Genetic, cellular, and connectomic characterization of the brain regions commonly plagued by glioma

Using neuroimaging and genetic methods, Mandal et al. show that gliomas localize to brain regions characterized by functional hubness, stem-like cells, and transcription of genes involved in tumour development. The results illustrate how factors of diverse scale can collectively determine the location of brain dysfunction.


Spin Test Methodology
The "spin test" involves comparing the observed inter-parcel correlation between maps of two measures with a distribution of the correlations calculated after one of these maps has been spatially permuted in a way that preserves contiguity between brain regions. Spatial permutation was accomplished by projecting the centroid coordinates for each parcel onto an inflation of the pial surface as a sphere (Fischl, 2012), applying a random rotation to that sphere, and then projecting the new coordinates back onto the pial surface and assigning them to the nearest centroid coordinates of the original parcellation. The result is a shuffled parcellation where most parcels remain contiguous.
Past studies using the spin test have focused on comparisons between cortical brain maps.
However, subcortical regions were also of interest in this study. Subcortical regions cannot be projected onto the inflated spherical pial surface, so an alternative approach was needed. We incorporated the subcortex into our null models by shuffling the eight subcortical regions with respect to one another, whereas the cortical regions were shuffled using the spin test.
After each spin permutation, two correlations were calculated; one between measures estimated from parcels in their original configuration and the other in its permuted configuration, and vice versa. These two correlations were averaged to form one of the 10000 values forming a null distribution to which the observed correlation was compared to determine statistical significance, as the proportion of null correlations greater than the observed correlation (i.e. Pspin).

AHBA Preprocessing
Custom microarrays were used to measure the expression of all genes in the genome in 3702 brain sample locations across cortex, subcortex, and cerebellum (Hawrylycz et al., 2012). Preprocessing of these data followed a similar pipeline to previous work from our group (Romero-Garcia et al., 2018. Microarray probes were mapped to genes using the genome assembly hg19 (UCSC GenomeBrowser; http://sourceforge.net/projects/reannotator/; Arloth et al., 2015).
In line with criteria from Richiardi and colleagues (Richiardi et al., 2015), probes were matched to a gene only if there were less than three mismatches between the probe and reference sequence. When a gene matched with multiple probes, the probe with the highest average expression across samples was selected to represent the expression patterns of that gene. A recent study demonstrated the effectiveness of this preprocessing step in increasing the correspondence between microarray and RNA-seq expression (Arnatkevic̆iūtė et al., 2019). In total, the expression patterns of 20647 genes across each sample location were evaluated.
Samples which were collected from the brain stem and cerebellum were excluded from the analysis, leading to a final number of 2748 samples.

Estimation of OPC Distribution
The estimation of OPC distribution followed three steps: (i) selection of a set of genes associated with OPC identity, (ii) filtering of this gene set to allow for integration with the AHBA, and (iii) assessment of median regional enrichment of OPC associated genes. First, an OPC gene set was derived from a single cell RNA sequencing study performed on adult postmortem cortical tissue (Lake et al., 2018) that determined genes with transcription patterns distinguishing cells by canonical cell types, including excitatory and inhibitory neurons, astrocytes, oligodendrocytes, and OPCs. The set of 132 genes that distinguished OPCs from other canonical cell classes across the cortex was downloaded from previously published material (Lake et al., 2018).
Next, we determined regional OPC enrichment in the adult brain using the publicly available Allen Human Brain Atlas (AHBA; Hawrylycz et al., 2012). Transcription patterns of 20,647 genes were aligned to the 159 left hemisphere cortical regions in our parcellation, using prior methods (Romero-Garcia et al., 2018 with code available for download (https://github.com/RafaelRomeroGarcia/geneExpression_Repository). The resulting 159 x 20,647 regional gene expression matrix was z-scored by parcel. Because the OPC gene set was derived from sequencing performed on cortical brain tissue, we decided to exclude subcortical regions from this part of the analysis. 13 genes in the OPC gene set were not matched to any AHBA probe and were consequently excluded from the analysis. We evaluated the spatial specificity of the remaining 119 OPC genes by comparing their co-expression pattern with 1000 identically-sized sets of randomly chosen genes. OPC genes that did not share a positive coexpression pattern with the overall group of genes were filtered out. Concretely, the 24 genes which had, on average, negative correlations with other genes in the set were removed from the 4 OPC gene set. We estimated OPC distribution by calculating the median regional enrichment of the filtered OPC gene set across cortical parcels. OPC distribution across 159 cortical parcels was then correlated with tumor frequency and tested for significance using the spin test.

Influence of spatial vicinity on brain maps
To determine the degree to which spatial proximity explained the variance in brain maps, we computed the Moran's I of each key variable considered in the study. Moran's I is a normalized ratio of the covariance among spatially vicinal regions to the variance of the total brain map (Moran, 1950). Adjacent brain regions within the 167-region parcellation were categorized as vicinal, using a spatial weight matrix W, where element Wij = 1 if parcels i and j are neighbors and 0 otherwise. Moran's I for each variable is displayed in Supplementary Table 2. Each of these variables exhibited significant (non-random) spatial autocorrelation (Permutation test; p < 0.001; Supplementary Figure 2).

Replication of main findings
To determine the robustness of the results, major findings were internally replicated by using  Supplementary Figure 2. Permutation tests confirm non-random spatial autocorrelation structure of brain maps. Moran's I for each brain map was compared to a distribution of Moran's I values for 1000 randomly permuted maps.