed as described in Helms et al. (2008)⁴⁷. The MTsat measurement was extracted from the equation:

$${MTsat}={M}_{0}\left(B1+\right) \cdot \alpha \frac{R1{TR}}{{S}_{{MT}}}-\frac{{((B1+\!) \cdot \alpha )}^{2}}{2\,}-R1{TR}$$

(10)

Where ${S}_{{MT}}$ is the signal of the SPGR scan with additional MT pulse, $\alpha$ is the flip angle and TR is the repetition time. ${M}_{0}$ (the equilibrium magnetization parameter), $B1+$ (the transmit inhomogeneity) and $R1$ estimations were computed from the non-MT weighted SPGR scans, during the pipeline described under “Quantitative R₁ & MTV mapping”. Registration of the ${S}_{{MT}}$ image to the imaging space of the R₁ map was done using a rigid-body alignment (R₁, B1+ and ${M}_{0}$ are all in the same space).

Brain segmentation in healthy subjects

Whole-brain segmentation was computed automatically using the FreeSurfer segmentation algorithm (v6.0)⁸⁵. For subjects with MPRAGE scan, we used that as a reference; for the other subjects the MP2RAGE scan was used. These anatomical images were registered to the R₁ space prior to the segmentation process, using a rigid-body alignment. FreeSurfer’s estimates of subcortical gray-matter structures were replaced with estimates from FSL’s FIRST tool⁸⁶.

Brain segmentation in meningioma patients

Tumor contouring was performed by the neurosurgeon (T.S.) using BrainLab’s Elements software, Cranial Navigation 3.1.5 (BrainLab AG, Munich, Germany) over the Gd-enhanced MPRAGE images, and exported as a DICOM file for further analysis. The contours of the tumors were registered to the R₁ space using rigid-body segmentation. Cases that required manual adjustment were examined and approved for accuracy by the neurosurgeon (T.S.).

Whole-brain segmentation was computed automatically using the FreeSurfer segmentation algorithm (v6.0)⁸⁵. We used the synthetic T1w image generated with mrQ as the reference image, from which we removed the skull and the tumor. We then ran FreeSurfer with the “-noskullstrip” flag. For each patient, we used FreeSurfer’s segmentation in the tumor-free hemisphere. Estimates for the entire white-matter and gray-matter tissues were averaged across the different FreeSurfer parcellations in these regions.

r₁-r₂ ^* relaxivity computation for ROIs in the human brain

We used MATLAB to compute the r₁-r₂^* relaxivity in different brain areas. For each ROI, we extracted the R₂^* and R₁ values from all voxels. R₂^* values were pooled into 36 bins spaced equally between 0 and 50 s⁻¹. This was done so that the linear fit would not be heavily affected by the density of the voxels in different R₂^* regimes. We removed any bins in which the number of voxels was smaller than 4% of the total voxel count in the ROI. The median R₂^* of each bin was computed, along with the R₁ median. We used these data points to fit the following linear model across bins using simple linear regression Eq. (8). 

$$R1=a\cdot{R_2}^ \ast+b$$

(11)

The slope of this linear model ($a$) represents the r₁-r₂^* relaxivity. $b$ is the intercept (the residual R₁ not explained by R₂^*).

Generating voxel-wise r₁-r₂ ^* relaxivity visualizations

In order to generate a voxel-wise visualization of the r₁-r₂^* relaxivity in the brain, we calculated the local linear dependency of R₁ on R₂^* using a moving-window approach. For each voxel within the brain mask, we extracted R₁ and R₂^* values of that voxel and all its neighboring voxels (a box of 125 voxels total). If at least 10 of these voxels were inside the brain mask, we fit the following linear model across these voxels Eq. (8). 

$$R1=a\cdot{R_2}^ \ast+b$$

(12)

The slope of this linear model ($a$) represents the local r₁-r₂^* relaxivity of the voxel. $b$ is constant.

Group-level comparison of qMRI parameters and histological measurements

We aggregated data published in different papers^5,7,9 that describe ferritin, transferrin and iron concentrations in 11 brain regions. One of the papers⁹ described the concentration of L-rich ferritin and H-rich ferritin independently and we combined these estimates for each ROI to get the total ferritin concentration. One of the papers⁵ reported the iron level in units of [µmol Fe/ g protein] and we converted these measurements to units of [µg Fe/g protein]. In order to use this data for our analysis we matched the brain regions reported in the literature with their corresponding FreeSurfer labels. Table 2 summarizes the aggregated data.

