Detection of allele-specific expression in spatial transcriptomics with spASE

Spatial transcriptomics technologies permit the study of the spatial distribution of RNA at near-single-cell resolution genome-wide. However, the feasibility of studying spatial allele-specific expression (ASE) from these data remains uncharacterized. Here, we introduce spASE, a computational framework for detecting and estimating spatial ASE. To tackle the challenges presented by cell type mixtures and a low signal to noise ratio, we implement a hierarchical model involving additive mixtures of spatial smoothing splines. We apply our method to allele-resolved Visium and Slide-seq from the mouse cerebellum and hippocampus and report new insight into the landscape of spatial and cell type-specific ASE therein. We used female CAST/EiJ x 129S1/SvImJ obtained from Jackson Laboratories. Slide-seqV2 was performed as described previously on 10um-thick coronal slices of the hippocampus from three mice (Mice 1-3) and cerebellum (Mouse 3). For Mice 1 and 3, two serial sections were sequenced and reads were aggregated downstream, while for Mouse 2, only one section was sequenced. Slide-seq libraries were processed with the slideseq-tools pipeline (https://github.com/MacoskoLab/slideseq-tools) and then re-aligned to a custom transcriptome. For longer read libraries (Mouse 3 Slide-seq), we used Atropos to trim adapter sequences and low quality bases. 10x Genomics Visium was performed on two mice (Mice 4,5) on 10umn-thick coronal slices of the cerebellum. Visium libraries were processed with the 10x Genomics Space Ranger 1.1.0 pipeline and then re-aligned to a custom transcriptome. We generated a pooled CASTx129 transcriptome using the command create-hybrid from the EMASE software on the CAST and 129 transcript fasta files downloaded from ftp://churchill-lab.jax.org/software/g2gtools/mouse/R84-REL1505/. We then aligned reads to this pooled transcriptome with bowtie2 using the parameters -k 100 -p 16 --very-sensitive. This method of alignment ensures there is no reference bias. The multi-mapping parameter k=100 was chosen to report a large amount of multi-maps as bowtie2 randomly reports multi-mapping locations (in order of increasing number of mismatches). We used a custom script (https://github.com/lulizou/spASE/blob/master/scripts/processBowtie2.py) for processing the aligned BAM file to create a gene UMI count matrix only from reads that uniquely aligned to one gene and one allele. We conservatively restricted attention to alignments with 3 or fewer mismatches and only considered alignments that had the fewest number of mismatches for that read. We merged the reads from samples that had two serial sections: Mouse 1 hippocampus, Mouse 3 hippocampus, and Mouse 3 cerebellum.