Fig. 2: Transcriptomic decoding of high-resolution surface-based imaging-derived phenotypes (IDPs).

a Singular Value Decomposition (SVD) was employed to reduce the spatially-interpolated mRNA expression signatures of all N = 15,633 genes in cortical brain tissue to a smaller subset of nine spatially-dense co-expression gradients (G1 to G9), which together captured ~41% of total variability in gene expression across the surface (see Supplementary Data Fig. 13a,b). b A total of N = 1000 spatial autocorrelation (α)-preserving null models (so-called surrogates¹⁸) were generated for each gradient pattern using the optimal k-nearest-neighbor (knn) parameter for each pattern (also see Supplementary Data Fig. 13c). These surrogate maps exhibit a similar degree of smoothness as the original patterns and were employed to testing the hypothesis of a significant spatial association between a target pattern and the predicted cortical expression signature of a gene. Data are presented as mean values +/- standard deviation (across N = 100 surrogate fits). c Genes were allocated to co-expression gradients based on their maximum absolute loadings (L) on gradient (G) patterns (G1 to G9). For each observed spatial correlation between a target IDP and a gene’s cortical expression profile, a non-parametric α-corrected p-value (p_perm, two-tailed) was identified based on the distribution of spatial null correlations (r₀) with the respective gradient pattern. To correct for multiple comparisons, p_perm-values were adjusted for (p_perm,adj) based on the empirical cumulative density function of the extreme value distribution (i.e., maximal spatial correlation or maxR) across gradient null-models (also see MaxT algorithm³⁸).