Interindividual variation in human cortical cell type abundance and expression

Abstract

Single-cell transcriptomic studies have identified a conserved set of neocortical cell types from small postmortem cohorts. We extended these efforts by assessing cell type variation across 75 adult individuals undergoing epilepsy and tumor surgeries. Nearly all nuclei map to one of 125 robust cell types identified in the middle temporal gyrus. However, we found interindividual variance in abundances and gene expression signatures, particularly in deep-layer glutamatergic neurons and microglia. A minority of donor variance is explainable by age, sex, ancestry, disease state, and cell state. Genomic variation was associated with expression of 150 to 250 genes for most cell types. This characterization of cellular variation provides a baseline for cell typing in health and disease.

Figure 1:. Study design and overall gene expression variation in neurosurgical cohort.

(A) Summary of data types, individuals, demographics, and quality metrics in this study. (B) Schematic of quality control and cell type assignment for single nuclei collected in this study. (C) (top) Dendrogram of reference supertypes from this study (identical to the taxonomy from sea-ad.org). (bottom) Number of cells from this study mapped to each supertype.

Figure 2:. Cell type-specific differences in abundance across donors.

A) (top) Fraction of cells per donor from each subclass, scaled to the total number of cells per class. For all bar plots, points represent values per donor, with bars showing mean+/−standard error. (bottom) Significant associations between abundances and metadata based on a linear model with covariates for listed metadata and batch. B) Abundances after dividing donors by medical condition and brain region for select subclasses with significant changes. C) Bar plots as in B but showing the (excitatory-inhibitory) E-I ratio per donor (E-I ratio; y-axis).

Figure 3:. Donor-specific gene signatures differ by cell type and sex.

A-B) Low dimensional (UMAP) representation of all cells from this study (matching Fig 1B) color-coded by donor (A) or donor entropy (B). C) Fraction of expressed genes with significantly higher distance between cells from different donors than cells from the same donor (x-axis) for each subclass (y-axis; FDR<0.00494). D) Scatterplot showing median inter-donor (x-axis) vs. intra-donor (y-axis) distance for each gene expressed (points; top 15 genes are labeled). E) Random forest (RF) predictability of cells differs by cell type and donor characteristics. Heatmap shows fraction of nuclei of a given subclass (x-axis) correctly assigned to a given donor (y-axis) using RF classification with 75% training / 25% test strategy. Bar plots on the x- and y- axes show the median fraction of cells correctly classified by donor and subtype, respectively. Significant associations between classification accuracies and the listed metadata when performing the same linear model as in Fig 2A are shown, additionally color-coded by direction of maximum accuracy. F) Bar plots showing RF accuracies after dividing donors or cell types by listed metrics (#: p<0.05). G) Correlation between the fraction of high variance genes per subclass (y-axis in C) and the median RF predication accuracy.

Figure 4:. Variation partitioning explains differences in cell type-specific gene expression.

A) Interaction graph visualizing the estimated contribution of each covariate, purple node, to the variance in expression for each gene, green node, (covariate-gene edge) and the associated subclass, red node, (gene-subclass edge). The covariate-gene edge weight represents the precent variance explained for a gene by a covariate and the gene-subclass edge weight represents the non-residual contributions to variance explained for the gene. B) Beeswarm plots showing for each subclass the amount of variation explained per-gene, represented by a point, by donor on the top row and supertype on the bottom row. C) Violin plots showing the percentage of variation explained per-gene by each covariate for the Pvalb and L6b subclass. D) Top genes whose variance can be associated with donor, supertype, ancestry and residuals for the Pvalb and L6b subclasses are shown in the paired boxplots. E) Top ranked genes with respect to variance explained by donor across all subclasses are shown in the heatmap. Each row is a subclass, and each column is a gene which are colored based on rank determined from the percent variance explained by donor per-subclass. F) Stratification of gene expression by donor for the top, median and 3^rd quantile donor-associated genes colored by red, purple and blue, respectively. Donors are ordered by median expression for each gene. G) Stratification of gene expression by supertype for the top, median and 3^rd quantile donor-associated genes colored by red, purple and blue, respectively. Supertypes are ordered by median expression of each gene.

Figure 5:. Cell-type specific gene expression variation is associated with genetic effects:

A) Barplots showing the amount of eGenes (FDR < 0.05) identified for each subclass, grouped by glutamatergic, GABAergic and non-neuronal B) Scatterplot showing the relationship between numbers of cis-eQTLs (FDR < 0.05) on the y-axis and number of nuclei on the x-axis for each subclass. C) Enrichment of cis-eQTLs around the transcription start site of the proximal gene for each subclass. D) Examples of cis-eQTLs, FDR < 0.05 denoted by *, with restricted to deep excitatory types and L6b specific effects. E) Examples of cis-eQTLs, FDR < 0.05 denoted by *, with cell type-agnostic and associated effects on gene expression.