Fig. 3: Imaging-transcriptomic associations.
a Schematic of the approaches for statistical testing of imaging-transcriptomic associations. Gene expression data for a subset of phenotype- or disease-specific genes are averaged and spatially compared to the patterns of multivariate topological changes in TLE and IGE independently. Spatial correlations are statistically assessed using one-tailed, non-parametric tests: (i) spatial permutation models, which preserve the spatial autocorrelation of brain maps (pspin; 10,000 permutations), and (ii) permutation models, which generate null distributions from randomised gene expression data with identical length as the original gene set (prand; 10,000 permutations). b Gene expression levels associated with two distinct epilepsy subtypes (focal epilepsy with hippocampal sclerosis and generalized epilepsy) were mapped to cortical and subcortical surface templates and spatially compared to patterns of multivariate topological alterations (which combined clustering and path length; see Fig. 2) across cortical and subcortical regions (n = 82) using one-tailed, non-parametric tests. In TLE, spatial associations between microarray data and multivariate topological changes were strongest for expression levels of hippocampal sclerosis genes (r = 0.33, pspin = 0.0028). On the other hand, in IGE, spatial associations were strongest for expression levels of generalized epilepsy genes (r = 0.31, pspin = 0.0032). Both TLE- and IGE-specific imaging-transcriptomic associations were robust against null distributions of effects based on selecting random genes from the full gene set (TLE: prand = 0.0030, IGE: prand = 0.018). HC = healthy control, IGE = idiopathic generalized epilepsy, TLE = temporal lobe epilepsy, pspin = p-value corrected against a null distribution of effects using a spatial permutation model, prand = p-value corrected against a null distribution of effects using a “random-gene” permutation model.
