Disentangling molecular alterations from water-content changes in the aging human brain using quantitative MRI

It is an open question whether aging-related changes throughout the brain are driven by a common factor or result from several distinct molecular mechanisms. Quantitative magnetic resonance imaging (qMRI) provides biophysical parametric measurements allowing for non-invasive mapping of the aging human brain. However, qMRI measurements change in response to both molecular composition and water content. Here, we present a tissue relaxivity approach that disentangles these two tissue components and decodes molecular information from the MRI signal. Our approach enables us to reveal the molecular composition of lipid samples and predict lipidomics measurements of the brain. It produces unique molecular signatures across the brain, which are correlated with specific gene-expression profiles. We uncover region-specific molecular changes associated with brain aging. These changes are independent from other MRI aging markers. Our approach opens the door to a quantitative characterization of the biological sources for aging, that until now was possible only post-mortem.

Supplementary Note 1-Modeling lipid mixtures through the MDM approach In order to predict the qMRI parameters of a lipid mixture from MDM measurements (Figure 1d) we used the following model (shown here for MTsat. The predictions for R1 and R2 was done similarly): Where L is the number of lipids in the mixture and is the fraction of the i'th lipid from the total lipid volume.
′ and are the MDM measurements of the pure lipids. These measures were estimated from the samples prepared exclusively with each i'th lipid and were extracted from the linear fit of these samples: (2) = ′ * + Importantly, this model implies that qMRI parameters of a lipid mixture can be computed as a linear sum of qMRI parameters of the pure lipids.
Where ′ and ss were extracted from the linear fit of the mixture. As this linear equation holds for large range of MTV values and due to the uniqueness of the interpolating polynomial: Supplementary Note 3-MDM-based Prediction for the molecular composition of the human brain The human brain is far more complex than a lipid mixture. As it is more difficult to a priori estimate the matrix for brain tissue, we show that it is possible to evaluate it through cross-validation.
First, we validated this approach on lipid samples. In each iteration, we fitted using Next, we used the MDM approach to predict the molecular composition of the human brain ( Figure 3c). For this, we used Supplementary Equation 5 and the same cross-validation process; prediction for each brain area was computed by removing it from the system and solving for the other brain areas.
The calculation involved 7 human brain molecular features (%PE, %PS, %PtdCho %PI, %Spg, phospholipids/proteins, phospholipids/cholesterol), and 4 MDM measurements (dR1/dMTV, dMTsat/dMTV, dR2dMTV, dMD/dMTV). PCA was used to reduce the dimensionality of the system and avoid over-fitting. We identified the three molecular features with the largest loadings on the first PC of molecular variability. The fractions of the lipids PE and PtdCho had the largest loadings, and we used the ratio between them as it was found to better characterize individual brain regions 1 .
The two other features with large loadings were the fraction of the lipid Spg and the phospholipids/proteins ratio. We predicted these 3 human brain molecular features using the MTV derivatives that account for most of the MDM variability. The two measurements with the largest loadings on the first PC of MDM were dR1/dMTV and dMTsat/dMTV. Therefore, in the case of the human brain, F is a 3X7 matrix of molecular composition estimated from the literature 1 .
The columns of F are 7 different brain areas. The rows of F are the 3 molecular features with largest loadings. M is a 2X7 matrix of the MDM measurements of 7 different brain regions. This data was calculated from the MRI   More evidence for the biological interpretability of the MDM signatures was provided by their correlation with spatial gene expression patterns throughout the cortex. We compared the cortical MDM signatures to a gene co-expression network based on a widespread survey of gene expression in the human brain undertaken by the Allen Human Brain Atlas project 2 . Nineteen modules were derived from the gene network, each comprised of a group of genes that covaries in space. The projection of different cortical areas on the first PC of each module captures the "eigengene" of the module-the main axes of the module's variation across the brain. The eigengenes of the different modules were compared to the cortical projection on the first PC of the MDM signatures (PC 1 MDM−in−vivo , Figure 4a). In this case the PCA is calculated using the R1 and MTsat maps (without MD and T2 maps), to allow sufficient resolution for cortical parcellation. The correlations were corrected for multiple comparisons using the FDR method. Six out of the nineteen gene modules were significantly correlated with the first PC of MDM (adjusted R 2 ranges between 0.098 and 0.26). The two modules with the highest correlations are presented in Figure 4b Table 2). Right: Human brain. The average percentage of four phospholipids as measured post-mortem by Söderberg (1990). The percentages are averaged over seven brain samples. Both in the human brain and the porcine brain, PE constitute over 40% of the lipid composition.

