An overview of the methodological pipeline is illustrated in Fig. 1. Of the 11,074 genes initially included in our analysis, 1,257 matched to a Human Protein Atlas brain “cell type-enriched” or “low cell type specificity” gene set. The mean MWF-low cell type specificity gene set pairing (r-value) was 0.13, with a standard deviation of ± 0.43 (n = 1,094 genes). Seven cell type-enriched gene sets, with n > 5 genes, were included in our final statistical analysis: adipocytes (n = 8 genes), Muller glial cells (n = 11 genes), oligodendrocyte precursor cells (n = 13 genes), inhibitory neurons (n = 14 genes), excitatory neurons (n = 18 genes), astrocytes (n = 20 genes), and oligodendrocytes (n = 65 genes). Non-parametric statistical testing revealed significant differences across gene sets (Fig. 2; Kruskal-Wallis: test statistic = 155.0, df = 7, p < 0.05). Compared to a low cell type specificity gene set (control; n = 1,094), MWF-gene set pairings for oligodendrocytes and adipocytes were significantly higher, while inhibitory and excitatory neurons were significantly lower (Table 1). The MWF-gene set pairings for oligodendrocytes and adipocytes were also significantly higher compared to astrocytes, Muller glial cells, and oligodendrocyte precursor cells (Supplementary Table 1). There were no significant differences between MWF-astrocytes, MWF-Muller glial cells, and MWF-oligodendrocyte precursor cell gene set pairings and the MWF-low cell type specificity (control) gene set pairing.
Table 1
Comparison of RNA single cell types to a control group. Each row evaluates whether MWF-gene pairings in high cell type specificity genes sets differ from the MWF-control gene set pairing, using the Mann-Whitney U statistic. The control set consisted of genes identified as having low cell type specificity in the Human Protein Atlas and were present in roughly similar levels across all cell types. Adjustments for multiple comparisons were made using the Bonferroni correction method. *: p < 0.05, **: p < 0.01, ***: p < 0.001, ****: p < 0.0001, *****: p is approximating 0.
High Cell Type Specificity Gene Set | Test Statistic | Std. Error | Std. Test Statistic | P value | Adjusted p value |
Oligodendrocytes | 439.48 | 45.83 | 9.59 | 0.00 ***** | 0.00 ***** |
Excitatory neurons | -472.61 | 85.30 | -5.54 | 3.02E-8 **** | 8.45E-7 **** |
Inhibitory neurons | -350.88 | 96.55 | -3.63 | 2.79E-4 *** | 7.81E-3 ** |
Oligodendrocyte precursor cells | -152.15 | 100.15 | -1.52 | 0.13 | 1.00 |
Muller glial cells | -138.11 | 108.78 | -1.27 | 0.20 | 1.00 |
Astrocytes | -23.18 | 81.00 | -0.29 | 0.77 | 1.00 |
Adipocytes | 406.62 | 127.38 | 3.19 | 1.41E-3 ** | 3.95E-2 * |
Aligning with our hypothesis, the MWF-oligodendrocyte gene set pairing was: a) positive and b) significantly higher than the pairing for control genes (i.e., the low cell type specificity gene set), in addition to all other gene set pairings. The opposite was the case for MWF-inhibitory and MWF-excitatory neuron gene set pairings, which were: a) negative and b) significantly lower than the pairing for control genes, suggesting that areas with higher gray matter content are associated with lower MWF. MWF-astrocyte, MWF-Muller glial cell, and MWF-oligodendrocyte precursor cell gene pairings were no different than the pairing for the control gene set. Interestingly, MWF-adipocyte gene set pairings were also significantly higher than control – a novel observation that may reflect adipocytes as the main constituent of the myelin lipid bilayer (21, 22).
To the best of our knowledge, we are the first to have employed a neuroimaging transcriptomic approach to understand spatial variations in MWF in white matter tracts. One earlier study examined mixed white and gray matter regions (14), where it is possible that associations between gene expression and MWF could reflect, to a large degree, differences in gray matter composition, where myelin levels are ~ 10x lower than white matter. The focus on mixed gray and white matter regions may also explain the significant negative MWF-astrocyte gene set pairing in another previous study (13), which we found no evidence of. Moreover, substantial differences across white matter tracts have been previously reported for MWF, with nearly a two-fold difference between the internal capsule and the corpus collosum (15). Based on our results, these MWF variations reflect normal phenotypic variations in white matter gene expression in the brain. Such conclusion can be drawn from the association between MWF and gene expression patterns of cells responsible for forming white matter structures, specifically oligodendrocytes and adipocytes.
In addition to examining mixed white and gray matter regions, previous neuroimaging transcriptomic studies have employed newer methods of estimating MWF. The combined gradient and spin-echo (GRASE) acquisition method used in the current study is a well-established myelin water imaging approach, having been extensively validated as an objective marker of myelin in the human brain (23, 24). Conversely, the multicomponent driven equilibrium single pulse observation of T1 and T2 (mcDESPOT) acquisition method is a newer approach, whose primary advantage is faster acquisition time and higher resolution (25). However, as noted by Patel and others (13, 26, 27), mcDESPOT grossly overestimates MWF. Even more problematic in the context of neuroimaging transcriptomics, mcDESPOT gives rise to smaller regional variations in MWF (28). Given that regional differences across white matter tracts are rather negligible with mcDESPOT, while GRASE shows substantial white matter variations, it is improbable that our GRASE-derived MWF approach would yield a similar outcome as mcDESPOT-derived MWF. In this regard, demonstrating an association between GRASE derived MWF and myelin related genes (i.e., oligodendrocytes) represents important and necessary validation of MWF as a myelin measure using neuroimaging transcriptomics.
Our observations are also important in that MWF was estimated in both men and women, whereas Patel and colleagues only sampled MWF from males (13). Methodologically, we incorporated a novel neuroimaging transcriptomic approach, characterizing regional variations in gene expression and MWF across white matter tracts as Cohen’s d effect sizes before estimating their spatial relationship. This was primarily done to further account for donor-level differences in regional gene expression that are often overlooked in neuroimaging transcriptomics studies. Further validation of this approach, including a direct comparison with other neuroimaging transcriptomic methods, is warranted.