Identification of spatial expression trends in single-cell gene expression data

Abstract

As methods for measuring spatial gene expression at single-cell resolution become available, there is a need for computational analysis strategies. We present trendsceek, a method based on marked point processes that identifies genes with statistically significant spatial expression trends. trendsceek finds these genes in spatial transcriptomic and sequential fluorescence in situ hybridization data, and also reveals significant gene expression gradients and hot spots in low-dimensional projections of dissociated single-cell RNA-seq data.

Figure 1. Illustrating trendsceek on simulated data.

(A) Simulated mark distributions with local hotspot, step gradients, non-radial streaks and linear gradient patters. Expression values were sampled from empirical seqFISH data and cells in certain regions were spiked by sampling from the upper quantile of the expression distribution (cells n=500, spiked cells n=˜50, mean expression spiked cells / mean expression background = ˜10). (B) Marked point pattern statistics (Methods) for the simulated hotspot shown in (A). The mark correlation and mark variogram clearly indicate a significant spatial pattern as the band, which indicates the 5% significance level of a null distribution based on resampling of the mark distribution, is exceeded at certain radii. (C) 3D-representations of the spatial expression trend, where the green surfaces show weighted kernel density estimation (wKDE) of the simulated datasets. The blue surfaces indicate the upper 5% quantile of a null distribution generated by wKDE of the resampled mark distribution for each dataset. (D) Density plot of the simulated datasets (A) with cells colored red if they exceeded a 5% significance level based on wKDE, indicated by the blue surface in (C). Scale bars in (A) and (D) apply to all items of the figure, including the radius shown in (B).

Figure 2. Applications of trendsceek on spatial and single-cell gene expression data.

(A) Spatial transcriptomics data from mouse olfactory bulb (replicate 3, n=269 array-spots). Left: hematoxylin and eosin stained tissue-sections (from Ståhl et al4), followed by examples of genes with significant expression trends. Expression was scaled to the range zero to one by unity-based normalization. (B) Density plots of gene expression with cells in regions of significantly elevated expression coloured red. (C) Spatial transcriptomics data from mouse olfactory bulb (replicate 12, n=280 array-spots), as in (A). (D) As in (C) for mouse olfactory bulb, replicate 12. (E) Spatial transcriptomics data from breast cancer biopsy (histological section “Layer 2”, n=251 array-spots), with examples of genes with significant expression trends. The distance between array-spots is 200μm, each spot covering multiple cells. (F) Examples of distinct spatial expression patterns identified by trendsceek within E6.5 mouse epiblast cells (cluster 3 in Scialdone et al., n=481 cells). (G) Identification of spatial patterns related to the positions of male and female cells within the cluster, with mutually exclusive expression of Xist (expressed in female cells) and Eif2s3y (located on the Y-chromosome, only expressed in male cells). (H) Examples of spatial expression patterns identified in mouse hippocampus seqFISH data (cells imaged; H=93, O=89, T=208). Left: Cartoon of hippocampus with the 21 imaged regions labelled according to previous publication. P-values represent (A-E) mark-correlation (F-G) mark-variogram and (H) Emark (two-sided, Benjamini-Hochberg adjusted).