Comparison of the MDM signatures of 18 older adults (aged 67±6 years, marked in gray) and 23 younger adults (aged 27±2 years, marked with different colors). Multidimensional aging-related changes revealed by the MDM approach are presented for different brain regions (young subjects are represented by different colors; older subjects are in gray). Each axis is the MTV derivative of a different qMRI parameter. The range of each axis was determined by the 5 and 95
Percentiles. Traces extends between these derivatives; shaded areas represent the variation across subjects. The statistical significance of the differences between the groups was estimated using a two-sample t-test and was corrected for multiple comparisons using the FDR method. * P < 0.05; ** P < 0.01; *** P < 0.001.
Supplementary Figure 14 -separating molecular and water related contributions.

Supplementary Figure 14: separating R1 measurement to its molecular and water related constituents. Comparison of MRI-driven measurements of 18 older adults (aged 67±6 years, marked in gray) and 23 younger adults (aged 27±2 years, marked with different colors) in different brain regions (see legend)
. Aging-related changes revealed by R1 measurement are presented in the R1 columns. The separation of chemophysical and water related constituents estimated by the MTV derivative and MTV respectively is shown in the dR1/dMTV and MTV columns. The statistical significance of the differences between the groups was estimated using a two-sample t-test and was corrected for multiple comparisons using the FDR method. * P < 0.05;** P < 0.01; *** P < 0.001.
Supplementary Figure 15-separating molecular and water related contributions.

Supplementary Figure 15: separating MTsat measurement to its molecular and water related contributions. Comparison of MRI-driven measurements of 18 older adults (aged 67±6 years, marked in gray) and 23 younger adults (aged 27±2 years, marked with different colors) in different brain regions (see legend). Aging-related changes revealed by MTsat measurement are presented in the MTsat columns. The separation of chemophysical and water related contributions estimated by the MTV derivative and MTV respectively is shown in the dMTsat/dMTV and MTV columns.
The statistical significance of the differences between the groups was estimated using a two-sample t-test and was corrected for multiple comparisons using the FDR method. * P < 0.05; ** P < 0.01; *** P < 0.001.
Supplementary Figure 16-separating molecular and water related contributions.

Supplementary Figure 16: separating MD measurement to its molecular and water related contributions. Comparison of MRI-driven measurements of 17 older adults (aged 68±5 years, marked in gray) and 19 younger adults (aged 27±2 years, marked with different colors) in different brain regions (see legend). Aging-related changes revealed by MD measurement are presented in the MD columns. The separation of chemophysical and water related contributions estimated by the MTV derivative and MTV respectively is shown in the dMD/dMTV and MTV columns.
The statistical significance of the differences between the groups was estimated using a two-sample t-test and was corrected for multiple comparisons using the FDR method. * P < 0.05; ** P < 0.01; *** P < 0.001.
Supplementary Figure 17-separating molecular and water related contributions.
Supplementary Figure 17: separating R2 measurement to its molecular and water related contributions.

