Surrogate maps of cortical and subcortical structures

[Kaonashi22]

Good morning,

I have a question regarding the generation of surrogate maps of brain atrophy while combining cortical and subcortical regions (including basal ganglia, brainstem, cerebellar lobules and small brainstem nuclei), and belonging to four different atlases.
The goal is then to use these surrogate maps for spatial null models, to test the significance of the empirical variance of principal components of a partial least square analysis between gene expression data (X) and atrophy values (Y).

I have already seen the discussion in issue #22, but I'm still unsure about the approach to use to generate surrogate maps and about the accuracy of combining cortical and subcortical regions.
Here is the approach that I used:

1. I extracted centroids for my different cortical and subcortical regions with netneurotools using this command
   netneurotools.utils.get_centroids(img, labels=labels, image_space=True)
2. Then I converted these coordinates into a pairwise distance matrix with the Brain Connectivity toolbox to fit Brainsmash input requirements.
3. Finally, I ran this piece of code from Brainsmash to generate the surrogate maps.
   base = Base(x=brain_map_file, D=dist_mat_file)
   surrogates = base(n=1000)

The spatial null models based on these surrogate maps are not significant (p=0.09) while the random null models showed significance (p<0.001). When I run the model only on cortical regions, the spin test becomes significant, which is not the case with subcortical regions only. This is surprising given that the atrophy involves mainly subcortical regions while the cortex is relatively spared.

If you have any insights or suggestions regarding these findings or alternative approaches I could explore, I would greatly appreciate your input.
Thanks!

Lydia

[jbburt]

Hi Lydia, my apologies but I don't fully understand what you've done or what you're asking. You claim that

"When I run the model only on cortical regions, the spin test becomes significant, which is not the case with subcortical regions only." 

But the only spin test I'm familiar with is only defined in cortex. How did you perform a subcortex-only spin test? And what null hypothesis are you specifically trying to test?

[Kaonashi22]

Hi Josh,

Thank you for your reply, and apologies for the misleading terminology.
I want to test the significance of my PLS between atrophy values from various brain regions (cortex and subcortex- including basal ganglia, brainstem, and cerebellum) and gene expression data.I ran a spatial null model on all these regions using Euclidian distances on Brainsmash. My null hypothesis is that my surrogate maps could give a similar variance to my empirical simply by chance.The test gives nonsignificant results whereas the random null model did show significance. When I run the spatial null model using either Euclidian or geodesic distances on the cortical parcels only, the p value falls below 0.05.But, when I combine them with the subcortex, or when I analyze the subcortex separately, the tests are not significant anymore.Before concluding that my results are negative, I would like to make sure whether my approach is appropriate or not.

The centroids of the cerebellar lobules are very close to the centroids of the posterior temporal and occipital regions, numerically speaking. The distance between the occipital cortex and the cerebellum is very short numerically, but these regions do have different functions. When we permute, we end up preserving a spatial proximity/autocorrelation between the cerebellum and the occipital regions that does not exist... In a previous issue, you suggested generating surrogate maps separately for cortex (geodesic distances) and subcortex (euclidian or geodesic distances), demeaning them, then combining. This approach would account for the spatial autocorrelation between cortex and subcortex, right?
Best.