Library size confounds biology in spatial transcriptomics data

Abstract

Spatial molecular technologies have revolutionised the study of disease microenvironments by providing spatial context to tissue heterogeneity. Recent spatial technologies are increasing the throughput and spatial resolution of measurements, resulting in larger datasets. The added spatial dimension and volume of measurements poses an analytics challenge that has, in the short-term, been addressed by adopting methods designed for the analysis of single-cell RNA-seq data. Though these methods work well in some cases, not all necessarily translate appropriately to spatial technologies. A common assumption is that total sequencing depth, also known as library size, represents technical variation in single-cell RNA-seq technologies, and this is often normalised out during analysis. Through analysis of several different spatial datasets, we noted that this assumption does not necessarily hold in spatial molecular data. To formally assess this, we explore the relationship between library size and independently annotated spatial regions, across 23 samples from 4 different spatial technologies with varying throughput and spatial resolution. We found that library size confounded biology across all technologies, regardless of the tissue being investigated. Statistical modelling of binned total transcripts shows that tissue region is strongly associated with library size across all technologies, even after accounting for cell density of the bins. Through a benchmarking experiment, we show that normalising out library size leads to sub-optimal spatial domain identification using common graph-based clustering algorithms. On average, better clustering was achieved when library size effects were not normalised out explicitly, especially with data from the newer sub-cellular localised technologies. Taking these results into consideration, we recommend that spatial data should not be specifically corrected for library size prior to analysis unless strongly motivated. We also emphasise that spatial data are different to single-cell RNA-seq and care should be taken when adopting algorithms designed for single cell data.