Comparison of MRI-driven measurements of 15 older adults (aged 69±5 years, marked in gray) and 22 younger adults (aged 27±2 years, marked with different colors) in different brain regions (see legend)
. Aging-related changes revealed by R2 measurement are presented in the R2 columns. The separation of chemophysical and water related contributions estimated by the MTV derivative and MTV respectively is shown in the dR2/dMTV and MTV columns. The statistical significance of the differences between the groups was estimated using a two-sample t-test and was corrected for multiple comparisons using the FDR method. * P < 0.05;** P < 0.01;*** P < 0.001.
Supplementary Figure 18-aging related changes revealed by the 1st principle component of MDM.
Supplementary Figure 18:  Supplementary Note 5-The effect of iron content on MDM signatures. We find that the MDM signatures correlate with the molecular composition of the human brain ( Figure 3). However, R1 and R2, which are used to generate the MDM signatures, are not sensitive only to the molecular environment and the water content but also to the iron content 6,7 . Iron is an important component of human brain tissue. Iron seems to accumulate in different brain regions during aging, and in a variety of CNS disorders 8 . We validated that the aging-related changes observed in the MTV derivative of R1 and R2 cannot be fully explained by alterations in the iron content with age. For this aim we used R2*, as it was shown to be a good surrogate marker of iron concentration 7 . We compared R2* values between the two age groups on a subset of our subjects (15 older adults and 17 younger adults) for which we acquired R2* map.
Supplementary Figure 19 shows the age-related differences in R2* in different brain regions. In the parietal white matter, we did find significant age-related changes in the MTV derivative of R1 and in MTV (Figure 6a). However, there were no significant age-related changes in R2* (Supplementary Figure 19). We therefore assume that the aging-related changes we found in the parietal white matter are not related to alterations in iron content. In other white matter regions, there were no R2* differences between the age groups as well (Supplementary Figure 19).
In the thalamus, we found a significant age-related change in the MTV derivative of R1 (Figure 6b). However, there was no significant change in the thalamus R2* values between the age groups (Supplementary Figure 19). This indicates that the age-related changes in the thalamus revealed by the MDM method are probably not related to iron effects. In contrast, for the caudate, putamen, hippocampus and pallidum, we found significant age-related changes in the MTV derivatives of R1, along with significant changes in R2* values (Supplementary Figure 19). In the parietal cortex, we found a significant age-related change in the MTV derivative of R1 (Figure 6c). However, in all cortical areas we found a significant change in iron content with age, as estimated by R2* (Supplementary Figure 19). This implies that some of the age-related changes in R1 can be explained by alterations in the iron content. Notably, in the frontal cortex we also found significant aging-related change in the MTsat dependency on MTV. As iron is less likely to affect MTsat measurements, it is probable that there are molecular alterations in this brain regions that cannot be explained by iron.
The MTV derivative of R2 revealed significant aging-related changes in three grey matter cortical regions (frontal cortex, parietal cortex and occipital cortex). These three brain regions also showed significant changes in R2* values with age.
The R2* analysis demonstrates that in some brain regions aging-related changes in the MTV derivative coincide with aging related changes in the iron content. In other brain regions aging-related changes in these two measures don't overlap. Therefore, the differences between the age groups revealed by MDM measurements probably incorporate some iron-related changes. However, these iron-related changes do not fully account for our findings. Hence MDM and R2* provide complementary information.
Supplementary Figure 19-Iron-related changes between the two age groups.
Supplementary Figure 19: Iron-related changes between the two age groups. Comparison of R2* measurements which are related to iron content in 15 older adults (marked in gray) and 17 younger adults (marked with different colors) in different brain regions (see legend). The statistical significance of the differences between the groups was estimated using a two-sample t-test and was corrected for multiple comparisons using the FDR method. * P < 0.05; ** P < 0.01; *** P < 0.001.
Supplementary Note 6-Correction for R2* weighting in MTV estimation. While our MTV estimation in the human brain was computed from MRI scans with relatively short TE (3.34 ms), it could still incorporate some R2* weighting.
In order to validate that our MTV estimates are not affected by this R2*-contamination, we corrected our MTV estimation for R2* in a subset of our subjects (18 old, 17 young, for further details see "R2* correction for MTV" in the methods). Supplementary Figure 20 shows the variation across subjects in the regional MTV derivatives of R1, MTsat, MD and R2 after this correction. Compared to the uncorrected case (Figure 2c & Supplementary Figure 6) we did observe some minor differences. Nonetheless, the general regional trend was preserved. The significant aging related changes in the MTV derivatives of R1 and MTsat we presented in Figure 6-7 survived the R2* correction (Supplementary Figure 21-22). However, after correcting MTV for R2* we could no longer identify the aging-related increase in MTV values of the frontal cortex and the hippocampus. Across all brain regions except for the frontal cortex and the amygdala, the significant changes in the MTV derivative of R1 survived the R2* correction (Supplementary Figure 23). For the MTV derivative of MTsat, significant age-related changes in the putamen diminished after the R2* correction, while the aging-related changes in the other brain regions remained (Supplementary Figure 24). The significance of the aging-related changes observed in the MTV derivative of MD in the temporal cortex (Supplementary Figure 25) and in the MTV derivative of R2 in the thalamus and temporal cortex did not survive the correction (Supplementary Figure 26). Interestingly, the R2* correction revealed significant aging-related changes in some brain regions that were not observed previously. Importantly, the agreement between MDM and histology presented in figure 3 was still valid after MTV was corrected for R2* (Supplementary Figure 27).
In conclusion, the R2* correction did not dramatically affect our results. Importantly, the correction introduced additional noise to the MTV maps, as we divided by the R2* contribution to the signal (see methods). Moreover, the comparison between the corrected and uncorrected MTV was done on 35 of our 41 subjects.
Supplementary Figure 21-Aging-related changes revealed by the R1 dependency on MTV-R2* correction.     Supplementary Figure 27-The biological interpretation of the MDM approach -following R2* correction.
Supplementary Figure 27: The biological interpretation of the MDM approach -following R2* the correction. The agreement between MDM and histology presented in figure 3