Transcriptional and imaging-genetic association of cortical interneurons, brain function, and schizophrenia risk

Abstract

Inhibitory interneurons orchestrate information flow across the cortex and are implicated in psychiatric illness. Although interneuron classes have unique functional properties and spatial distributions, the influence of interneuron subtypes on brain function, cortical specialization, and illness risk remains elusive. Here, we demonstrate stereotyped negative correlation of somatostatin and parvalbumin transcripts within human and non-human primates. Cortical distributions of somatostatin and parvalbumin cell gene markers are strongly coupled to regional differences in functional MRI variability. In the general population (n = 9,713), parvalbumin-linked genes account for an enriched proportion of heritable variance in in-vivo functional MRI signal amplitude. Single-marker and polygenic cell deconvolution establish that this relationship is spatially dependent, following the topography of parvalbumin expression in post-mortem brain tissue. Finally, schizophrenia genetic risk is enriched among interneuron-linked genes and predicts cortical signal amplitude in parvalbumin-biased regions. These data indicate that the molecular-genetic basis of brain function is shaped by interneuron-related transcripts and may capture individual differences in schizophrenia risk.

Fig. 1. Cortical expression of SST and PVALB are negatively correlated across species and developmental stages.

a AHBA tissue samples mapped to the human cortical surface, and b an illustration of non-human primate tissue sample locations, colored by relative expression of SST (red) and PVALB (blue). Normalized expression difference reflects the sample-wise subtraction of z-transformed PVALB from SST. Relative SST–PVALB expression among anatomically defined groups from the c AHBA (n = 6 donors; n = 1683 samples; n = 41 regions) and d NIH Blueprint Non-Human Primate Atlas (n = 4 donors; n = 182 samples; n = 10 regions); circles = median, thick lines = interquartile range, thin lines = min and max values. e Sample-wise negative correlation of SST and PVALB in human cortex (r = −0.45, p < 2.2e−16; r_s = −0.40, p < 2.2e−16) and f non-human primates (r(34) = −0.74, p = 2.2e−7; r_s = −0.60, p = 0.0001). g Correlation of cortical SST and PVALB across nine developmental stages using data from the Brainspan Atlas of the Developing Human Brain (n = 42 donors; n = 362 samples; n = 9 developmental stages). h–j The SST to PVALB correlation is at the left tail of the distribution of all gene-wise correlations to SST (AUC = 0.009) and to PVALB (AUC = 0.033), as well as all possible two-gene spatial correlations (AUC = 0.001). *p ≤ 0.05, uncorrected; ^†p ≤ 0.10; error bars = standard error.

Fig. 2. Deconvolved cell type distributions are consistent with SST and PVALB single marker expression maps.

Using CibersortX, frontal and visual cortex snDrop-Seq data from Lake and colleagues were used to deconvolve cell type fractions from bulk AHBA microarray expression data. Deconvolved cell fractions of a somatostatin and b parvalbumin interneurons across cortex using single-cell data from frontal (left) and visual (right) cortex. c Somatostatin and parvalbumin cell fraction maps were z-transformed and subtracted (SST–PVALB) to illustrate the relative density of each subtype across cortex. d Spatial correlations of each deconvolved cell type across cortex using frontal (top-left triangle) and visual (bottom-right triangle) cell signatures. e Example marker genes for each cell class. f Deconvolved parvalbumin fractions across cortex were positively spatially correlated (Pearson’s) to single-gene PVALB expression (frontal cortex: r(337) = 0.81, p < 2.2e−16; visual cortex: r(337) = 0.48, p < 2.2e−16). g Deconvolved somatostatin fractions across cortex are positively spatially correlated (Pearson’s) to single-gene SST expression (frontal cortex: r(337) = 0.72, p < 2.2e−16; visual cortex: r(337) = 0.60, p < 2.2e−16).

Fig. 3. Subcortical SST and PVALB correlation and consistency in rodents.

a The correlation between SST and PVALB was estimated using AHBA data for each of seven subcortical territories. Red/blue circles denote Pearson correlations between SST and PVALB. Light and dark gray boxplots show the distribution of spatial correlations of all other genes (n = 17,447) to PVALB and SST, respectively (center = median, box = Q1–Q3, whiskers = 1.5*IQR, circles = outliers). The strength of the SST to PVALB relationship is quantified relative to gene-wide reference distributions. SST was more negatively correlated to PVALB than what is expected by chance in the hypothalamus (r(100) = −0.72, q = 1.9e−16, AUC_sst = 0.002, AUC_pvalb = 0.01), globus pallidus (r(37) = −0.39, q = 0.011), amygdala (r(65) = −0.30, q = 0.019), and thalamus (r(173) = −0.19, q = 0.019), but not the hippocampus (r(156) = −0.12, q = 0.14), ventral tegmentum/substantia nigra (r(63) = 0.09, q = 0.50), and striatum (r(168) = 0.23, q = 0.011). b Relative expression of SST and PVALB in human subcortical areas when compared to ground truth cell densities in rodent homologs from Kim et al. (Spearman’s r_s = 0.41, p = 0.025).