Table 2 Iron compounds concentrations reported in the literature for different brain regions (with references and sample sizes) and the corresponding FreeSurfer labels

Western blot analysis of meningioma tissue

Fresh frozen meningioma samples (40–50 mg) from 16 patients were homogenized in 200 µL of RIPA buffer (Thermo Fisher Scientific) supplemented with protease inhibitor (Sigma-Aldrich) using a Bioruptor Pico sonication device (Diagenode) and protein extraction beads (Diagenode, NJ, USA) according to the manufacturer instructions. Protein concentration was determined using the Pierce assay (Thermo Fisher Scientific, MA, USA). Samples containing 20 µg of protein were separated on 4–20% Tris-Glycine SDS-PAGE gel (Bio-Rad, CA, USA) and transferred to PVDF membrane using Trans-Blot Turbo transfer system and transfer packs (Bio-Rad, Hercules, CA, USA). Membranes were probed using Anti-Ferritin Light chain (#AB69090, Abcam, 1:1,000 dilution) and Anti-Transferrin (#AB82411, Abcam, 1:10,000 dilution) primary antibodies and appropriate horseradish peroxidase-conjugated secondary antibody (#AB6721, Abcam, 1:20000 dilution). Membranes were treated with EZ-ECL (Biological industries, Beit-Ha’emek, Israel) and visualized using ImageQuant LAS 4000 (GE Healthcare, IL, USA). Blot intensities were quantified using the FIJI ImageJ software (v2.35)⁸⁷. The ratio of transferrin/ferritin was based on the ratio in the blot intensities of transferrin and ferritin. Due to the noisy nature of the western-blot analysis, we averaged the estimates over six repetitions. We then used the median transferrin/ferritin ratio across subjects (which was equal to 1) as the threshold between the two groups (low and high transferrin/ferritin ratio).

RNA-sequencing of meningioma tissue

RNA-seq libraries

Tumor samples from 17 patients (samples from 16 patients and a replicate for one) were flash frozen and kept in −80c until processing. RNA isolation was done with the following steps: First, frozen tissue was chopped and transferred with a 2 ml lysis buffer (Macherey-Nagel, 740955) five times through a needle attached to a 0.9 mm syringe to achieve homogenization. Next, total RNA was extracted with NucleoSpin RNA kit (Macherey-Nagel, 740955), following the standard protocol. Finally, mRNA was isolated using the NEBNext Poly(A) mRNA Magnetic Isolation Module (NEB E7490S), using 5µg of total RNA as an input and following the standard protocol. The purified mRNA was used as input for cDNA library preparation, using NEBNext® Ultra™ II Directional RNA Library Prep Kit for Illumina (NEB E7760), and following the standard protocol. Quantification of the libraries was done by Qubit and TapeStation. Paired-end sequencing of the libraries was performed on Nextseq 550.

Data processing

The demultiplexing of the samples was done with Illumina’s bcl2fastq software (v2.19.1.403). The fastq files were next aligned to the human genome (human genome assembly GRCh38) using STAR (v2.7.3a) and the transcriptome alignment and gene counts were obtained with HTseq (v0.13.5). For quality control RNAseQC software was used (Picard v2.26.10). Quality control and data normalization were done in R using the DEseq2 package from Bioconductor (v3.13). The counts matrix per gene and sample were normalized using the Variance stabilizing transformation. Genes with less than 5 counts were filtered out of the analysis. The filtered and normalized matrix was used in all downstream analysis.

Gene set enrichment analysis (GSEA)

The final sequencing dataset included the expression of approximately 27,000 genes in 17 tumor samples. We then excluded unannotated genes based on the gene ontology resource (http://geneontology.org/) as well as genes with low (<6) expression levels, yielding 19,500 genes.

We used GSEA software (v4.1.0) to further validate that the subset of highly correlated genes is not random, but rather represents known biological pathways. For this aim, we calculated the correlations across patients between the expression of each of the genes and one of the qMRI parameters (R₁, R₂^* or r₁-r₂^* relaxivity). For each of the qMRI parameters, genes were ranked based on the r values of the correlations, and the ranked list was used in the GSEAPranked toolbox^59,60. The gene sets databases used for this analysis included go, biocarta, kegg, pid, reactome and wikipathways.

The primary result of the GSEA is the enrichment score (ES), which reflects the degree to which a gene set is overrepresented at the top or bottom of a ranked list of genes: a positive ES indicates gene set enrichment at the top of the list, while a negative ES indicates gene set enrichment at the bottom. The normalized ES (NES) accounts for differences in gene set size and in correlations between gene sets and the expression dataset.

One tumor sample was excluded from the analysis, as the R₁ and R₂^* values in the tumor were relatively high, which led to the fact that no significantly enriched pathway were found for R₁ and R₂^* (though we did detect significantly enriched pathways for the r₁-r₂^* relaxivity). Removing this outlier improved GSEA results for R₁ and R₂^* and we therefore excluded this subject.

Following the GSEA analysis, we found a total of 101 significantly enriched pathways for at least one of the R₁, R₂^* and r₁-r₂^* relaxivity. We then clustered those significantly enriched pathways using the “clustergram” function in MATLAB. In order to evaluate which genes are included within the top enrichment pathways for each MRI parameter, we used Leading Edge Analysis (as implemented in the GSEA toolbox).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.