Fig. 4. SST–PVALB difference tracks inter-regional variation in cortical brain function.

a RSFA across each of the 400 Schaefer atlas parcels, averaged across 9713 UKB subjects. b Between-subject hierarchical clustering of residualized RSFA reveals 7-clusters cortical partitions with similar amplitude signatures; Light beige = limbic A, dark beige = limbic B, teal = cingulo-opercular, orange = temporal-parietal, red = prefrontal, blue = somato/motor, and purple = visual. c Relative presence of SST–PVALB is negatively correlated (Pearson’s) with cortical RSFA (r(337) = −0.53, p < 2.2e−16). d Across all deconvolved cell types, RSFA is most negatively spatially correlated (Pearson’s) to SST cell fractions (frontal cortex signature: r(337) = −0.48, p < 2.2e−16; visual cortex signature: r(337) = −0.36, p = 1.2e−11) and most positively correlated to PVALB cell fraction (frontal cortex signature: r(337) = 0.47, p < 2.2e−16; visual cortex signature: r(337) = 0.34, p = 1.7e−10). e Across frontal cortex (n = 272) and visual cortex (n = 240) cell type pairs, the relative difference of SST and PVALB cell fractions is most spatially associated with cortical RSFA.

Fig. 5. PVALB-linked genetic variation explains patterns of heritable brain function.

a RSFA was significantly heritable across the seven empirically defined spatial clusters (n = 9713 UKB subjects). b Partitioned heritability analyses reveal that the 500-gene PVALB_SNP set accounted for a significant proportion of heritable variance in all seven clusters (n = 9713 UKB subjects). c Parcel-wise PVALB_SNP partitioned heritability of RSFA (left panel) tracks the PVALB single-gene expression across cortex (Pearson’s r(326) = 0.35, p = 1.04e−10; r_s = 0.40, p = 2.2e−14). d Across all genes, PVALB was the 115th most correlated gene to PVALB_SNP partitioned heritability (AUC = 0.007). e Across all deconvolved cell types, inferred PVALB cell fraction was the most positively correlated to PVALB_SNP partitioned heritability (frontal cortex: Pearson’s r(329) = 0.35, p = 2.27e−11; r_s = 0.39, p = 2.20e−13; visual cortex: Pearson’s r(329) = 0.31, p = 6.67e−9; r_s = 0.34, p = 1.50e−10). f SST_SNP partitioned heritability was not associated with SST single-marker expression (Pearson’s r(326) = −0.02, p = 0.75), a null finding that was consistent when put in context of all genes (g) and inferred cell types (h). Error bars = standard error.

Fig. 6. Schizophrenia polygenic risk predicts brain function and tracks PVALB expression.

a Genes were rank-ordered by cortical spatial correlation to SST and PVALB, then divided into 500-gene bins. MAGMA competitive gene set analysis revealed enrichment of polygenic risk for schizophrenia in the top PVALB (p = 0.022), but not the top SST (p = 0.51) set. Enrichment decreased across ordered bins for PVALB (Spearman’s r_s = −0.48, p = 0.03) but not for SST (Spearman’s r_s = −0.001, p = 0.51). b Schizophrenia polygenic risk negatively predicts RSFA within the visual (q_1.0 = 0.04) cluster, as well as somato/motor (q_1.0 = 0.08) and prefrontal (q_1.0 = 0.10) clusters at trend-levels (corrected for multiple-comparisons). c Parcel-wise prediction of RSFA by the schizophrenia PRS negatively correlated with cortical expression of PVALB (Pearson’s r = −0.33, p = 3.1e−10), which was also significant relative to all genes (PVALB = 145/17,448, AUC = 0.008). d Across all deconvolved cell type distributions, PVALB was the most negatively correlated to cortical SCZ-RSFA effects (frontal cortex: Pearson’s r(337) = −0.36, p = 9.5e−12; visual cortex: Pearson’s r(337) = −0.26, p = 9.0e−7). SCZ = schizophrenia; PRS = polygenic risk score; RSFA = resting state functional amplitude. *q ≤ 0.05. Error bars = standard error.